diff --git a/bindings/ruby/Rakefile b/bindings/ruby/Rakefile new file mode 100644 index 00000000..354d8ef2 --- /dev/null +++ b/bindings/ruby/Rakefile @@ -0,0 +1,12 @@ +require 'rake/clean' + require 'rubygems/package' + +desc 'Build gem' +task :package do + spec_source = File.read File.join(File.dirname(__FILE__),'whispercpp.gemspec') + spec = nil + # see: http://gist.github.com/16215 + Thread.new { spec = eval("#{spec_source}") }.join + spec.validate + Gem::Package.build(spec) +end diff --git a/bindings/ruby/ext/ggml-backend-impl.h b/bindings/ruby/ext/ggml-backend-impl.h index 31788cd6..f121e1de 100644 --- a/bindings/ruby/ext/ggml-backend-impl.h +++ b/bindings/ruby/ext/ggml-backend-impl.h @@ -12,31 +12,63 @@ extern "C" { // Backend buffer // + // buffer type + typedef void * ggml_backend_buffer_type_context_t; + + struct ggml_backend_buffer_type_i { + const char * (*GGML_CALL get_name) (ggml_backend_buffer_type_t buft); + ggml_backend_buffer_t (*GGML_CALL alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size); + size_t (*GGML_CALL get_alignment) (ggml_backend_buffer_type_t buft); // tensor alignment + size_t (*GGML_CALL get_max_size) (ggml_backend_buffer_type_t buft); // allocation max size + size_t (*GGML_CALL get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding + bool (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend + // check if tensor data is in host memory + // should be equivalent to supports_backend(buft, ggml_backend_cpu_init()) + bool (*GGML_CALL is_host) (ggml_backend_buffer_type_t buft); + }; + + struct ggml_backend_buffer_type { + struct ggml_backend_buffer_type_i iface; + ggml_backend_buffer_type_context_t context; + }; + + // buffer typedef void * ggml_backend_buffer_context_t; struct ggml_backend_buffer_i { - void (*free_buffer) (ggml_backend_buffer_t buffer); - void * (*get_base) (ggml_backend_buffer_t buffer); // get base pointer - size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback - void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback - void (*free_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback + const char * (*GGML_CALL get_name) (ggml_backend_buffer_t buffer); + void (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer); + void * (*GGML_CALL get_base) (ggml_backend_buffer_t buffer); + void (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + void (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + void (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + bool (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer + void (*GGML_CALL clear) (ggml_backend_buffer_t buffer, uint8_t value); + void (*GGML_CALL reset) (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras }; struct ggml_backend_buffer { - struct ggml_backend_buffer_i iface; - - ggml_backend_t backend; + struct ggml_backend_buffer_i iface; + ggml_backend_buffer_type_t buft; ggml_backend_buffer_context_t context; - size_t size; + enum ggml_backend_buffer_usage usage; }; - GGML_API ggml_backend_buffer_t ggml_backend_buffer_init( - struct ggml_backend * backend, + GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init( + ggml_backend_buffer_type_t buft, struct ggml_backend_buffer_i iface, ggml_backend_buffer_context_t context, size_t size); + // do not use directly, use ggml_backend_tensor_copy instead + bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst); + + // buffer that contains a collection of buffers + GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers); + GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer); + GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); + // // Backend // @@ -44,44 +76,66 @@ extern "C" { typedef void * ggml_backend_context_t; struct ggml_backend_i { - const char * (*get_name)(ggml_backend_t backend); + const char * (*GGML_CALL get_name)(ggml_backend_t backend); - void (*free)(ggml_backend_t backend); + void (*GGML_CALL free)(ggml_backend_t backend); // buffer allocation - ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size); + ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend); - // get buffer alignment - size_t (*get_alignment)(ggml_backend_t backend); + // (optional) asynchronous tensor data access + void (*GGML_CALL set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst); - // tensor data access - // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize - void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - void (*synchronize) (ggml_backend_t backend); + // (optional) complete all pending operations + void (*GGML_CALL synchronize)(ggml_backend_t backend); - // (optional) copy tensor between different backends, allow for single-copy tranfers - void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); - void (*cpy_tensor_to) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); + // compute graph with a plan (not used currently) + ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph); + void (*GGML_CALL graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); // compute graph with a plan - ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph); - void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); - - // compute graph without a plan - bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); + enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); + // compute graph without a plan (async) + enum ggml_status (*GGML_CALL graph_compute) (ggml_backend_t backend, struct ggml_cgraph * cgraph); // check if the backend supports an operation - bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); + bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); + + // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer + // these should be expensive operations with large batch sizes that may benefit from running on this backend + // even if the weight has to be copied from the CPU temporarily + bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op); + + // (optional) event synchronization + ggml_backend_event_t (*GGML_CALL event_new) (ggml_backend_t backend); + void (*GGML_CALL event_free) (ggml_backend_event_t event); + void (*GGML_CALL event_record) (ggml_backend_event_t event); + void (*GGML_CALL event_wait) (ggml_backend_t backend, ggml_backend_event_t event); + void (*GGML_CALL event_synchronize) (ggml_backend_event_t event); }; struct ggml_backend { - struct ggml_backend_i iface; + ggml_guid_t guid; + struct ggml_backend_i iface; ggml_backend_context_t context; }; + struct ggml_backend_event { + ggml_backend_t backend; + void * context; + }; + + // + // Backend registry + // + + typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data); + + GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data); + #ifdef __cplusplus } #endif diff --git a/bindings/ruby/ext/ggml-backend.c b/bindings/ruby/ext/ggml-backend.c index 128e33ce..402d86ef 100644 --- a/bindings/ruby/ext/ggml-backend.c +++ b/bindings/ruby/ext/ggml-backend.c @@ -9,31 +9,76 @@ #include #include -#define UNUSED GGML_UNUSED #define MAX(a, b) ((a) > (b) ? (a) : (b)) +// backend buffer type + +const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) { + return buft->iface.get_name(buft); +} + +GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + return buft->iface.alloc_buffer(buft, size); +} + +size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) { + return buft->iface.get_alignment(buft); +} + +size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) { + // get_max_size is optional, defaults to SIZE_MAX + if (buft->iface.get_max_size) { + return buft->iface.get_max_size(buft); + } + return SIZE_MAX; +} + +GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { + // get_alloc_size is optional, defaults to ggml_nbytes + if (buft->iface.get_alloc_size) { + size_t size = buft->iface.get_alloc_size(buft, tensor); + assert(size >= ggml_nbytes(tensor)); + return size; + } + return ggml_nbytes(tensor); +} + +bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + return buft->iface.supports_backend(buft, backend); +} + +bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) { + if (buft->iface.is_host) { + return buft->iface.is_host(buft); + } + return false; +} + // backend buffer -ggml_backend_buffer_t ggml_backend_buffer_init( - struct ggml_backend * backend, +GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init( + ggml_backend_buffer_type_t buft, struct ggml_backend_buffer_i iface, ggml_backend_buffer_context_t context, size_t size) { ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer)); - GGML_ASSERT(iface.get_base != NULL); - (*buffer) = (struct ggml_backend_buffer) { /* .interface = */ iface, - /* .backend = */ backend, + /* .buft = */ buft, /* .context = */ context, /* .size = */ size, + /* .usage = */ GGML_BACKEND_BUFFER_USAGE_ANY }; return buffer; } +const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) { + return buffer->iface.get_name(buffer); +} + void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) { if (buffer == NULL) { return; @@ -45,10 +90,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) { free(buffer); } -size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) { - return ggml_backend_get_alignment(buffer->backend); -} - size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) { return buffer->size; } @@ -61,32 +102,67 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { return base; } -size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { - // get_alloc_size is optional, defaults to ggml_nbytes - if (buffer->iface.get_alloc_size) { - return buffer->iface.get_alloc_size(buffer, tensor); - } - return ggml_nbytes(tensor); -} - -void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +GGML_CALL void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { // init_tensor is optional if (buffer->iface.init_tensor) { buffer->iface.init_tensor(buffer, tensor); } } -void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { - // free_tensor is optional - if (buffer->iface.free_tensor) { - buffer->iface.free_tensor(buffer, tensor); +size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) { + return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer)); +} + +size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) { + return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer)); +} + +size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor); +} + +void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + buffer->iface.clear(buffer, value); +} + +bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) { + return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer)); +} + +void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) { + buffer->usage = usage; + + // FIXME: add a generic callback to the buffer interface + if (ggml_backend_buffer_is_multi_buffer(buffer)) { + ggml_backend_multi_buffer_set_usage(buffer, usage); } } +ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) { + return buffer->buft; +} + +void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) { + if (buffer->iface.reset) { + buffer->iface.reset(buffer); + } +} + +bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) { + ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer; + if (dst_buf->iface.cpy_tensor) { + return src->buffer->iface.cpy_tensor(dst_buf, src, dst); + } + return false; +} + // backend -ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) { - return tensor->buffer ? tensor->buffer->backend : NULL; +ggml_guid_t ggml_backend_guid(ggml_backend_t backend) { + if (backend == NULL) { + return NULL; + } + return backend->guid; } const char * ggml_backend_name(ggml_backend_t backend) { @@ -104,59 +180,105 @@ void ggml_backend_free(ggml_backend_t backend) { backend->iface.free(backend); } +ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) { + return backend->iface.get_default_buffer_type(backend); +} + ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) { - return backend->iface.alloc_buffer(backend, size); + return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size); } size_t ggml_backend_get_alignment(ggml_backend_t backend) { - return backend->iface.get_alignment(backend); + return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend)); } -void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size); +size_t ggml_backend_get_max_size(ggml_backend_t backend) { + return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend)); } -void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size); -} - -void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - ggml_backend_t backend = ggml_get_backend(tensor); - +void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); - GGML_ASSERT(backend != NULL && "tensor backend not set"); + GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); - backend->iface.set_tensor_async(backend, tensor, data, offset, size); - backend->iface.synchronize(backend); + if (backend->iface.set_tensor_async == NULL) { + ggml_backend_tensor_set(tensor, data, offset, size); + } else { + backend->iface.set_tensor_async(backend, tensor, data, offset, size); + } } -void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - ggml_backend_t backend = ggml_get_backend(tensor); - +void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); - GGML_ASSERT(backend != NULL && "tensor backend not set"); + GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds"); - backend->iface.get_tensor_async(backend, tensor, data, offset, size); - backend->iface.synchronize(backend); + if (backend->iface.get_tensor_async == NULL) { + ggml_backend_tensor_get(tensor, data, offset, size); + } else { + backend->iface.get_tensor_async(backend, tensor, data, offset, size); + } +} + +GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; + + GGML_ASSERT(buf != NULL && "tensor buffer not set"); + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); + + if (!size) { + return; + } + + buf->iface.set_tensor(buf, tensor, data, offset, size); +} + +GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { + ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer; + + GGML_ASSERT(buf != NULL && "tensor buffer not set"); + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds"); + + if (!size) { + return; + } + + buf->iface.get_tensor(buf, tensor, data, offset, size); } void ggml_backend_synchronize(ggml_backend_t backend) { + if (backend->iface.synchronize == NULL) { + return; + } + backend->iface.synchronize(backend); } ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) { + GGML_ASSERT(backend->iface.graph_plan_create != NULL); + return backend->iface.graph_plan_create(backend, cgraph); } void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { + GGML_ASSERT(backend->iface.graph_plan_free != NULL); + backend->iface.graph_plan_free(backend, plan); } -void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { - backend->iface.graph_plan_compute(backend, plan); +enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { + GGML_ASSERT(backend->iface.graph_plan_compute != NULL); + + return backend->iface.graph_plan_compute(backend, plan); } -bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { + enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph); + ggml_backend_synchronize(backend); + return err; +} + +enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) { return backend->iface.graph_compute(backend, cgraph); } @@ -164,6 +286,13 @@ bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * return backend->iface.supports_op(backend, op); } +bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) { + if (backend->iface.offload_op != NULL) { + return backend->iface.offload_op(backend, op); + } + return false; +} + // backend copy static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) { @@ -182,27 +311,20 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml } void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) { - //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]); - //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]); GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts"); - // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src)); - if (src == dst) { return; } - // TODO: allow backends to support copy to/from same backend - - if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) { - ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst); - } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) { - ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst); - } else { - // shouldn't be hit when copying from/to CPU - #ifndef NDEBUG - fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend)); - #endif + if (ggml_backend_buffer_is_host(src->buffer)) { + ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src)); + } else if (ggml_backend_buffer_is_host(dst->buffer)) { + ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src)); + } else if (!ggml_backend_buffer_copy_tensor(src, dst)) { +#ifndef NDEBUG + fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer)); +#endif size_t nbytes = ggml_nbytes(src); void * data = malloc(nbytes); ggml_backend_tensor_get(src, data, 0, nbytes); @@ -211,101 +333,424 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst } } +void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts"); + + if (src == dst) { + return; + } + + if (backend_dst->iface.cpy_tensor_async != NULL) { + if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) { + return; + } + } + + // an async copy would normally happen after all the queued operations on both backends are completed + // sync src, set_async dst + if (ggml_backend_buffer_is_host(src->buffer)) { + ggml_backend_synchronize(backend_src); + ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src)); + } else { + ggml_backend_synchronize(backend_src); + ggml_backend_tensor_copy(src, dst); + ggml_backend_synchronize(backend_dst); + } +} + +// events + +ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) { + if (backend->iface.event_new == NULL) { + return NULL; + } + return backend->iface.event_new(backend); +} + +void ggml_backend_event_free(ggml_backend_event_t event) { + if (event == NULL) { + return; + } + event->backend->iface.event_free(event); +} + +void ggml_backend_event_record(ggml_backend_event_t event) { + GGML_ASSERT(event->backend->iface.event_record != NULL); + + event->backend->iface.event_record(event); +} + +void ggml_backend_event_synchronize(ggml_backend_event_t event) { + GGML_ASSERT(event->backend->iface.event_synchronize != NULL); + + event->backend->iface.event_synchronize(event); +} + +void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) { + GGML_ASSERT(backend->iface.event_wait != NULL); + + backend->iface.event_wait(backend, event); +} + +// backend registry + +#define GGML_REG_MAX_BACKENDS 16 + +struct ggml_backend_reg { + char name[128]; + ggml_backend_init_fn init_fn; + ggml_backend_buffer_type_t default_buffer_type; + void * user_data; +}; + +static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS]; +static size_t ggml_backend_registry_count = 0; + +GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data); + +GGML_CALL static void ggml_backend_registry_init(void) { + static bool initialized = false; + + if (initialized) { + return; + } + + initialized = true; + + ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL); + + // add forward decls here to avoid including the backend headers +#ifdef GGML_USE_CUDA + extern GGML_CALL void ggml_backend_cuda_reg_devices(void); + ggml_backend_cuda_reg_devices(); +#endif + +#ifdef GGML_USE_SYCL + extern void ggml_backend_sycl_reg_devices(void); + ggml_backend_sycl_reg_devices(); +#endif + +#ifdef GGML_USE_METAL + extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); + extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); + ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL); +#endif + +#ifdef GGML_USE_VULKAN + extern GGML_CALL int ggml_backend_vk_reg_devices(void); + ggml_backend_vk_reg_devices(); +#endif + +#ifdef GGML_USE_KOMPUTE + extern GGML_CALL void ggml_backend_kompute_reg_devices(void); + ggml_backend_kompute_reg_devices(); +#endif +} + +GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) { + GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS); + + size_t id = ggml_backend_registry_count; + + ggml_backend_registry[id] = (struct ggml_backend_reg) { + /* .name = */ {0}, + /* .fn = */ init_fn, + /* .default_buffer_type = */ default_buffer_type, + /* .user_data = */ user_data, + }; + + snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name); + +#ifndef NDEBUG + fprintf(stderr, "%s: registered backend %s\n", __func__, name); +#endif + + ggml_backend_registry_count++; +} + +size_t ggml_backend_reg_get_count(void) { + ggml_backend_registry_init(); + + return ggml_backend_registry_count; +} + +size_t ggml_backend_reg_find_by_name(const char * name) { + ggml_backend_registry_init(); + + for (size_t i = 0; i < ggml_backend_registry_count; i++) { + // TODO: case insensitive in a portable way + if (strcmp(ggml_backend_registry[i].name, name) == 0) { + return i; + } + } + + // not found + return SIZE_MAX; +} + +// init from backend:params string +ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) { + ggml_backend_registry_init(); + + const char * params = strchr(backend_str, ':'); + char backend_name[128]; + if (params == NULL) { + snprintf(backend_name, sizeof(backend_name), "%s", backend_str); + params = ""; + } else { + snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str); + params++; + } + + size_t backend_i = ggml_backend_reg_find_by_name(backend_name); + + if (backend_i == SIZE_MAX) { + fprintf(stderr, "%s: backend %s not found\n", __func__, backend_name); + return NULL; + } + + return ggml_backend_reg_init_backend(backend_i, params); +} + +const char * ggml_backend_reg_get_name(size_t i) { + ggml_backend_registry_init(); + + GGML_ASSERT(i < ggml_backend_registry_count); + return ggml_backend_registry[i].name; +} + +ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) { + ggml_backend_registry_init(); + + GGML_ASSERT(i < ggml_backend_registry_count); + return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data); +} + +ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) { + ggml_backend_registry_init(); + + GGML_ASSERT(i < ggml_backend_registry_count); + return ggml_backend_registry[i].default_buffer_type; +} + +ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) { + ggml_backend_registry_init(); + + GGML_ASSERT(i < ggml_backend_registry_count); + return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size); +} + // backend CPU +static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment + +GGML_CALL static const char * ggml_backend_cpu_buffer_name(ggml_backend_buffer_t buffer) { + return "CPU"; + + GGML_UNUSED(buffer); +} + +GGML_CALL static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) { + uintptr_t data = (uintptr_t)buffer->context; + + // align the buffer + if (data % TENSOR_ALIGNMENT != 0) { + data = GGML_PAD(data, TENSOR_ALIGNMENT); + } + + return (void *)data; +} + +GGML_CALL static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) { + free(buffer->context); +} + +GGML_CALL static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { + memcpy((char *)tensor->data + offset, data, size); + + GGML_UNUSED(buffer); +} + +GGML_CALL static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { + memcpy(data, (const char *)tensor->data + offset, size); + + GGML_UNUSED(buffer); +} + +GGML_CALL static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { + if (ggml_backend_buffer_is_host(src->buffer)) { + memcpy(dst->data, src->data, ggml_nbytes(src)); + return true; + } + return false; + + GGML_UNUSED(buffer); +} + +GGML_CALL static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + memset(buffer->context, value, buffer->size); +} + +static struct ggml_backend_buffer_i cpu_backend_buffer_i = { + /* .get_name = */ ggml_backend_cpu_buffer_name, + /* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer, + /* .get_base = */ ggml_backend_cpu_buffer_get_base, + /* .init_tensor = */ NULL, // no initialization required + /* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor, + /* .clear = */ ggml_backend_cpu_buffer_clear, + /* .reset = */ NULL, +}; + +// for buffers from ptr, free is not called +static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = { + /* .get_name = */ ggml_backend_cpu_buffer_name, + /* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed + /* .get_base = */ ggml_backend_cpu_buffer_get_base, + /* .init_tensor = */ NULL, // no initialization required + /* .set_tensor = */ ggml_backend_cpu_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_cpu_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_cpu_buffer_cpy_tensor, + /* .clear = */ ggml_backend_cpu_buffer_clear, + /* .reset = */ NULL, +}; + +GGML_CALL static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "CPU"; + + GGML_UNUSED(buft); +} + +GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned + void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h) + if (data == NULL) { + fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size); + return NULL; + } + + return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size); +} + +GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return TENSOR_ALIGNMENT; + + GGML_UNUSED(buft); +} + +GGML_CALL static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + return ggml_backend_is_cpu(backend); + + GGML_UNUSED(buft); +} + +GGML_CALL static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) { + return true; + + GGML_UNUSED(buft); +} + +GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) { + static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = { + /* .iface = */ { + /* .get_name = */ ggml_backend_cpu_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_cpu_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes + /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend, + /* .is_host = */ ggml_backend_cpu_buffer_type_is_host, + }, + /* .context = */ NULL, + }; + + return &ggml_backend_cpu_buffer_type; +} + +#ifdef GGML_USE_CPU_HBM + +// buffer type HBM + +#include + +GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "CPU_HBM"; + + GGML_UNUSED(buft); +} + +GGML_CALL static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) { + return "CPU_HBM"; + + GGML_UNUSED(buf); +} + +GGML_CALL static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) { + hbw_free(buffer->context); +} + +GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + //void * ptr = hbw_malloc(size); + void * ptr; + int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size); + if (result != 0) { + fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size); + return NULL; + } + + ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); + buffer->buft = buft; + buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name; + buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer; + + return buffer; +} + +ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) { + static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = { + /* .iface = */ { + /* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes + /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend, + /* .is_host = */ ggml_backend_cpu_buffer_type_is_host, + }, + /* .context = */ NULL, + }; + + return &ggml_backend_cpu_buffer_type_hbm; +} +#endif + struct ggml_backend_cpu_context { int n_threads; void * work_data; size_t work_size; + + ggml_abort_callback abort_callback; + void * abort_callback_data; }; -static const char * ggml_backend_cpu_name(ggml_backend_t backend) { +GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) { return "CPU"; - UNUSED(backend); + GGML_UNUSED(backend); } -static void ggml_backend_cpu_free(ggml_backend_t backend) { +GGML_CALL static void ggml_backend_cpu_free(ggml_backend_t backend) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; free(cpu_ctx->work_data); free(cpu_ctx); free(backend); } -static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) { - return (void *)buffer->context; -} +GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) { + return ggml_backend_cpu_buffer_type(); -static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) { - free(buffer->context); - UNUSED(buffer); -} - -static struct ggml_backend_buffer_i cpu_backend_buffer_i = { - /* .free_buffer = */ ggml_backend_cpu_buffer_free_buffer, - /* .get_base = */ ggml_backend_cpu_buffer_get_base, - /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes - /* .init_tensor = */ NULL, // no initialization required - /* .free_tensor = */ NULL, // no cleanup required -}; - -// for buffers from ptr, free is not called -static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = { - /* .free_buffer = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed - /* .get_base = */ ggml_backend_cpu_buffer_get_base, - /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes - /* .init_tensor = */ NULL, - /* .free_tensor = */ NULL, -}; - -static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512 - -static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) { - size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned - void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC? - - GGML_ASSERT(data != NULL && "failed to allocate buffer"); - - return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size); -} - -static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) { - return TENSOR_ALIGNMENT; - UNUSED(backend); -} - -static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds"); - GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); - - memcpy((char *)tensor->data + offset, data, size); - - UNUSED(backend); -} - -static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds"); - GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); - - memcpy(data, (const char *)tensor->data + offset, size); - - UNUSED(backend); -} - -static void ggml_backend_cpu_synchronize(ggml_backend_t backend) { - UNUSED(backend); -} - -static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) { - ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src)); - - UNUSED(backend); -} - -static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) { - ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src)); - - UNUSED(backend); + GGML_UNUSED(backend); } struct ggml_backend_plan_cpu { @@ -313,95 +758,134 @@ struct ggml_backend_plan_cpu { struct ggml_cgraph cgraph; }; -static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu)); cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); - cpu_plan->cgraph = *cgraph; + cpu_plan->cgraph = *cgraph; // FIXME: deep copy if (cpu_plan->cplan.work_size > 0) { cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size); + if (cpu_plan->cplan.work_data == NULL) { + free(cpu_plan); + return NULL; + } } + cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback; + cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data; + return cpu_plan; } -static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan; free(cpu_plan->cplan.work_data); free(cpu_plan); - UNUSED(backend); + GGML_UNUSED(backend); } -static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan; - ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan); + return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan); - UNUSED(backend); + GGML_UNUSED(backend); } -static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); if (cpu_ctx->work_size < cplan.work_size) { - // TODO: may be faster to free and use malloc to avoid the copy - cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size); + free(cpu_ctx->work_data); + cpu_ctx->work_data = malloc(cplan.work_size); + if (cpu_ctx->work_data == NULL) { + cpu_ctx->work_size = 0; + return GGML_STATUS_ALLOC_FAILED; + } cpu_ctx->work_size = cplan.work_size; } - cplan.work_data = cpu_ctx->work_data; - ggml_graph_compute(cgraph, &cplan); + cplan.abort_callback = cpu_ctx->abort_callback; + cplan.abort_callback_data = cpu_ctx->abort_callback_data; + + return ggml_graph_compute(cgraph, &cplan); } -static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { - return true; - UNUSED(backend); - UNUSED(op); +GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { + switch (op->op) { + case GGML_OP_CPY: + return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float + case GGML_OP_MUL_MAT: + return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type; + default: + return true; + } + + GGML_UNUSED(backend); } static struct ggml_backend_i cpu_backend_i = { - /* .get_name = */ ggml_backend_cpu_name, - /* .free = */ ggml_backend_cpu_free, - /* .alloc_buffer = */ ggml_backend_cpu_alloc_buffer, - /* .get_alignment = */ ggml_backend_cpu_get_alignment, - /* .set_tensor_async = */ ggml_backend_cpu_set_tensor_async, - /* .get_tensor_async = */ ggml_backend_cpu_get_tensor_async, - /* .synchronize = */ ggml_backend_cpu_synchronize, - /* .cpy_tensor_from = */ ggml_backend_cpu_cpy_tensor_from, - /* .cpy_tensor_to = */ ggml_backend_cpu_cpy_tensor_to, - /* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create, - /* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free, - /* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute, - /* .graph_compute = */ ggml_backend_cpu_graph_compute, - /* .supports_op = */ ggml_backend_cpu_supports_op, + /* .get_name = */ ggml_backend_cpu_name, + /* .free = */ ggml_backend_cpu_free, + /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type, + /* .set_tensor_async = */ NULL, + /* .get_tensor_async = */ NULL, + /* .cpy_tensor_async = */ NULL, + /* .synchronize = */ NULL, + /* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create, + /* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free, + /* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute, + /* .graph_compute = */ ggml_backend_cpu_graph_compute, + /* .supports_op = */ ggml_backend_cpu_supports_op, + /* .offload_op = */ NULL, + /* .event_new = */ NULL, + /* .event_free = */ NULL, + /* .event_record = */ NULL, + /* .event_wait = */ NULL, + /* .event_synchronize = */ NULL, }; +static ggml_guid_t ggml_backend_cpu_guid(void) { + static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 }; + return &guid; +} + ggml_backend_t ggml_backend_cpu_init(void) { struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context)); + if (ctx == NULL) { + return NULL; + } - ctx->n_threads = GGML_DEFAULT_N_THREADS; - ctx->work_data = NULL; - ctx->work_size = 0; + ctx->n_threads = GGML_DEFAULT_N_THREADS; + ctx->work_data = NULL; + ctx->work_size = 0; + ctx->abort_callback = NULL; + ctx->abort_callback_data = NULL; ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend)); + if (cpu_backend == NULL) { + free(ctx); + return NULL; + } *cpu_backend = (struct ggml_backend) { + /* .guid = */ ggml_backend_cpu_guid(), /* .interface = */ cpu_backend_i, /* .context = */ ctx }; return cpu_backend; } -bool ggml_backend_is_cpu(ggml_backend_t backend) { - return backend->iface.get_name == ggml_backend_cpu_name; +GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) { + return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid()); } void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) { @@ -411,159 +895,99 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) { ctx->n_threads = n_threads; } -ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) { - return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size); +void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) { + GGML_ASSERT(ggml_backend_is_cpu(backend_cpu)); + + struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context; + ctx->abort_callback = abort_callback; + ctx->abort_callback_data = abort_callback_data; } -// scheduler +GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) { + GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned"); + return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size); +} -#define GGML_MAX_BACKENDS 4 -#define GGML_MAX_SPLITS 256 -#define GGML_MAX_SPLIT_INPUTS 16 +GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) { + return ggml_backend_cpu_init(); -struct ggml_backend_sched_split { - ggml_tallocr_t tallocr; - int i_start; - int i_end; - struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS]; - int n_inputs; - struct ggml_cgraph * graph; + GGML_UNUSED(params); + GGML_UNUSED(user_data); +} + +// multi-buffer buffer + +struct ggml_backend_multi_buffer_context { + ggml_backend_buffer_t * buffers; + size_t n_buffers; }; -struct ggml_backend_sched { - int n_backends; - ggml_backend_t backends[GGML_MAX_BACKENDS]; - ggml_tallocr_t tallocs[GGML_MAX_BACKENDS]; +typedef struct ggml_backend_multi_buffer_context * ggml_backend_multi_buffer_context_t; - ggml_gallocr_t galloc; +GGML_CALL static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) { + ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; - struct ggml_hash_set hash_set; - ggml_tallocr_t * node_talloc; // [hash_set.size] - struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS] - - struct ggml_cgraph * graph; - struct ggml_backend_sched_split splits[GGML_MAX_SPLITS]; - int n_splits; - - struct ggml_context * ctx; - - // align context_buffer to GGML_MEM_ALIGN - #ifdef _MSC_VER - __declspec(align(GGML_MEM_ALIGN)) - #else - __attribute__((aligned(GGML_MEM_ALIGN))) - #endif - char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)]; -}; - -#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node) -#define node_allocr(node) sched->node_talloc[hash_id(node)] - -static bool ggml_is_view_op(enum ggml_op op) { - return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE; + return ctx->buffers[0]->iface.get_name(ctx->buffers[0]); } -// returns the priority of the backend, lower is better -static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) { - for (int i = 0; i < sched->n_backends; i++) { - if (sched->backends[i] == backend) { - return i; - } +GGML_CALL static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) { + ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; + for (size_t i = 0; i < ctx->n_buffers; i++) { + ggml_backend_buffer_free(ctx->buffers[i]); } - return INT_MAX; + + free(ctx->buffers); + free(ctx); } -static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) { - for (int i = 0; i < sched->n_backends; i++) { - if (sched->tallocs[i] == allocr) { - return i; - } +GGML_CALL static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; + for (size_t i = 0; i < ctx->n_buffers; i++) { + ggml_backend_buffer_clear(ctx->buffers[i], value); } - return INT_MAX; } -// returns the backend that should be used for the node based on the current locations -char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove -static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) { - // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there - // ie. kv cache updates - // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend. - // dst - ggml_backend_t cur_backend = ggml_get_backend(node); - if (cur_backend != NULL) { - sprintf(causes[hash_id(node)], "1.dst"); - return cur_backend; - } +static struct ggml_backend_buffer_i ggml_backend_multi_buffer_context_interface(void) { + static struct ggml_backend_buffer_i multi_backend_buffer_i = { + /* .get_name = */ ggml_backend_multi_buffer_get_name, + /* .free_buffer = */ ggml_backend_multi_buffer_free_buffer, + /* .get_base = */ NULL, + /* .init_tensor = */ NULL, + /* .set_tensor = */ NULL, + /* .get_tensor = */ NULL, + /* .cpy_tensor = */ NULL, + /* .clear = */ ggml_backend_multi_buffer_clear, + /* .reset = */ NULL, + }; - // view_src - if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) { - sprintf(causes[hash_id(node)], "1.vsrc"); - return ggml_get_backend(node->view_src); - } - - // src - int cur_prio = INT_MAX; - size_t cur_size = 0; - - for (int i = 0; i < GGML_MAX_SRC; i++) { - const struct ggml_tensor * src = node->src[i]; - if (src == NULL) { - break; - } - ggml_backend_t src_backend = ggml_get_backend(src); - if (src_backend != NULL) { - int src_prio = sched_backend_prio(sched, src_backend); - size_t src_size = ggml_nbytes(src); - if (src_prio < cur_prio && src_size >= cur_size) { - cur_prio = src_prio; - cur_size = src_size; - cur_backend = src_backend; - sprintf(causes[hash_id(node)], "1.src%d", i); - } - } - } - return cur_backend; + return multi_backend_buffer_i; } -static char * fmt_size(size_t size) { - static char buffer[128]; - if (size >= 1024*1024) { - sprintf(buffer, "%zuM", size/1024/1024); - } else { - sprintf(buffer, "%zuK", size/1024); +GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) { + ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) malloc(sizeof(struct ggml_backend_multi_buffer_context)); + ctx->n_buffers = n_buffers; + ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t)); + + GGML_ASSERT(ctx->buffers != NULL); + + size_t total_size = 0; + for (size_t i = 0; i < n_buffers; i++) { + ctx->buffers[i] = buffers[i]; + total_size += ggml_backend_buffer_get_size(buffers[i]); } - return buffer; + + return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_context_interface(), ctx, total_size); } -static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { - int cur_split = 0; - for (int i = 0; i < graph->n_nodes; i++) { - if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) { - ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend; - fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs); - for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) { - fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j]))); - } - fprintf(stderr, "\n"); - cur_split++; - } - struct ggml_tensor * node = graph->nodes[i]; - if (ggml_is_view_op(node->op)) { - continue; - } - ggml_tallocr_t node_allocr = node_allocr(node); - ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL; - fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]); - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * src = node->src[j]; - if (src == NULL) { - break; - } - ggml_tallocr_t src_allocr = node_allocr(src); - ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL; - fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]); - } - fprintf(stderr, "\n"); +GGML_CALL bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) { + return buffer->iface.get_name == ggml_backend_multi_buffer_get_name; +} + +GGML_CALL void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) { + GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer)); + ggml_backend_multi_buffer_context_t ctx = (ggml_backend_multi_buffer_context_t) buffer->context; + for (size_t i = 0; i < ctx->n_buffers; i++) { + ggml_backend_buffer_set_usage(ctx->buffers[i], usage); } } @@ -576,315 +1000,760 @@ static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, co return dup; } -// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend -// TODO: merge passes -static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { - // reset state - size_t hash_size = sched->hash_set.size; - memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size); - memset(sched->node_talloc, 0, sizeof(sched->node_talloc[0]) * hash_size); - memset(sched->node_copies, 0, sizeof(sched->node_copies[0]) * hash_size); - sched->n_splits = 0; +static bool ggml_is_view_op(enum ggml_op op) { + return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE; +} - struct ggml_init_params params = { - /*.mem_size = */ sizeof(sched->context_buffer), - /*.mem_buffer = */ sched->context_buffer, - /*.no_alloc = */ true - }; +// scheduler - if (sched->ctx != NULL) { - ggml_free(sched->ctx); +#ifndef GGML_SCHED_MAX_BACKENDS +#define GGML_SCHED_MAX_BACKENDS 16 +#endif + +#ifndef GGML_SCHED_MAX_SPLITS +#define GGML_SCHED_MAX_SPLITS 2048 +#endif + +#ifndef GGML_SCHED_MAX_SPLIT_INPUTS +#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC +#endif + +#ifndef GGML_SCHED_MAX_COPIES +#define GGML_SCHED_MAX_COPIES 4 +#endif + +struct ggml_backend_sched_split { + int backend_id; + int i_start; + int i_end; + struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS]; + int n_inputs; + // graph view of this split + struct ggml_cgraph graph; +}; + +struct ggml_backend_sched { + bool is_reset; // true if the scheduler has been reset since the last graph split + bool is_alloc; + + int n_backends; + + ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS]; + ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS]; + ggml_gallocr_t galloc; + + // hash keys of the nodes in the graph + struct ggml_hash_set hash_set; + // hash values + int * tensor_backend_id; + struct ggml_tensor * (* tensor_copies)[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES]; + + int * node_backend_ids; // [graph_size] + int * leaf_backend_ids; // [graph_size] + + // copy of the graph with modified inputs + struct ggml_cgraph * graph; + + // graph splits + struct ggml_backend_sched_split * splits; + int n_splits; + int splits_capacity; + + // pipeline parallelism support + int n_copies; + int cur_copy; + ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES]; + struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS]; + int n_graph_inputs; + + struct ggml_context * ctx; + + ggml_backend_sched_eval_callback callback_eval; + void * callback_eval_user_data; + + // align context_buffer to GGML_MEM_ALIGN +#ifdef _MSC_VER + __declspec(align(GGML_MEM_ALIGN)) +#else + __attribute__((aligned(GGML_MEM_ALIGN))) +#endif + char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)]; +}; + +#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor) +#define tensor_backend_id(tensor) sched->tensor_backend_id[hash_id(tensor)] + +// returns the priority of the backend, lower id is higher priority +static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) { + for (int i = 0; i < sched->n_backends; i++) { + if (sched->backends[i] == backend) { + return i; + } + } + return -1; +} + +static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor) { + ggml_backend_buffer_t buffer = tensor->buffer; + if (buffer == NULL) { + return -1; } - sched->ctx = ggml_init(params); + // find highest prio backend that supports the buffer type + for (int i = 0; i < sched->n_backends; i++) { + if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) { + return i; + } + } - // pass 1: assign backends to ops with allocated inputs - for (int i = 0; i < graph->n_leafs; i++) { - struct ggml_tensor * leaf = graph->leafs[i]; - if (node_allocr(leaf) != NULL) { - // do not overwrite user assignments + fprintf(stderr, "%s: error: no backend supports buffer type %s used in tensor %s\n", + __func__, ggml_backend_buffer_name(buffer), tensor->name); + GGML_ASSERT(false); + + return -1; +} + +#if 0 +static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only +#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__) +#define GET_CAUSE(node) causes[hash_id(node)] +#else +#define SET_CAUSE(node, ...) +#define GET_CAUSE(node) "" +#endif + +// returns the backend that should be used for the node based on the current locations +static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) { + // TODO: use supports_op to check if the backend supports the op + + // assign pre-allocated nodes to their backend + int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor); + if (cur_backend_id != -1) { + SET_CAUSE(tensor, "1.dst"); + return cur_backend_id; + } + + // view_src + if (tensor->view_src != NULL) { + cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src); + if (cur_backend_id != -1) { + SET_CAUSE(tensor, "1.vsrc"); + return cur_backend_id; + } + } + + // graph input + if (tensor->flags & GGML_TENSOR_FLAG_INPUT) { + cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU) + SET_CAUSE(tensor, "1.inp"); + return cur_backend_id; + } + + // assign nodes that use weights to the backend of the weights + // operations with weights are preferably run on the same backend as the weights + for (int i = 0; i < GGML_MAX_SRC; i++) { + const struct ggml_tensor * src = tensor->src[i]; + if (src == NULL) { continue; } - ggml_backend_t leaf_backend = ggml_get_backend(leaf); - if (leaf_backend == NULL && leaf->view_src != NULL) { - leaf_backend = ggml_get_backend(leaf->view_src); - } - if (leaf_backend != NULL) { - node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend); - } - } - - for (int i = 0; i < graph->n_nodes; i++) { - struct ggml_tensor * node = graph->nodes[i]; - if (node_allocr(node) != NULL) { - // do not overwrite user assignments - continue; - } - ggml_backend_t node_backend = sched_backend_from_cur(sched, node); - if (node_backend != NULL) { - node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend); - } - } - //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); - - // pass 2: assign backends to ops from current assignments - // TODO: - // - reuse sched_backend_from_cur - for (int i = 0; i < graph->n_nodes; i++) { - struct ggml_tensor * node = graph->nodes[i]; - ggml_tallocr_t node_allocr = node_allocr(node); - if (node_allocr == NULL) { - int cur_prio = INT_MAX; - size_t cur_size = 0; - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * src = node->src[j]; - if (src == NULL) { - break; - } - ggml_tallocr_t src_allocr = node_allocr(src); - if (src_allocr != NULL) { - int src_prio = sched_allocr_prio(sched, src_allocr); - size_t src_size = ggml_nbytes(src); - if (src_prio < cur_prio && src_size >= cur_size) { - cur_prio = src_prio; - cur_size = src_size; - node_allocr = src_allocr; - sprintf(causes[hash_id(node)], "2.src%d", j); + if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) { + int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src); + // check if a backend with higher prio wants to offload the op + if (src_backend_id == sched->n_backends - 1) { + for (int b = 0; b < src_backend_id; b++) { + if (ggml_backend_offload_op(sched->backends[b], tensor)) { + SET_CAUSE(tensor, "1.off"); + return b; } } } - if (node_allocr != NULL) { - node_allocr(node) = node_allocr; - } + SET_CAUSE(tensor, "1.wgt%d", i); + return src_backend_id; } } - //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); - // pass 3: assign backends to remaining src from dst (should only be leafs) - for (int i = 0; i < graph->n_nodes; i++) { - struct ggml_tensor * node = graph->nodes[i]; - ggml_tallocr_t node_allocr = node_allocr(node); - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * src = node->src[j]; - if (src == NULL) { - break; - } - ggml_tallocr_t src_allocr = node_allocr(src); - if (src_allocr == NULL) { - node_allocr(src) = node_allocr; - } - } + return -1; +} + +static char * fmt_size(size_t size) { + static char buffer[128]; + if (size >= 1024*1024) { + sprintf(buffer, "%zuM", size/1024/1024); + } else { + sprintf(buffer, "%zuK", size/1024); } - //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + return buffer; +} - // pass 4: split graph, find tensors that need to be copied - // TODO: - // - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost - // find first backend +static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { int cur_split = 0; for (int i = 0; i < graph->n_nodes; i++) { - struct ggml_tensor * node = graph->nodes[i]; - if (node->view_src == NULL) { - sched->splits[0].tallocr = node_allocr(node); - break; + if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) { + ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id]; + fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), + sched->splits[cur_split].n_inputs); + for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) { + fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, + fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j]))); + } + fprintf(stderr, "\n"); + cur_split++; } - } - sched->splits[0].i_start = 0; - sched->splits[0].n_inputs = 0; - memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK - ggml_tallocr_t cur_allocr = sched->splits[0].tallocr; - size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr); - for (int i = 0; i < graph->n_nodes; i++) { struct ggml_tensor * node = graph->nodes[i]; - if (ggml_is_view_op(node->op)) { continue; } - - ggml_tallocr_t node_allocr = node_allocr(node); - - if (node_allocr != cur_allocr) { - sched->splits[cur_split].i_end = i; - cur_split++; - GGML_ASSERT(cur_split < GGML_MAX_SPLITS); - sched->splits[cur_split].tallocr = node_allocr; - sched->splits[cur_split].i_start = i; - sched->splits[cur_split].n_inputs = 0; - memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK - cur_allocr = node_allocr; - cur_backend_id = sched_allocr_prio(sched, cur_allocr); - } - - // find inputs that are not on the same backend + ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node); + fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name, + fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node)); for (int j = 0; j < GGML_MAX_SRC; j++) { struct ggml_tensor * src = node->src[j]; if (src == NULL) { - break; - } - ggml_tallocr_t src_allocr = node_allocr(src); - if (src_allocr != node_allocr) { - int n_inputs = sched->splits[cur_split].n_inputs++; - GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); - sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src; - - // create copies - size_t id = hash_id(src); - if (sched->node_copies[id][cur_backend_id] == NULL) { - struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); - sched->node_copies[id][cur_backend_id] = tensor_copy; - node_allocr(tensor_copy) = cur_allocr; - ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend; - ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); - } - node->src[j] = sched->node_copies[id][cur_backend_id]; + continue; } + ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src); + fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name, + fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src)); } + fprintf(stderr, "\n"); } - sched->splits[cur_split].i_end = graph->n_nodes; - sched->n_splits = cur_split + 1; +} - //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout); +//#define DEBUG_PASS1 +//#define DEBUG_PASS2 +//#define DEBUG_PASS3 +//#define DEBUG_PASS4 + +// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend +static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + // reset splits + sched->n_splits = 0; + sched->n_graph_inputs = 0; + sched->is_reset = false; + + struct ggml_init_params params = { + /* .mem_size = */ sizeof(sched->context_buffer), + /* .mem_buffer = */ sched->context_buffer, + /* .no_alloc = */ true + }; + + ggml_free(sched->ctx); + + sched->ctx = ggml_init(params); + if (sched->ctx == NULL) { + fprintf(stderr, "%s: failed to initialize context\n", __func__); + GGML_ASSERT(false); + } + + // pass 1: assign backends to ops with pre-allocated inputs + for (int i = 0; i < graph->n_leafs; i++) { + struct ggml_tensor * leaf = graph->leafs[i]; + int * leaf_backend_id = &tensor_backend_id(leaf); + if (*leaf_backend_id != -1) { + // do not overwrite user assignments + continue; + } + *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf); + } -#if 1 - // sanity check: all sources should have the same backend as the node for (int i = 0; i < graph->n_nodes; i++) { struct ggml_tensor * node = graph->nodes[i]; - ggml_tallocr_t node_allocr = node_allocr(node); - if (node_allocr == NULL) { - fprintf(stderr, "!!!!!!! %s has no backend\n", node->name); + int * node_backend_id = &tensor_backend_id(node); + if (*node_backend_id != -1) { + // do not overwrite user assignments + continue; + } + *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node); + // src + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + continue; + } + int * src_backend_id = &tensor_backend_id(src); + if (*src_backend_id == -1) { + *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src); + } + } + } +#ifdef DEBUG_PASS1 + fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph); +#endif + + // pass 2: expand current backend assignments + // assign the same backend to adjacent nodes + // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend) + // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops + + + // pass 2.2 expand gpu down + { + int cur_backend_id = -1; + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (ggml_is_view_op(node->op)) { + continue; + } + int * node_backend_id = &tensor_backend_id(node); + if (*node_backend_id != -1) { + if (*node_backend_id == sched->n_backends - 1) { + // skip cpu (lowest prio backend) + cur_backend_id = -1; + } else { + cur_backend_id = *node_backend_id; + } + } else { + *node_backend_id = cur_backend_id; + SET_CAUSE(node, "2.2"); + } + } + } + // pass 2.1 expand gpu up + { + int cur_backend_id = -1; + for (int i = graph->n_nodes - 1; i >= 0; i--) { + struct ggml_tensor * node = graph->nodes[i]; + if (ggml_is_view_op(node->op)) { + continue; + } + int * node_backend_id = &tensor_backend_id(node); + if (*node_backend_id != -1) { + if (*node_backend_id == sched->n_backends - 1) { + // skip cpu (lowest prio backend) + cur_backend_id = -1; + } else { + cur_backend_id = *node_backend_id; + } + } else { + *node_backend_id = cur_backend_id; + SET_CAUSE(node, "2.1"); + } + } + } + // pass 2.4 expand rest down + { + int cur_backend_id = -1; + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (ggml_is_view_op(node->op)) { + continue; + } + int * node_backend_id = &tensor_backend_id(node); + if (*node_backend_id != -1) { + cur_backend_id = *node_backend_id; + } else { + *node_backend_id = cur_backend_id; + SET_CAUSE(node, "2.4"); + } + } + } + // pass 2.3 expand rest up + { + int cur_backend_id = -1; + for (int i = graph->n_nodes - 1; i >= 0; i--) { + struct ggml_tensor * node = graph->nodes[i]; + if (ggml_is_view_op(node->op)) { + continue; + } + int * node_backend_id = &tensor_backend_id(node); + if (*node_backend_id != -1) { + cur_backend_id = *node_backend_id; + } else { + *node_backend_id = cur_backend_id; + SET_CAUSE(node, "2.3"); + } + } + } + +#ifdef DEBUG_PASS2 + fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph); +#endif + + // pass 3: assign backends to remaining src from dst and view_src + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + int * cur_backend_id = &tensor_backend_id(node); + if (node->view_src != NULL && *cur_backend_id == -1) { + *cur_backend_id = tensor_backend_id(node->view_src); + SET_CAUSE(node, "3.vsrc"); } for (int j = 0; j < GGML_MAX_SRC; j++) { struct ggml_tensor * src = node->src[j]; if (src == NULL) { - break; + continue; } - ggml_tallocr_t src_allocr = node_allocr(src); - if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now - fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n", - node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL", - j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL"); + int * src_backend_id = &tensor_backend_id(src); + if (*src_backend_id == -1) { + if (src->view_src != NULL) { + // views are always on the same backend as the source + *src_backend_id = tensor_backend_id(src->view_src); + SET_CAUSE(src, "3.vsrc"); + } else { + *src_backend_id = *cur_backend_id; + SET_CAUSE(src, "3.cur"); + } } } } +#ifdef DEBUG_PASS3 + fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph); +#endif + + // pass 4: split graph, find tensors that need to be copied + { + int i_split = 0; + struct ggml_backend_sched_split * split = &sched->splits[0]; + // find the backend of the first split, skipping view ops + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (!ggml_is_view_op(node->op)) { + split->backend_id = tensor_backend_id(node); + break; + } + } + split->i_start = 0; + split->n_inputs = 0; + memset(split->inputs, 0, sizeof(split->inputs)); //HACK + int cur_backend_id = split->backend_id; + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + + if (ggml_is_view_op(node->op)) { + continue; + } + + const int node_backend_id = tensor_backend_id(node); + + GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now + + // check if we should start a new split based on the sources of the current node + bool need_new_split = false; + if (node_backend_id == cur_backend_id && split->n_inputs > 0) { + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + continue; + } + // check if a weight is on a different backend + // by starting a new split, the memory of the previously offloaded weights can be reused + if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) { + int src_backend_id = tensor_backend_id(src); + if (src_backend_id != -1 && src_backend_id != cur_backend_id) { + need_new_split = true; + break; + } + } + // check if the split has too many inputs + if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) { + const size_t id = hash_id(src); + int src_backend_id = sched->tensor_backend_id[id]; + if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL) { + //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name); + need_new_split = true; + break; + } + } + } + } + + if (node_backend_id != cur_backend_id || need_new_split) { + split->i_end = i; + i_split++; + if (i_split >= sched->splits_capacity) { + sched->splits_capacity *= 2; + sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split)); + GGML_ASSERT(sched->splits != NULL); + } + GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS); + split = &sched->splits[i_split]; + split->backend_id = node_backend_id; + split->i_start = i; + split->n_inputs = 0; + cur_backend_id = node_backend_id; + } + + // find inputs that are not on the same backend + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + continue; + } + + const int src_backend_id = tensor_backend_id(src); + assert(src_backend_id != -1); // all inputs should be assigned by now + + if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) { + size_t id = hash_id(src); + if (sched->tensor_copies[id][src_backend_id][0] == NULL) { + ggml_backend_t backend = sched->backends[src_backend_id]; + for (int c = 0; c < sched->n_copies; c++) { + struct ggml_tensor * tensor_copy; + if (c == sched->cur_copy) { + tensor_copy = src; // use the original tensor as the current copy + } else { + tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); + ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c); + } + if (sched->n_copies > 1) { + ggml_set_input(tensor_copy); + ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor + } + sched->tensor_copies[id][src_backend_id][c] = tensor_copy; + SET_CAUSE(tensor_copy, "4.cpy"); + } + int n_graph_inputs = sched->n_graph_inputs++; + GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS); + sched->graph_inputs[n_graph_inputs] = src; + } + } + + if (src_backend_id != node_backend_id) { + // create a copy of the input in the split's backend + const size_t id = hash_id(src); + if (sched->tensor_copies[id][cur_backend_id][0] == NULL) { + ggml_backend_t backend = sched->backends[cur_backend_id]; + for (int c = 0; c < sched->n_copies; c++) { + struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); + ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c); + if (sched->n_copies > 1) { + ggml_set_input(tensor_copy); + ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor + } + sched->tensor_copies[id][cur_backend_id][c] = tensor_copy; + SET_CAUSE(tensor_copy, "4.cpy"); + } + int n_inputs = split->n_inputs++; + GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS); + split->inputs[n_inputs] = src; + } + node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy]; + } + } + } + split->i_end = graph->n_nodes; + sched->n_splits = i_split + 1; + } +#ifdef DEBUG_PASS4 + fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph); #endif // create copies of the graph for each split - // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way - struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false); + // TODO: avoid this copy + struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2, false); for (int i = 0; i < sched->n_splits; i++) { struct ggml_backend_sched_split * split = &sched->splits[i]; - split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end); + split->graph = ggml_graph_view(graph, split->i_start, split->i_end); // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split for (int j = 0; j < split->n_inputs; j++) { + assert(graph_copy->size > (graph_copy->n_nodes + 1)); + struct ggml_tensor * input = split->inputs[j]; - struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)]; - input_cpy->src[0] = input; + const size_t input_id = hash_id(input); + struct ggml_tensor * input_cpy = sched->tensor_copies[input_id][split->backend_id][sched->cur_copy]; + + // add a dependency to the input source so that it is not freed before the copy is done + struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input); + input_dep->src[0] = input; + sched->node_backend_ids[graph_copy->n_nodes] = sched->tensor_backend_id[input_id]; + graph_copy->nodes[graph_copy->n_nodes++] = input_dep; + + // add a dependency to the input copy so that it is allocated at the start of the split + sched->node_backend_ids[graph_copy->n_nodes] = split->backend_id; graph_copy->nodes[graph_copy->n_nodes++] = input_cpy; } for (int j = split->i_start; j < split->i_end; j++) { + assert(graph_copy->size > graph_copy->n_nodes); + sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]); graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j]; } } + + if (sched->n_copies > 1) { + // add input copies as leafs so that they are allocated first + for (int i = 0; i < sched->n_graph_inputs; i++) { + struct ggml_tensor * input = sched->graph_inputs[i]; + size_t id = hash_id(input); + int backend_id = tensor_backend_id(input); + for (int c = 0; c < sched->n_copies; c++) { + struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c]; + sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id; + graph_copy->leafs[graph_copy->n_leafs++] = input_cpy; + } + } + + for (int i = 0; i < sched->n_splits; i++) { + struct ggml_backend_sched_split * split = &sched->splits[i]; + int backend_id = split->backend_id; + for (int j = 0; j < split->n_inputs; j++) { + struct ggml_tensor * input = split->inputs[j]; + size_t id = hash_id(input); + for (int c = 0; c < sched->n_copies; c++) { + struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c]; + sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id; + graph_copy->leafs[graph_copy->n_leafs++] = input_cpy; + } + } + } + } + + // add leafs from the original graph + for (int i = 0; i < graph->n_leafs; i++) { + struct ggml_tensor * leaf = graph->leafs[i]; + sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf); + graph_copy->leafs[graph_copy->n_leafs++] = leaf; + } + sched->graph = graph_copy; } -static void sched_alloc_splits(ggml_backend_sched_t sched) { - ggml_gallocr_alloc_graph_n( - sched->galloc, - sched->graph, - sched->hash_set, - sched->node_talloc); +static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) { + // allocate graph + if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) { + // the re-allocation may cause the split inputs to be moved to a different address + ggml_backend_sched_synchronize(sched); +#ifndef NDEBUG + fprintf(stderr, "%s: failed to allocate graph, reserving\n", __func__); +#endif + ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids); + if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) { + fprintf(stderr, "%s: failed to allocate graph\n", __func__); + return false; + } + } + + return true; } -static void sched_compute_splits(ggml_backend_sched_t sched) { - uint64_t copy_us[GGML_MAX_BACKENDS] = {0}; - uint64_t compute_us[GGML_MAX_BACKENDS] = {0}; - +static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) { struct ggml_backend_sched_split * splits = sched->splits; for (int i = 0; i < sched->n_splits; i++) { struct ggml_backend_sched_split * split = &splits[i]; - ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend; - int split_backend_id = sched_backend_prio(sched, split_backend); + int split_backend_id = split->backend_id; + ggml_backend_t split_backend = sched->backends[split_backend_id]; // copy the input tensors to the split backend - uint64_t copy_start_us = ggml_time_us(); for (int j = 0; j < split->n_inputs; j++) { - struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)]; - if (split->inputs[j]->buffer == NULL) { - if (split->inputs[j]->view_src == NULL) { - fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name); - exit(1); + ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]); + struct ggml_tensor * input = split->inputs[j]; + struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy]; + + if (input->flags & GGML_TENSOR_FLAG_INPUT) { + // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done + if (sched->events[split_backend_id][sched->cur_copy] != NULL) { + ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]); + } else { + ggml_backend_synchronize(split_backend); } - struct ggml_tensor * view = split->inputs[j]; - view->backend = view->view_src->backend; - view->buffer = view->view_src->buffer; - view->data = (char *)view->view_src->data + view->view_offs; - ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view); + ggml_backend_tensor_copy(input, input_cpy); + } else { + // wait for the split backend to finish using the input before overwriting it + if (sched->events[split_backend_id][sched->cur_copy] != NULL) { + ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]); + } else { + ggml_backend_synchronize(split_backend); + } + ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy); } - if (input_cpy->buffer == NULL) { - fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name); - exit(1); - } - GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend); - GGML_ASSERT(input_cpy->buffer->backend == split_backend); - ggml_backend_tensor_copy(split->inputs[j], input_cpy); } - // ggml_backend_synchronize(split_backend); - int64_t copy_end_us = ggml_time_us(); - copy_us[split_backend_id] += copy_end_us - copy_start_us; -#if 0 - char split_filename[GGML_MAX_NAME]; - snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend)); - ggml_graph_dump_dot(split->graph, NULL, split_filename); -#endif + if (!sched->callback_eval) { + enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph); + if (ec != GGML_STATUS_SUCCESS) { + return ec; + } + } else { + // similar to ggml_backend_compare_graph_backend + for (int j0 = 0; j0 < split->graph.n_nodes; j0++) { + struct ggml_tensor * t = split->graph.nodes[j0]; - uint64_t compute_start_us = ggml_time_us(); - ggml_backend_graph_compute(split_backend, split->graph); - // ggml_backend_synchronize(split_backend); - uint64_t compute_end_us = ggml_time_us(); - compute_us[split_backend_id] += compute_end_us - compute_start_us; - } + // check if the user needs data from this node + bool need = sched->callback_eval(t, true, sched->callback_eval_user_data); -#if 0 - // per-backend timings - fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits); - for (int i = 0; i < sched->n_backends; i++) { - if (copy_us[i] > 0 || compute_us[i] > 0) { - fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]); + int j1 = j0; + + // determine the range [j0, j1] of nodes that can be computed together + while (!need && j1 < split->graph.n_nodes - 1) { + t = split->graph.nodes[++j1]; + need = sched->callback_eval(t, true, sched->callback_eval_user_data); + } + + struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1); + + enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv); + if (ec != GGML_STATUS_SUCCESS) { + return ec; + } + + // TODO: pass backend to the callback, then the user can decide if they want to synchronize + ggml_backend_synchronize(split_backend); + + if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) { + break; + } + + j0 = j1; + } + } + + // record the event of this copy + if (split->n_inputs > 0) { + if (sched->events[split_backend_id][sched->cur_copy] != NULL) { + ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]); + } } } -#endif + + sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies; + + return GGML_STATUS_SUCCESS; } -static void sched_reset(ggml_backend_sched_t sched) { - for (int i = 0; i < sched->n_backends; i++) { - ggml_tallocr_reset(sched->tallocs[i]); - } -} +ggml_backend_sched_t ggml_backend_sched_new( + ggml_backend_t * backends, + ggml_backend_buffer_type_t * bufts, + int n_backends, + size_t graph_size, + bool parallel) { + GGML_ASSERT(n_backends > 0); + GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS); + GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU -ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) { - GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS); + struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1); - struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched)); - memset(sched, 0, sizeof(struct ggml_backend_sched)); + // initialize hash table + sched->hash_set = ggml_hash_set_new(graph_size); + sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size); + sched->tensor_copies = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size); - fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024); + const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2; + sched->node_backend_ids = calloc(sizeof(sched->node_backend_ids[0]), nodes_size); + sched->leaf_backend_ids = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size); sched->n_backends = n_backends; - for (int i = 0; i < n_backends; i++) { - sched->backends[i] = backends[i]; + + sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1; + + const int initial_splits_capacity = 16; + sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity); + sched->splits_capacity = initial_splits_capacity; + + for (int b = 0; b < n_backends; b++) { + sched->backends[b] = backends[b]; + sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]); + GGML_ASSERT(ggml_backend_buft_supports_backend(sched->bufts[b], backends[b])); + if (sched->n_copies > 1) { + for (int c = 0; c < sched->n_copies; c++) { + sched->events[b][c] = ggml_backend_event_new(backends[b]); + } + } } - sched->galloc = ggml_gallocr_new(); + sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends); - // init measure allocs for each backend - for (int i = 0; i < n_backends; i++) { - sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]); - } + ggml_backend_sched_reset(sched); return sched; } @@ -893,58 +1762,334 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) { if (sched == NULL) { return; } - for (int i = 0; i < sched->n_backends; i++) { - ggml_tallocr_free(sched->tallocs[i]); + for (int b = 0; b < sched->n_backends; b++) { + for (int c = 0; c < sched->n_copies; c++) { + ggml_backend_event_free(sched->events[b][c]); + } } ggml_gallocr_free(sched->galloc); + ggml_free(sched->ctx); + free(sched->splits); free(sched->hash_set.keys); - free(sched->node_talloc); - free(sched->node_copies); + free(sched->tensor_backend_id); + free(sched->tensor_copies); + free(sched->node_backend_ids); + free(sched->leaf_backend_ids); free(sched); } -void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) { - // initialize hash tables - size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS; - sched->hash_set.size = hash_size; - sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size); - sched->node_talloc = malloc(sizeof(sched->node_talloc[0]) * hash_size); - sched->node_copies = malloc(sizeof(sched->node_copies[0]) * hash_size); +void ggml_backend_sched_reset(ggml_backend_sched_t sched) { + // reset state for the next run + size_t hash_size = sched->hash_set.size; + memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size); // NOLINT + memset(sched->tensor_backend_id, -1, sizeof(sched->tensor_backend_id[0]) * hash_size); + memset(sched->tensor_copies, 0, sizeof(sched->tensor_copies[0]) * hash_size); - sched_split_graph(sched, measure_graph); - sched_alloc_splits(sched); + sched->is_reset = true; + sched->is_alloc = false; +} - // allocate buffers and reset allocators - for (int i = 0; i < sched->n_backends; i++) { - size_t size = ggml_tallocr_max_size(sched->tallocs[i]); - ggml_tallocr_free(sched->tallocs[i]); - sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size); +bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) { + GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes); + + ggml_backend_sched_split_graph(sched, measure_graph); + + // TODO: extract this to a separate function + if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) { + return false; } - sched_reset(sched); + ggml_backend_sched_reset(sched); + ggml_backend_sched_synchronize(sched); + + return true; } -void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { - GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS); +bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes); - sched_split_graph(sched, graph); - sched_alloc_splits(sched); - sched_compute_splits(sched); - sched_reset(sched); + ggml_backend_sched_split_graph(sched, graph); + + if (!ggml_backend_sched_alloc_splits(sched)) { + return false; + } + + sched->is_alloc = true; + + return true; } -ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) { - int backend_index = sched_backend_prio(sched, backend); - return sched->tallocs[backend_index]; +enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph); + ggml_backend_sched_synchronize(sched); + return err; } -ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) { - int backend_index = sched_backend_prio(sched, backend); - return ggml_tallocr_get_buffer(sched->tallocs[backend_index]); +enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + if (!sched->is_reset && !sched->is_alloc) { + ggml_backend_sched_reset(sched); + } + + if (!sched->is_alloc) { + if (!ggml_backend_sched_alloc_graph(sched, graph)) { + return GGML_STATUS_ALLOC_FAILED; + } + } + + return ggml_backend_sched_compute_splits(sched); } -void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) { - int backend_index = sched_backend_prio(sched, backend); +void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) { + for (int i = 0; i < sched->n_backends; i++) { + ggml_backend_synchronize(sched->backends[i]); + } +} + +void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) { + sched->callback_eval = callback; + sched->callback_eval_user_data = user_data; +} + +int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) { + return sched->n_splits; +} + +int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) { + return sched->n_copies; +} + +size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) { + int backend_index = ggml_backend_sched_backend_id(sched, backend); GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends); - node_allocr(node) = sched->tallocs[backend_index]; + + return ggml_gallocr_get_buffer_size(sched->galloc, backend_index); +} + +void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) { + int backend_index = ggml_backend_sched_backend_id(sched, backend); + GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends); + tensor_backend_id(node) = backend_index; +} + +ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) { + int backend_index = tensor_backend_id(node); + if (backend_index == -1) { + return NULL; + } + return sched->backends[backend_index]; +} + +// utils + +void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + GGML_ASSERT(tensor->buffer == NULL); + GGML_ASSERT(tensor->view_src != NULL); + GGML_ASSERT(tensor->view_src->buffer != NULL); + GGML_ASSERT(tensor->view_src->data != NULL); + + tensor->buffer = buffer; + tensor->data = (char *)tensor->view_src->data + tensor->view_offs; + tensor->backend = tensor->view_src->backend; + ggml_backend_buffer_init_tensor(buffer, tensor); +} + +void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) { + GGML_ASSERT(tensor->buffer == NULL); + GGML_ASSERT(tensor->data == NULL); + GGML_ASSERT(tensor->view_src == NULL); + GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer)); + GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <= + (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer)); + + tensor->buffer = buffer; + tensor->data = addr; + ggml_backend_buffer_init_tensor(buffer, tensor); +} + +static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, + struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) { + + GGML_ASSERT(src != NULL); + GGML_ASSERT(src->data && "graph must be allocated"); + + size_t id = ggml_hash_insert(hash_set, src); + if (id == GGML_HASHTABLE_ALREADY_EXISTS) { + return node_copies[ggml_hash_find(hash_set, src)]; + } + + struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src); + if (src->view_src != NULL) { + dst->view_src = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src); + dst->view_offs = src->view_offs; + } + dst->op = src->op; + memcpy(dst->op_params, src->op_params, sizeof(dst->op_params)); + ggml_set_name(dst, src->name); + + // copy src + for (int i = 0; i < GGML_MAX_SRC; i++) { + struct ggml_tensor * s = src->src[i]; + if (s == NULL) { + continue; + } + dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s); + } + + node_copies[id] = dst; + return dst; +} + +static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) { + size_t id = ggml_hash_find(hash_set, src); + if (node_init[id]) { + return; + } + node_init[id] = true; + + struct ggml_tensor * dst = node_copies[id]; + if (dst->view_src != NULL) { + graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src); + ggml_backend_view_init(dst->view_src->buffer, dst); + } + else { + ggml_backend_tensor_copy(src, dst); + } + + // init src + for (int i = 0; i < GGML_MAX_SRC; i++) { + struct ggml_tensor * s = src->src[i]; + if (s == NULL) { + continue; + } + graph_copy_init_tensor(hash_set, node_copies, node_init, s); + } +} + +struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) { + struct ggml_hash_set hash_set = { + /* .size = */ graph->visited_hash_table.size, + /* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT + }; + struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT + bool * node_init = calloc(sizeof(node_init[0]), hash_set.size); + + struct ggml_init_params params = { + /* .mem_size = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false), + /* .mem_buffer = */ NULL, + /* .no_alloc = */ true + }; + + struct ggml_context * ctx_allocated = ggml_init(params); + struct ggml_context * ctx_unallocated = ggml_init(params); + + if (ctx_allocated == NULL || ctx_unallocated == NULL) { + fprintf(stderr, "failed to allocate context for graph copy\n"); + free(hash_set.keys); + free(node_copies); + free(node_init); + ggml_free(ctx_allocated); + ggml_free(ctx_unallocated); + return (struct ggml_backend_graph_copy) { + /* .buffer = */ NULL, + /* .ctx_allocated = */ NULL, + /* .ctx_unallocated = */ NULL, + /* .graph = */ NULL, + }; + } + + // dup nodes + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node); + } + + // allocate nodes + ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend); + if (buffer == NULL) { + fprintf(stderr, "failed to allocate buffer for graph copy\n"); + free(hash_set.keys); + free(node_copies); + free(node_init); + ggml_free(ctx_allocated); + ggml_free(ctx_unallocated); + return (struct ggml_backend_graph_copy) { + /* .buffer = */ NULL, + /* .ctx_allocated = */ NULL, + /* .ctx_unallocated = */ NULL, + /* .graph = */ NULL, + }; + } + + //printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024); + + // copy data and init views + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + graph_copy_init_tensor(hash_set, node_copies, node_init, node); + } + + // build graph copy + struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false); + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + struct ggml_tensor * node_copy = node_copies[ggml_hash_find(hash_set, node)]; + graph_copy->nodes[i] = node_copy; + } + graph_copy->n_nodes = graph->n_nodes; + + free(hash_set.keys); + free(node_copies); + free(node_init); + + return (struct ggml_backend_graph_copy) { + /* .buffer = */ buffer, + /* .ctx_allocated = */ ctx_allocated, + /* .ctx_unallocated = */ ctx_unallocated, + /* .graph = */ graph_copy, + }; +} + +void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) { + ggml_backend_buffer_free(copy.buffer); + ggml_free(copy.ctx_allocated); + ggml_free(copy.ctx_unallocated); +} + +bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) { + struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph); + if (copy.buffer == NULL) { + return false; + } + + struct ggml_cgraph * g1 = graph; + struct ggml_cgraph * g2 = copy.graph; + + assert(g1->n_nodes == g2->n_nodes); + + for (int i = 0; i < g1->n_nodes; i++) { + //printf("eval %d/%d\n", i, g1->n_nodes); + struct ggml_tensor * t1 = g1->nodes[i]; + struct ggml_tensor * t2 = g2->nodes[i]; + + assert(t1->op == t2->op && ggml_are_same_layout(t1, t2)); + + struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1); + struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1); + + ggml_backend_graph_compute(backend1, &g1v); + ggml_backend_graph_compute(backend2, &g2v); + + if (ggml_is_view_op(t1->op)) { + continue; + } + + // compare results, calculate rms etc + if (!callback(i, t1, t2, user_data)) { + break; + } + } + + ggml_backend_graph_copy_free(copy); + + return true; } diff --git a/bindings/ruby/ext/ggml-backend.h b/bindings/ruby/ext/ggml-backend.h index 793a0a9d..744b6a77 100644 --- a/bindings/ruby/ext/ggml-backend.h +++ b/bindings/ruby/ext/ggml-backend.h @@ -7,69 +7,123 @@ extern "C" { #endif + typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t; + typedef struct ggml_backend_buffer * ggml_backend_buffer_t; + typedef struct ggml_backend_event * ggml_backend_event_t; + typedef struct ggml_backend * ggml_backend_t; + typedef void * ggml_backend_graph_plan_t; + // // Backend buffer // - struct ggml_backend_buffer; - typedef struct ggml_backend_buffer * ggml_backend_buffer_t; + // buffer type + GGML_API const char * ggml_backend_buft_name (ggml_backend_buffer_type_t buft); + GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size); + GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft); + GGML_API size_t ggml_backend_buft_get_max_size (ggml_backend_buffer_type_t buft); + GGML_API GGML_CALL size_t ggml_backend_buft_get_alloc_size (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); + GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend); + GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft); - // backend buffer functions - GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); - GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - GGML_API void ggml_backend_buffer_free_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + // buffer + enum ggml_backend_buffer_usage { + GGML_BACKEND_BUFFER_USAGE_ANY = 0, + GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1, + }; + + GGML_API const char * ggml_backend_buffer_name (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); + GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); + GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); + GGML_API GGML_CALL void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); + GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer); + GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value); + GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); + GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_get_type (ggml_backend_buffer_t buffer); + GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer); // // Backend // - struct ggml_backend; - typedef struct ggml_backend * ggml_backend_t; - typedef void * ggml_backend_graph_plan_t; - - GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor); - + GGML_API ggml_guid_t ggml_backend_guid(ggml_backend_t backend); GGML_API const char * ggml_backend_name(ggml_backend_t backend); GGML_API void ggml_backend_free(ggml_backend_t backend); - GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size); + GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend); + GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size); + GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend); + GGML_API size_t ggml_backend_get_max_size(ggml_backend_t backend); - GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend); + GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - GGML_API void ggml_backend_tensor_set_async( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - - GGML_API void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + GGML_API GGML_CALL void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); GGML_API void ggml_backend_synchronize(ggml_backend_t backend); - GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph); + GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph); + GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API bool ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); - GGML_API bool ggml_backend_supports_op (ggml_backend_t backend, const struct ggml_tensor * op); + GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan); + GGML_API enum ggml_status ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); + GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph); + GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op); + GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op); // tensor copy between different backends GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst); + // asynchronous copy + // the copy is performed after all the currently queued operations in backend_src + // backend_dst will wait for the copy to complete before performing other operations + // automatic fallback to sync copy if async is not supported + GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst); + + // events + GGML_API ggml_backend_event_t ggml_backend_event_new (ggml_backend_t backend); + GGML_API void ggml_backend_event_free (ggml_backend_event_t event); + GGML_API void ggml_backend_event_record (ggml_backend_event_t event); + GGML_API void ggml_backend_event_synchronize(ggml_backend_event_t event); + GGML_API void ggml_backend_event_wait (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event + // // CPU backend // GGML_API ggml_backend_t ggml_backend_cpu_init(void); - GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend); - GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads); + GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend); + GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads); + GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data); // Create a backend buffer from an existing pointer - GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size); + GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size); + GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void); + +#ifdef GGML_USE_CPU_HBM + GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void); +#endif + + // + // Backend registry + // + + // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way + + GGML_API size_t ggml_backend_reg_get_count(void); + GGML_API size_t ggml_backend_reg_find_by_name(const char * name); + GGML_API ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params] + GGML_API const char * ggml_backend_reg_get_name(size_t i); + GGML_API ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific + GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i); + GGML_API ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size); // // Backend scheduler @@ -83,53 +137,96 @@ extern "C" { /* Example usage: - sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends); - // sched is initialized with measure allocators and cannot be used until allocated with a measure graph + // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned + // preferrably to run on the same backend as the buffer + ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS); - // initialize buffers from a measure graph - measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed + sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false); - // in build_graph: - build_graph(...) { - // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer) - alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu); - ggml_allocr_alloc(alloc_cpu, tensor); + // initialize buffers from a max size graph (optional) + reserve_graph = build_graph(sched, max_batch_size); - // manually assigning nodes to a backend (optional, shouldn't be needed in most cases) - struct ggml_tensor * node = ggml_mul_mat(ctx, ...); - ggml_backend_sched_set_node_backend(sched, node, backend_gpu); - } + // manually assign nodes to a backend (optional, should not be needed in most cases) + struct ggml_tensor * node = ggml_mul_mat(ctx, ...); + ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu); - // allocate backend buffers from measure graph - ggml_backend_sched_init_measure(sched, measure_graph); - - // the scheduler is now ready to compute graphs + ggml_backend_sched_reserve(sched, reserve_graph); // compute graph = build_graph(sched); ggml_backend_sched_graph_compute(sched, graph); + + // if there are graph inputs: + ggml_backend_sched_reset(sched); + ggml_backend_sched_alloc_graph(sched, graph); + ggml_backend_tensor_set(input_tensor, ...); + ggml_backend_sched_graph_compute(sched, graph); + } */ struct ggml_backend_sched; typedef struct ggml_backend_sched * ggml_backend_sched_t; - // Initialize a backend scheduler - GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends); + // when ask == true, the scheduler wants to know if the user wants to observe this node + // this allows the scheduler to batch nodes together in order to evaluate them in a single call + // + // when ask == false, the scheduler is passing the node tensor to the user for observation + // if the user returns false, the scheduler will cancel the graph compute + // + typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); - GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); + // Initialize a backend scheduler + GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel); + GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); // Initialize backend buffers from a measure graph - GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); + GGML_API bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); - GGML_API ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend); - GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend); + // Get the number of splits of the last graph + GGML_API int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched); + GGML_API int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched); - GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend); + GGML_API size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend); + + GGML_API void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend); + GGML_API ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node); + + // Allocate and compute graph on the backend scheduler + GGML_API bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); + GGML_API enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph); + GGML_API enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph); + GGML_API void ggml_backend_sched_synchronize(ggml_backend_sched_t sched); + + // Reset all assignments and allocators - must be called before changing the node backends + GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched); + + // Set a callback to be called for each resulting node during graph compute + GGML_API void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data); + + // + // Utils + // + + struct ggml_backend_graph_copy { + ggml_backend_buffer_t buffer; + struct ggml_context * ctx_allocated; + struct ggml_context * ctx_unallocated; + struct ggml_cgraph * graph; + }; + + // Copy a graph to a different backend + GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph); + GGML_API void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy); + + typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data); + + // Compare the output of two backends + GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data); + + // Tensor initialization + GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr); + GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); - // Allocate a graph on the backend scheduler - GGML_API void ggml_backend_sched_graph_compute( - ggml_backend_sched_t sched, - struct ggml_cgraph * graph); #ifdef __cplusplus } diff --git a/bindings/ruby/ext/ggml-common.h b/bindings/ruby/ext/ggml-common.h new file mode 100644 index 00000000..43c7978a --- /dev/null +++ b/bindings/ruby/ext/ggml-common.h @@ -0,0 +1,1853 @@ +#ifndef GGML_COMMON_DECL + +#if defined(GGML_COMMON_DECL_C) +#include + +typedef uint16_t ggml_half; +typedef uint32_t ggml_half2; + +#define GGML_COMMON_AGGR + +#define GGML_COMMON_DECL +#elif defined(GGML_COMMON_DECL_METAL) +#include + +typedef half ggml_half; +typedef half2 ggml_half2; + +#define GGML_COMMON_AGGR + +#define GGML_COMMON_DECL +#elif defined(GGML_COMMON_DECL_CUDA) +#include +#include + +typedef half ggml_half; +typedef half2 ggml_half2; + +#define GGML_COMMON_AGGR data + +#define GGML_COMMON_DECL +#elif defined(GGML_COMMON_DECL_HIP) +#include +#include + +typedef half ggml_half; +typedef half2 ggml_half2; + +#define GGML_COMMON_AGGR data + +#define GGML_COMMON_DECL +#elif defined(GGML_COMMON_DECL_SYCL) +#include +#include + +typedef sycl::half ggml_half; +typedef sycl::half2 ggml_half2; + +#define GGML_COMMON_AGGR data + +#define GGML_COMMON_DECL +#endif + +#if defined(GGML_COMMON_DECL) + +#ifndef __cplusplus +#ifndef static_assert +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L) +#define static_assert(cond, msg) _Static_assert(cond, msg) +#else +#define static_assert(cond, msg) struct global_scope_noop_trick +#endif +#endif +#endif // __cplusplus + +// QK = number of values after dequantization +// QK_K = super-block size + +#ifdef GGML_QKK_64 +#define QK_K 64 +#define K_SCALE_SIZE 4 +#else +#define QK_K 256 +#define K_SCALE_SIZE 12 +#endif // GGML_QKK_64 + +#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL) +// QR = QK / number of values before dequantization +// QI = number of 32 bit integers before dequantization + +#define QI4_0 (QK4_0 / (4 * QR4_0)) +#define QR4_0 2 + +#define QI4_1 (QK4_1 / (4 * QR4_1)) +#define QR4_1 2 + +#define QI5_0 (QK5_0 / (4 * QR5_0)) +#define QR5_0 2 + +#define QI5_1 (QK5_1 / (4 * QR5_1)) +#define QR5_1 2 + +#define QI8_0 (QK8_0 / (4 * QR8_0)) +#define QR8_0 1 + +#define QI8_1 (QK8_1 / (4 * QR8_1)) +#define QR8_1 1 + +#define QI2_K (QK_K / (4*QR2_K)) +#define QR2_K 4 + +#define QI3_K (QK_K / (4*QR3_K)) +#define QR3_K 4 + +#define QI4_K (QK_K / (4*QR4_K)) +#define QR4_K 2 + +#define QI5_K (QK_K / (4*QR5_K)) +#define QR5_K 2 + +#define QI6_K (QK_K / (4*QR6_K)) +#define QR6_K 2 + +#define QI2_XXS (QK_K / (4*QR2_XXS)) +#define QR2_XXS 8 + +#define QI2_XS (QK_K / (4*QR2_XS)) +#define QR2_XS 8 + +#define QI2_S (QK_K / (4*QR2_S)) +#define QR2_S 8 + +#define QI3_XXS (QK_K / (4*QR3_XXS)) +#define QR3_XXS 8 + +#define QI3_XS (QK_K / (4*QR3_XS)) +#define QR3_XS 8 + +#define QI1_S (QK_K / (4*QR1_S)) +#define QR1_S 8 + +#define QI4_NL (QK4_NL / (4*QR4_NL)) +#define QR4_NL 2 + +#if QK_K == 64 +#define QI4_XS QI4_NL +#define QR4_XS QR4_NL +#else +#define QI4_XS (QK_K / (4*QR4_XS)) +#define QR4_XS 8 +#endif + +#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP + +#define QK4_0 32 +typedef struct { + ggml_half d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +typedef struct { + union { + struct { + ggml_half d; // delta + ggml_half m; // min + } GGML_COMMON_AGGR; + ggml_half2 dm; + }; + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding"); + +#define QK5_0 32 +typedef struct { + ggml_half d; // delta + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_0 / 2]; // nibbles / quants +} block_q5_0; +static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); + +#define QK5_1 32 +typedef struct { + union { + struct { + ggml_half d; // delta + ggml_half m; // min + } GGML_COMMON_AGGR; + ggml_half2 dm; + }; + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_1 / 2]; // nibbles / quants +} block_q5_1; +static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); + +#define QK8_0 32 +typedef struct { + ggml_half d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding"); + +#define QK8_1 32 +typedef struct { + union { + struct { + ggml_half d; // delta + ggml_half s; // d * sum(qs[i]) + } GGML_COMMON_AGGR; + ggml_half2 ds; + }; + int8_t qs[QK8_1]; // quants +} block_q8_1; +static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding"); + +// +// Super-block quantization structures +// + +// 2-bit quantization +// weight is represented as x = a * q + b +// 16 blocks of 16 elements each +// Effectively 2.625 bits per weight +typedef struct { + uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits + uint8_t qs[QK_K/4]; // quants + union { + struct { + ggml_half d; // super-block scale for quantized scales + ggml_half dmin; // super-block scale for quantized mins + } GGML_COMMON_AGGR; + ggml_half2 dm; + }; +} block_q2_K; +static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding"); + +// 3-bit quantization +// weight is represented as x = a * q +// 16 blocks of 16 elements each +// Effectively 3.4375 bits per weight +#ifdef GGML_QKK_64 +typedef struct { + uint8_t hmask[QK_K/8]; // quants - high bit + uint8_t qs[QK_K/4]; // quants - low 2 bits + uint8_t scales[2]; + ggml_half d; // super-block scale +} block_q3_K; +static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding"); +#else +typedef struct { + uint8_t hmask[QK_K/8]; // quants - high bit + uint8_t qs[QK_K/4]; // quants - low 2 bits + uint8_t scales[12]; // scales, quantized with 6 bits + ggml_half d; // super-block scale +} block_q3_K; +static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding"); +#endif + +// 4-bit quantization +// 8 blocks of 32 elements each +// weight is represented as x = a * q + b +// Effectively 4.5 bits per weight +#ifdef GGML_QKK_64 +typedef struct { + ggml_half d[2]; // super-block scales/mins + uint8_t scales[2]; // 4-bit block scales/mins + uint8_t qs[QK_K/2]; // 4--bit quants +} block_q4_K; +static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding"); +#else +typedef struct { + union { + struct { + ggml_half d; // super-block scale for quantized scales + ggml_half dmin; // super-block scale for quantized mins + } GGML_COMMON_AGGR; + ggml_half2 dm; + }; + uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits + uint8_t qs[QK_K/2]; // 4--bit quants +} block_q4_K; +static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding"); +#endif + +// 5-bit quantization +// 8 blocks of 32 elements each +// weight is represented as x = a * q + b +// Effectively 5.5 bits per weight +#ifdef GGML_QKK_64 +typedef struct { + ggml_half d; // super-block scale + int8_t scales[QK_K/16]; // 8-bit block scales + uint8_t qh[QK_K/8]; // quants, high bit + uint8_t qs[QK_K/2]; // quants, low 4 bits +} block_q5_K; +static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding"); +#else +typedef struct { + union { + struct { + ggml_half d; // super-block scale for quantized scales + ggml_half dmin; // super-block scale for quantized mins + } GGML_COMMON_AGGR; + ggml_half2 dm; + }; + uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits + uint8_t qh[QK_K/8]; // quants, high bit + uint8_t qs[QK_K/2]; // quants, low 4 bits +} block_q5_K; +static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding"); +#endif + +// 6-bit quantization +// weight is represented as x = a * q +// 16 blocks of 16 elements each +// Effectively 6.5625 bits per weight +typedef struct { + uint8_t ql[QK_K/2]; // quants, lower 4 bits + uint8_t qh[QK_K/4]; // quants, upper 2 bits + int8_t scales[QK_K/16]; // scales, quantized with 8 bits + ggml_half d; // super-block scale +} block_q6_K; +static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding"); + +// This is only used for intermediate quantization and dot products +typedef struct { + float d; // delta + int8_t qs[QK_K]; // quants + int16_t bsums[QK_K/16]; // sum of quants in groups of 16 +} block_q8_K; +static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding"); + +// (Almost) "true" 2-bit quantization. +// Due to the need to use blocks as per ggml design, it ends up using +// 2.0625 bpw because of the 16-bit scale for each block of 256. +typedef struct { + ggml_half d; + uint16_t qs[QK_K/8]; +} block_iq2_xxs; +static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding"); + +// 2.3125 bpw quants +typedef struct { + ggml_half d; + uint16_t qs[QK_K/8]; + uint8_t scales[QK_K/32]; +} block_iq2_xs; +static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding"); + +// 2.5625 bpw quants +typedef struct { + ggml_half d; + uint8_t qs[QK_K/4]; + uint8_t qh[QK_K/32]; + uint8_t scales[QK_K/32]; +} block_iq2_s; +static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding"); + +// (Almost) "true" 3-bit quantization. +// Due to the need to use blocks as per ggml design, it ends up using +// 3.0625 bpw because of the 16-bit scale for each block of 256. +typedef struct { + ggml_half d; + uint8_t qs[3*QK_K/8]; +} block_iq3_xxs; +static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding"); + +// 3.4375 bpw +#if QK_K == 64 +#define IQ3S_N_SCALE 2 +#else +#define IQ3S_N_SCALE QK_K/64 +#endif +typedef struct { + ggml_half d; + uint8_t qs[QK_K/4]; + uint8_t qh[QK_K/32]; + uint8_t signs[QK_K/8]; + uint8_t scales[IQ3S_N_SCALE]; +} block_iq3_s; +static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding"); + +typedef struct { + ggml_half d; + uint8_t qs[QK_K/8]; + uint16_t qh[QK_K/32]; +} block_iq1_s; +static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding"); + +// 1.75 bpw +typedef struct { + uint8_t qs[QK_K/8]; // grid index, low 8 bits + uint8_t qh[QK_K/16]; // grid index, high 3 bits + grid shift bit (for two groups of 8) +#if QK_K == 64 + ggml_half d; +#endif + uint8_t scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64) +} block_iq1_m; +#if QK_K == 64 +static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding"); +#else +static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding"); +#endif + +// Used by IQ1_M quants +typedef union { + ggml_half f16; + uint16_t u16; +} iq1m_scale_t; + +// Non-linear quants +#define QK4_NL 32 +typedef struct { + ggml_half d; + uint8_t qs[QK4_NL/2]; +} block_iq4_nl; +static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding"); + +#if QK_K == 64 +#define block_iq4_xs block_iq4_nl +#else +typedef struct { + ggml_half d; + uint16_t scales_h; + uint8_t scales_l[QK_K/64]; + uint8_t qs[QK_K/2]; +} block_iq4_xs; +static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding"); +#endif + +#endif // GGML_COMMON_DECL +#endif // GGML_COMMON_DECL + +//////////////////////////////////////////////////////////////////////////////// + +#ifndef GGML_COMMON_IMPL + +#if defined(GGML_COMMON_IMPL_C) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_METAL) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_SYCL) + +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#endif + +#if defined(GGML_COMMON_IMPL) + +GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8) + 1, 2, 4, 8, 16, 32, 64, 128 +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128) + 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, + 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, + 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, + 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, + 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, + 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, + 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, + 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, +GGML_TABLE_END() + +//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +GGML_TABLE_BEGIN(uint64_t, ksigns64, 128) + 0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff, + 0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff, + 0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff, + 0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff, + 0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff, + 0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff, + 0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff, + 0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff, + 0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff, + 0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff, + 0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff, + 0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff, + 0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff, + 0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff, + 0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff, + 0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff, + 0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff, + 0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff, + 0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff, + 0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff, + 0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff, + 0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff, + 0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff, + 0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff, + 0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff, + 0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff, + 0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff, + 0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff, + 0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff, + 0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff, + 0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff, + 0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff, +GGML_TABLE_END() +//#endif + + +GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256) + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, + 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, + 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819, + 0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b, + 0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, + 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08, + 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b, + 0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, + 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08, + 0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, + 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, + 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808, + 0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808, + 0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, + 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08, + 0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, + 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819, + 0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808, + 0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, + 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908, + 0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808, + 0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, + 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819, + 0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, + 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908, + 0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19, + 0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, + 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b, + 0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808, + 0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, + 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08, + 0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08, + 0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, + 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819, + 0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808, + 0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, + 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19, + 0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819, + 0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, + 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b, + 0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08, + 0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, + 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908, + 0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b, + 0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, + 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08, + 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08, + 0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, + 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b, + 0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b, + 0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, + 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819, + 0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808, + 0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, + 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b, + 0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808, + 0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, + 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b, + 0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808, + 0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, + 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512) + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, + 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, + 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, + 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, + 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, + 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, + 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, + 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, + 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, + 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, + 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, + 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, + 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, + 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, + 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, + 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, + 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, + 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, + 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, + 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, + 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, + 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, + 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, + 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, + 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, + 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, + 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, + 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, + 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, + 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, + 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, + 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, + 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, + 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, + 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, + 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, + 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, + 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, + 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, + 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, + 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, + 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, + 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, + 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, + 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, + 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, + 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, + 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, + 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, + 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, + 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, + 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, + 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, + 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, + 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, + 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, + 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, + 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, + 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, + 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, + 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, + 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, + 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, + 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, + 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, + 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, + 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, + 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, + 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, + 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, + 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, + 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, + 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, + 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, + 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, + 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, + 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, + 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, + 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, + 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, + 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, + 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, + 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, + 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, + 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, + 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, + 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, + 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, + 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, + 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, + 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, + 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, + 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, + 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, + 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, + 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, + 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, + 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, + 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, + 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, + 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, + 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, + 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, + 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, + 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, + 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, + 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, + 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, + 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, + 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, + 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, + 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, + 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024) + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b, + 0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919, + 0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808, + 0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908, + 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b, + 0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908, + 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08, + 0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19, + 0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819, + 0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919, + 0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b, + 0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, + 0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908, + 0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908, + 0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, + 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919, + 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b, + 0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, + 0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908, + 0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b, + 0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b, + 0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08, + 0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, + 0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819, + 0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808, + 0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908, + 0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b, + 0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908, + 0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08, + 0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808, + 0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08, + 0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819, + 0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908, + 0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919, + 0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b, + 0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919, + 0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808, + 0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819, + 0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919, + 0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919, + 0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808, + 0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819, + 0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b, + 0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908, + 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, + 0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, + 0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919, + 0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b, + 0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919, + 0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b, + 0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819, + 0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919, + 0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908, + 0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b, + 0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908, + 0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b, + 0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908, + 0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08, + 0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908, + 0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819, + 0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819, + 0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808, + 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, + 0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19, + 0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819, + 0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808, + 0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819, + 0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919, + 0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808, + 0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19, + 0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08, + 0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b, + 0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908, + 0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808, + 0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819, + 0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908, + 0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819, + 0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808, + 0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808, + 0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819, + 0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908, + 0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08, + 0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819, + 0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b, + 0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b, + 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08, + 0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19, + 0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819, + 0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919, + 0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908, + 0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808, + 0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808, + 0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908, + 0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808, + 0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08, + 0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08, + 0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908, + 0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919, + 0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808, + 0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819, + 0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908, + 0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08, + 0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819, + 0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808, + 0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808, + 0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819, + 0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808, + 0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908, + 0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b, + 0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, + 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, + 0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b, + 0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808, + 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b, + 0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19, + 0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819, + 0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08, + 0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b, + 0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, + 0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b, + 0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b, + 0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919, + 0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808, + 0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819, + 0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908, + 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08, + 0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08, + 0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819, + 0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919, + 0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908, + 0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b, + 0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908, + 0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b, + 0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908, + 0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08, + 0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819, + 0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808, + 0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819, + 0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919, + 0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808, + 0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808, + 0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08, + 0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819, + 0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919, + 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808, + 0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819, + 0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919, + 0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808, + 0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b, + 0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908, + 0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808, + 0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908, + 0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b, + 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908, + 0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b, + 0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908, + 0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b, + 0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908, + 0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08, + 0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908, + 0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b, + 0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908, + 0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08, + 0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819, + 0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919, + 0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808, + 0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19, + 0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b, + 0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919, + 0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808, + 0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819, + 0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908, + 0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919, + 0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808, + 0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808, + 0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b, + 0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919, + 0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808, + 0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b, + 0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808, + 0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919, + 0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b, + 0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08, + 0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919, + 0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808, + 0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b, + 0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908, + 0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808, + 0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808, + 0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808, + 0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908, + 0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808, + 0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808, + 0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b, + 0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908, + 0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808, + 0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808, + 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, + 0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b, + 0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808, + 0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819, + 0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b, + 0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908, + 0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08, + 0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908, + 0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919, + 0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819, + 0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908, + 0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808, + 0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819, + 0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908, + 0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919, + 0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808, + 0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808, + 0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808, + 0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919, + 0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908, + 0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908, + 0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08, + 0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819, + 0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b, + 0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808, + 0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819, + 0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908, + 0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819, + 0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808, + 0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808, + 0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b, + 0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908, + 0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808, + 0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908, + 0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819, + 0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819, + 0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808, + 0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b, + 0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b, + 0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819, + 0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b, + 0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b, + 0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b, + 0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819, + 0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19, + 0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819, + 0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908, + 0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808, + 0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256) + 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414, + 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, + 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, + 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e, + 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, + 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, + 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34, + 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, + 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, + 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04, + 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, + 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, + 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434, + 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, + 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, + 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24, + 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, + 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, + 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c, + 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, + 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, + 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e, + 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, + 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, + 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c, + 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, + 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, + 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c, + 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, + 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, + 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c, + 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512) + 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, + 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, + 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, + 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, + 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, + 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, + 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, + 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, + 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, + 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, + 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, + 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, + 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, + 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, + 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, + 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, + 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, + 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, + 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, + 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, + 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, + 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, + 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, + 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, + 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, + 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, + 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, + 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, + 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, + 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, + 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, + 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, + 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, + 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, + 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, + 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, + 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, + 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, + 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, + 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, + 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, + 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, + 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, + 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, + 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, + 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, + 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, + 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, + 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, + 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, + 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, + 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, + 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, + 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, + 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, + 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, + 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, + 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, + 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, + 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, + 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, + 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, + 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, + 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, +GGML_TABLE_END() + +#define NGRID_IQ1S 2048 +#define IQ1S_DELTA 0.125f +#define IQ1M_DELTA 0.125f +#if defined(GGML_COMMON_IMPL_C) +GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S) + 0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff, + 0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff, + 0xffffffffff01ff01, 0xffffffffff0101ff, 0xffffffffff010101, 0xffffffff00ff0000, + 0xffffffff0000ff00, 0xffffffff000000ff, 0xffffffff00000001, 0xffffffff00010000, + 0xffffffff01ffffff, 0xffffffff01ffff01, 0xffffffff01ff01ff, 0xffffffff01ff0101, + 0xffffffff01000000, 0xffffffff0101ffff, 0xffffffff0101ff01, 0xffffffff010101ff, + 0xffffffff01010101, 0xffffff00ffff00ff, 0xffffff00ffff0000, 0xffffff00ff00ff00, + 0xffffff00ff0000ff, 0xffffff00ff000001, 0xffffff00ff000100, 0xffffff00ff000101, + 0xffffff00ff010000, 0xffffff0000ffff00, 0xffffff0000ff0001, 0xffffff0000ff0100, + 0xffffff000000ff01, 0xffffff0000000000, 0xffffff0000000101, 0xffffff000001ff00, + 0xffffff00000100ff, 0xffffff0000010001, 0xffffff00000101ff, 0xffffff0001ff0000, + 0xffffff000100ff00, 0xffffff00010000ff, 0xffffff0001000001, 0xffffff0001010000, + 0xffffff01ffffffff, 0xffffff01ffffff01, 0xffffff01ffff01ff, 0xffffff01ffff0101, + 0xffffff01ff000000, 0xffffff01ff01ffff, 0xffffff01ff01ff01, 0xffffff01ff0101ff, + 0xffffff01ff010101, 0xffffff0100ff0000, 0xffffff010000ff00, 0xffffff0100000100, + 0xffffff01000100ff, 0xffffff0100010100, 0xffffff0101ffffff, 0xffffff0101ffff01, + 0xffffff0101ff01ff, 0xffffff0101ff0101, 0xffffff010100ff00, 0xffffff0101000000, + 0xffffff0101000100, 0xffffff010101ffff, 0xffffff010101ff01, 0xffffff01010101ff, + 0xffffff0101010101, 0xffff00ffff00ff00, 0xffff00ffff0000ff, 0xffff00ffff000001, + 0xffff00ffff010000, 0xffff00ff00ffff00, 0xffff00ff00ff0100, 0xffff00ff00000000, + 0xffff00ff00000101, 0xffff00ff000100ff, 0xffff00ff00010000, 0xffff00ff0100ff00, + 0xffff00ff01000100, 0xffff00ff01010000, 0xffff0000ffffff00, 0xffff0000ffff00ff, + 0xffff0000ffff0000, 0xffff0000ffff0001, 0xffff0000ff000000, 0xffff0000ff0001ff, + 0xffff0000ff000101, 0xffff0000ff010100, 0xffff000000ffffff, 0xffff000000ff0000, + 0xffff000000ff0101, 0xffff00000000ffff, 0xffff00000000ff00, 0xffff0000000000ff, + 0xffff000000000000, 0xffff000000000001, 0xffff000000000100, 0xffff00000001ffff, + 0xffff00000001ff01, 0xffff000000010000, 0xffff0000000101ff, 0xffff000000010101, + 0xffff000001ffff00, 0xffff00000100ff00, 0xffff000001000000, 0xffff0000010001ff, + 0xffff000001000101, 0xffff00000101ff00, 0xffff0000010100ff, 0xffff000001010000, + 0xffff000001010001, 0xffff000001010100, 0xffff0001ff0000ff, 0xffff0001ff000100, + 0xffff000100ffff00, 0xffff000100ff00ff, 0xffff00010000ffff, 0xffff00010000ff01, + 0xffff000100000000, 0xffff0001000001ff, 0xffff00010001ffff, 0xffff00010001ff00, + 0xffff000100010001, 0xffff000100010100, 0xffff000101ff0000, 0xffff00010100ff00, + 0xffff0001010000ff, 0xffff000101000100, 0xffff01ffffffffff, 0xffff01ffffffff01, + 0xffff01ffffff01ff, 0xffff01ffffff0101, 0xffff01ffff000000, 0xffff01ffff01ffff, + 0xffff01ffff01ff01, 0xffff01ffff0101ff, 0xffff01ffff010101, 0xffff01ff00ff0000, + 0xffff01ff0000ff00, 0xffff01ff00000001, 0xffff01ff00010000, 0xffff01ff01ffffff, + 0xffff01ff01ffff01, 0xffff01ff01ff01ff, 0xffff01ff01ff0101, 0xffff01ff01000000, + 0xffff01ff0101ffff, 0xffff01ff0101ff01, 0xffff01ff010101ff, 0xffff01ff01010101, + 0xffff0100ffff0000, 0xffff0100ff00ff00, 0xffff0100ff0000ff, 0xffff0100ff000100, + 0xffff0100ff0100ff, 0xffff0100ff010000, 0xffff010000ffff00, 0xffff01000000ffff, + 0xffff01000000ff00, 0xffff010000000000, 0xffff01000001ff00, 0xffff0100000100ff, + 0xffff010000010100, 0xffff01000100ff00, 0xffff0100010000ff, 0xffff010001000001, + 0xffff010001000100, 0xffff010001010000, 0xffff0101ffffffff, 0xffff0101ffffff01, + 0xffff0101ffff01ff, 0xffff0101ffff0101, 0xffff0101ff000000, 0xffff0101ff01ffff, + 0xffff0101ff01ff01, 0xffff0101ff0101ff, 0xffff0101ff010101, 0xffff010100ff0000, + 0xffff01010000ff00, 0xffff010100000100, 0xffff01010001ff00, 0xffff010100010000, + 0xffff010101ffffff, 0xffff010101ffff01, 0xffff010101ff0000, 0xffff010101ff01ff, + 0xffff010101ff0101, 0xffff010101000000, 0xffff01010101ffff, 0xffff01010101ff01, + 0xffff0101010101ff, 0xffff010101010101, 0xff00ffffff00ffff, 0xff00ffffff00ff00, + 0xff00ffffff0000ff, 0xff00ffffff000100, 0xff00ffffff0100ff, 0xff00ffffff010000, + 0xff00ffff00ffff00, 0xff00ffff00ff00ff, 0xff00ffff0000ffff, 0xff00ffff00000000, + 0xff00ffff000001ff, 0xff00ffff0001ff00, 0xff00ffff000100ff, 0xff00ffff00010000, + 0xff00ffff00010100, 0xff00ffff0100ff00, 0xff00ffff010000ff, 0xff00ffff01000001, + 0xff00ffff0101ff00, 0xff00ffff01010000, 0xff00ff00ffffff00, 0xff00ff00ffff00ff, + 0xff00ff00ffff0001, 0xff00ff00ffff0100, 0xff00ff00ff00ffff, 0xff00ff00ff00ff01, + 0xff00ff00ff000000, 0xff00ff00ff0001ff, 0xff00ff00ff01ff00, 0xff00ff00ff0100ff, + 0xff00ff00ff010100, 0xff00ff0000ff0000, 0xff00ff0000ff0101, 0xff00ff000000ffff, + 0xff00ff000000ff00, 0xff00ff000000ff01, 0xff00ff00000000ff, 0xff00ff0000000000, + 0xff00ff0000000001, 0xff00ff0000000100, 0xff00ff000001ffff, 0xff00ff0000010000, + 0xff00ff0001ff00ff, 0xff00ff000100ff01, 0xff00ff0001000000, 0xff00ff000101ff00, + 0xff00ff00010100ff, 0xff00ff01ff00ff00, 0xff00ff01ff0000ff, 0xff00ff01ff000001, + 0xff00ff01ff010000, 0xff00ff0100ffffff, 0xff00ff0100ff0001, 0xff00ff0100ff0100, + 0xff00ff010000ff01, 0xff00ff0100000000, 0xff00ff01000001ff, 0xff00ff0100000101, + 0xff00ff01000100ff, 0xff00ff0100010001, 0xff00ff0101ff0000, 0xff00ff010100ff00, + 0xff00ff01010000ff, 0xff00ff0101000001, 0xff00ff0101010000, 0xff0000ffffffff00, + 0xff0000ffffff0001, 0xff0000ffffff0100, 0xff0000ffff0000ff, 0xff0000ffff000000, + 0xff0000ffff0001ff, 0xff0000ffff000100, 0xff0000ffff01ff00, 0xff0000ffff010001, + 0xff0000ff00ffff00, 0xff0000ff00ff0000, 0xff0000ff00ff0001, 0xff0000ff00ff01ff, + 0xff0000ff00ff0101, 0xff0000ff0000ff00, 0xff0000ff000000ff, 0xff0000ff00000000, + 0xff0000ff00000001, 0xff0000ff00000100, 0xff0000ff0001ff01, 0xff0000ff00010000, + 0xff0000ff000101ff, 0xff0000ff01ff00ff, 0xff0000ff01ff0100, 0xff0000ff0100ffff, + 0xff0000ff010000ff, 0xff0000ff01000000, 0xff0000ff010001ff, 0xff0000ff01000100, + 0xff0000ff01000101, 0xff0000ff0101ff00, 0xff0000ff010100ff, 0xff0000ff01010000, + 0xff0000ff01010100, 0xff000000ffffff01, 0xff000000ffff0000, 0xff000000ffff0101, + 0xff000000ff00ff00, 0xff000000ff0000ff, 0xff000000ff000000, 0xff000000ff000001, + 0xff000000ff000100, 0xff000000ff01ffff, 0xff000000ff01ff01, 0xff000000ff010000, + 0xff000000ff0101ff, 0xff000000ff010101, 0xff00000000ffff00, 0xff00000000ff00ff, + 0xff00000000ff0000, 0xff00000000ff0001, 0xff0000000000ff00, 0xff0000000000ff01, + 0xff000000000000ff, 0xff00000000000000, 0xff00000000000001, 0xff00000000000100, + 0xff00000000000101, 0xff0000000001ff00, 0xff000000000100ff, 0xff00000000010000, + 0xff00000000010001, 0xff00000000010100, 0xff00000001ffffff, 0xff00000001ffff01, + 0xff00000001ff00ff, 0xff00000001ff0000, 0xff00000001ff01ff, 0xff00000001ff0101, + 0xff0000000100ffff, 0xff0000000100ff00, 0xff000000010000ff, 0xff00000001000000, + 0xff00000001000001, 0xff00000001000100, 0xff00000001000101, 0xff0000000101ffff, + 0xff0000000101ff01, 0xff00000001010000, 0xff000001ffffff00, 0xff000001ffff00ff, + 0xff000001ffff0000, 0xff000001ffff0001, 0xff000001ff000000, 0xff000001ff000001, + 0xff000001ff0001ff, 0xff000001ff000101, 0xff000001ff01ff00, 0xff000001ff010001, + 0xff00000100ffffff, 0xff00000100ffff01, 0xff00000100ff00ff, 0xff00000100ff0000, + 0xff00000100ff01ff, 0xff00000100ff0101, 0xff0000010000ff00, 0xff00000100000000, + 0xff00000100000001, 0xff000001000001ff, 0xff00000100000100, 0xff0000010001ff00, + 0xff000001000100ff, 0xff00000100010000, 0xff000001000101ff, 0xff00000100010100, + 0xff00000100010101, 0xff00000101ff0001, 0xff00000101ff0101, 0xff0000010100ff01, + 0xff00000101000000, 0xff000001010100ff, 0xff00000101010100, 0xff0001ffff00ff00, + 0xff0001ffff000001, 0xff0001ffff010000, 0xff0001ff00ffff00, 0xff0001ff00ff00ff, + 0xff0001ff00ff0001, 0xff0001ff00ff0100, 0xff0001ff0000ffff, 0xff0001ff00000000, + 0xff0001ff000001ff, 0xff0001ff00000101, 0xff0001ff0001ffff, 0xff0001ff0001ff00, + 0xff0001ff000100ff, 0xff0001ff00010001, 0xff0001ff00010100, 0xff0001ff01ff0000, + 0xff0001ff0100ff00, 0xff0001ff010000ff, 0xff0001ff01010000, 0xff000100ff00ffff, + 0xff000100ff00ff01, 0xff000100ff000000, 0xff000100ff000101, 0xff000100ff01ff00, + 0xff000100ff010000, 0xff00010000ffff01, 0xff00010000ff00ff, 0xff00010000ff0000, + 0xff00010000ff01ff, 0xff0001000000ff00, 0xff000100000000ff, 0xff00010000000000, + 0xff00010000000001, 0xff00010000000100, 0xff00010000000101, 0xff0001000001ffff, + 0xff00010000010000, 0xff00010000010101, 0xff00010001ff0100, 0xff0001000100ff00, + 0xff0001000100ff01, 0xff00010001000000, 0xff000100010001ff, 0xff0001000101ff00, + 0xff00010001010001, 0xff00010001010100, 0xff000101ffff0100, 0xff000101ff000001, + 0xff000101ff0100ff, 0xff000101ff010001, 0xff00010100ff00ff, 0xff00010100ff0001, + 0xff00010100ff0100, 0xff0001010000ffff, 0xff0001010000ff01, 0xff00010100000000, + 0xff000101000001ff, 0xff0001010001ff00, 0xff00010100010001, 0xff00010100010100, + 0xff00010101ff0000, 0xff0001010100ff00, 0xff00010101000001, 0xff00010101000101, + 0xff01ffffffffffff, 0xff01ffffffffff01, 0xff01ffffffff01ff, 0xff01ffffffff0101, + 0xff01ffffff000000, 0xff01ffffff01ffff, 0xff01ffffff01ff01, 0xff01ffffff010000, + 0xff01ffffff0101ff, 0xff01ffffff010101, 0xff01ffff00ff0000, 0xff01ffff0000ff00, + 0xff01ffff00000100, 0xff01ffff0001ff00, 0xff01ffff00010000, 0xff01ffff01ffffff, + 0xff01ffff01ffff01, 0xff01ffff01ff01ff, 0xff01ffff01ff0101, 0xff01ffff01000000, + 0xff01ffff0101ffff, 0xff01ffff0101ff01, 0xff01ffff01010000, 0xff01ffff010101ff, + 0xff01ffff01010101, 0xff01ff00ffff0000, 0xff01ff00ff00ff00, 0xff01ff00ff0000ff, + 0xff01ff00ff000100, 0xff01ff00ff010000, 0xff01ff0000ffff01, 0xff01ff0000ff00ff, + 0xff01ff0000ff0100, 0xff01ff0000000000, 0xff01ff00000001ff, 0xff01ff0000000101, + 0xff01ff000001ff00, 0xff01ff00000100ff, 0xff01ff0000010000, 0xff01ff0000010001, + 0xff01ff0001ff0000, 0xff01ff000100ffff, 0xff01ff0001000001, 0xff01ff0001000100, + 0xff01ff0001010000, 0xff01ff01ffffff00, 0xff01ff01ffff01ff, 0xff01ff01ffff0101, + 0xff01ff01ff00ff00, 0xff01ff01ff000000, 0xff01ff01ff01ffff, 0xff01ff01ff01ff01, + 0xff01ff01ff0101ff, 0xff01ff01ff010101, 0xff01ff0100ff0000, 0xff01ff010000ff00, + 0xff01ff0100000001, 0xff01ff0100000100, 0xff01ff0100010000, 0xff01ff0101ffff00, + 0xff01ff0101ff01ff, 0xff01ff0101ff0101, 0xff01ff010100ff00, 0xff01ff0101000000, + 0xff01ff010101ffff, 0xff01ff010101ff01, 0xff01ff01010101ff, 0xff01ff0101010101, + 0xff0100ffffff0000, 0xff0100ffff0000ff, 0xff0100ffff000001, 0xff0100ffff000100, + 0xff0100ffff010000, 0xff0100ff00ff00ff, 0xff0100ff00ff0000, 0xff0100ff00ff0001, + 0xff0100ff00ff0100, 0xff0100ff0000ff01, 0xff0100ff00000000, 0xff0100ff000001ff, + 0xff0100ff00000101, 0xff0100ff00010001, 0xff0100ff01ff0000, 0xff0100ff0100ff00, + 0xff0100ff010000ff, 0xff0100ff01000100, 0xff0100ff0101ff00, 0xff0100ff01010000, + 0xff010000ffff0100, 0xff010000ff000000, 0xff010000ff01ff00, 0xff010000ff010100, + 0xff01000000ffffff, 0xff01000000ff0000, 0xff01000000ff01ff, 0xff0100000000ff00, + 0xff010000000000ff, 0xff01000000000000, 0xff01000000000100, 0xff0100000001ff01, + 0xff01000000010000, 0xff010000000101ff, 0xff01000001ff0100, 0xff0100000100ffff, + 0xff010000010000ff, 0xff01000001000000, 0xff010000010001ff, 0xff01000001000101, + 0xff0100000101ff00, 0xff010000010100ff, 0xff01000001010001, 0xff01000001010100, + 0xff010001ffff0000, 0xff010001ff00ffff, 0xff010001ff00ff01, 0xff010001ff000100, + 0xff010001ff010000, 0xff01000100ffff00, 0xff01000100ff0100, 0xff01000100000000, + 0xff0100010001ffff, 0xff0100010001ff00, 0xff01000100010100, 0xff01000101ff00ff, + 0xff01000101ff0001, 0xff0100010100ffff, 0xff01000101000101, 0xff0101ffffffffff, + 0xff0101ffffffff01, 0xff0101ffffff01ff, 0xff0101ffffff0101, 0xff0101ffff000000, + 0xff0101ffff01ffff, 0xff0101ffff01ff01, 0xff0101ffff0101ff, 0xff0101ffff010101, + 0xff0101ff00ff0000, 0xff0101ff0000ff00, 0xff0101ff000000ff, 0xff0101ff00010000, + 0xff0101ff01ffffff, 0xff0101ff01ffff01, 0xff0101ff01ff01ff, 0xff0101ff01ff0101, + 0xff0101ff0101ffff, 0xff0101ff0101ff01, 0xff0101ff010101ff, 0xff0101ff01010101, + 0xff010100ffff0100, 0xff010100ff00ff00, 0xff010100ff0000ff, 0xff010100ff000100, + 0xff010100ff010000, 0xff01010000ff0001, 0xff01010000ff0100, 0xff0101000000ff01, + 0xff01010000000000, 0xff0101000001ff00, 0xff010100000100ff, 0xff01010000010001, + 0xff01010000010100, 0xff01010001ff0000, 0xff0101000100ffff, 0xff01010001000001, + 0xff01010001000100, 0xff010100010100ff, 0xff01010001010000, 0xff010101ffffffff, + 0xff010101ffffff01, 0xff010101ffff01ff, 0xff010101ffff0101, 0xff010101ff01ffff, + 0xff010101ff01ff01, 0xff010101ff0101ff, 0xff010101ff010101, 0xff01010100ff0000, + 0xff0101010000ff00, 0xff01010100000001, 0xff01010100000100, 0xff01010100010000, + 0xff01010101ffffff, 0xff01010101ffff01, 0xff01010101ff01ff, 0xff01010101ff0101, + 0xff01010101000000, 0xff0101010101ffff, 0xff0101010101ff01, 0xff010101010101ff, + 0xff01010101010101, 0x00ffffffffff0000, 0x00ffffffff00ff00, 0x00ffffffff000001, + 0x00ffffffff010000, 0x00ffffff00ff0100, 0x00ffffff0000ff01, 0x00ffffff00000000, + 0x00ffffff000001ff, 0x00ffffff00000101, 0x00ffffff0001ff00, 0x00ffffff000100ff, + 0x00ffffff00010001, 0x00ffffff010000ff, 0x00ffffff01000100, 0x00ffffff0101ff00, + 0x00ffffff01010001, 0x00ffff00ffffffff, 0x00ffff00ffffff00, 0x00ffff00ffff00ff, + 0x00ffff00ffff0001, 0x00ffff00ffff0100, 0x00ffff00ff00ff01, 0x00ffff00ff000000, + 0x00ffff00ff000001, 0x00ffff00ff0001ff, 0x00ffff00ff000101, 0x00ffff00ff01ff00, + 0x00ffff00ff010001, 0x00ffff00ff010100, 0x00ffff0000ff0000, 0x00ffff0000ff01ff, + 0x00ffff0000ff0101, 0x00ffff000000ff00, 0x00ffff00000000ff, 0x00ffff0000000000, + 0x00ffff0000000001, 0x00ffff0000000100, 0x00ffff0000000101, 0x00ffff0000010000, + 0x00ffff00000101ff, 0x00ffff0000010101, 0x00ffff0001ffff00, 0x00ffff0001ff00ff, + 0x00ffff0001ff0001, 0x00ffff000100ffff, 0x00ffff000100ff01, 0x00ffff0001000000, + 0x00ffff000101ffff, 0x00ffff000101ff00, 0x00ffff000101ff01, 0x00ffff01ffff0000, + 0x00ffff01ff00ff00, 0x00ffff01ff0000ff, 0x00ffff01ff000001, 0x00ffff01ff010000, + 0x00ffff0100ffff00, 0x00ffff010000ff01, 0x00ffff0100000000, 0x00ffff0100000101, + 0x00ffff01000100ff, 0x00ffff0100010100, 0x00ffff0101ff0100, 0x00ffff01010000ff, + 0x00ffff0101010000, 0x00ff00ffffffff00, 0x00ff00ffff000000, 0x00ff00ffff000100, + 0x00ff00ffff010100, 0x00ff00ff00ff0000, 0x00ff00ff00ff01ff, 0x00ff00ff00ff0101, + 0x00ff00ff0000ff00, 0x00ff00ff000000ff, 0x00ff00ff00000000, 0x00ff00ff00000001, + 0x00ff00ff0001ff00, 0x00ff00ff0001ff01, 0x00ff00ff00010000, 0x00ff00ff000101ff, + 0x00ff00ff00010101, 0x00ff00ff01ffff00, 0x00ff00ff01ff0001, 0x00ff00ff01ff0100, + 0x00ff00ff0100ffff, 0x00ff00ff0100ff01, 0x00ff00ff01000000, 0x00ff00ff0101ffff, + 0x00ff00ff0101ff00, 0x00ff00ff01010100, 0x00ff0000ffffff00, 0x00ff0000ffffff01, + 0x00ff0000ffff0000, 0x00ff0000ffff0101, 0x00ff0000ff00ff00, 0x00ff0000ff0000ff, + 0x00ff0000ff000000, 0x00ff0000ff000001, 0x00ff0000ff000100, 0x00ff0000ff01ffff, + 0x00ff0000ff010000, 0x00ff0000ff010101, 0x00ff000000ffff00, 0x00ff000000ff00ff, + 0x00ff000000ff0000, 0x00ff000000ff0001, 0x00ff000000ff0100, 0x00ff00000000ffff, + 0x00ff00000000ff00, 0x00ff0000000000ff, 0x00ff000000000000, 0x00ff000000000001, + 0x00ff0000000001ff, 0x00ff000000000100, 0x00ff00000001ff00, 0x00ff0000000100ff, + 0x00ff000000010000, 0x00ff000000010001, 0x00ff000000010100, 0x00ff000001ffff01, + 0x00ff000001ff00ff, 0x00ff000001ff0000, 0x00ff000001ff01ff, 0x00ff00000100ff00, + 0x00ff0000010000ff, 0x00ff000001000000, 0x00ff000001000001, 0x00ff000001000100, + 0x00ff000001000101, 0x00ff000001010000, 0x00ff0000010101ff, 0x00ff000001010101, + 0x00ff0001ffffff00, 0x00ff0001ffff0000, 0x00ff0001ffff0100, 0x00ff0001ff0000ff, + 0x00ff0001ff000000, 0x00ff0001ff0001ff, 0x00ff0001ff000101, 0x00ff0001ff01ff00, + 0x00ff0001ff0100ff, 0x00ff0001ff010100, 0x00ff000100ffffff, 0x00ff000100ffff01, + 0x00ff000100ff0000, 0x00ff000100ff01ff, 0x00ff00010000ffff, 0x00ff00010000ff00, + 0x00ff00010000ff01, 0x00ff000100000000, 0x00ff000100000001, 0x00ff000100000100, + 0x00ff00010001ff01, 0x00ff000100010000, 0x00ff0001000101ff, 0x00ff000101ffff00, + 0x00ff000101ff0000, 0x00ff000101ff0101, 0x00ff0001010000ff, 0x00ff000101000000, + 0x00ff00010101ff00, 0x00ff0001010100ff, 0x00ff000101010001, 0x00ff01ffffff0000, + 0x00ff01ffff00ff00, 0x00ff01ffff000000, 0x00ff01ffff000101, 0x00ff01ffff010000, + 0x00ff01ff00ffff01, 0x00ff01ff00ff0100, 0x00ff01ff0000ffff, 0x00ff01ff00000000, + 0x00ff01ff000001ff, 0x00ff01ff0001ff00, 0x00ff01ff000100ff, 0x00ff01ff00010001, + 0x00ff01ff00010100, 0x00ff01ff01ff0000, 0x00ff01ff0100ff00, 0x00ff01ff010000ff, + 0x00ff01ff01000001, 0x00ff01ff01000100, 0x00ff01ff01010000, 0x00ff0100ffffff00, + 0x00ff0100ffff0000, 0x00ff0100ffff0001, 0x00ff0100ffff0101, 0x00ff0100ff00ffff, + 0x00ff0100ff0000ff, 0x00ff0100ff000000, 0x00ff0100ff0001ff, 0x00ff0100ff01ff00, + 0x00ff0100ff0100ff, 0x00ff0100ff010001, 0x00ff010000ffffff, 0x00ff010000ff0000, + 0x00ff010000ff0101, 0x00ff01000000ff00, 0x00ff01000000ff01, 0x00ff0100000000ff, + 0x00ff010000000000, 0x00ff010000000001, 0x00ff010000000100, 0x00ff01000001ffff, + 0x00ff01000001ff01, 0x00ff010000010000, 0x00ff010000010001, 0x00ff010000010101, + 0x00ff010001ff0001, 0x00ff010001ff0100, 0x00ff01000100ff01, 0x00ff010001000000, + 0x00ff010001000001, 0x00ff0100010001ff, 0x00ff01000101ff00, 0x00ff0100010100ff, + 0x00ff010001010001, 0x00ff010001010100, 0x00ff0101ff000001, 0x00ff010100ff00ff, + 0x00ff010100ff0001, 0x00ff010100ff0100, 0x00ff010100000000, 0x00ff0101000001ff, + 0x00ff010100000101, 0x00ff0101000100ff, 0x00ff010100010100, 0x00ff0101010000ff, + 0x00ff010101010000, 0x0000ffffffffff00, 0x0000ffffffff00ff, 0x0000ffffffff0000, + 0x0000ffffffff0001, 0x0000ffffffff0100, 0x0000ffffff00ff01, 0x0000ffffff000000, + 0x0000ffffff000101, 0x0000ffffff01ff00, 0x0000ffffff0100ff, 0x0000ffffff010100, + 0x0000ffff00ffffff, 0x0000ffff00ff0000, 0x0000ffff00ff01ff, 0x0000ffff0000ff00, + 0x0000ffff000000ff, 0x0000ffff00000000, 0x0000ffff00000001, 0x0000ffff00000100, + 0x0000ffff00010000, 0x0000ffff000101ff, 0x0000ffff01ff0001, 0x0000ffff01ff0100, + 0x0000ffff01000000, 0x0000ffff010001ff, 0x0000ffff0101ffff, 0x0000ffff0101ff00, + 0x0000ffff01010001, 0x0000ffff01010100, 0x0000ff00ffff0000, 0x0000ff00ffff01ff, + 0x0000ff00ffff0100, 0x0000ff00ffff0101, 0x0000ff00ff00ff00, 0x0000ff00ff0000ff, + 0x0000ff00ff000000, 0x0000ff00ff000001, 0x0000ff00ff0001ff, 0x0000ff00ff000100, + 0x0000ff00ff01ffff, 0x0000ff00ff010000, 0x0000ff00ff010001, 0x0000ff00ff0101ff, + 0x0000ff00ff010101, 0x0000ff0000ffff00, 0x0000ff0000ff00ff, 0x0000ff0000ff0000, + 0x0000ff0000ff0001, 0x0000ff0000ff0100, 0x0000ff000000ffff, 0x0000ff000000ff00, + 0x0000ff000000ff01, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001, + 0x0000ff00000001ff, 0x0000ff0000000100, 0x0000ff0000000101, 0x0000ff000001ff00, + 0x0000ff00000100ff, 0x0000ff0000010000, 0x0000ff0000010001, 0x0000ff0000010100, + 0x0000ff0001ffff01, 0x0000ff0001ff0000, 0x0000ff000100ff00, 0x0000ff00010000ff, + 0x0000ff0001000000, 0x0000ff0001000001, 0x0000ff0001000100, 0x0000ff000101ffff, + 0x0000ff0001010000, 0x0000ff0001010101, 0x0000ff01ffffff00, 0x0000ff01ffff0001, + 0x0000ff01ff00ff01, 0x0000ff01ff000000, 0x0000ff01ff000101, 0x0000ff01ff01ff00, + 0x0000ff01ff0100ff, 0x0000ff0100ffff01, 0x0000ff0100ff0000, 0x0000ff0100ff0101, + 0x0000ff010000ff00, 0x0000ff01000000ff, 0x0000ff0100000000, 0x0000ff0100000001, + 0x0000ff0100000100, 0x0000ff010001ff01, 0x0000ff0100010000, 0x0000ff0101ff0000, + 0x0000ff010100ffff, 0x0000ff010100ff01, 0x0000ff0101000000, 0x0000ff0101000100, + 0x0000ff0101000101, 0x0000ff01010100ff, 0x000000ffffff00ff, 0x000000ffffff0000, + 0x000000ffff00ff00, 0x000000ffff0000ff, 0x000000ffff000000, 0x000000ffff000001, + 0x000000ffff0001ff, 0x000000ffff000100, 0x000000ffff01ff00, 0x000000ffff010000, + 0x000000ffff0101ff, 0x000000ffff010101, 0x000000ff00ffff00, 0x000000ff00ff00ff, + 0x000000ff00ff0000, 0x000000ff00ff0001, 0x000000ff00ff0100, 0x000000ff00ff0101, + 0x000000ff0000ffff, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000, + 0x000000ff00000001, 0x000000ff000001ff, 0x000000ff00000100, 0x000000ff00000101, + 0x000000ff0001ff00, 0x000000ff0001ff01, 0x000000ff000100ff, 0x000000ff00010000, + 0x000000ff00010001, 0x000000ff00010100, 0x000000ff01ffffff, 0x000000ff01ff01ff, + 0x000000ff01ff0101, 0x000000ff0100ff00, 0x000000ff010000ff, 0x000000ff01000000, + 0x000000ff01000001, 0x000000ff01000100, 0x000000ff0101ff00, 0x000000ff010100ff, + 0x000000ff01010000, 0x000000ff01010101, 0x00000000ffffff00, 0x00000000ffffff01, + 0x00000000ffff00ff, 0x00000000ffff0000, 0x00000000ffff0001, 0x00000000ffff0100, + 0x00000000ff00ffff, 0x00000000ff00ff00, 0x00000000ff00ff01, 0x00000000ff0000ff, + 0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff000101, + 0x00000000ff01ff00, 0x00000000ff0100ff, 0x00000000ff010000, 0x00000000ff010001, + 0x00000000ff010100, 0x0000000000ffffff, 0x0000000000ffff00, 0x0000000000ffff01, + 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, 0x0000000000ff01ff, + 0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01, + 0x00000000000000ff, 0x0000000000000000, 0x0000000000000001, 0x00000000000001ff, + 0x0000000000000100, 0x0000000000000101, 0x000000000001ffff, 0x000000000001ff00, + 0x00000000000100ff, 0x0000000000010000, 0x0000000000010001, 0x00000000000101ff, + 0x0000000000010100, 0x0000000000010101, 0x0000000001ffff00, 0x0000000001ff00ff, + 0x0000000001ff0000, 0x0000000001ff0100, 0x0000000001ff0101, 0x000000000100ffff, + 0x000000000100ff00, 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, + 0x00000000010001ff, 0x0000000001000100, 0x000000000101ff00, 0x00000000010100ff, + 0x0000000001010000, 0x0000000001010001, 0x0000000001010100, 0x00000001ffffffff, + 0x00000001ffffff00, 0x00000001ffffff01, 0x00000001ffff00ff, 0x00000001ffff0001, + 0x00000001ffff01ff, 0x00000001ffff0100, 0x00000001ff00ff00, 0x00000001ff0000ff, + 0x00000001ff000000, 0x00000001ff0001ff, 0x00000001ff000100, 0x00000001ff01ffff, + 0x00000001ff01ff00, 0x00000001ff01ff01, 0x00000001ff0100ff, 0x00000001ff010000, + 0x00000001ff010001, 0x00000001ff0101ff, 0x00000001ff010100, 0x0000000100ffff00, + 0x0000000100ff0000, 0x0000000100ff0001, 0x0000000100ff01ff, 0x0000000100ff0100, + 0x0000000100ff0101, 0x000000010000ffff, 0x000000010000ff00, 0x000000010000ff01, + 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, 0x00000001000001ff, + 0x0000000100000100, 0x0000000100000101, 0x000000010001ff00, 0x00000001000100ff, + 0x0000000100010000, 0x0000000100010100, 0x0000000101ffff01, 0x0000000101ff0000, + 0x0000000101ff0001, 0x0000000101ff01ff, 0x0000000101ff0100, 0x0000000101ff0101, + 0x000000010100ff00, 0x0000000101000000, 0x0000000101000101, 0x000000010101ff01, + 0x0000000101010000, 0x0000000101010001, 0x00000001010101ff, 0x0000000101010100, + 0x000001ffffff00ff, 0x000001ffffff0000, 0x000001ffffff0001, 0x000001ffffff0100, + 0x000001ffff00ffff, 0x000001ffff000000, 0x000001ffff0001ff, 0x000001ffff01ff00, + 0x000001ffff010101, 0x000001ff00ff0000, 0x000001ff00ff01ff, 0x000001ff00ff0101, + 0x000001ff0000ff00, 0x000001ff000000ff, 0x000001ff00000000, 0x000001ff00000001, + 0x000001ff000001ff, 0x000001ff00000100, 0x000001ff0001ffff, 0x000001ff0001ff01, + 0x000001ff000100ff, 0x000001ff00010000, 0x000001ff01ffff01, 0x000001ff01ff0100, + 0x000001ff0100ffff, 0x000001ff0100ff01, 0x000001ff01000000, 0x000001ff010001ff, + 0x000001ff0101ff00, 0x000001ff01010100, 0x00000100ffffff00, 0x00000100ffffff01, + 0x00000100ffff0000, 0x00000100ffff0101, 0x00000100ff00ff00, 0x00000100ff0000ff, + 0x00000100ff000000, 0x00000100ff000001, 0x00000100ff000100, 0x00000100ff010000, + 0x0000010000ffff00, 0x0000010000ff00ff, 0x0000010000ff0000, 0x0000010000ff0001, + 0x0000010000ff0100, 0x000001000000ffff, 0x000001000000ff00, 0x000001000000ff01, + 0x00000100000000ff, 0x0000010000000000, 0x0000010000000001, 0x00000100000001ff, + 0x0000010000000100, 0x0000010000000101, 0x000001000001ff00, 0x00000100000100ff, + 0x0000010000010000, 0x0000010000010001, 0x0000010000010100, 0x0000010001ffff00, + 0x0000010001ff0000, 0x0000010001ff0100, 0x000001000100ff00, 0x00000100010000ff, + 0x0000010001000000, 0x0000010001000001, 0x00000100010001ff, 0x0000010001000100, + 0x0000010001010000, 0x00000101ffff00ff, 0x00000101ffff01ff, 0x00000101ff000000, + 0x00000101ff000101, 0x00000101ff01ffff, 0x00000101ff010000, 0x00000101ff010001, + 0x00000101ff010100, 0x0000010100ff0000, 0x0000010100ff01ff, 0x0000010100ff0100, + 0x000001010000ff00, 0x0000010100000000, 0x0000010100000001, 0x00000101000001ff, + 0x0000010100000100, 0x000001010001ff01, 0x0000010100010000, 0x00000101000101ff, + 0x0000010100010101, 0x0000010101ffff00, 0x0000010101ff0101, 0x000001010100ff01, + 0x0000010101000000, 0x0000010101000001, 0x00000101010001ff, 0x0000010101000101, + 0x000001010101ff00, 0x0001ffffffff0000, 0x0001ffffff0000ff, 0x0001ffffff000001, + 0x0001ffffff000100, 0x0001ffffff010000, 0x0001ffff00ff00ff, 0x0001ffff0000ffff, + 0x0001ffff00000000, 0x0001ffff00000001, 0x0001ffff000001ff, 0x0001ffff00000101, + 0x0001ffff0001ff00, 0x0001ffff000100ff, 0x0001ffff00010001, 0x0001ffff00010100, + 0x0001ffff01ffff00, 0x0001ffff01000001, 0x0001ffff01010000, 0x0001ff00ffffff00, + 0x0001ff00ffff00ff, 0x0001ff00ffff0001, 0x0001ff00ffff0100, 0x0001ff00ff00ff01, + 0x0001ff00ff000000, 0x0001ff00ff01ff00, 0x0001ff00ff01ff01, 0x0001ff00ff010001, + 0x0001ff00ff010100, 0x0001ff0000ff0000, 0x0001ff0000ff0100, 0x0001ff000000ff00, + 0x0001ff0000000000, 0x0001ff0000000001, 0x0001ff0000000100, 0x0001ff0000010000, + 0x0001ff0000010001, 0x0001ff0000010101, 0x0001ff0001ff00ff, 0x0001ff0001ff0101, + 0x0001ff000100ff01, 0x0001ff0001000000, 0x0001ff000101ff00, 0x0001ff0001010001, + 0x0001ff0001010100, 0x0001ff01ff00ff00, 0x0001ff01ff000001, 0x0001ff01ff000100, + 0x0001ff0100ffffff, 0x0001ff0100ffff00, 0x0001ff0100ff0001, 0x0001ff0100000000, + 0x0001ff0100000001, 0x0001ff01000001ff, 0x0001ff010001ffff, 0x0001ff0101ff0000, + 0x0001ff010100ff00, 0x0001ff0101000001, 0x0001ff0101010000, 0x000100ffff00ff00, + 0x000100ffff00ff01, 0x000100ffff000000, 0x000100ffff000001, 0x000100ffff000101, + 0x000100ffff01ff00, 0x000100ffff010001, 0x000100ffff010100, 0x000100ff00ffffff, + 0x000100ff00ffff01, 0x000100ff00ff0000, 0x000100ff00ff01ff, 0x000100ff00ff0101, + 0x000100ff0000ff00, 0x000100ff000000ff, 0x000100ff00000000, 0x000100ff00000001, + 0x000100ff00000100, 0x000100ff00000101, 0x000100ff0001ffff, 0x000100ff0001ff01, + 0x000100ff00010000, 0x000100ff01ff00ff, 0x000100ff01ff0000, 0x000100ff01ff0100, + 0x000100ff0100ffff, 0x000100ff0100ff01, 0x000100ff010000ff, 0x000100ff01000000, + 0x000100ff01000001, 0x000100ff010001ff, 0x000100ff01000101, 0x000100ff0101ff00, + 0x000100ff010100ff, 0x000100ff01010100, 0x00010000ffff0000, 0x00010000ffff01ff, + 0x00010000ffff0101, 0x00010000ff00ff00, 0x00010000ff000000, 0x00010000ff000001, + 0x00010000ff000100, 0x0001000000ff00ff, 0x0001000000ff0000, 0x0001000000ff0001, + 0x0001000000ff0100, 0x000100000000ffff, 0x000100000000ff00, 0x00010000000000ff, + 0x0001000000000000, 0x0001000000000001, 0x0001000000000100, 0x000100000001ff00, + 0x00010000000100ff, 0x0001000000010000, 0x0001000000010001, 0x0001000000010100, + 0x0001000001ff0001, 0x0001000001ff0100, 0x0001000001ff0101, 0x000100000100ff00, + 0x0001000001000000, 0x0001000001000001, 0x0001000001000100, 0x0001000001000101, + 0x000100000101ff01, 0x0001000001010000, 0x0001000001010001, 0x00010000010101ff, + 0x00010001ffffff01, 0x00010001ffff0100, 0x00010001ff000000, 0x00010001ff01ffff, + 0x00010001ff010001, 0x00010001ff0101ff, 0x00010001ff010100, 0x0001000100ffffff, + 0x0001000100ff0000, 0x0001000100ff01ff, 0x0001000100ff0101, 0x000100010000ff00, + 0x00010001000000ff, 0x0001000100000000, 0x0001000100000001, 0x00010001000001ff, + 0x0001000100000101, 0x000100010001ffff, 0x0001000100010000, 0x00010001000101ff, + 0x0001000101ffffff, 0x0001000101ffff01, 0x0001000101ff0000, 0x0001000101ff0101, + 0x00010001010000ff, 0x0001000101000001, 0x00010001010001ff, 0x0001000101000100, + 0x000100010101ffff, 0x00010001010100ff, 0x0001000101010001, 0x0001000101010101, + 0x000101ffff000001, 0x000101ffff000100, 0x000101ffff010000, 0x000101ff00ffff00, + 0x000101ff0000ff01, 0x000101ff00000000, 0x000101ff00000101, 0x000101ff0001ff00, + 0x000101ff00010100, 0x000101ff01ff0000, 0x000101ff0100ff00, 0x000101ff010001ff, + 0x000101ff01010001, 0x00010100ffffff00, 0x00010100ffff00ff, 0x00010100ff00ffff, + 0x00010100ff000000, 0x00010100ff01ff00, 0x00010100ff0100ff, 0x00010100ff010001, + 0x00010100ff010100, 0x0001010000ffffff, 0x0001010000ffff00, 0x0001010000ff0000, + 0x0001010000ff0001, 0x0001010000ff01ff, 0x000101000000ff00, 0x00010100000000ff, + 0x0001010000000000, 0x0001010000000001, 0x0001010000000100, 0x000101000001ffff, + 0x0001010000010000, 0x0001010000010101, 0x0001010001ffff01, 0x0001010001ff00ff, + 0x0001010001ff0101, 0x0001010001000000, 0x000101000101ff00, 0x00010100010100ff, + 0x0001010001010000, 0x0001010001010100, 0x00010101ff00ff00, 0x00010101ff000001, + 0x00010101ff0001ff, 0x0001010100ffff00, 0x0001010100ff00ff, 0x0001010100ff0100, + 0x000101010000ffff, 0x0001010100000000, 0x00010101000001ff, 0x0001010100000101, + 0x00010101000100ff, 0x0001010100010000, 0x0001010100010100, 0x0001010101ff0001, + 0x00010101010000ff, 0x00010101010001ff, 0x0001010101000101, 0x0001010101010001, + 0x01ffffffffffffff, 0x01ffffffffffff01, 0x01ffffffffff01ff, 0x01ffffffffff0101, + 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, 0x01ffffffff010101, + 0x01ffffff00ff0000, 0x01ffffff0000ffff, 0x01ffffff0000ff00, 0x01ffffff000000ff, + 0x01ffffff00000001, 0x01ffffff00000100, 0x01ffffff00010000, 0x01ffffff01ffffff, + 0x01ffffff01ffff01, 0x01ffffff01ff01ff, 0x01ffffff01ff0101, 0x01ffffff01000000, + 0x01ffffff0101ffff, 0x01ffffff0101ff01, 0x01ffffff010101ff, 0x01ffffff01010101, + 0x01ffff00ffff0000, 0x01ffff00ff00ff00, 0x01ffff00ff0000ff, 0x01ffff00ff000001, + 0x01ffff00ff000100, 0x01ffff00ff010000, 0x01ffff0000ffff00, 0x01ffff0000ff00ff, + 0x01ffff0000ff0100, 0x01ffff000000ffff, 0x01ffff000000ff01, 0x01ffff0000000000, + 0x01ffff0000000001, 0x01ffff00000001ff, 0x01ffff0000000100, 0x01ffff00000100ff, + 0x01ffff0000010001, 0x01ffff0000010100, 0x01ffff0001ff0000, 0x01ffff0001ff0100, + 0x01ffff00010000ff, 0x01ffff0001000001, 0x01ffff0001000100, 0x01ffff0001010000, + 0x01ffff01ffffffff, 0x01ffff01ffffff01, 0x01ffff01ffff01ff, 0x01ffff01ffff0101, + 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff01ff01, 0x01ffff01ff0101ff, + 0x01ffff01ff010101, 0x01ffff010000ff00, 0x01ffff01000000ff, 0x01ffff0100000100, + 0x01ffff0100010000, 0x01ffff0101ffffff, 0x01ffff0101ffff01, 0x01ffff0101ff01ff, + 0x01ffff0101ff0101, 0x01ffff0101000000, 0x01ffff010101ffff, 0x01ffff010101ff01, + 0x01ffff01010101ff, 0x01ffff0101010101, 0x01ff00ffff0000ff, 0x01ff00ffff000100, + 0x01ff00ff00ffff00, 0x01ff00ff00ff00ff, 0x01ff00ff0000ff00, 0x01ff00ff00000000, + 0x01ff00ff00000101, 0x01ff00ff0001ff00, 0x01ff00ff000100ff, 0x01ff00ff00010100, + 0x01ff00ff010000ff, 0x01ff00ff01000100, 0x01ff0000ffffff00, 0x01ff0000ffff0100, + 0x01ff0000ff00ff01, 0x01ff0000ff000000, 0x01ff0000ff000101, 0x01ff0000ff010001, + 0x01ff0000ff010100, 0x01ff000000ffffff, 0x01ff000000ffff00, 0x01ff000000ff0000, + 0x01ff000000ff01ff, 0x01ff00000000ff00, 0x01ff0000000000ff, 0x01ff000000000000, + 0x01ff000000000001, 0x01ff000000000100, 0x01ff000000000101, 0x01ff000000010000, + 0x01ff000000010001, 0x01ff0000000101ff, 0x01ff000000010101, 0x01ff000001ffff00, + 0x01ff000001ff00ff, 0x01ff000001ff0001, 0x01ff000001ff0100, 0x01ff00000100ffff, + 0x01ff00000100ff01, 0x01ff000001000000, 0x01ff0000010001ff, 0x01ff000001010001, + 0x01ff0001ff00ff00, 0x01ff0001ff000001, 0x01ff0001ff000100, 0x01ff0001ff010000, + 0x01ff000100ffff00, 0x01ff000100ff00ff, 0x01ff000100ff0100, 0x01ff000100ff0101, + 0x01ff00010000ffff, 0x01ff000100000000, 0x01ff000100000100, 0x01ff000100000101, + 0x01ff00010001ff00, 0x01ff000100010001, 0x01ff000100010101, 0x01ff000101ff0000, + 0x01ff00010100ff00, 0x01ff000101000101, 0x01ff0001010100ff, 0x01ff01ffffffffff, + 0x01ff01ffffffff01, 0x01ff01ffffff01ff, 0x01ff01ffffff0101, 0x01ff01ffff000000, + 0x01ff01ffff01ffff, 0x01ff01ffff01ff01, 0x01ff01ffff0101ff, 0x01ff01ffff010101, + 0x01ff01ff00ffff00, 0x01ff01ff00ff0000, 0x01ff01ff0000ff00, 0x01ff01ff000000ff, + 0x01ff01ff00000100, 0x01ff01ff00010000, 0x01ff01ff00010100, 0x01ff01ff01ffffff, + 0x01ff01ff01ffff01, 0x01ff01ff01ff01ff, 0x01ff01ff01ff0101, 0x01ff01ff01000000, + 0x01ff01ff0101ffff, 0x01ff01ff0101ff01, 0x01ff01ff010101ff, 0x01ff01ff01010101, + 0x01ff0100ffff0000, 0x01ff0100ffff0001, 0x01ff0100ff00ff00, 0x01ff0100ff0000ff, + 0x01ff0100ff000001, 0x01ff0100ff010000, 0x01ff010000ffff00, 0x01ff010000ff00ff, + 0x01ff010000ff0001, 0x01ff010000ff0100, 0x01ff01000000ffff, 0x01ff01000000ff01, + 0x01ff010000000000, 0x01ff010000000101, 0x01ff01000001ff00, 0x01ff0100000100ff, + 0x01ff010001ff0000, 0x01ff010001000001, 0x01ff010001000100, 0x01ff010001010000, + 0x01ff0101ffffffff, 0x01ff0101ffffff01, 0x01ff0101ffff01ff, 0x01ff0101ffff0101, + 0x01ff0101ff000000, 0x01ff0101ff01ffff, 0x01ff0101ff01ff01, 0x01ff0101ff0101ff, + 0x01ff0101ff010101, 0x01ff010100ff0000, 0x01ff01010000ff00, 0x01ff0101000000ff, + 0x01ff010100000001, 0x01ff010101ffffff, 0x01ff010101ffff01, 0x01ff010101ff01ff, + 0x01ff010101ff0101, 0x01ff010101000000, 0x01ff01010101ffff, 0x01ff01010101ff01, + 0x01ff0101010101ff, 0x01ff010101010101, 0x0100ffffffff0000, 0x0100ffffff00ff00, + 0x0100ffffff000001, 0x0100ffffff0001ff, 0x0100ffffff000100, 0x0100ffffff010000, + 0x0100ffff00ffff00, 0x0100ffff00ff0001, 0x0100ffff00ff0100, 0x0100ffff00000000, + 0x0100ffff000001ff, 0x0100ffff00000101, 0x0100ffff00010100, 0x0100ffff00010101, + 0x0100ffff01ff0000, 0x0100ffff0100ff00, 0x0100ffff010000ff, 0x0100ffff01000001, + 0x0100ffff01000100, 0x0100ffff01010000, 0x0100ff00ffffff00, 0x0100ff00ffff00ff, + 0x0100ff00ffff0001, 0x0100ff00ffff0100, 0x0100ff00ff00ffff, 0x0100ff00ff000000, + 0x0100ff00ff0001ff, 0x0100ff00ff000101, 0x0100ff00ff01ff00, 0x0100ff00ff0100ff, + 0x0100ff00ff010001, 0x0100ff00ff010100, 0x0100ff0000ffffff, 0x0100ff0000ff0000, + 0x0100ff000000ffff, 0x0100ff000000ff00, 0x0100ff00000000ff, 0x0100ff0000000000, + 0x0100ff0000000001, 0x0100ff0000000100, 0x0100ff000001ff01, 0x0100ff0000010000, + 0x0100ff0001ff00ff, 0x0100ff0001ff0001, 0x0100ff000100ff01, 0x0100ff0001000000, + 0x0100ff00010001ff, 0x0100ff000101ff00, 0x0100ff00010100ff, 0x0100ff0001010001, + 0x0100ff0001010100, 0x0100ff01ffff0000, 0x0100ff01ff00ff00, 0x0100ff01ff0000ff, + 0x0100ff01ff000100, 0x0100ff01ff010000, 0x0100ff0100ff00ff, 0x0100ff0100ff0001, + 0x0100ff0100ff0100, 0x0100ff010000ffff, 0x0100ff010000ff01, 0x0100ff0100000000, + 0x0100ff01000001ff, 0x0100ff0100010001, 0x0100ff0100010100, 0x0100ff0101ff0000, + 0x0100ff01010000ff, 0x0100ff0101000001, 0x0100ff0101010100, 0x010000ffffffff00, + 0x010000ffffff00ff, 0x010000ffffff0001, 0x010000ffff00ffff, 0x010000ffff000000, + 0x010000ffff0001ff, 0x010000ffff010001, 0x010000ff00ffffff, 0x010000ff00ff0101, + 0x010000ff0000ff00, 0x010000ff000000ff, 0x010000ff00000000, 0x010000ff00000001, + 0x010000ff000001ff, 0x010000ff00000100, 0x010000ff0001ffff, 0x010000ff0001ff00, + 0x010000ff0001ff01, 0x010000ff00010000, 0x010000ff01ff00ff, 0x010000ff01ff0001, + 0x010000ff0100ff01, 0x010000ff010000ff, 0x010000ff01000000, 0x010000ff010001ff, + 0x010000ff0101ff00, 0x010000ff01010100, 0x01000000ffffffff, 0x01000000ffff0000, + 0x01000000ffff01ff, 0x01000000ffff0101, 0x01000000ff00ffff, 0x01000000ff00ff00, + 0x01000000ff0000ff, 0x01000000ff000000, 0x01000000ff000001, 0x01000000ff000100, + 0x01000000ff01ff00, 0x01000000ff010000, 0x01000000ff010100, 0x01000000ff010101, + 0x0100000000ffff00, 0x0100000000ff00ff, 0x0100000000ff0000, 0x0100000000ff0001, + 0x0100000000ff0100, 0x010000000000ffff, 0x010000000000ff00, 0x010000000000ff01, + 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, 0x01000000000001ff, + 0x0100000000000100, 0x0100000000000101, 0x010000000001ff00, 0x01000000000100ff, + 0x0100000000010000, 0x0100000000010001, 0x0100000000010100, 0x0100000001ffff00, + 0x0100000001ff0000, 0x0100000001ff01ff, 0x010000000100ff00, 0x010000000100ff01, + 0x01000000010000ff, 0x0100000001000000, 0x0100000001000001, 0x0100000001000100, + 0x0100000001000101, 0x010000000101ffff, 0x010000000101ff01, 0x0100000001010000, + 0x01000000010101ff, 0x0100000001010101, 0x01000001ffffff00, 0x01000001ffff00ff, + 0x01000001ff00ffff, 0x01000001ff000000, 0x01000001ff000100, 0x01000001ff01ffff, + 0x01000001ff010001, 0x01000001ff010100, 0x0100000100ff0000, 0x0100000100ff01ff, + 0x0100000100ff0100, 0x010000010000ff00, 0x010000010000ff01, 0x0100000100000000, + 0x0100000100000001, 0x0100000100000100, 0x0100000100010000, 0x01000001000101ff, + 0x0100000101ffff01, 0x0100000101ff00ff, 0x0100000101ff0100, 0x0100000101ff0101, + 0x010000010100ff01, 0x01000001010000ff, 0x0100000101000000, 0x01000001010100ff, + 0x0100000101010001, 0x0100000101010100, 0x010001ffffff0000, 0x010001ffff000001, + 0x010001ffff000100, 0x010001ffff010000, 0x010001ff00ffff00, 0x010001ff00ff0001, + 0x010001ff0000ffff, 0x010001ff0000ff01, 0x010001ff00000000, 0x010001ff00000001, + 0x010001ff00000101, 0x010001ff000100ff, 0x010001ff00010000, 0x010001ff01ff0000, + 0x010001ff0100ff00, 0x010001ff01000001, 0x010001ff01000100, 0x010001ff01010000, + 0x01000100ffff00ff, 0x01000100ffff0001, 0x01000100ffff0100, 0x01000100ff00ffff, + 0x01000100ff00ff01, 0x01000100ff000000, 0x01000100ff0001ff, 0x01000100ff000101, + 0x01000100ff01ffff, 0x01000100ff01ff00, 0x01000100ff0100ff, 0x01000100ff010001, + 0x0100010000ffffff, 0x0100010000ffff01, 0x0100010000ff0000, 0x0100010000ff01ff, + 0x0100010000ff0101, 0x010001000000ff00, 0x01000100000000ff, 0x0100010000000000, + 0x0100010000000001, 0x0100010000000100, 0x010001000001ff01, 0x0100010000010000, + 0x0100010000010001, 0x0100010000010101, 0x0100010001ffff00, 0x0100010001ff00ff, + 0x010001000100ffff, 0x010001000100ff01, 0x0100010001000000, 0x0100010001000101, + 0x010001000101ff00, 0x0100010001010001, 0x01000101ffff0000, 0x01000101ff000000, + 0x01000101ff010000, 0x0100010100ff00ff, 0x0100010100ff0001, 0x0100010100ff0100, + 0x010001010000ffff, 0x0100010100000000, 0x01000101000001ff, 0x010001010001ff00, + 0x0100010101ff0000, 0x010001010100ff00, 0x01000101010000ff, 0x0100010101000000, + 0x0100010101000001, 0x0101ffffffffffff, 0x0101ffffffffff01, 0x0101ffffffff01ff, + 0x0101ffffffff0101, 0x0101ffffff000000, 0x0101ffffff01ffff, 0x0101ffffff01ff01, + 0x0101ffffff0101ff, 0x0101ffffff010101, 0x0101ffff00ff0000, 0x0101ffff0000ff00, + 0x0101ffff000000ff, 0x0101ffff00000001, 0x0101ffff00000100, 0x0101ffff01ffffff, + 0x0101ffff01ffff01, 0x0101ffff01ff01ff, 0x0101ffff01ff0101, 0x0101ffff01000000, + 0x0101ffff0101ffff, 0x0101ffff0101ff01, 0x0101ffff010101ff, 0x0101ffff01010101, + 0x0101ff00ffff0000, 0x0101ff00ffff0100, 0x0101ff00ff00ff00, 0x0101ff00ff0000ff, + 0x0101ff00ff000001, 0x0101ff00ff000100, 0x0101ff00ff000101, 0x0101ff0000ff0001, + 0x0101ff0000ff0100, 0x0101ff000000ff00, 0x0101ff0000000000, 0x0101ff00000001ff, + 0x0101ff0000000101, 0x0101ff000001ff00, 0x0101ff00000100ff, 0x0101ff0001ff0000, + 0x0101ff000100ffff, 0x0101ff000100ff01, 0x0101ff0001000001, 0x0101ff0001000100, + 0x0101ff01ffffff01, 0x0101ff01ffff01ff, 0x0101ff01ffff0101, 0x0101ff01ff00ffff, + 0x0101ff01ff000100, 0x0101ff01ff01ff01, 0x0101ff01ff0101ff, 0x0101ff01ff010101, + 0x0101ff0100ff0000, 0x0101ff010000ff00, 0x0101ff0100000001, 0x0101ff0100000100, + 0x0101ff0100010000, 0x0101ff0101ffffff, 0x0101ff0101ffff01, 0x0101ff0101ff01ff, + 0x0101ff0101ff0101, 0x0101ff0101000000, 0x0101ff010101ffff, 0x0101ff010101ff01, + 0x0101ff01010101ff, 0x0101ff0101010101, 0x010100ffff000100, 0x010100ffff010000, + 0x010100ff00ffff00, 0x010100ff00ff00ff, 0x010100ff0000ffff, 0x010100ff000000ff, + 0x010100ff00000000, 0x010100ff000001ff, 0x010100ff00000101, 0x010100ff0001ff00, + 0x010100ff00010000, 0x010100ff00010001, 0x010100ff000101ff, 0x010100ff00010100, + 0x010100ff01ff0000, 0x01010000ffff0001, 0x01010000ffff0100, 0x01010000ff00ffff, + 0x01010000ff00ff01, 0x01010000ff000000, 0x01010000ff0001ff, 0x01010000ff010001, + 0x01010000ff010100, 0x0101000000ffff01, 0x0101000000ff0000, 0x010100000000ff00, + 0x01010000000000ff, 0x0101000000000000, 0x0101000000000001, 0x0101000000000100, + 0x0101000000010000, 0x0101000000010101, 0x0101000001ffff00, 0x0101000001ff00ff, + 0x0101000001ff0000, 0x0101000001ff0001, 0x0101000001ff0100, 0x010100000100ff01, + 0x0101000001000000, 0x01010000010001ff, 0x01010001ffff0000, 0x01010001ff00ff00, + 0x01010001ff000001, 0x01010001ff000101, 0x01010001ff01ff00, 0x01010001ff010000, + 0x0101000100ff00ff, 0x0101000100ff0001, 0x0101000100ff0101, 0x010100010000ff01, + 0x0101000100000000, 0x0101000100000001, 0x01010001000001ff, 0x010100010001ffff, + 0x010100010001ff01, 0x0101000101ff0001, 0x010100010100ffff, 0x0101000101000000, + 0x0101000101000001, 0x0101000101000100, 0x010100010101ff00, 0x01010001010100ff, + 0x0101000101010001, 0x010101ffffffffff, 0x010101ffffffff01, 0x010101ffffff01ff, + 0x010101ffffff0101, 0x010101ffff01ffff, 0x010101ffff01ff01, 0x010101ffff0101ff, + 0x010101ffff010101, 0x010101ff0000ff00, 0x010101ff000000ff, 0x010101ff00000001, + 0x010101ff00000100, 0x010101ff01ffffff, 0x010101ff01ffff01, 0x010101ff01ff01ff, + 0x010101ff01ff0101, 0x010101ff01000000, 0x010101ff0101ffff, 0x010101ff0101ff01, + 0x010101ff010101ff, 0x010101ff01010101, 0x01010100ffff0000, 0x01010100ff0000ff, + 0x01010100ff000100, 0x01010100ff01ff00, 0x01010100ff010000, 0x0101010000ffff00, + 0x010101000000ffff, 0x0101010000000000, 0x0101010000000101, 0x010101000001ff00, + 0x0101010000010001, 0x0101010000010100, 0x010101000100ffff, 0x0101010001000001, + 0x01010101ffffffff, 0x01010101ffffff01, 0x01010101ffff01ff, 0x01010101ffff0101, + 0x01010101ff01ffff, 0x01010101ff01ff01, 0x01010101ff0101ff, 0x01010101ff010101, + 0x010101010000ff00, 0x01010101000000ff, 0x0101010100000001, 0x0101010101ffffff, + 0x0101010101ffff01, 0x0101010101ff01ff, 0x0101010101ff0101, 0x0101010101000000, + 0x010101010101ffff, 0x010101010101ff01, 0x01010101010101ff, 0x0101010101010101, +GGML_TABLE_END() +#else +GGML_TABLE_BEGIN(uint32_t, iq1s_grid_gpu, NGRID_IQ1S) + 0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000, + 0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101, + 0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200, + 0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212, + 0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011, + 0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111, + 0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220, + 0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022, + 0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220, + 0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101, + 0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110, + 0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111, + 0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010, + 0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210, + 0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221, + 0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021, + 0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002, + 0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101, + 0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101, + 0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211, + 0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110, + 0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022, + 0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121, + 0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220, + 0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001, + 0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101, + 0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102, + 0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012, + 0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010, + 0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111, + 0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122, + 0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222, + 0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001, + 0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102, + 0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101, + 0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000, + 0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101, + 0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112, + 0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110, + 0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211, + 0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012, + 0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111, + 0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120, + 0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122, + 0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121, + 0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221, + 0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001, + 0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101, + 0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101, + 0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011, + 0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111, + 0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011, + 0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122, + 0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121, + 0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222, + 0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101, + 0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000, + 0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200, + 0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110, + 0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112, + 0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222, + 0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021, + 0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121, + 0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201, + 0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200, + 0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101, + 0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011, + 0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010, + 0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211, + 0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121, + 0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000, + 0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202, + 0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202, + 0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211, + 0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112, + 0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020, + 0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121, + 0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222, + 0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102, + 0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100, + 0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110, + 0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011, + 0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111, + 0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110, + 0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121, + 0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222, + 0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201, + 0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102, + 0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201, + 0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012, + 0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010, + 0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010, + 0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110, + 0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011, + 0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212, + 0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021, + 0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021, + 0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021, + 0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101, + 0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101, + 0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100, + 0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010, + 0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111, + 0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010, + 0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111, + 0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120, + 0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120, + 0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101, + 0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001, + 0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201, + 0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210, + 0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211, + 0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111, + 0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112, + 0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211, + 0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010, + 0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021, + 0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122, + 0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221, + 0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102, + 0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100, + 0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101, + 0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101, + 0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101, + 0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012, + 0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110, + 0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112, + 0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210, + 0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210, + 0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210, + 0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010, + 0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110, + 0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122, + 0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020, + 0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021, + 0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022, + 0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120, + 0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222, + 0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221, + 0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001, + 0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102, + 0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201, + 0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012, + 0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111, + 0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012, + 0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110, + 0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110, + 0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121, + 0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221, + 0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220, + 0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222, + 0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000, + 0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201, + 0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012, + 0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011, + 0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212, + 0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221, + 0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121, + 0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202, + 0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202, + 0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002, + 0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101, + 0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210, + 0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112, + 0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011, + 0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011, + 0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210, + 0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020, + 0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220, + 0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222, + 0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222, + 0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001, + 0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010, + 0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111, + 0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010, + 0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110, + 0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221, + 0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122, + 0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202, + 0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100, + 0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101, + 0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112, + 0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111, + 0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211, + 0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222, + 0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221, + 0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022, + 0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101, + 0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211, + 0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111, + 0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111, + 0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010, + 0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121, + 0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222, + 0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000, + 0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202, + 0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000, + 0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202, + 0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110, + 0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110, + 0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222, + 0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120, + 0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022, + 0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101, + 0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202, + 0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110, + 0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110, + 0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111, + 0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111, + 0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120, + 0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121, + 0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001, + 0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202, + 0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001, + 0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200, + 0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011, + 0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212, + 0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012, + 0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110, + 0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012, + 0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111, + 0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020, + 0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121, + 0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222, + 0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102, + 0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102, + 0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101, + 0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212, + 0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210, + 0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111, + 0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212, + 0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221, + 0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121, + 0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002, + 0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000, + 0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202, + 0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112, + 0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111, + 0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020, + 0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221, + 0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022, + 0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100, + 0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201, + 0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112, + 0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211, + 0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012, + 0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121, + 0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020, + 0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120, + 0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200, + 0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200, + 0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110, + 0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011, + 0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222, + 0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020, + 0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222, +GGML_TABLE_END() +#endif + +#endif // GGML_COMMON_IMPL +#endif // GGML_COMMON_IMPL diff --git a/bindings/ruby/ext/ggml-cuda.h b/bindings/ruby/ext/ggml-cuda.h new file mode 100644 index 00000000..5eb4af40 --- /dev/null +++ b/bindings/ruby/ext/ggml-cuda.h @@ -0,0 +1,43 @@ +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#ifdef GGML_USE_HIPBLAS +#define GGML_CUDA_NAME "ROCm" +#define GGML_CUBLAS_NAME "hipBLAS" +#else +#define GGML_CUDA_NAME "CUDA" +#define GGML_CUBLAS_NAME "cuBLAS" +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +#define GGML_CUDA_MAX_DEVICES 16 + +// backend API +GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device); + +GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend); + +// device buffer +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device); + +// split tensor buffer that splits matrices by rows across multiple devices +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split); + +// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void); + +GGML_API GGML_CALL int ggml_backend_cuda_get_device_count(void); +GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size); +GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total); + +GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size); +GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer); + +#ifdef __cplusplus +} +#endif diff --git a/bindings/ruby/ext/ggml-impl.h b/bindings/ruby/ext/ggml-impl.h index d88f2614..93a4f1a2 100644 --- a/bindings/ruby/ext/ggml-impl.h +++ b/bindings/ruby/ext/ggml-impl.h @@ -5,6 +5,7 @@ // GGML internal header #include +#include // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/ #include #include #include // memcpy @@ -18,6 +19,7 @@ extern "C" { // fall back to the _Static_assert C11 keyword. // if C99 - static_assert is noop // ref: https://stackoverflow.com/a/53923785/4039976 +#ifndef __cplusplus #ifndef static_assert #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L) #define static_assert(cond, msg) _Static_assert(cond, msg) @@ -25,6 +27,7 @@ extern "C" { #define static_assert(cond, msg) struct global_scope_noop_trick #endif #endif +#endif // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512 #if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__)) @@ -34,16 +37,17 @@ extern "C" { #ifndef __F16C__ #define __F16C__ #endif +#endif + +// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available +#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)) #ifndef __SSE3__ #define __SSE3__ #endif +#ifndef __SSSE3__ +#define __SSSE3__ +#endif #endif - -#undef MIN -#undef MAX - -#define MIN(a, b) ((a) < (b) ? (a) : (b)) -#define MAX(a, b) ((a) > (b) ? (a) : (b)) // 16-bit float // on Arm, we use __fp16 @@ -56,14 +60,30 @@ extern "C" { // #include -#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x)) -#define GGML_COMPUTE_FP32_TO_FP16(x) (x) +typedef __fp16 ggml_fp16_internal_t; -#define GGML_FP16_TO_FP32(x) ((float) (x)) -#define GGML_FP32_TO_FP16(x) (x) +#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + +#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) + +static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + ggml_fp16_internal_t tmp; + memcpy(&tmp, &h, sizeof(ggml_fp16_t)); + return (float)tmp; +} + +static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + ggml_fp16_t res; + ggml_fp16_internal_t tmp = f; + memcpy(&res, &tmp, sizeof(ggml_fp16_t)); + return res; +} #else +typedef uint16_t ggml_fp16_internal_t; + #ifdef __wasm_simd128__ #include #else @@ -217,8 +237,7 @@ extern float ggml_table_f32_f16[1 << 16]; // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. // This is also true for POWER9. -#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16) - +#if !defined(GGML_FP16_TO_FP32) inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { uint16_t s; memcpy(&s, &f, sizeof(uint16_t)); @@ -226,19 +245,23 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { } #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) +#endif +#if !defined(GGML_FP32_TO_FP16) +#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) #endif #define GGML_HASHTABLE_FULL ((size_t)-1) #define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2) +struct ggml_hash_set ggml_hash_set_new(size_t size); + bool ggml_hash_contains (const struct ggml_hash_set hash_set, struct ggml_tensor * key); // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted size_t ggml_hash_find (const struct ggml_hash_set hash_set, struct ggml_tensor * key); -// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full +// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full size_t ggml_hash_insert ( struct ggml_hash_set hash_set, struct ggml_tensor * key); // return index, asserts if table is full diff --git a/bindings/ruby/ext/ggml-kompute.h b/bindings/ruby/ext/ggml-kompute.h new file mode 100644 index 00000000..17146545 --- /dev/null +++ b/bindings/ruby/ext/ggml-kompute.h @@ -0,0 +1,46 @@ +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#include +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +struct ggml_vk_device { + int index; + int type; // same as VkPhysicalDeviceType + size_t heapSize; + const char * name; + const char * vendor; + int subgroupSize; + uint64_t bufferAlignment; + uint64_t maxAlloc; +}; + +struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count); +bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name); +bool ggml_vk_has_vulkan(void); +bool ggml_vk_has_device(void); +struct ggml_vk_device ggml_vk_current_device(void); + +// +// backend API +// + +// forward declaration +typedef struct ggml_backend * ggml_backend_t; + +GGML_API ggml_backend_t ggml_backend_kompute_init(int device); + +GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend); + +GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device); + +#ifdef __cplusplus +} +#endif diff --git a/bindings/ruby/ext/ggml-metal.h b/bindings/ruby/ext/ggml-metal.h new file mode 100644 index 00000000..a5c54218 --- /dev/null +++ b/bindings/ruby/ext/ggml-metal.h @@ -0,0 +1,66 @@ +// An interface allowing to compute ggml_cgraph with Metal +// +// This is a fully functional interface that extends ggml with GPU support for Apple devices. +// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.) +// +// How it works? +// +// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this +// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you +// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.) +// +// You only need to make sure that all memory buffers that you used during the graph creation +// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is +// used during the graph evaluation to determine the arguments of the compute kernels. +// +// Synchronization between device and host memory (for example for input and output tensors) +// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions. +// + +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#include +#include + +// max memory buffers that can be mapped to the device +#define GGML_METAL_MAX_BUFFERS 64 + +struct ggml_tensor; +struct ggml_cgraph; + +#ifdef __cplusplus +extern "C" { +#endif + +// +// backend API +// user-code should use only these functions +// + +GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data); + +GGML_API ggml_backend_t ggml_backend_metal_init(void); + +GGML_API bool ggml_backend_is_metal(ggml_backend_t backend); + +GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size); + +GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb); + +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); + +// helper to check if the device supports a specific family +// ideally, the user code should be doing these checks +// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf +GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family); + +// capture all command buffers committed the next time `ggml_backend_graph_compute` is called +GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend); + +#ifdef __cplusplus +} +#endif + diff --git a/bindings/ruby/ext/ggml-opencl.h b/bindings/ruby/ext/ggml-opencl.h new file mode 100644 index 00000000..257a6be6 --- /dev/null +++ b/bindings/ruby/ext/ggml-opencl.h @@ -0,0 +1,36 @@ +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + +GGML_API void ggml_cl_init(void); + +GGML_API void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +GGML_API void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +GGML_API bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst); +GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +GGML_API void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize); + +// GGML_API void * ggml_cl_host_malloc(size_t size); +// GGML_API void ggml_cl_host_free(void * ptr); + +GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor); + +GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor); + +// backend API + +// GGML_API ggml_backend_t ggml_backend_opencl_init(void); + +// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend); + +GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void); +// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void); + +#ifdef __cplusplus +} +#endif diff --git a/bindings/ruby/ext/ggml-quants.c b/bindings/ruby/ext/ggml-quants.c index 740be6dc..32e84434 100644 --- a/bindings/ruby/ext/ggml-quants.c +++ b/bindings/ruby/ext/ggml-quants.c @@ -1,10 +1,18 @@ +#define GGML_COMMON_IMPL_C +#include "ggml-common.h" + #include "ggml-quants.h" #include "ggml-impl.h" +#define GGML_COMMON_IMPL_C +#include "ggml-common.h" + #include #include #include #include +#include // for qsort +#include // for GGML_ASSERT #ifdef __ARM_NEON @@ -14,32 +22,12 @@ // #include -#if !defined(__aarch64__) -inline static int32_t vaddvq_s16(int16x8_t v) { - return - (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + - (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + - (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + - (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); -} - -inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { - int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); - int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); - return vcombine_s16(a0, b0); -} - -inline static int32_t vaddvq_s32(int32x4_t v) { - return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); -} -#endif - #else #ifdef __wasm_simd128__ #include #else -#ifdef __POWER9_VECTOR__ +#if defined(__POWER9_VECTOR__) || defined(__powerpc64__) #include #undef bool #define bool _Bool @@ -47,13 +35,15 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #if defined(_MSC_VER) || defined(__MINGW32__) #include #else -#if !defined(__riscv) && !defined(__s390__) +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) +#if !defined(__riscv) #include #endif #endif #endif #endif #endif +#endif #ifdef __riscv_v_intrinsic #include @@ -61,9 +51,13 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #undef MIN #undef MAX + #define MIN(a, b) ((a) < (b) ? (a) : (b)) #define MAX(a, b) ((a) > (b) ? (a) : (b)) +#define UNUSED GGML_UNUSED + +// some compilers don't provide _mm256_set_m128i, e.g. gcc 7 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1) #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) @@ -138,7 +132,7 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) { } static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { -#if __AVXVNNI__ +#if defined(__AVXVNNI__) || defined(__AVX512VNNI__) const __m256i zero = _mm256_setzero_si256(); const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); return _mm256_cvtepi32_ps(summed_pairs); @@ -284,8 +278,50 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 #if defined(__ARM_NEON) +#ifdef _MSC_VER + +#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) } + +#else + +#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) } + +#endif + #if !defined(__aarch64__) +// 64-bit compatibility + +// vaddvq_s16 +// vpaddq_s16 +// vpaddq_s32 +// vaddvq_s32 +// vaddvq_f32 +// vmaxvq_f32 +// vcvtnq_s32_f32 +// vzip1_u8 +// vzip2_u8 + +inline static int32_t vaddvq_s16(int16x8_t v) { + return + (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + + (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + + (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + + (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); +} + +inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { + int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); + int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); + return vcombine_s16(a0, b0); +} + +inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) { + int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a)); + int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b)); + return vcombine_s32(a0, b0); +} + inline static int32_t vaddvq_s32(int32x4_t v) { return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); } @@ -311,7 +347,185 @@ inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) { return res; } +inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) { + uint8x8_t res; + + res[0] = a[0]; res[1] = b[0]; + res[2] = a[1]; res[3] = b[1]; + res[4] = a[2]; res[5] = b[2]; + res[6] = a[3]; res[7] = b[3]; + + return res; +} + +inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) { + uint8x8_t res; + + res[0] = a[4]; res[1] = b[4]; + res[2] = a[5]; res[3] = b[5]; + res[4] = a[6]; res[5] = b[6]; + res[6] = a[7]; res[7] = b[7]; + + return res; +} + +// vld1q_s16_x2 +// vld1q_u8_x2 +// vld1q_u8_x4 +// vld1q_s8_x2 +// vld1q_s8_x4 +// TODO: double-check these work correctly + +typedef struct ggml_int16x8x2_t { + int16x8_t val[2]; +} ggml_int16x8x2_t; + +inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) { + ggml_int16x8x2_t res; + + res.val[0] = vld1q_s16(ptr + 0); + res.val[1] = vld1q_s16(ptr + 8); + + return res; +} + +typedef struct ggml_uint8x16x2_t { + uint8x16_t val[2]; +} ggml_uint8x16x2_t; + +inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) { + ggml_uint8x16x2_t res; + + res.val[0] = vld1q_u8(ptr + 0); + res.val[1] = vld1q_u8(ptr + 16); + + return res; +} + +typedef struct ggml_uint8x16x4_t { + uint8x16_t val[4]; +} ggml_uint8x16x4_t; + +inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) { + ggml_uint8x16x4_t res; + + res.val[0] = vld1q_u8(ptr + 0); + res.val[1] = vld1q_u8(ptr + 16); + res.val[2] = vld1q_u8(ptr + 32); + res.val[3] = vld1q_u8(ptr + 48); + + return res; +} + +typedef struct ggml_int8x16x2_t { + int8x16_t val[2]; +} ggml_int8x16x2_t; + +inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) { + ggml_int8x16x2_t res; + + res.val[0] = vld1q_s8(ptr + 0); + res.val[1] = vld1q_s8(ptr + 16); + + return res; +} + +typedef struct ggml_int8x16x4_t { + int8x16_t val[4]; +} ggml_int8x16x4_t; + +inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) { + ggml_int8x16x4_t res; + + res.val[0] = vld1q_s8(ptr + 0); + res.val[1] = vld1q_s8(ptr + 16); + res.val[2] = vld1q_s8(ptr + 32); + res.val[3] = vld1q_s8(ptr + 48); + + return res; +} + +// NOTE: not tested +inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) { + int8x16_t res; + + res[ 0] = a[b[ 0]]; + res[ 1] = a[b[ 1]]; + res[ 2] = a[b[ 2]]; + res[ 3] = a[b[ 3]]; + res[ 4] = a[b[ 4]]; + res[ 5] = a[b[ 5]]; + res[ 6] = a[b[ 6]]; + res[ 7] = a[b[ 7]]; + res[ 8] = a[b[ 8]]; + res[ 9] = a[b[ 9]]; + res[10] = a[b[10]]; + res[11] = a[b[11]]; + res[12] = a[b[12]]; + res[13] = a[b[13]]; + res[14] = a[b[14]]; + res[15] = a[b[15]]; + + return res; +} + +// NOTE: not tested +inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) { + uint8x16_t res; + + res[ 0] = a[b[ 0]]; + res[ 1] = a[b[ 1]]; + res[ 2] = a[b[ 2]]; + res[ 3] = a[b[ 3]]; + res[ 4] = a[b[ 4]]; + res[ 5] = a[b[ 5]]; + res[ 6] = a[b[ 6]]; + res[ 7] = a[b[ 7]]; + res[ 8] = a[b[ 8]]; + res[ 9] = a[b[ 9]]; + res[10] = a[b[10]]; + res[11] = a[b[11]]; + res[12] = a[b[12]]; + res[13] = a[b[13]]; + res[14] = a[b[14]]; + res[15] = a[b[15]]; + + return res; +} + +#else + +#define ggml_int16x8x2_t int16x8x2_t +#define ggml_uint8x16x2_t uint8x16x2_t +#define ggml_uint8x16x4_t uint8x16x4_t +#define ggml_int8x16x2_t int8x16x2_t +#define ggml_int8x16x4_t int8x16x4_t + +#define ggml_vld1q_s16_x2 vld1q_s16_x2 +#define ggml_vld1q_u8_x2 vld1q_u8_x2 +#define ggml_vld1q_u8_x4 vld1q_u8_x4 +#define ggml_vld1q_s8_x2 vld1q_s8_x2 +#define ggml_vld1q_s8_x4 vld1q_s8_x4 +#define ggml_vqtbl1q_s8 vqtbl1q_s8 +#define ggml_vqtbl1q_u8 vqtbl1q_u8 + #endif + +#if !defined(__ARM_FEATURE_DOTPROD) + +inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) { + const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b)); + const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b)); + + return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1))); +} + +#else + +#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c) + +#endif + #endif #if defined(__ARM_NEON) || defined(__wasm_simd128__) @@ -330,7 +544,7 @@ static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4 #endif // reference implementation for deterministic creation of model files -void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { +void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int64_t k) { static const int qk = QK4_0; assert(k % qk == 0); @@ -367,11 +581,12 @@ void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict } } -void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) { +void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) { quantize_row_q4_0_reference(x, y, k); } -void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) { + +void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int64_t k) { const int qk = QK4_1; assert(k % qk == 0); @@ -408,11 +623,11 @@ void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict } } -void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) { +void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) { quantize_row_q4_1_reference(x, y, k); } -void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { +void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int64_t k) { static const int qk = QK5_0; assert(k % qk == 0); @@ -456,11 +671,11 @@ void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict } } -void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) { +void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) { quantize_row_q5_0_reference(x, y, k); } -void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { +void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int64_t k) { const int qk = QK5_1; assert(k % qk == 0); @@ -504,12 +719,12 @@ void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict } } -void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) { +void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) { quantize_row_q5_1_reference(x, y, k); } // reference implementation for deterministic creation of model files -void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) { +void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int64_t k) { assert(k % QK8_0 == 0); const int nb = k / QK8_0; @@ -534,7 +749,7 @@ void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict } } -void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; @@ -723,7 +938,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { } // reference implementation for deterministic creation of model files -void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) { +void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int64_t k) { assert(QK8_1 == 32); assert(k % QK8_1 == 0); const int nb = k / QK8_1; @@ -739,7 +954,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); int sum = 0; @@ -754,11 +969,11 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict sum += y[i].qs[QK8_1/2 + j]; } - y[i].s = sum*d; + y[i].s = GGML_FP32_TO_FP16(sum*d); } } -void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) { assert(k % QK8_1 == 0); const int nb = k / QK8_1; @@ -782,7 +997,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); int32x4_t accv = vdupq_n_s32(0); @@ -798,7 +1013,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { accv = vaddq_s32(accv, vi); } - y[i].s = d * vaddvq_s32(accv); + y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv)); } #elif defined(__wasm_simd128__) for (int i = 0; i < nb; i++) { @@ -821,7 +1036,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); v128_t accv = wasm_i32x4_splat(0); @@ -837,10 +1052,11 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { accv = wasm_i32x4_add(accv, vi); } - y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) + - wasm_i32x4_extract_lane(accv, 1) + - wasm_i32x4_extract_lane(accv, 2) + - wasm_i32x4_extract_lane(accv, 3)); + y[i].s = GGML_FP32_TO_FP16( + d * (wasm_i32x4_extract_lane(accv, 0) + + wasm_i32x4_extract_lane(accv, 1) + + wasm_i32x4_extract_lane(accv, 2) + + wasm_i32x4_extract_lane(accv, 3))); } #elif defined(__AVX2__) || defined(__AVX__) for (int i = 0; i < nb; i++) { @@ -865,7 +1081,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { // Quantize these floats const float d = maxScalar / 127.f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; const __m256 mul = _mm256_set1_ps( id ); @@ -889,7 +1105,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { #if defined(__AVX2__) // Compute the sum of the quants and set y[i].s - y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); + y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)))); // Convert int32 to int16 i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 @@ -919,7 +1135,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { // Compute the sum of the quants and set y[i].s const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); - y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1)); + y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1))); // Convert int32 to int16 ni0 = _mm_packs_epi32( ni0, ni1 ); @@ -950,7 +1166,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); @@ -967,7 +1183,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { // set y[i].s int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); - y[i].s = sum*d; + y[i].s = GGML_FP32_TO_FP16(sum*d); } #else GGML_UNUSED(nb); @@ -976,7 +1192,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { #endif } -void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) { +void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) { static const int qk = QK4_0; assert(k % qk == 0); @@ -996,7 +1212,7 @@ void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int } } -void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) { +void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int64_t k) { static const int qk = QK4_1; assert(k % qk == 0); @@ -1017,7 +1233,7 @@ void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int } } -void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) { +void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int64_t k) { static const int qk = QK5_0; assert(k % qk == 0); @@ -1043,7 +1259,7 @@ void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int } } -void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) { +void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int64_t k) { static const int qk = QK5_1; assert(k % qk == 0); @@ -1070,7 +1286,7 @@ void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int } } -void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k) { +void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int64_t k) { static const int qk = QK8_0; assert(k % qk == 0); @@ -1100,7 +1316,8 @@ static inline int nearest_int(float fval) { return (i & 0x007fffff) - 0x00400000; } -static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type) { +static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type, + const float * restrict qw) { float max = 0; float amax = 0; for (int i = 0; i < n; ++i) { @@ -1126,14 +1343,18 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * rmse_type = -rmse_type; return_early = true; } - int weight_type = rmse_type%2; float sumlx = 0; float suml2 = 0; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 0; i < n; ++i) { +#else for (int i = 0; i < n; ++i) { +#endif int l = nearest_int(iscale * x[i]); l = MAX(-nmax, MIN(nmax-1, l)); L[i] = l + nmax; - float w = weight_type == 1 ? x[i] * x[i] : 1; + float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i])); sumlx += w*x[i]*l; suml2 += w*l*l; } @@ -1149,7 +1370,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * for (int i = 0; i < n; ++i) { int l = nearest_int(iscale * x[i]); l = MAX(-nmax, MIN(nmax-1, l)); - float w = weight_type == 1 ? x[i] * x[i] : 1; + float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i])); sumlx += w*x[i]*l; suml2 += w*l*l; } @@ -1273,7 +1494,12 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f float max = x[0]; float sum_w = weights[0]; float sum_x = sum_w * x[0]; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 1; i < n; ++i) { +#else for (int i = 1; i < n; ++i) { +#endif if (x[i] < min) min = x[i]; if (x[i] > max) max = x[i]; float w = weights[i]; @@ -1355,7 +1581,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * //========================- 2-bit (de)-quantization -void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k) { +void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1432,7 +1658,7 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict } } -void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k) { +void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1478,23 +1704,281 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int } } -void quantize_row_q2_K(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) { quantize_row_q2_K_reference(x, vy, k); } -size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - (void)hist; // TODO: collect histograms - - for (int j = 0; j < n; j += k) { - block_q2_K * restrict y = (block_q2_K *)dst + j/QK_K; - quantize_row_q2_K_reference(src + j, y, k); +static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights, + uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, + float rmin, float rdelta, int nstep, bool use_mad) { + float min = x[0]; + float max = x[0]; + float sum_w = weights ? weights[0] : x[0]*x[0]; + float sum_x = sum_w * x[0]; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 1; i < n; ++i) { +#else + for (int i = 1; i < n; ++i) { +#endif + if (x[i] < min) min = x[i]; + if (x[i] > max) max = x[i]; + float w = weights ? weights[i] : x[i]*x[i]; + sum_w += w; + sum_x += w * x[i]; } - return (n/QK_K*sizeof(block_q2_K)); + if (min > 0) { + min = 0; + } + if (max <= min) { + memset(L, 0, n); + *the_min = -min; + return 0.f; + } + float iscale = nmax/(max - min); + float scale = 1/iscale; + float best_mad = 0; + for (int i = 0; i < n; ++i) { + int l = nearest_int(iscale*(x[i] - min)); + L[i] = MAX(0, MIN(nmax, l)); + float diff = scale * L[i] + min - x[i]; + diff = use_mad ? fabsf(diff) : diff*diff; + float w = weights ? weights[i] : x[i]*x[i]; + best_mad += w * diff; + } + if (nstep < 1) { + *the_min = -min; + return scale; + } + for (int is = 0; is <= nstep; ++is) { + iscale = (rmin + rdelta*is + nmax)/(max - min); + float sum_l = 0, sum_l2 = 0, sum_xl = 0; + for (int i = 0; i < n; ++i) { + int l = nearest_int(iscale*(x[i] - min)); + l = MAX(0, MIN(nmax, l)); + Laux[i] = l; + float w = weights ? weights[i] : x[i]*x[i]; + sum_l += w*l; + sum_l2 += w*l*l; + sum_xl += w*l*x[i]; + } + float D = sum_w * sum_l2 - sum_l * sum_l; + if (D > 0) { + float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D; + float this_min = (sum_l2 * sum_x - sum_l * sum_xl)/D; + if (this_min > 0) { + this_min = 0; + this_scale = sum_xl / sum_l2; + } + float mad = 0; + for (int i = 0; i < n; ++i) { + float diff = this_scale * Laux[i] + this_min - x[i]; + diff = use_mad ? fabsf(diff) : diff*diff; + float w = weights ? weights[i] : x[i]*x[i]; + mad += w * diff; + } + if (mad < best_mad) { + for (int i = 0; i < n; ++i) { + L[i] = Laux[i]; + } + best_mad = mad; + scale = this_scale; + min = this_min; + } + } + } + *the_min = -min; + return scale; +} + +static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, const float * quant_weights) { + float max = 0; + for (int i = 0; i < n; ++i) { + max = MAX(max, x[i]); + } + if (!max) { // all zero + for (int i = 0; i < n; ++i) { L[i] = 0; } + return 0.f; + } + float iscale = nmax / max; + for (int i = 0; i < n; ++i) { + L[i] = nearest_int(iscale * x[i]); + } + float scale = 1/iscale; + float best_mse = 0; + for (int i = 0; i < n; ++i) { + float diff = x[i] - scale*L[i]; + float w = quant_weights[i]; + best_mse += w*diff*diff; + } + for (int is = -4; is <= 4; ++is) { + if (is == 0) continue; + float iscale_is = (0.1f*is + nmax)/max; + float scale_is = 1/iscale_is; + float mse = 0; + for (int i = 0; i < n; ++i) { + int l = nearest_int(iscale_is*x[i]); + l = MIN(nmax, l); + float diff = x[i] - scale_is*l; + float w = quant_weights[i]; + mse += w*diff*diff; + } + if (mse < best_mse) { + best_mse = mse; + iscale = iscale_is; + } + } + float sumlx = 0; + float suml2 = 0; + for (int i = 0; i < n; ++i) { + int l = nearest_int(iscale * x[i]); + l = MIN(nmax, l); + L[i] = l; + float w = quant_weights[i]; + sumlx += w*x[i]*l; + suml2 += w*l*l; + } + for (int itry = 0; itry < 5; ++itry) { + int n_changed = 0; + for (int i = 0; i < n; ++i) { + float w = quant_weights[i]; + float slx = sumlx - w*x[i]*L[i]; + float sl2 = suml2 - w*L[i]*L[i]; + if (slx > 0 && sl2 > 0) { + int new_l = nearest_int(x[i] * sl2 / slx); + new_l = MIN(nmax, new_l); + if (new_l != L[i]) { + slx += w*x[i]*new_l; + sl2 += w*new_l*new_l; + if (slx*slx*suml2 > sumlx*sumlx*sl2) { + L[i] = new_l; sumlx = slx; suml2 = sl2; + ++n_changed; + } + } + } + } + if (!n_changed) { + break; + } + } + return sumlx / suml2; +} + +static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) { + GGML_ASSERT(quant_weights); + assert(k % QK_K == 0); + const int nb = k / QK_K; + const bool requantize = true; + + uint8_t L[QK_K]; + uint8_t Laux[16]; + float mins[QK_K/16]; + float scales[QK_K/16]; + float sw[QK_K/16]; + float weight[16]; + uint8_t Ls[QK_K/16], Lm[QK_K/16]; + + for (int i = 0; i < nb; i++) { + memset(sw, 0, QK_K/16*sizeof(float)); + float sumx2 = 0; + for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j]; + float sigma2 = sumx2/QK_K; + for (int j = 0; j < QK_K/16; ++j) { + const float * restrict qw = quant_weights + QK_K * i + 16*j; + for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]); + for (int l = 0; l < QK_K/16; ++l) sw[j] += weight[l]; + scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false); + } + + float dm, mm; +#if QK_K == 64 + float max_scale = 0, max_min = 0; + for (int j = 0; j < QK_K/16; ++j) { + max_scale = MAX(max_scale, scales[j]); + max_min = MAX(max_min, mins[j]); + } + dm = max_scale/15; + mm = max_min/15; + if (max_scale) { + float id = 1/dm; + for (int j = 0; j < QK_K/16; ++j) { + int l = nearest_int(id*scales[j]); + Ls[j] = MAX(0, MIN(15, l)); + } + } else { + memset(Ls, 0, QK_K/16); + } + if (max_min) { + float id = 1/mm; + for (int j = 0; j < QK_K/16; ++j) { + int l = nearest_int(id*mins[j]); + Lm[j] = MAX(0, MIN(15, l)); + } + } else { + memset(Lm, 0, QK_K/16); + } +#else + dm = make_qp_quants(QK_K/16, 15, scales, Ls, sw); + mm = make_qp_quants(QK_K/16, 15, mins, Lm, sw); +#endif + y[i].d = GGML_FP32_TO_FP16(dm); + y[i].dmin = GGML_FP32_TO_FP16(mm); + dm = GGML_FP16_TO_FP32(y[i].d); + mm = GGML_FP16_TO_FP32(y[i].dmin); + + for (int j = 0; j < QK_K/16; ++j) { + y[i].scales[j] = Ls[j] | (Lm[j] << 4); + } + + if (requantize) { + for (int j = 0; j < QK_K/16; ++j) { + const float d = dm * (y[i].scales[j] & 0xF); + if (!d) continue; + const float m = mm * (y[i].scales[j] >> 4); + for (int ii = 0; ii < 16; ++ii) { + int l = nearest_int((x[16*j + ii] + m)/d); + l = MAX(0, MIN(3, l)); + L[16*j + ii] = l; + } + } + } + +#if QK_K == 256 + for (int j = 0; j < QK_K; j += 128) { + for (int l = 0; l < 32; ++l) { + y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6); + } + } +#else + for (int l = 0; l < 16; ++l) { + y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6); + } +#endif + + x += QK_K; + + } +} + +size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row); + if (!quant_weights) { + quantize_row_q2_K_reference(src, dst, (int64_t)nrow*n_per_row); + } + else { + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q2_K_impl(src, (block_q2_K*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + } + return nrow * row_size; } //========================= 3-bit (de)-quantization -void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k) { +void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1608,7 +2092,7 @@ void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict } #if QK_K == 256 -void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) { +void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1658,7 +2142,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int } } #else -void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) { +void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); assert(QK_K == 64); const int nb = k / QK_K; @@ -1691,23 +2175,118 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int } #endif -void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) { quantize_row_q3_K_reference(x, vy, k); } -size_t ggml_quantize_q3_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - (void)hist; // TODO: collect histograms +static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) { +#if QK_K != 256 + (void)quant_weights; + quantize_row_q3_K_reference(x, y, n_per_row); +#else + assert(n_per_row % QK_K == 0); + const int nb = n_per_row / QK_K; - for (int j = 0; j < n; j += k) { - block_q3_K * restrict y = (block_q3_K *)dst + j/QK_K; - quantize_row_q3_K_reference(src + j, y, k); + int8_t L[QK_K]; + float scales[QK_K / 16]; + float weight[16]; + float sw[QK_K / 16]; + int8_t Ls[QK_K / 16]; + + for (int i = 0; i < nb; i++) { + + float sumx2 = 0; + for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j]; + float sigma2 = 2*sumx2/QK_K; + + for (int j = 0; j < QK_K/16; ++j) { + if (quant_weights) { + const float * qw = quant_weights ? quant_weights + QK_K * i + 16*j : NULL; + for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j+l]*x[16*j+l]); + } else { + for (int l = 0; l < 16; ++l) weight[l] = x[16*j+l]*x[16*j+l]; + } + float sumw = 0; + for (int l = 0; l < 16; ++l) sumw += weight[l]; + sw[j] = sumw; + + scales[j] = make_qx_quants(16, 4, x + 16*j, L + 16*j, 1, weight); + + } + + memset(y[i].scales, 0, 12); + + float d_block = make_qx_quants(QK_K/16, 32, scales, Ls, 1, sw); + for (int j = 0; j < QK_K/16; ++j) { + int l = Ls[j]; + if (j < 8) { + y[i].scales[j] = l & 0xF; + } else { + y[i].scales[j-8] |= ((l & 0xF) << 4); + } + l >>= 4; + y[i].scales[j%4 + 8] |= (l << (2*(j/4))); + } + y[i].d = GGML_FP32_TO_FP16(d_block); + + int8_t sc; + for (int j = 0; j < QK_K/16; ++j) { + sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4; + sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32; + float d = GGML_FP16_TO_FP32(y[i].d) * sc; + if (!d) { + continue; + } + for (int ii = 0; ii < 16; ++ii) { + int l = nearest_int(x[16*j + ii]/d); + l = MAX(-4, MIN(3, l)); + L[16*j + ii] = l + 4; + } + } + + memset(y[i].hmask, 0, QK_K/8); + // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc. + int m = 0; + uint8_t hm = 1; + for (int j = 0; j < QK_K; ++j) { + if (L[j] > 3) { + y[i].hmask[m] |= hm; + L[j] -= 4; + } + if (++m == QK_K/8) { + m = 0; hm <<= 1; + } + } + for (int j = 0; j < QK_K; j += 128) { + for (int l = 0; l < 32; ++l) { + y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6); + } + } + + x += QK_K; } - return (n/QK_K*sizeof(block_q3_K)); +#endif +} + +size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row); + if (!quant_weights) { + quantize_row_q3_K_reference(src, dst, (int64_t)nrow*n_per_row); + } + else { + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q3_K_impl(src, (block_q3_K*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + } + return nrow * row_size; } // ====================== 4-bit (de)-quantization -void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k) { +void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1814,7 +2393,7 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict } } -void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k) { +void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1853,28 +2432,111 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int } } -void quantize_row_q4_K(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) { assert(k % QK_K == 0); block_q4_K * restrict y = vy; quantize_row_q4_K_reference(x, y, k); } -size_t ggml_quantize_q4_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - assert(k % QK_K == 0); - (void)hist; // TODO: collect histograms +static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) { +#if QK_K != 256 + (void)quant_weights; + quantize_row_q4_K_reference(x, y, n_per_row); +#else + assert(n_per_row % QK_K == 0); + const int64_t nb = n_per_row / QK_K; + + uint8_t L[QK_K]; + uint8_t Laux[32]; + uint8_t Ls[QK_K/32]; + uint8_t Lm[QK_K/32]; + float weights[32]; + float sw[QK_K/32]; + float mins[QK_K/32]; + float scales[QK_K/32]; + + for (int i = 0; i < nb; i++) { + + float sum_x2 = 0; + for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l]; + float sigma2 = 2*sum_x2/QK_K; + float av_x = sqrtf(sigma2); + + for (int j = 0; j < QK_K/32; ++j) { + if (quant_weights) { + const float * qw = quant_weights + QK_K*i + 32*j; + for (int l = 0; l < 32; ++l) weights[l] = qw[l] * sqrtf(sigma2 + x[32*j + l]*x[32*j + l]); + } else { + for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]); + } + float sumw = 0; + for (int l = 0; l < 32; ++l) sumw += weights[l]; + sw[j] = sumw; + scales[j] = make_qkx3_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false); + } + + float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw); + float m_block = make_qp_quants(QK_K/32, 63, mins, Lm, sw); + for (int j = 0; j < QK_K/32; ++j) { + uint8_t ls = Ls[j]; + uint8_t lm = Lm[j]; + if (j < 4) { + y[i].scales[j] = ls; + y[i].scales[j+4] = lm; + } else { + y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4); + y[i].scales[j-4] |= ((ls >> 4) << 6); + y[i].scales[j-0] |= ((lm >> 4) << 6); + } + } + y[i].d = GGML_FP32_TO_FP16(d_block); + y[i].dmin = GGML_FP32_TO_FP16(m_block); + + uint8_t sc, m; + for (int j = 0; j < QK_K/32; ++j) { + get_scale_min_k4(j, y[i].scales, &sc, &m); + const float d = GGML_FP16_TO_FP32(y[i].d) * sc; + if (!d) continue; + const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m; + for (int ii = 0; ii < 32; ++ii) { + int l = nearest_int((x[32*j + ii] + dm)/d); + l = MAX(0, MIN(15, l)); + L[32*j + ii] = l; + } + } + uint8_t * q = y[i].qs; + for (int j = 0; j < QK_K; j += 64) { + for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4); + q += 32; + } + + x += QK_K; - for (int j = 0; j < n; j += k) { - block_q4_K * restrict y = (block_q4_K *)dst + j/QK_K; - quantize_row_q4_K_reference(src + j, y, k); } - return (n/QK_K*sizeof(block_q4_K)); +#endif +} + +size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row); + if (!quant_weights) { + quantize_row_q4_K_reference(src, dst, (int64_t)nrow*n_per_row); + } + else { + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q4_K_impl(src, (block_q4_K*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + } + return nrow * row_size; } // ====================== 5-bit (de)-quantization -void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k) { +void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int64_t k) { assert(k % QK_K == 0); - const int nb = k / QK_K; + const int64_t nb = k / QK_K; #if QK_K == 256 uint8_t L[QK_K]; @@ -1965,7 +2627,7 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict #else float max_scale = 0, amax = 0; for (int j = 0; j < QK_K/16; ++j) { - scales[j] = make_qx_quants(16, 16, x + 16*j, L + 16*j, 1); + scales[j] = make_qx_quants(16, 16, x + 16*j, L + 16*j, 1, NULL); float abs_scale = fabsf(scales[j]); if (abs_scale > amax) { amax = abs_scale; @@ -2014,9 +2676,9 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict } } -void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k) { +void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); - const int nb = k / QK_K; + const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { @@ -2059,28 +2721,131 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int } } -void quantize_row_q5_K(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) { assert(k % QK_K == 0); block_q5_K * restrict y = vy; quantize_row_q5_K_reference(x, y, k); } -size_t ggml_quantize_q5_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - assert(k % QK_K == 0); - (void)hist; // TODO: collect histograms +static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) { +#if QK_K != 256 + (void)quant_weights; + quantize_row_q5_K_reference(x, y, n_per_row); +#else + assert(n_per_row % QK_K == 0); + const int64_t nb = n_per_row / QK_K; + + uint8_t L[QK_K]; + uint8_t Laux[32]; + uint8_t Ls[QK_K/32]; + uint8_t Lm[QK_K/32]; + float mins[QK_K/32]; + float scales[QK_K/32]; + float sw[QK_K/32]; + float weights[32]; + + for (int i = 0; i < nb; i++) { + + float sum_x2 = 0; + for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l]; + float sigma2 = 2*sum_x2/QK_K; + float av_x = sqrtf(sigma2); + + for (int j = 0; j < QK_K/32; ++j) { + if (quant_weights) { + const float * qw = quant_weights + QK_K*i + 32*j; + for (int l = 0; l < 32; ++l) weights[l] = qw[l] * sqrtf(sigma2 + x[32*j + l]*x[32*j + l]); + } else { + for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]); + } + float sumw = 0; + for (int l = 0; l < 32; ++l) sumw += weights[l]; + sw[j] = sumw; + + scales[j] = make_qkx3_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false); + } + + float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw); + float m_block = make_qp_quants(QK_K/32, 63, mins, Lm, sw); + + for (int j = 0; j < QK_K/32; ++j) { + uint8_t ls = Ls[j]; + uint8_t lm = Lm[j]; + ls = MIN(63, ls); + lm = MIN(63, lm); + if (j < 4) { + y[i].scales[j] = ls; + y[i].scales[j+4] = lm; + } else { + y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4); + y[i].scales[j-4] |= ((ls >> 4) << 6); + y[i].scales[j-0] |= ((lm >> 4) << 6); + } + } + y[i].d = GGML_FP32_TO_FP16(d_block); + y[i].dmin = GGML_FP32_TO_FP16(m_block); + + uint8_t sc, m; + for (int j = 0; j < QK_K/32; ++j) { + get_scale_min_k4(j, y[i].scales, &sc, &m); + const float d = GGML_FP16_TO_FP32(y[i].d) * sc; + if (!d) continue; + const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m; + for (int ii = 0; ii < 32; ++ii) { + int l = nearest_int((x[32*j + ii] + dm)/d); + l = MAX(0, MIN(31, l)); + L[32*j + ii] = l; + } + } + + uint8_t * restrict qh = y[i].qh; + uint8_t * restrict ql = y[i].qs; + memset(qh, 0, QK_K/8); + + uint8_t m1 = 1, m2 = 2; + for (int n = 0; n < QK_K; n += 64) { + for (int j = 0; j < 32; ++j) { + int l1 = L[n + j]; + if (l1 > 15) { + l1 -= 16; qh[j] |= m1; + } + int l2 = L[n + j + 32]; + if (l2 > 15) { + l2 -= 16; qh[j] |= m2; + } + ql[j] = l1 | (l2 << 4); + } + m1 <<= 2; m2 <<= 2; + ql += 32; + } + + x += QK_K; - for (int j = 0; j < n; j += k) { - block_q5_K * restrict y = (block_q5_K *)dst + j/QK_K; - quantize_row_q5_K_reference(src + j, y, k); } - return (n/QK_K*sizeof(block_q5_K)); +#endif +} + +size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row); + if (!quant_weights) { + quantize_row_q5_K_reference(src, dst, (int64_t)nrow*n_per_row); + } + else { + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q5_K_impl(src, (block_q5_K*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + } + return nrow * row_size; } // ====================== 6-bit (de)-quantization -void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k) { +void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int64_t k) { assert(k % QK_K == 0); - const int nb = k / QK_K; + const int64_t nb = k / QK_K; int8_t L[QK_K]; float scales[QK_K/16]; @@ -2092,7 +2857,7 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict for (int ib = 0; ib < QK_K/16; ++ib) { - const float scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1); + const float scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, NULL); scales[ib] = scale; const float abs_scale = fabsf(scale); @@ -2160,9 +2925,9 @@ void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict } } -void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k) { +void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); - const int nb = k / QK_K; + const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { @@ -2207,28 +2972,619 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int } } -void quantize_row_q6_K(const float * restrict x, void * restrict vy, int k) { +void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) { assert(k % QK_K == 0); block_q6_K * restrict y = vy; quantize_row_q6_K_reference(x, y, k); } -size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK_K == 0); - (void)hist; // TODO: collect histograms +static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) { +#if QK_K != 256 + (void)quant_weights; + quantize_row_q6_K_reference(x, y, n_per_row); +#else + assert(n_per_row % QK_K == 0); + const int64_t nb = n_per_row / QK_K; + + int8_t L[QK_K]; + float scales[QK_K/16]; + //float weights[16]; + + for (int i = 0; i < nb; i++) { + + //float sum_x2 = 0; + //for (int j = 0; j < QK_K; ++j) sum_x2 += x[j]*x[j]; + //float sigma2 = sum_x2/QK_K; + + float max_scale = 0; + float max_abs_scale = 0; + + for (int ib = 0; ib < QK_K/16; ++ib) { + + float scale; + if (quant_weights) { + const float * qw = quant_weights + QK_K*i + 16*ib; + //for (int j = 0; j < 16; ++j) weights[j] = qw[j] * sqrtf(sigma2 + x[16*ib + j]*x[16*ib + j]); + //scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, weights); + scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, qw); + } else { + scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, NULL); + } + scales[ib] = scale; + + const float abs_scale = fabsf(scale); + if (abs_scale > max_abs_scale) { + max_abs_scale = abs_scale; + max_scale = scale; + } + + } + + if (!max_abs_scale) { + memset(&y[i], 0, sizeof(block_q6_K)); + y[i].d = GGML_FP32_TO_FP16(0.f); + x += QK_K; + continue; + } + + float iscale = -128.f/max_scale; + y[i].d = GGML_FP32_TO_FP16(1/iscale); + for (int ib = 0; ib < QK_K/16; ++ib) { + y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib])); + } + + for (int j = 0; j < QK_K/16; ++j) { + float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j]; + if (!d) { + continue; + } + for (int ii = 0; ii < 16; ++ii) { + int l = nearest_int(x[16*j + ii]/d); + l = MAX(-32, MIN(31, l)); + L[16*j + ii] = l + 32; + } + } + + uint8_t * restrict ql = y[i].ql; + uint8_t * restrict qh = y[i].qh; + for (int j = 0; j < QK_K; j += 128) { + for (int l = 0; l < 32; ++l) { + const uint8_t q1 = L[j + l + 0] & 0xF; + const uint8_t q2 = L[j + l + 32] & 0xF; + const uint8_t q3 = L[j + l + 64] & 0xF; + const uint8_t q4 = L[j + l + 96] & 0xF; + ql[l+ 0] = q1 | (q3 << 4); + ql[l+32] = q2 | (q4 << 4); + qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6); + } + ql += 64; + qh += 32; + } + + x += QK_K; - for (int j = 0; j < n; j += k) { - block_q6_K * restrict y = (block_q6_K *)dst + j/QK_K; - quantize_row_q6_K_reference(src + j, y, k); } - return (n/QK_K*sizeof(block_q6_K)); +#endif +} + +size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row); + if (!quant_weights) { + quantize_row_q6_K_reference(src, dst, (int64_t)nrow*n_per_row); + } + else { + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q6_K_impl(src, (block_q6_K*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + } + return nrow * row_size; +} + +static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restrict y, int64_t n_per_row, const float * quant_weights) { + static_assert(QK4_0 == 32, "QK4_0 must be 32"); + + if (!quant_weights) { + quantize_row_q4_0_reference(x, y, n_per_row); + return; + } + + float weight[QK4_0]; + int8_t L[QK4_0]; + + float sum_x2 = 0; + for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j]; + float sigma2 = sum_x2/n_per_row; + + const int64_t nb = n_per_row/QK4_0; + for (int ib = 0; ib < nb; ++ib) { + const float * xb = x + QK4_0 * ib; + const float * qw = quant_weights + QK4_0 * ib; + for (int j = 0; j < QK4_0; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]); + float d = make_qx_quants(QK4_0, 8, xb, L, 1, weight); + y[ib].d = GGML_FP32_TO_FP16(d); + for (int j = 0; j < 16; ++j) { + y[ib].qs[j] = L[j] | (L[j+16] << 4); + } + } +} + +size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + quantize_row_q4_0_reference(src, dst, (int64_t)nrow*n_per_row); + return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row); + } + size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row); + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q4_0_impl(src, (block_q4_0*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + return nrow * row_size; +} + +static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restrict y, int64_t n_per_row, const float * quant_weights) { + static_assert(QK4_1 == 32, "QK4_1 must be 32"); + + if (!quant_weights) { + quantize_row_q4_1_reference(x, y, n_per_row); + return; + } + + float weight[QK4_1]; + uint8_t L[QK4_1], Laux[QK4_1]; + + float sum_x2 = 0; + for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j]; + float sigma2 = sum_x2/n_per_row; + + const int64_t nb = n_per_row/QK4_1; + for (int ib = 0; ib < nb; ++ib) { + const float * xb = x + QK4_1 * ib; + const float * qw = quant_weights + QK4_1 * ib; + for (int j = 0; j < QK4_1; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]); + float min; + float d = make_qkx3_quants(QK4_1, 15, xb, weight, L, &min, Laux, -0.9f, 0.05f, 36, false); + y[ib].d = GGML_FP32_TO_FP16(d); + y[ib].m = GGML_FP32_TO_FP16(-min); + for (int j = 0; j < 16; ++j) { + y[ib].qs[j] = L[j] | (L[j+16] << 4); + } + } +} + +size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + quantize_row_q4_1_reference(src, dst, (int64_t)nrow*n_per_row); + return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row); + } + size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row); + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q4_1_impl(src, (block_q4_1*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + return nrow * row_size; +} + +static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restrict y, int64_t n_per_row, const float * quant_weights) { + static_assert(QK5_0 == 32, "QK5_0 must be 32"); + + if (!quant_weights) { + quantize_row_q5_0_reference(x, y, n_per_row); + return; + } + + float weight[QK5_0]; + int8_t L[QK5_0]; + + float sum_x2 = 0; + for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j]; + float sigma2 = sum_x2/n_per_row; + + const int64_t nb = n_per_row/QK5_0; + for (int ib = 0; ib < nb; ++ib) { + const float * xb = x + QK5_0 * ib; + const float * qw = quant_weights + QK5_0 * ib; + for (int j = 0; j < QK5_0; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]); + float d = make_qx_quants(QK5_0, 16, xb, L, 1, weight); + y[ib].d = GGML_FP32_TO_FP16(d); + + uint32_t qh = 0; + + for (int j = 0; j < 16; ++j) { + const uint8_t xi0 = L[j]; + const uint8_t xi1 = L[j+16]; + y[ib].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2); + } + + memcpy(&y[ib].qh, &qh, sizeof(qh)); + } +} + +size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + quantize_row_q5_0_reference(src, dst, (int64_t)nrow*n_per_row); + return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row); + } + size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row); + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q5_0_impl(src, (block_q5_0*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + return nrow * row_size; +} + +static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restrict y, int64_t n_per_row, const float * quant_weights) { + static_assert(QK5_1 == 32, "QK5_1 must be 32"); + + if (!quant_weights) { + quantize_row_q5_1_reference(x, y, n_per_row); + return; + } + + float weight[QK5_1]; + uint8_t L[QK5_1], Laux[QK5_1]; + + float sum_x2 = 0; + for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j]; + float sigma2 = sum_x2/n_per_row; + + const int64_t nb = n_per_row/QK5_1; + for (int ib = 0; ib < nb; ++ib) { + const float * xb = x + QK5_1 * ib; + const float * qw = quant_weights + QK5_1 * ib; + for (int j = 0; j < QK5_1; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]); + float min; + float d = make_qkx3_quants(QK5_1, 31, xb, weight, L, &min, Laux, -0.9f, 0.05f, 36, false); + y[ib].d = GGML_FP32_TO_FP16(d); + y[ib].m = GGML_FP32_TO_FP16(-min); + + uint32_t qh = 0; + for (int j = 0; j < 16; ++j) { + const uint8_t xi0 = L[j]; + const uint8_t xi1 = L[j+16]; + y[ib].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2); + } + memcpy(&y[ib].qh, &qh, sizeof(qh)); + } +} + +size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + if (!quant_weights) { + quantize_row_q5_1_reference(src, dst, (int64_t)nrow*n_per_row); + return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row); + } + size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row); + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_q5_1_impl(src, (block_q5_1*)qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += row_size; + } + return nrow * row_size; +} + +size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + (void)quant_weights; // not used + const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row); + quantize_row_q8_0_reference(src, dst, (int64_t)nrow*n_per_row); + return nrow * row_size; +} + +// ====================== "True" 2-bit (de)-quantization + +void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + uint32_t aux32[2]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + memcpy(aux32, x[i].qs + 4*ib32, 2*sizeof(uint32_t)); + const float db = d * (0.5f + (aux32[1] >> 28)) * 0.25f; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; + for (int j = 0; j < 8; ++j) { + y[j] = db * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + } + y += 8; + } + } + } +} + +// ====================== 2.3125 bpw (de)-quantization + +void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + float db[2]; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f; + db[1] = d * (0.5f + (x[i].scales[ib32] >> 4)) * 0.25f; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (x[i].qs[4*ib32 + l] & 511)); + const uint8_t signs = ksigns_iq2xs[x[i].qs[4*ib32 + l] >> 9]; + for (int j = 0; j < 8; ++j) { + y[j] = db[l/2] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + } + y += 8; + } + } + } +} + +// ====================== 2.5625 bpw (de)-quantization + +void dequantize_row_iq2_s(const block_iq2_s * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + float db[2]; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + const uint8_t * signs = qs + QK_K/8; + + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f; + db[1] = d * (0.5f + (x[i].scales[ib32] >> 4)) * 0.25f; + for (int l = 0; l < 4; ++l) { + const float dl = db[l/2]; + const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); + for (int j = 0; j < 8; ++j) { + y[j] = dl * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1.f : 1.f); + } + y += 8; + } + qs += 4; + signs += 4; + } + } +} + +// ====================== 3.0625 bpw (de)-quantization + +void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + uint32_t aux32; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + const uint8_t * qs = x[i].qs; + const uint8_t * scales_and_signs = qs + QK_K/4; + + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + memcpy(&aux32, scales_and_signs + 4*ib32, sizeof(uint32_t)); + const float db = d * (0.5f + (aux32 >> 28)) * 0.5f; + for (int l = 0; l < 4; ++l) { + const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; + const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + qs[2*l+0]); + const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + qs[2*l+1]); + for (int j = 0; j < 4; ++j) { + y[j+0] = db * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = db * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + } + y += 8; + } + qs += 8; + } + } +} + +// ====================== 3.3125 bpw (de)-quantization + +void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + const uint8_t * signs = x[i].signs; + + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const float db1 = d * (1 + 2*(x[i].scales[ib32/2] & 0xf)); + const float db2 = d * (1 + 2*(x[i].scales[ib32/2] >> 4)); + for (int l = 0; l < 4; ++l) { + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256))); + for (int j = 0; j < 4; ++j) { + y[j+0] = db1 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = db1 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f); + } + y += 8; + } + qs += 8; + signs += 4; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256))); + for (int j = 0; j < 4; ++j) { + y[j+0] = db2 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = db2 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f); + } + y += 8; + } + qh += 2; + qs += 8; + signs += 4; + } + } +} + +// ====================== 1.5625 bpw (de)-quantization + +void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + const uint8_t * qs = x[i].qs; + const uint16_t * qh = x[i].qh; + + for (int ib = 0; ib < QK_K/32; ++ib) { + const float dl = d * (2*((qh[ib] >> 12) & 7) + 1); + const float delta = qh[ib] & 0x8000 ? -IQ1S_DELTA : IQ1S_DELTA; + for (int l = 0; l < 4; ++l) { + const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); + for (int j = 0; j < 8; ++j) { + y[j] = dl * (grid[j] + delta); + } + y += 8; + } + qs += 4; + } + } +} + +void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + + float delta[4]; + uint16_t idx[4]; + +#if QK_K != 64 + iq1m_scale_t scale; +#endif + + for (int i = 0; i < nb; i++) { + + const uint16_t * sc = (const uint16_t *)x[i].scales; +#if QK_K == 64 + const float d = GGML_FP16_TO_FP32(x[i].d); +#else + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); + const float d = GGML_FP16_TO_FP32(scale.f16); +#endif + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + + for (int ib = 0; ib < QK_K/32; ++ib) { +#if QK_K == 64 + const float dl1 = d * (2*((sc[ib/2] >> (8*(ib%2)+0)) & 0xf) + 1); + const float dl2 = d * (2*((sc[ib/2] >> (8*(ib%2)+4)) & 0xf) + 1); +#else + const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1); + const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1); +#endif + idx[0] = qs[0] | ((qh[0] << 8) & 0x700); + idx[1] = qs[1] | ((qh[0] << 4) & 0x700); + idx[2] = qs[2] | ((qh[1] << 8) & 0x700); + idx[3] = qs[3] | ((qh[1] << 4) & 0x700); + delta[0] = qh[0] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA; + delta[1] = qh[0] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA; + delta[2] = qh[1] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA; + delta[3] = qh[1] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA; + for (int l = 0; l < 2; ++l) { + const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]); + for (int j = 0; j < 8; ++j) { + y[j] = dl1 * (grid[j] + delta[l]); + } + y += 8; + } + for (int l = 2; l < 4; ++l) { + const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]); + for (int j = 0; j < 8; ++j) { + y[j] = dl2 * (grid[j] + delta[l]); + } + y += 8; + } + qs += 4; + qh += 2; + } + } +} + +static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; + +void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int64_t k) { + assert(k % QK4_NL == 0); + const int64_t nb = k / QK4_NL; + + for (int i = 0; i < nb; i++) { + + const uint8_t * qs = x[i].qs; + + const float d = GGML_FP16_TO_FP32(x[i].d); + for (int j = 0; j < QK4_NL/2; ++j) { + y[j+ 0] = d * kvalues_iq4nl[qs[j] & 0xf]; + y[j+QK4_NL/2] = d * kvalues_iq4nl[qs[j] >> 4]; + } + y += QK4_NL; + qs += QK4_NL/2; + } +} + +void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) { + assert(k % QK_K == 0); +#if QK_K == 64 + dequantize_row_iq4_nl((const block_iq4_nl *)x, y, k); +#else + const int64_t nb = k / QK_K; + + for (int i = 0; i < nb; i++) { + + const uint8_t * qs = x[i].qs; + + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int ib = 0; ib < QK_K/32; ++ib) { + const int ls = ((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4); + const float dl = d * (ls - 32); + for (int j = 0; j < 16; ++j) { + y[j+ 0] = dl * kvalues_iq4nl[qs[j] & 0xf]; + y[j+16] = dl * kvalues_iq4nl[qs[j] >> 4]; + } + y += 32; + qs += 16; + } + } +#endif } //===================================== Q8_K ============================================== -void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) { +void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int64_t k) { assert(k % QK_K == 0); - const int nb = k / QK_K; + const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { @@ -2246,7 +3602,9 @@ void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict x += QK_K; continue; } - const float iscale = -128.f/max; + //const float iscale = -128.f/max; + // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward + const float iscale = -127.f/max; for (int j = 0; j < QK_K; ++j) { int v = nearest_int(iscale*x[j]); y[i].qs[j] = MIN(127, v); @@ -2263,9 +3621,9 @@ void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict } } -void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k) { +void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int64_t k) { assert(k % QK_K == 0); - const int nb = k / QK_K; + const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { for (int j = 0; j < QK_K; ++j) { @@ -2274,7 +3632,7 @@ void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int } } -void quantize_row_q8_K(const float * restrict x, void * restrict y, int k) { +void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) { quantize_row_q8_K_reference(x, y, k); } @@ -2323,15 +3681,92 @@ static inline __m128i get_scale_shuffle(int i) { } #endif -void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; assert(n % qk == 0); +#if defined(__ARM_FEATURE_MATMUL_INT8) + assert((nrc == 2) || (nrc == 1)); +#else + assert(nrc == 1); +#endif + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q4_0 * restrict x = vx; const block_q8_0 * restrict y = vy; +#if defined(__ARM_FEATURE_MATMUL_INT8) + if (nrc == 2) { + const block_q4_0 * restrict vx0 = vx; + const block_q4_0 * restrict vx1 = vx + bx; + + const block_q8_0 * restrict vy0 = vy; + const block_q8_0 * restrict vy1 = vy + by; + + float32x4_t sumv0 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i++) { + const block_q4_0 * restrict b_x0 = &vx0[i]; + const block_q4_0 * restrict b_x1 = &vx1[i]; + const block_q8_0 * restrict b_y0 = &vy0[i]; + const block_q8_0 * restrict b_y1 = &vy1[i]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + const int8x16_t s8b = vdupq_n_s8(0x8); + + const uint8x16_t v0_0 = vld1q_u8(b_x0->qs); + const uint8x16_t v0_1 = vld1q_u8(b_x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // sub 8 + const int8x16_t x0_l = vsubq_s8(v0_0l, s8b); + const int8x16_t x0_h = vsubq_s8(v0_0h, s8b); + const int8x16_t x1_l = vsubq_s8(v0_1l, s8b); + const int8x16_t x1_h = vsubq_s8(v0_1h, s8b); + + // load y + const int8x16_t y0_l = vld1q_s8(b_y0->qs); + const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16); + const int8x16_t y1_l = vld1q_s8(b_y1->qs); + const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); + + float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)}; + + int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); + int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); + + int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h))); + int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h))); + + int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l))); + int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l))); + + int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); + int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); + + sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)), + l1, r1)), l2, r2)), l3, r3))), scale); + } + float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2); + float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1); + + vst1_f32(s, vget_low_f32(sumv2)); + vst1_f32(s + bs, vget_high_f32(sumv2)); + return; + } +#endif #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -2368,32 +3803,12 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); -#if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); + const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); + const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#endif } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); @@ -2406,15 +3821,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, /* Compute combined scale for the block */ const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); - __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i qx = bytes_from_nibbles_32(x[i].qs); // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. const __m256i off = _mm256_set1_epi8( 8 ); - bx = _mm256_sub_epi8( bx, off ); + qx = _mm256_sub_epi8( qx, off ); - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_i8_pairs_float(bx, by); + const __m256 q = mul_sum_i8_pairs_float(qx, qy); /* Multiply q with scale and accumulate */ acc = _mm256_fmadd_ps( d, q, acc ); @@ -2435,15 +3850,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs); - __m128i bx = _mm_and_si128(lowMask, tmp); - __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs); - bx = _mm_sub_epi8(bx, off); - const __m128i i32_0 = mul_sum_i8_pairs(bx, by); + __m128i bx_0 = _mm_and_si128(lowMask, tmp); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); - bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); - by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); - bx = _mm_sub_epi8(bx, off); - const __m128i i32_1 = mul_sum_i8_pairs(bx, by); + bx_0 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); + by_0 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_0, by_0); // Convert int32_t to float __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1)); @@ -2633,15 +4048,93 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, #endif } -void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { const int qk = QK8_1; const int nb = n / qk; assert(n % qk == 0); +#if defined(__ARM_FEATURE_MATMUL_INT8) + assert((nrc == 2) || (nrc == 1)); +#else + assert(nrc == 1); +#endif + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q4_1 * restrict x = vx; const block_q8_1 * restrict y = vy; +#if defined(__ARM_FEATURE_MATMUL_INT8) + if (nrc == 2) { + const block_q4_1 * restrict vx0 = vx; + const block_q4_1 * restrict vx1 = vx + bx; + const block_q8_1 * restrict vy0 = vy; + const block_q8_1 * restrict vy1 = vy + by; + + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t summs0 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i++) { + const block_q4_1 * restrict b_x0 = &vx0[i]; + const block_q4_1 * restrict b_x1 = &vx1[i]; + const block_q8_1 * restrict b_y0 = &vy0[i]; + const block_q8_1 * restrict b_y1 = &vy1[i]; + + float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s), + GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s), + GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s), + GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)}; + summs0 += summs_t; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + const uint8x16_t v0_0 = vld1q_u8(b_x0->qs); + const uint8x16_t v0_1 = vld1q_u8(b_x1->qs); + + // 4-bit -> 8-bit + const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // load y + const int8x16_t y0_l = vld1q_s8(b_y0->qs); + const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16); + const int8x16_t y1_l = vld1q_s8(b_y1->qs); + const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); + + // mmla into int32x4_t + float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d, + GGML_FP16_TO_FP32(b_x0->d)*b_y1->d, + GGML_FP16_TO_FP32(b_x1->d)*b_y0->d, + GGML_FP16_TO_FP32(b_x1->d)*b_y1->d}; + + int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); + int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); + + int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h))); + int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h))); + + int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l))); + int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l))); + + int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); + int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); + sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)), + l1, r1)), l2, r2)), l3, r3))), scale); + } + + float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2); + float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1); + sumv2 = sumv2 + summs0; + + vst1_f32(s, vget_low_f32(sumv2)); + vst1_f32(s + bs, vget_high_f32(sumv2)); + return; + } +#endif // TODO: add WASM SIMD #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); @@ -2657,7 +4150,7 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri const block_q8_1 * restrict y0 = &y[i + 0]; const block_q8_1 * restrict y1 = &y[i + 1]; - summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s; + summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -2676,32 +4169,12 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); -#if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); + const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); + const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); -#endif + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; @@ -2714,9 +4187,9 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri // Main loop for (int i = 0; i < nb; ++i) { const float d0 = GGML_FP16_TO_FP32(x[i].d); - const float d1 = y[i].d; + const float d1 = GGML_FP16_TO_FP32(y[i].d); - summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s); const __m256 d0v = _mm256_set1_ps( d0 ); const __m256 d1v = _mm256_set1_ps( d1 ); @@ -2725,10 +4198,10 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes - const __m256i bx = bytes_from_nibbles_32(x[i].qs); - const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); + const __m256i qx = bytes_from_nibbles_32(x[i].qs); + const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[i].qs ); - const __m256 xy = mul_sum_us8_pairs_float(bx, by); + const __m256 xy = mul_sum_us8_pairs_float(qx, qy); // Accumulate d0*d1*x*y #if defined(__AVX2__) @@ -2768,7 +4241,7 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s); } *s = sumf; @@ -2786,19 +4259,24 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s); } *s = sumf; #endif } -void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; assert(n % qk == 0); assert(qk == QK5_0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q5_0 * restrict x = vx; const block_q8_0 * restrict y = vy; @@ -2863,32 +4341,12 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); -#if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#endif + ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); @@ -2962,14 +4420,14 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri /* Compute combined scale for the block */ const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); - __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i qx = bytes_from_nibbles_32(x[i].qs); __m256i bxhi = bytes_from_bits_32(x[i].qh); bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0)); - bx = _mm256_or_si256(bx, bxhi); + qx = _mm256_or_si256(qx, bxhi); - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_i8_pairs_float(bx, by); + const __m256 q = mul_sum_i8_pairs_float(qx, qy); /* Multiply q with scale and accumulate */ acc = _mm256_fmadd_ps(d, q, acc); @@ -2986,21 +4444,21 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri /* Compute combined scale for the block */ const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); - __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bx_0 = bytes_from_nibbles_32(x[i].qs); const __m256i bxhi = bytes_from_bits_32(x[i].qh); __m128i bxhil = _mm256_castsi256_si128(bxhi); __m128i bxhih = _mm256_extractf128_si256(bxhi, 1); bxhil = _mm_andnot_si128(bxhil, mask); bxhih = _mm_andnot_si128(bxhih, mask); - __m128i bxl = _mm256_castsi256_si128(bx); - __m128i bxh = _mm256_extractf128_si256(bx, 1); + __m128i bxl = _mm256_castsi256_si128(bx_0); + __m128i bxh = _mm256_extractf128_si256(bx_0, 1); bxl = _mm_or_si128(bxl, bxhil); bxh = _mm_or_si128(bxh, bxhih); - bx = MM256_SET_M128I(bxh, bxl); + bx_0 = MM256_SET_M128I(bxh, bxl); - const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_i8_pairs_float(bx, by); + const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0); /* Multiply q with scale and accumulate */ acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc); @@ -3014,7 +4472,7 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri size_t vl = __riscv_vsetvl_e8m1(qk/2); - // These tempory registers are for masking and shift operations + // These temporary registers are for masking and shift operations vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl); @@ -3099,12 +4557,17 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri #endif } -void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { const int qk = QK8_1; const int nb = n / qk; assert(n % qk == 0); assert(qk == QK5_1); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q5_1 * restrict x = vx; const block_q8_1 * restrict y = vy; @@ -3132,8 +4595,8 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri const uint8x16_t m4b = vdupq_n_u8(0x0F); - summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s; - summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s; + summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); + summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); // extract the 5th bit via lookup table ((b) << 4) memcpy(&qh0, x0->qh, sizeof(qh0)); @@ -3175,32 +4638,12 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); -#if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d); + ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); - - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); - const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); -#endif + ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; @@ -3217,7 +4660,7 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri const block_q5_1 * restrict x0 = &x[i]; const block_q8_1 * restrict y0 = &y[i]; - summs += GGML_FP16_TO_FP32(x0->m) * y0->s; + summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); const v128_t m4b = wasm_i8x16_splat(0x0F); @@ -3264,7 +4707,7 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri wasm_i32x4_dot_i16x8(v0lfh, v1lh)), wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), - wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d))); + wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); } *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + @@ -3279,17 +4722,17 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; i++) { const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); - summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s); - __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i qx = bytes_from_nibbles_32(x[i].qs); __m256i bxhi = bytes_from_bits_32(x[i].qh); bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); - bx = _mm256_or_si256(bx, bxhi); + qx = _mm256_or_si256(qx, bxhi); - const __m256 dy = _mm256_set1_ps(y[i].d); - const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d)); + const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_us8_pairs_float(bx, by); + const __m256 q = mul_sum_us8_pairs_float(qx, qy); acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); } @@ -3306,24 +4749,24 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; i++) { const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); - summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; + summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s); - __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bx_0 = bytes_from_nibbles_32(x[i].qs); const __m256i bxhi = bytes_from_bits_32(x[i].qh); __m128i bxhil = _mm256_castsi256_si128(bxhi); __m128i bxhih = _mm256_extractf128_si256(bxhi, 1); bxhil = _mm_and_si128(bxhil, mask); bxhih = _mm_and_si128(bxhih, mask); - __m128i bxl = _mm256_castsi256_si128(bx); - __m128i bxh = _mm256_extractf128_si256(bx, 1); + __m128i bxl = _mm256_castsi256_si128(bx_0); + __m128i bxh = _mm256_extractf128_si256(bx_0, 1); bxl = _mm_or_si128(bxl, bxhil); bxh = _mm_or_si128(bxh, bxhih); - bx = MM256_SET_M128I(bxh, bxl); + bx_0 = MM256_SET_M128I(bxh, bxl); - const __m256 dy = _mm256_set1_ps(y[i].d); - const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d)); + const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_us8_pairs_float(bx, by); + const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0); acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc); } @@ -3387,7 +4830,7 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s); } *s = sumf; @@ -3411,22 +4854,86 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); } - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s); } *s = sumf; #endif } -void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; assert(n % qk == 0); +#if defined(__ARM_FEATURE_MATMUL_INT8) + assert((nrc == 2) || (nrc == 1)); +#else + assert(nrc == 1); +#endif + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q8_0 * restrict x = vx; const block_q8_0 * restrict y = vy; +#if defined(__ARM_FEATURE_MATMUL_INT8) + if (nrc == 2) { + const block_q8_0 * restrict vx0 = vx; + const block_q8_0 * restrict vx1 = vx + bx; + const block_q8_0 * restrict vy0 = vy; + const block_q8_0 * restrict vy1 = vy + by; + + float32x4_t sumv0 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i++) { + const block_q8_0 * restrict b_x0 = &vx0[i]; + const block_q8_0 * restrict b_y0 = &vy0[i]; + + const block_q8_0 * restrict b_x1 = &vx1[i]; + const block_q8_0 * restrict b_y1 = &vy1[i]; + + const int8x16_t x0_l = vld1q_s8(b_x0->qs); + const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16); + const int8x16_t x1_l = vld1q_s8(b_x1->qs); + const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16); + + // load y + const int8x16_t y0_l = vld1q_s8(b_y0->qs); + const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16); + const int8x16_t y1_l = vld1q_s8(b_y1->qs); + const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); + + float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), + GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)}; + + int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); + int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); + + int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h))); + int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h))); + + int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l))); + int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l))); + + int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); + int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h))); + + sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)), + l1, r1)), l2, r2)), l3, r3))), scale); + } + float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2); + float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1); + + vst1_f32(s, vget_low_f32(sumv2)); + vst1_f32(s + bs, vget_high_f32(sumv2)); + return; + } +#endif #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -3450,34 +4957,13 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri const int8x16_t y1_0 = vld1q_s8(y1->qs); const int8x16_t y1_1 = vld1q_s8(y1->qs + 16); -#if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), - vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), + ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), - vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); - -#else - const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0)); - const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0)); - const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1)); - const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1)); - - const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0)); - const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0)); - const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1)); - const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1)); - - const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1)); - const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3)); - const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1)); - const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3)); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); -#endif + ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), + ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); @@ -3489,10 +4975,10 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { // Compute combined scale for the block const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); - __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs); - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + __m256i qx = _mm256_loadu_si256((const __m256i *)x[i].qs); + __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_i8_pairs_float(bx, by); + const __m256 q = mul_sum_i8_pairs_float(qx, qy); // Multiply q with scale and accumulate #if defined(__AVX2__) @@ -3509,10 +4995,10 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; i++) { // load elements - vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl); - vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl); + vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[i].qs, vl); + vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[i].qs, vl); - vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl); + vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl); vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl); vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl); @@ -3542,7 +5028,12 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri } #if QK_K == 256 -void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q2_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -3550,20 +5041,17 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - const uint8x16_t m3 = vdupq_n_u8(0x3); const uint8x16_t m4 = vdupq_n_u8(0xF); -#if defined(__ARM_FEATURE_DOTPROD) - const int32x4_t vzero = vdupq_n_s32(0); -#endif - int8x16x2_t q2bytes; + const int32x4_t vzero = vdupq_n_s32(0); + + ggml_int8x16x2_t q2bytes; uint8_t aux[16]; float sum = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); @@ -3576,8 +5064,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri vst1q_u8(aux, scales); const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4); - const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums); - const int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}; + const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); + const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}}; const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])), vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0]))); const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])), @@ -3589,47 +5077,33 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri // We use this macro instead of a function call because for some reason // the code runs 2-3% slower, even if the function is declared inline -#if defined(__ARM_FEATURE_DOTPROD) #define MULTIPLY_ACCUM_WITH_SCALE(index)\ - isum += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\ - isum += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)]; -#else -#define MULTIPLY_ACCUM_WITH_SCALE(index)\ - {\ - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])),\ - vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0])));\ - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])),\ - vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1])));\ - isum += vaddvq_s16(p1) * aux[is+(index)] + vaddvq_s16(p2) * aux[is+1+(index)];\ - } -#endif + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\ + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)]; #define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\ - q8bytes = vld1q_s8_x2(q8); q8 += 32;\ + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\ q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\ q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\ MULTIPLY_ACCUM_WITH_SCALE((index)); - for (int j = 0; j < QK_K/128; ++j) { + const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32; - const uint8x16x2_t q2bits = vld1q_u8_x2(q2); q2 += 32; - - int8x16x2_t q8bytes = vld1q_s8_x2(q8); q8 += 32; + ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3)); q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3)); + MULTIPLY_ACCUM_WITH_SCALE(0); SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2); - SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4); - SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6); is += 8; } - sum += d * isum; + sum += d * isum; } *s = sum; @@ -3935,7 +5409,12 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri #else -void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q2_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -3943,13 +5422,11 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - const uint8x16_t m3 = vdupq_n_u8(0x3); -#if defined(__ARM_FEATURE_DOTPROD) - const int32x4_t vzero = vdupq_n_s32(0); -#endif - int8x16x4_t q2bytes; + const int32x4_t vzero = vdupq_n_s32(0); + + ggml_int8x16x4_t q2bytes; uint32_t aux32[2]; const uint8_t * scales = (const uint8_t *)aux32; @@ -3958,8 +5435,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * (float)x[i].d; - const float dmin = -y[i].d * (float)x[i].dmin; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -3974,35 +5451,19 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t q2bits = vld1q_u8(q2); - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3)); q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3)); q2bytes.val[2] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 4), m3)); q2bytes.val[3] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 6), m3)); -#if defined(__ARM_FEATURE_DOTPROD) - isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0]; - isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1]; - isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2]; - isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3]; -#else - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - isum1 += vaddvq_s16(p1) * scales[0]; - isum2 += vaddvq_s16(p2) * scales[1]; + isum1 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0]; + isum2 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1]; + isum1 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2]; + isum2 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3]; - const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q2bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - const int16x8_t p4 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q2bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - isum1 += vaddvq_s16(p3) * scales[2]; - isum2 += vaddvq_s16(p4) * scales[3]; -#endif sum += d * (isum1 + isum2); - } *s = sum; @@ -4124,8 +5585,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * (float)x[i].d; - const float dmin = -y[i].d * (float)x[i].dmin; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -4177,7 +5638,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri float sumf = 0; - int isum[4]; + int isum[QK_K/16]; for (int i = 0; i < nb; ++i) { @@ -4193,14 +5654,14 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - isum[0] = isum[1] = isum[2] = isum[3] = 0; + memset(isum, 0, (QK_K/16)*sizeof(int)); for (int l = 0; l < 16; ++l) { isum[0] += q8[l+ 0] * ((q2[l] >> 0) & 3); isum[1] += q8[l+16] * ((q2[l] >> 2) & 3); isum[2] += q8[l+32] * ((q2[l] >> 4) & 3); isum[3] += q8[l+48] * ((q2[l] >> 6) & 3); } - for (int l = 0; l < 4; ++l) { + for (int l = 0; l < QK_K/16; ++l) { isum[l] *= (sc[l] & 0xF); } sumf += dall * (isum[0] + isum[1] + isum[2] + isum[3]) - dmin * summs; @@ -4211,8 +5672,13 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri #endif #if QK_K == 256 -void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const uint32_t kmask1 = 0x03030303; const uint32_t kmask2 = 0x0f0f0f0f; @@ -4228,9 +5694,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri uint32_t utmp[4]; const uint8x16_t m3b = vdupq_n_u8(0x3); -#ifdef __ARM_FEATURE_DOTPROD const int32x4_t vzero = vdupq_n_s32(0); -#endif const uint8x16_t m0 = vdupq_n_u8(1); const uint8x16_t m1 = vshlq_n_u8(m0, 1); @@ -4238,7 +5702,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t m3 = vshlq_n_u8(m0, 3); const int8_t m32 = 32; - int8x16x4_t q3bytes; + ggml_int8x16x4_t q3bytes; float sum = 0; @@ -4250,9 +5714,9 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict qh = x[i].hmask; const int8_t * restrict q8 = y[i].qs; - uint8x16x2_t qhbits = vld1q_u8_x2(qh); + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); - uint8x16x4_t q3h; + ggml_uint8x16x4_t q3h; int32_t isum = 0; @@ -4268,9 +5732,9 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32; - const int8x16x4_t q8bytes_1 = vld1q_s8_x4(q8); q8 += 64; - const int8x16x4_t q8bytes_2 = vld1q_s8_x4(q8); q8 += 64; + const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32; + const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64; + const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64; q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2); q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2); @@ -4282,22 +5746,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2])); q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3]; -#else - int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_1.val[0])), - vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_1.val[0]))); - int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_1.val[1])), - vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_1.val[1]))); - int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_1.val[2])), - vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_1.val[2]))); - int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_1.val[3])), - vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_1.val[3]))); - isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3]; -#endif + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3]; + scale += 4; q3h.val[0] = vbicq_u8(m2, qhbits.val[0]); @@ -4310,22 +5763,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2])); q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3]; -#else - p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_2.val[0])), - vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_2.val[0]))); - p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_2.val[1])), - vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_2.val[1]))); - p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_2.val[2])), - vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_2.val[2]))); - p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_2.val[3])), - vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_2.val[3]))); - isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3]; -#endif + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3]; + scale += 4; if (j == 0) { @@ -4657,7 +6099,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri vl = 16; - // retreive lane to multiply with scale + // retrieve lane to multiply with scale vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl); vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl); vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl); @@ -4755,8 +6197,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri #else -void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q3_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -4764,15 +6211,12 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - -#ifdef __ARM_FEATURE_DOTPROD - const int32x4_t vzero = vdupq_n_s32(0); -#endif + const int32x4_t vzero = vdupq_n_s32(0); const uint8x16_t m3b = vdupq_n_u8(0x3); const uint8x16_t mh = vdupq_n_u8(4); - int8x16x4_t q3bytes; + ggml_int8x16x4_t q3bytes; uint16_t aux16[2]; int8_t * scales = (int8_t *)aux16; @@ -4781,11 +6225,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - uint8x16x4_t q3h; + ggml_uint8x16x4_t q3h; const uint8x8_t hbits = vld1_u8(x[i].hmask); const uint8x16_t q3bits = vld1q_u8(x[i].qs); - const int8x16x4_t q8bytes = vld1q_s8_x4(y[i].qs); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(y[i].qs); const uint16_t a = *(const uint16_t *)x[i].scales; aux16[0] = a & 0x0f0f; @@ -4795,7 +6239,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]); - const float d = y[i].d * (float)x[i].d; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1)); q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2)); @@ -4808,22 +6252,10 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri q3bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 4), m3b), q3h.val[2])); q3bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q3bits, 6), q3h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1]; - isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3]; -#else - const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - isum += vaddvq_s16(p0) * scales[0] + vaddvq_s16(p1) * scales[2] + vaddvq_s16(p2) * scales[1] + vaddvq_s16(p3) * scales[3]; -#endif + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3]; sum += d * isum; @@ -5009,7 +6441,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]); - const float d = y[i].d * (float)x[i].d; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); @@ -5113,8 +6545,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri #endif #if QK_K == 256 -void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q4_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -5128,14 +6565,11 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri uint32_t utmp[4]; #ifdef __ARM_NEON - const uint8x16_t m4b = vdupq_n_u8(0xf); -#ifdef __ARM_FEATURE_DOTPROD const int32x4_t mzero = vdupq_n_s32(0); -#endif - int8x16x2_t q4bytes; - int8x16x2_t q8bytes; + ggml_int8x16x2_t q4bytes; + ggml_int8x16x2_t q8bytes; float sumf = 0; @@ -5169,44 +6603,22 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri int32_t sumi2 = 0; for (int j = 0; j < QK_K/64; ++j) { + const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32; - const uint8x16x2_t q4bits = vld1q_u8_x2(q4); q4 += 32; - -#ifdef __ARM_FEATURE_DOTPROD - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); - const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); + const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); sumi1 += vaddvq_s32(p1) * scales[2*j+0]; - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); - const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); + const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); sumi2 += vaddvq_s32(p2) * scales[2*j+1]; -#else - q8bytes = vld1q_s8_x2(q8); q8 += 32; - q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); - q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); - const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) * scales[2*j+0]; - - q8bytes = vld1q_s8_x2(q8); q8 += 32; - q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); - q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - sumi2 += vaddvq_s16(vaddq_s16(p2, p3)) * scales[2*j+1]; - -#endif } sumf += d * (sumi1 + sumi2); @@ -5494,8 +6906,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri #endif } #else -void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q4_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -5503,17 +6920,14 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - const uint8x16_t m4b = vdupq_n_u8(0xf); -#ifdef __ARM_FEATURE_DOTPROD const int32x4_t mzero = vdupq_n_s32(0); -#endif float sumf = 0; - int8x16x2_t q4bytes; - int8x16x4_t q8bytes; + ggml_int8x16x2_t q4bytes; + ggml_int8x16x4_t q8bytes; float sum_mins = 0.f; @@ -5530,47 +6944,26 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri aux16[1] = (a[0] >> 4) & 0x0f0f; const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]); - sum_mins += y[i].d * (float)x[i].d[1] * summi; + sum_mins += y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * summi; - const float d = y[i].d * (float)x[i].d[0]; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]); - const uint8x16x2_t q4bits = vld1q_u8_x2(q4); + const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); -#ifdef __ARM_FEATURE_DOTPROD - q8bytes = vld1q_s8_x4(q8); + q8bytes = ggml_vld1q_s8_x4(q8); q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); - const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); + const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); const int32_t sumi1 = vaddvq_s32(p1) * scales[0]; q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); - const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]); + const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]); const int32_t sumi2 = vaddvq_s32(p2) * scales[1]; -#else - q8bytes = vld1q_s8_x4(q8); - q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); - q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); - const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - int32_t sumi1 = vaddvq_s16(vaddq_s16(p0, p1)) * scales[0]; - - q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); - q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[2]))); - const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[3]))); - int32_t sumi2 = vaddvq_s16(vaddq_s16(p2, p3)) * scales[1]; - -#endif sumf += d * (sumi1 + sumi2); - } *s = sumf - sum_mins; @@ -5761,8 +7154,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri #endif #if QK_K == 256 -void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q5_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -5775,17 +7173,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri uint32_t utmp[4]; - #ifdef __ARM_NEON - const uint8x16_t m4b = vdupq_n_u8(0xf); const uint8x16_t mone = vdupq_n_u8(1); const uint8x16_t mtwo = vdupq_n_u8(2); -#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t mzero = vdupq_n_s32(0); -#endif - int8x16x4_t q5bytes; + ggml_int8x16x4_t q5bytes; float sumf = 0; @@ -5815,16 +7209,16 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict qh = x[i].qh; const int8_t * restrict q8 = y[i].qs; - uint8x16x2_t qhbits = vld1q_u8_x2(qh); + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); - uint8x16x4_t q5h; + ggml_uint8x16x4_t q5h; int32_t sumi = 0; for (int j = 0; j < QK_K/64; ++j) { - const uint8x16x2_t q5bits = vld1q_u8_x2(q5); q5 += 32; - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64; + const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32; + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); @@ -5838,28 +7232,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2])); q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) - - sumi += vaddvq_s32(vdotq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++; - sumi += vaddvq_s32(vdotq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++; -#else - - const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - sumi += vaddvq_s16(vaddq_s16(p0, p1)) * *scales++; - - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - sumi += vaddvq_s16(vaddq_s16(p2, p3)) * *scales++; -#endif + sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++; + sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++; } sumf += d * sumi - dmin * sumi_mins; - } *s = sumf; @@ -6202,8 +7579,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri #else -void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q5_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -6211,21 +7593,18 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - const uint8x16_t m4b = vdupq_n_u8(0xf); const uint8x16_t mh = vdupq_n_u8(16); -#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t mzero = vdupq_n_s32(0); -#endif - int8x16x4_t q5bytes; - uint8x16x4_t q5h; + ggml_int8x16x4_t q5bytes; + ggml_uint8x16x4_t q5h; float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * (float)x[i].d; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const int8_t * sc = x[i].scales; const uint8_t * restrict q5 = x[i].qs; @@ -6234,8 +7613,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8x8_t qhbits = vld1_u8(qh); - const uint8x16x2_t q5bits = vld1q_u8_x2(q5); - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); + const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1)); q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4)); @@ -6248,32 +7627,12 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri q5bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[0], 4)), vreinterpretq_s8_u8(q5h.val[2])); q5bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[1], 4)), vreinterpretq_s8_u8(q5h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) - - int32_t sumi1 = sc[0] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0])); - int32_t sumi2 = sc[1] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1])); - int32_t sumi3 = sc[2] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2])); - int32_t sumi4 = sc[3] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3])); + int32_t sumi1 = sc[0] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0])); + int32_t sumi2 = sc[1] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1])); + int32_t sumi3 = sc[2] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2])); + int32_t sumi4 = sc[3] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3])); sumf += d * (sumi1 + sumi2 + sumi3 + sumi4); - -#else - - const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - int32_t sumi = sc[0] * vaddvq_s16(p0) + sc[1] * vaddvq_s16(p1); - - const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - sumi += sc[2] * vaddvq_s16(p2) + sc[3] * vaddvq_s16(p3); - - sumf += d*sumi; -#endif - } *s = sumf; @@ -6387,7 +7746,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d = y[i].d * (float)x[i].d; + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const int8_t * sc = x[i].scales; const uint8_t * restrict q5 = x[i].qs; @@ -6491,8 +7850,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri #if QK_K == 256 -void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q6_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -6500,19 +7864,16 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - float sum = 0; const uint8x16_t m4b = vdupq_n_u8(0xF); -#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t vzero = vdupq_n_s32(0); -#endif //const int8x16_t m32s = vdupq_n_s8(32); const uint8x16_t mone = vdupq_n_u8(3); - int8x16x4_t q6bytes; - uint8x16x4_t q6h; + ggml_int8x16x4_t q6bytes; + ggml_uint8x16x4_t q6h; for (int i = 0; i < nb; ++i) { @@ -6524,9 +7885,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const int8_t * restrict scale = x[i].scales; - const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums); + const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); const int8x16_t scales = vld1q_s8(scale); - const int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}; + const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}}; const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])), vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))), @@ -6538,9 +7899,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - uint8x16x2_t qhbits = vld1q_u8_x2(qh); qh += 32; - uint8x16x4_t q6bits = vld1q_u8_x4(q6); q6 += 64; - int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64; + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32; + ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64; + ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); @@ -6558,32 +7919,14 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])); q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; - isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; scale += 4; -#else - - int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1]; - scale += 2; - - int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - isum += vaddvq_s16(p2) * scale[0] + vaddvq_s16(p3) * scale[1]; - scale += 2; -#endif - - q8bytes = vld1q_s8_x4(q8); q8 += 64; + q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; shifted = vshrq_n_u8(qhbits.val[0], 4); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4); @@ -6603,34 +7946,11 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])); q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])); -#if defined(__ARM_FEATURE_DOTPROD) - - isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; scale += 4; - - //for (int l = 0; l < 4; ++l) { - // const int32x4_t p = vdotq_s32(vzero, q6bytes.val[l], q8bytes.val[l]); - // isum += vaddvq_s32(p) * *scale++; - //} -#else - p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1]; - scale += 2; - - p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - isum += vaddvq_s16(p2) * scale[0] + vaddvq_s16(p3) * scale[1]; - scale += 2; -#endif - } //sum += isum * d_all * y[i].d; sum += d_all * y[i].d * (isum - 32 * isum_mins); @@ -6967,8 +8287,13 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri #else -void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { +void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); const block_q6_K * restrict x = vx; const block_q8_K * restrict y = vy; @@ -6976,23 +8301,20 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const int nb = n / QK_K; #ifdef __ARM_NEON - float sum = 0; const uint8x16_t m4b = vdupq_n_u8(0xF); const int8x16_t m32s = vdupq_n_s8(32); -#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t vzero = vdupq_n_s32(0); -#endif const uint8x16_t mone = vdupq_n_u8(3); - int8x16x4_t q6bytes; - uint8x16x4_t q6h; + ggml_int8x16x4_t q6bytes; + ggml_uint8x16x4_t q6h; for (int i = 0; i < nb; ++i) { - const float d_all = (float)x[i].d; + const float d_all = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -7002,9 +8324,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri int32_t isum = 0; - uint8x16_t qhbits = vld1q_u8(qh); - uint8x16x2_t q6bits = vld1q_u8_x2(q6); - int8x16x4_t q8bytes = vld1q_s8_x4(q8); + uint8x16_t qhbits = vld1q_u8(qh); + ggml_uint8x16x2_t q6bits = ggml_vld1q_u8_x2(q6); + ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4); uint8x16_t shifted = vshrq_n_u8(qhbits, 2); @@ -7019,26 +8341,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[2])), m32s); q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[3])), m32s); -#if defined(__ARM_FEATURE_DOTPROD) - - isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + - vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; -#else - - int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])), - vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0]))); - int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])), - vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1]))); - isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1]; - - int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])), - vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2]))); - int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])), - vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3]))); - isum += vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3]; -#endif + isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] + + vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3]; sum += isum * d_all * y[i].d; @@ -7179,7 +8485,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - const float d_all = (float)x[i].d; + const float d_all = GGML_FP16_TO_FP32(x[i].d); const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -7280,3 +8586,4093 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri } #endif + +#if defined (__AVX2__) || defined (__ARM_NEON) +static const int8_t keven_signs_q2xs[1024] = { + 1, 1, 1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, 1, + 1, 1, -1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, 1, 1, -1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, -1, + 1, 1, 1, -1, 1, 1, 1, -1, -1, 1, 1, -1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, -1, + 1, 1, -1, -1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, -1, 1, -1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, 1, + 1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, 1, -1, 1, 1, 1, 1, -1, 1, 1, -1, 1, 1, 1, -1, -1, 1, 1, -1, 1, 1, -1, + 1, 1, -1, 1, -1, 1, 1, 1, -1, 1, -1, 1, -1, 1, 1, -1, 1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, 1, + 1, 1, 1, -1, -1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, -1, 1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, 1, + 1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, 1, 1, -1, -1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, -1, + 1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, 1, -1, 1, 1, 1, -1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, 1, -1, 1, -1, + 1, 1, -1, 1, 1, -1, 1, 1, -1, 1, -1, 1, 1, -1, 1, -1, 1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, 1, 1, + 1, 1, 1, -1, 1, -1, 1, 1, -1, 1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, 1, 1, + 1, 1, -1, -1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, 1, 1, 1, -1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, -1, + 1, 1, 1, 1, -1, -1, 1, 1, -1, 1, 1, 1, -1, -1, 1, -1, 1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, -1, -1, 1, 1, + 1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, 1, -1, -1, 1, 1, 1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, -1, + 1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, 1, 1, -1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, -1, + 1, 1, -1, -1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, -1, 1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1, + 1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, -1, 1, 1, -1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, 1, -1, -1, + 1, 1, -1, 1, 1, 1, -1, 1, -1, 1, -1, 1, 1, 1, -1, -1, 1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, -1, 1, + 1, 1, 1, -1, 1, 1, -1, 1, -1, 1, 1, -1, 1, 1, -1, -1, 1, -1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, 1, + 1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, 1, 1, -1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, -1, + 1, 1, 1, 1, -1, 1, -1, 1, -1, 1, 1, 1, -1, 1, -1, -1, 1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, 1, -1, 1, + 1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, 1, -1, 1, 1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, -1, + 1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, 1, 1, -1, 1, -1, -1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, -1, + 1, 1, -1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, -1, 1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, 1, + 1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, -1, -1, -1, 1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, -1, -1, 1, + 1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, -1, 1, 1, -1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, -1, -1, -1, + 1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, 1, -1, -1, 1, 1, -1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, 1, -1, -1, -1, + 1, 1, -1, -1, 1, -1, -1, 1, -1, 1, -1, -1, 1, -1, -1, -1, 1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, 1, + 1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, -1, -1, -1, 1, 1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, -1, + 1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1, -1, 1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, 1, + 1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, -1, 1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, 1, + 1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, -1, +}; +#endif + +void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq2_xxs * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined(__ARM_NEON) + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + uint32_t aux32[4]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + ggml_int8x16x4_t q2u; + ggml_int8x16x4_t q2s; + ggml_int8x16x4_t q8b; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + float sumf1 = 0, sumf2 = 0; + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8; + q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1]))); + q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3]))); + q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9]))); + q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11]))); + q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 7) & 127)))); + q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127)))); + q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 7) & 127)))); + q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127)))); + q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]); + q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]); + q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]); + q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]); + const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]); + const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]); + sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28)); + sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28)); + } + sumf += d*(sumf1 + sumf2); + } + *s = 0.25f * sumf; + +#elif defined(__AVX2__) + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + uint32_t aux32[4]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8; + const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]); + const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]); + const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127], + signs64[(aux32[1] >> 7) & 127], signs64[(aux32[1] >> 0) & 127]); + const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127], + signs64[(aux32[3] >> 7) & 127], signs64[(aux32[3] >> 0) & 127]); + const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1); + const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2); + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); + const uint16_t ls1 = aux32[1] >> 28; + const uint16_t ls2 = aux32[3] >> 28; + const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1)); + const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1)); + sumi1 = _mm256_add_epi32(sumi1, p1); + sumi2 = _mm256_add_epi32(sumi2, p2); + } + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf); + + } + + *s = 0.125f * hsum_float_8(accumf); + +#else + + uint32_t aux32[2]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + float sumf = 0.f; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + int32_t bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + memcpy(aux32, q2, 2*sizeof(uint32_t)); + q2 += 4; + const uint32_t ls = 2*(aux32[1] >> 28) + 1; + int32_t sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; + for (int j = 0; j < 8; ++j) { + sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += sumi * ls; + } + sumf += d * bsum; + } + *s = 0.125f * sumf; +#endif +} + +void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq2_xs * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined(__ARM_NEON) + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + ggml_int8x16x4_t q2u; + ggml_int8x16x4_t q2s; + ggml_int8x16x4_t q8b; + + int32x4x4_t scales32; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + const uint8x8_t scales8 = vld1_u8(x[i].scales); + const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf)); + const uint8x8_t scales_h = vshr_n_u8(scales8, 4); + uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h)); + scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1)); + const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales)); + const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales)); + scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1))); + scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1))); + scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2))); + scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2))); + int32x4_t sumi = vdupq_n_s32(0); + for (int ib64 = 0; ib64 < QK_K/64; ++ib64) { + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511)))); + q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511)))); + q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511)))); + q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511)))); + q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9)))); + q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9)))); + q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9)))); + q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9)))); + q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]); + q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]); + q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]); + q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]); + const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]); + const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]); + const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]); + const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]); + const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4)); + sumi = vmlaq_s32(sumi, p, scales32.val[ib64]); + q2 += 8; + } + sumf += d*vaddvq_s32(sumi); + } + *s = 0.125f * sumf; + +#elif defined(__AVX2__) + + const __m256i mone = _mm256_set1_epi8(1); + static const char block_sign_shuffle_mask_1[32] = { + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, + 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, + }; + static const char block_sign_shuffle_mask_2[32] = { + 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, + 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, + }; + static const uint8_t bit_selector_mask_bytes[32] = { + 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, + 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, + }; + + const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes); + const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1); + const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2); + +#if QK_K == 64 + static const uint8_t k_bit_helper[16] = { + 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, + }; + const __m128i bit_helper = _mm_loadu_si128((const __m128i*)k_bit_helper); + const __m128i m511 = _mm_set1_epi16(511); + typedef union { + __m128i vec_index; + uint16_t index[8]; + } index_t; + + index_t idx; + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const __m128i q2_data = _mm_loadu_si128((const __m128i*)x[i].qs); + idx.vec_index = _mm_and_si128(q2_data, m511); + + const __m128i partial_sign_bits = _mm_srli_epi16(q2_data, 9); + const __m128i partial_sign_bits_upper = _mm_srli_epi16(q2_data, 13); + const __m128i partial_sign_bits_for_counting = _mm_xor_si128(partial_sign_bits, partial_sign_bits_upper); + + const __m128i odd_bits = _mm_shuffle_epi8(bit_helper, partial_sign_bits_for_counting); + const __m128i full_sign_bits = _mm_or_si128(partial_sign_bits, odd_bits); + const __m256i full_signs = MM256_SET_M128I(full_sign_bits, full_sign_bits); + + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)y[i].qs); + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)(y[i].qs+32)); + + const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[idx.index[3]], iq2xs_grid[idx.index[2]], + iq2xs_grid[idx.index[1]], iq2xs_grid[idx.index[0]]); + const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[idx.index[7]], iq2xs_grid[idx.index[6]], + iq2xs_grid[idx.index[5]], iq2xs_grid[idx.index[4]]); + + __m256i signs; + signs = _mm256_shuffle_epi8(full_signs, block_sign_shuffle_1); + signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); + const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone)); + + signs = _mm256_shuffle_epi8(full_signs, block_sign_shuffle_2); + signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); + const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone)); + + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); + + const __m256i sc1 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1)); + const __m256i sc2 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1)); + + const __m256i sum = _mm256_add_epi32(_mm256_madd_epi16(sc1, dot1), _mm256_madd_epi16(sc2, dot2)); + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sum), accumf); + + } + + *s = 0.125f * hsum_float_8(accumf); +#else + + static const uint8_t k_bit_helper[32] = { + 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, + 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00, + }; + const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper); + const __m256i m511 = _mm256_set1_epi16(511); + const __m128i m4 = _mm_set1_epi8(0xf); + const __m128i m1 = _mm_set1_epi8(1); + + uint64_t aux64; + + // somewhat hacky, but gives a significant boost in performance + __m256i aux_gindex; + const uint16_t * gindex = (const uint16_t *)&aux_gindex; + + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + memcpy(&aux64, x[i].scales, 8); + __m128i stmp = _mm_set1_epi64x(aux64); + stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4)); + const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1); + + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) { + + const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2); q2 += 16; + aux_gindex = _mm256_and_si256(q2_data, m511); + + const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9); + const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13); + const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper); + + const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting); + const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits); + + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + + const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]], + iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]); + const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]], + iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]); + const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]], + iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]); + const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]], + iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]); + + const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits); + const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1); + const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l); + const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h); + + __m256i signs; + signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1); + signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); + const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone)); + + signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2); + signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); + const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone)); + + signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1); + signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); + const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone)); + + signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2); + signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask); + const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone)); + + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); + const __m256i dot3 = _mm256_maddubs_epi16(q2_3, q8s_3); + const __m256i dot4 = _mm256_maddubs_epi16(q2_4, q8s_4); + + const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0))); + const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1))); + const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2))); + const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3))); + + sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1)); + sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2)); + sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3)); + sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4)); + } + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf); + + } + + *s = 0.125f * hsum_float_8(accumf); +#endif + +#else + + float sumf = 0.f; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const uint8_t * restrict sc = x[i].scales; + const int8_t * restrict q8 = y[i].qs; + int32_t bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; + const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; + int32_t sumi = 0; + for (int l = 0; l < 2; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); + const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; + for (int j = 0; j < 8; ++j) { + sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += sumi * ls1; + sumi = 0; + for (int l = 2; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); + const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; + for (int j = 0; j < 8; ++j) { + sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += sumi * ls2; + q2 += 4; + } + sumf += d * bsum; + } + *s = 0.125f * sumf; +#endif +} + +void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq2_s * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined(__ARM_NEON) + + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 + }; + + static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,}; + + const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1); + const uint8x16_t mask2 = vld1q_u8(k_mask2); + const uint8x16_t m1 = vdupq_n_u8(1); + const int32x4_t vzero = vdupq_n_s32(0); + + uint8x16x2_t vs; + ggml_int8x16x4_t q2s; + ggml_int8x16x4_t q8b; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + + const uint8_t * restrict qs = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); + const int8_t * restrict q8 = y[i].qs; + + int sumi1 = 0, sumi2 = 0; + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + q2s.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[0] | ((qh[ib32+0] << 8) & 0x300)))), + vld1_s8((const int8_t *)(iq2s_grid + (qs[1] | ((qh[ib32+0] << 6) & 0x300))))); + q2s.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[2] | ((qh[ib32+0] << 4) & 0x300)))), + vld1_s8((const int8_t *)(iq2s_grid + (qs[3] | ((qh[ib32+0] << 2) & 0x300))))); + q2s.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[4] | ((qh[ib32+1] << 8) & 0x300)))), + vld1_s8((const int8_t *)(iq2s_grid + (qs[5] | ((qh[ib32+1] << 6) & 0x300))))); + q2s.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[6] | ((qh[ib32+1] << 4) & 0x300)))), + vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300))))); + qs += 8; + + vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16))); + vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); + vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); + vs.val[0] = vceqq_u8(vs.val[0], mask2); + vs.val[1] = vceqq_u8(vs.val[1], mask2); + + q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]); + q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]); + + vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16))); + vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); + vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); + vs.val[0] = vceqq_u8(vs.val[0], mask2); + vs.val[1] = vceqq_u8(vs.val[1], mask2); + + signs += 4; + + q2s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[2]); + q2s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[3]); + + const int32x4_t p1 = ggml_vdotq_s32(vzero, q2s.val[0], q8b.val[0]); + const int32x4_t p2 = ggml_vdotq_s32(vzero, q2s.val[1], q8b.val[1]); + const int32x4_t p3 = ggml_vdotq_s32(vzero, q2s.val[2], q8b.val[2]); + const int32x4_t p4 = ggml_vdotq_s32(vzero, q2s.val[3], q8b.val[3]); + + sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32+0] & 0xf)); + sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32+0] >> 4)); + sumi1 += vaddvq_s32(p3) * (1 + 2*(x[i].scales[ib32+1] & 0xf)); + sumi2 += vaddvq_s32(p4) * (1 + 2*(x[i].scales[ib32+1] >> 4)); + } + sumf += d*(sumi1 + sumi2); + } + + *s = 0.125f * sumf; + +#elif defined(__AVX2__) + + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 + }; + + static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, + 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, + }; + + const __m128i m4 = _mm_set1_epi8(0xf); + const __m128i m1 = _mm_set1_epi8(1); + + const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1); + const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2); + + uint64_t aux64; + + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict qs = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); + const int8_t * restrict q8 = y[i].qs; + + memcpy(&aux64, x[i].scales, 8); + const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1); + const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15 + + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)], + iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)], + iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)], + iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]); + const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)], + iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)], + iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)], + iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]); + qs += 8; + + __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16)); + aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2); + const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2); + const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1); + + aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16)); + aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2); + const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2); + const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2); + + signs += 4; + + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1 + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3 + + const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0))); + const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1))); + sumi1 = _mm256_add_epi32(sumi1, p1); + sumi2 = _mm256_add_epi32(sumi2, p2); + } + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf); + + } + + *s = 0.125f * hsum_float_8(accumf); + +#else + + float sumf = 0; + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + const uint8_t * signs = qs + QK_K/8; + + int bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf); + int ls2 = 1 + 2*(x[i].scales[ib32] >> 4); + int sumi1 = 0, sumi2 = 0; + for (int l = 0; l < 2; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); + for (int j = 0; j < 8; ++j) { + sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + for (int l = 2; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); + for (int j = 0; j < 8; ++j) { + sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += ls1 * sumi1 + ls2 * sumi2; + qs += 4; + signs += 4; + } + + sumf += d * bsum; + } + + *s = 0.125f * sumf; + +#endif + +} + +void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq3_xxs * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined(__ARM_NEON) + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + uint32_t aux32[2]; + + ggml_int8x16x4_t q3s; + ggml_int8x16x4_t q8b; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict q3 = x[i].qs; + const uint8_t * restrict gas = x[i].qs + QK_K/4; + const int8_t * restrict q8 = y[i].qs; + float sumf1 = 0, sumf2 = 0; + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t); + const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]); + const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]); + const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]); + const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]); + q3 += 16; + q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 7) & 127)))); + q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127)))); + q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 7) & 127)))); + q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127)))); + q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0)); + q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1)); + q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2)); + q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3)); + const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]); + const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]); + sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28)); + sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28)); + } + sumf += d*(sumf1 + sumf2); + } + *s = 0.5f * sumf; + +#elif defined(__AVX2__) + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + uint32_t aux32[2]; + + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict q3 = x[i].qs; + const uint8_t * restrict gas = x[i].qs + QK_K/4; + const int8_t * restrict q8 = y[i].qs; + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]], + iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]); + q3 += 8; + const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]], + iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]); + q3 += 8; + memcpy(aux32, gas, 8); gas += 8; + const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127], + signs64[(aux32[0] >> 7) & 127], signs64[(aux32[0] >> 0) & 127]); + const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127], + signs64[(aux32[1] >> 7) & 127], signs64[(aux32[1] >> 0) & 127]); + const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1); + const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2); + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); + const uint16_t ls1 = aux32[0] >> 28; + const uint16_t ls2 = aux32[1] >> 28; + const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1)); + const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1)); + sumi1 = _mm256_add_epi32(sumi1, p1); + sumi2 = _mm256_add_epi32(sumi2, p2); + } + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf); + + } + + *s = 0.25f * hsum_float_8(accumf); + +#else + + uint32_t aux32; + + float sumf = 0.f; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict q3 = x[i].qs; + const uint8_t * restrict gas = x[i].qs + QK_K/4; + const int8_t * restrict q8 = y[i].qs; + int32_t bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t); + const uint32_t ls = 2*(aux32 >> 28) + 1; + int32_t sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]); + const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]); + const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; + for (int j = 0; j < 4; ++j) { + sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1); + sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1); + } + q8 += 8; + } + q3 += 8; + bsum += sumi * ls; + } + sumf += d * bsum; + } + *s = 0.25f * sumf; +#endif +} + +void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq3_s * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined(__ARM_NEON) + + typedef union { + uint16x8_t vec_index; + uint16_t index[8]; + } vec_index_t; + + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 + }; + + static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,}; + + static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1}; + + const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1); + const uint8x16_t mask2 = vld1q_u8(k_mask2); + + const int16x8_t hshift = vld1q_s16(k_shift); + const uint16x8_t m256 = vdupq_n_u16(256); + const uint8x16_t m1 = vdupq_n_u8(1); + + uint8x16x2_t vs; + ggml_int8x16x4_t q3s; + ggml_int8x16x4_t q8b; + vec_index_t idx; + +#if QK_K == 256 + uint32_t scales32[2]; + const uint8_t * scales8 = (const uint8_t *)scales32; +#endif + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict qs = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint16_t * restrict signs = (const uint16_t *)x[i].signs; + const int8_t * restrict q8 = y[i].qs; + +#if QK_K == 256 + memcpy(scales32, x[i].scales, 4); + scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101; + scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101; +#endif + + int sumi1 = 0, sumi2 = 0; + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + + const uint8x16_t idx_l = vld1q_u8(qs); qs += 16; + idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256)); + const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]], + iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]); + const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]], + iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]); + idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256)); + const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]], + iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]); + const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]], + iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]); + + + vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16))); + vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); + vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); + vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1); + vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1); + + q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0)); + q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1)); + + vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16))); + vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); + vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); + vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1); + vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1); + + signs += 4; + + q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2)); + q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3)); + + const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]); + const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]); +#if QK_K == 256 + sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0]; + sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4]; +#else + sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32/2] & 0xf)); + sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32/2] >> 4)); +#endif + } + sumf += d*(sumi1 + sumi2); + } + *s = sumf; + +#elif defined(__AVX2__) + + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 + }; + + static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, + 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, + }; + + const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1); + const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2); + + const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8); + const __m256i idx_mask = _mm256_set1_epi32(256); + + typedef union { + __m256i vec[2]; + uint32_t index[16]; + } index_t; + + index_t idx; + + __m256 accumf = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict qs = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint16_t * restrict signs = (const uint16_t *)x[i].signs; + const int8_t * restrict q8 = y[i].qs; + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16; + idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]); + idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]); + idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask); + idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask); + idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l))); + idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1))); + + // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange. + //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4); + //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4); + const __m256i q2_1 = _mm256_set_epi32( + iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]], + iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]] + ); + const __m256i q2_2 = _mm256_set_epi32( + iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]], + iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]] + ); + + __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16)); + aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2); + const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2); + const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1); + + aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16)); + aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2); + const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2); + const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2); + + signs += 4; + + const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); + const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); + const uint16_t ls1 = x[i].scales[ib32/2] & 0xf; + const uint16_t ls2 = x[i].scales[ib32/2] >> 4; + const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1)); + const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1)); + sumi1 = _mm256_add_epi32(sumi1, p1); + sumi2 = _mm256_add_epi32(sumi2, p2); + } + + accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf); + + } + + *s = hsum_float_8(accumf); + +#else + + float sumf = 0.f; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint8_t * restrict qs = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint8_t * restrict signs = x[i].signs; + const int8_t * restrict q8 = y[i].qs; + int32_t bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { + const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; + const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; + int32_t sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); + for (int j = 0; j < 4; ++j) { + sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); + sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); + } + q8 += 8; + } + qs += 8; + signs += 4; + bsum += sumi * ls1; + sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); + for (int j = 0; j < 4; ++j) { + sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); + sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); + } + q8 += 8; + } + qs += 8; + signs += 4; + bsum += sumi * ls2; + } + sumf += d * bsum; + } + *s = sumf; +#endif +} + + +#ifdef __AVX2__ +static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) { + const __m256i ax = _mm256_sign_epi8(x, x); + const __m256i sy = _mm256_sign_epi8(y, x); + return _mm256_maddubs_epi16(ax, sy); +} +#endif + +void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq1_s * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined __ARM_NEON + + ggml_int8x16x4_t q1b; + ggml_int8x16x4_t q8b; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint16_t * qh = x[i].qh; + + int sumi1 = 0, sumi2 = 0, sumi3 = 0; + + for (int ib = 0; ib < QK_K/32; ib += 2) { + + q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[ib+0] << 8) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[ib+0] << 5) & 0x700))))); + q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[ib+0] << 2) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[ib+0] >> 1) & 0x700))))); + q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[ib+1] << 8) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[ib+1] << 5) & 0x700))))); + q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[ib+1] << 2) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[ib+1] >> 1) & 0x700))))); + qs += 8; + + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + + const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]); + const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]); + + const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1; + const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1; + sumi1 += vaddvq_s32(p1) * ls1; + sumi2 += vaddvq_s32(p2) * ls2; + sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1) + + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1); + + } + + sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3); + } + + *s = sumf; + +#elif defined __AVX2__ + + __m256 accum = _mm256_setzero_ps(); + float accum1 = 0; + for (int i = 0; i < nb; ++i) { + + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint16_t * qh = x[i].qh; + + __m256i sumi = _mm256_setzero_si256(); + int sumi1 = 0; + for (int ib = 0; ib < QK_K/32; ib += 2) { + const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], + iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]); + const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], + iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]); + qs += 8; + const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; + const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; + + const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1); + const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2); + const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1; + const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1; + const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1)); + const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2)); + + sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2)); + sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1 + + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; + } + + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum); + accum1 += d * sumi1; + + } + + *s = hsum_float_8(accum) + IQ1S_DELTA * accum1; + +#else + + float sumf = 0; + for (int i = 0; i < nb; i++) { + + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint16_t * qh = x[i].qh; + + int sumi = 0, sumi1 = 0; + for (int ib = 0; ib < QK_K/32; ++ib) { + const int ls = 2*((qh[ib] >> 12) & 7) + 1; + const int delta = qh[ib] & 0x8000 ? -1 : 1; + int lsum = 0; + for (int l = 0; l < 4; ++l) { + const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); + for (int j = 0; j < 8; ++j) { + lsum += q8[j] * grid[j]; + } + q8 += 8; + } + sumi += ls * lsum; + sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]); + qs += 4; + } + + sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); + } + + *s = sumf; + +#endif +} + +void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(n % QK_K == 0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_iq1_m * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if QK_K != 64 + iq1m_scale_t scale; +#endif + +#if defined __ARM_NEON + +#if QK_K == 64 + const int32x4_t mask = vdupq_n_s32(0xf); +#else + const int32x4_t mask = vdupq_n_s32(0x7); +#endif + const int32x4_t mone = vdupq_n_s32(1); + const int32x4_t mzero = vdupq_n_s32(0); + + ggml_int8x16x4_t deltas; + deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1)); + deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1)); + deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1)); + deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1)); + + ggml_int8x16x4_t q1b; + ggml_int8x16x4_t q8b; + + uint32_t aux32; + const uint8_t * aux8 = (const uint8_t *)&aux32; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + const uint16_t * sc = (const uint16_t *)x[i].scales; + +#if QK_K != 64 + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); +#endif + + int32x4_t sumi1 = mzero; + int32x4_t sumi2 = mzero; + + for (int ib = 0; ib < QK_K/32; ib += 2) { + + q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700))))); + q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700))))); + q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700))))); + q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))), + vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700))))); + + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + + const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1])); + const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3])); + const int32x4_t p12 = vpaddq_s32(p1, p2); + + const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that + aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202); + + const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1])); + const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3])); + const int32x4_t p34 = vpaddq_s32(p3, p4); + +#if QK_K == 64 + int32x4_t scales_4 = ggml_vld1q_u32(sc[0] >> 0, sc[0] >> 4, sc[0] >> 8, sc[0] >> 12); +#else + int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9); +#endif + scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone); + + sumi1 = vmlaq_s32(sumi1, scales_4, p12); + sumi2 = vmlaq_s32(sumi2, scales_4, p34); + + qs += 8; qh += 4; + + } + +#if QK_K == 64 + sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); +#else + sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); +#endif + } + + *s = sumf; + +#elif defined __AVX2__ + +#if QK_K == 64 + const __m256i mask = _mm256_set1_epi16(0xf); +#else + const __m256i mask = _mm256_set1_epi16(0x7); +#endif + const __m256i mone = _mm256_set1_epi16(1); + + __m256 accum1 = _mm256_setzero_ps(); + __m256 accum2 = _mm256_setzero_ps(); + for (int i = 0; i < nb; ++i) { + + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + const uint16_t * sc = (const uint16_t *)x[i].scales; + +#if QK_K != 64 + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); +#endif + + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib = 0; ib < QK_K/32; ib += 2) { + const __m256i q1b_1 = _mm256_set_epi64x( + iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)], + iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)] + ); + const __m256i q1b_2 = _mm256_set_epi64x( + iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)], + iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)] + ); + const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; + const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; + + const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1); + const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2); + + const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, + qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101, + qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, + qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101); + const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, + qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101, + qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, + qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101); + + const __m256i dot3 = mul_add_epi8(delta1, q8b_1); + const __m256i dot4 = mul_add_epi8(delta2, q8b_2); +#if QK_K == 64 + __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[0] >> 4), _mm_set1_epi16(sc[0] >> 0)); + __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[0] >> 12), _mm_set1_epi16(sc[0] >> 8)); +#else + __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0)); + __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6)); +#endif + scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone); + scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone); + const __m256i p1 = _mm256_madd_epi16(dot1, scale1); + const __m256i p2 = _mm256_madd_epi16(dot2, scale2); + const __m256i p3 = _mm256_madd_epi16(dot3, scale1); + const __m256i p4 = _mm256_madd_epi16(dot4, scale2); + + sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2)); + sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4)); + + qs += 8; qh += 4; + } + +#if QK_K == 64 + const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)); +#else + const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16)); +#endif + accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1); + accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2); + + } + + *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2); + +#else + + int sum1[2], sum2[2], delta[4]; + + float sumf = 0; + for (int i = 0; i < nb; i++) { + + const int8_t * q8 = y[i].qs; + const uint8_t * qs = x[i].qs; + const uint8_t * qh = x[i].qh; + const uint16_t * sc = (const uint16_t *)x[i].scales; + +#if QK_K != 64 + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); +#endif + + int sumi1 = 0, sumi2 = 0; + for (int ib = 0; ib < QK_K/32; ++ib) { + delta[0] = qh[0] & 0x08 ? -1 : 1; + delta[1] = qh[0] & 0x80 ? -1 : 1; + delta[2] = qh[1] & 0x08 ? -1 : 1; + delta[3] = qh[1] & 0x80 ? -1 : 1; + sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0; + for (int l = 0; l < 4; ++l) { + const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700))); + int lsum1 = 0, lsum2 = 0; + for (int j = 0; j < 8; ++j) { + lsum1 += q8[j] * grid[j]; + lsum2 += q8[j]; + } + q8 += 8; + sum1[l/2] += lsum1; + sum2[l/2] += lsum2*delta[l]; + } +#if QK_K == 64 + const int ls1 = 2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1; + const int ls2 = 2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1; +#else + const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1; + const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1; +#endif + sumi1 += sum1[0] * ls1 + sum1[1] * ls2; + sumi2 += sum2[0] * ls1 + sum2[1] * ls2; + qs += 4; + qh += 2; + } + +#if QK_K == 64 + sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); +#else + sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); +#endif + } + + *s = sumf; + +#endif +} + +void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + assert(n % QK4_NL == 0); + static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same"); + + const block_iq4_nl * restrict x = vx; + const block_q8_0 * restrict y = vy; + + const int nb = n / QK4_NL; + +#if defined __ARM_NEON + const int8x16_t values = vld1q_s8(kvalues_iq4nl); + const uint8x16_t m4b = vdupq_n_u8(0x0f); + uint8x16x2_t q4bits; + int8x16x4_t q4b; + int8x16x4_t q8b; + int32x4_t prod_1, prod_2; + + float sumf = 0; + + for (int ib = 0; ib < nb; ib += 2) { + + q4bits.val[0] = vld1q_u8(x[ib+0].qs); + q4bits.val[1] = vld1q_u8(x[ib+1].qs); + q8b.val[0] = vld1q_s8(y[ib+0].qs); + q8b.val[1] = vld1q_s8(y[ib+0].qs + 16); + q8b.val[2] = vld1q_s8(y[ib+1].qs); + q8b.val[3] = vld1q_s8(y[ib+1].qs + 16); + + q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[0], m4b)); + q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4)); + q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[1], m4b)); + q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4)); + + prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]); + prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]); + + sumf += + GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib+0].d) * vaddvq_s32(prod_1) + + GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib+1].d) * vaddvq_s32(prod_2); + } + + *s = sumf; + +#elif defined __AVX2__ + + const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl); + const __m128i m4b = _mm_set1_epi8(0x0f); + const __m256i mone = _mm256_set1_epi16(1); + + __m256 accum1 = _mm256_setzero_ps(); + __m256 accum2 = _mm256_setzero_ps(); + for (int ib = 0; ib < nb; ib += 2) { + const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[0].qs); + const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[1].qs); + const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[0].qs); + const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[1].qs); + const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b))); + const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b))); + const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1); + const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); + const __m256i p_1 = _mm256_madd_epi16(p16_1, mone); + const __m256i p_2 = _mm256_madd_epi16(p16_2, mone); + accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[0].d)*GGML_FP16_TO_FP32(x[0].d)), + _mm256_cvtepi32_ps(p_1), accum1); + accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[1].d)*GGML_FP16_TO_FP32(x[1].d)), + _mm256_cvtepi32_ps(p_2), accum2); + + y += 2; + x += 2; + } + + *s = hsum_float_8(_mm256_add_ps(accum1, accum2)); + +#else + float sumf = 0; + for (int ib = 0; ib < nb; ++ib) { + const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < QK4_NL/2; ++j) { + sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; + sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4]; + } + sumf += d * (sumi1 + sumi2); + } + *s = sumf; +#endif +} + +void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + assert(n % QK_K == 0); +#if QK_K == 64 + ggml_vec_dot_iq4_nl_q8_0(n, s, bs, vx, bx, vy, by, nrc); +#else + + const block_iq4_xs * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#if defined __ARM_NEON + const int8x16_t values = vld1q_s8(kvalues_iq4nl); + const uint8x16_t m4b = vdupq_n_u8(0x0f); + ggml_uint8x16x2_t q4bits; + ggml_int8x16x4_t q4b; + ggml_int8x16x4_t q8b; + int32x4_t prod_1, prod_2; + + float sumf = 0; + + for (int ibl = 0; ibl < nb; ++ibl) { + + const int8_t * q8 = y[ibl].qs; + const uint8_t * q4 = x[ibl].qs; + uint16_t h = x[ibl].scales_h; + + int sumi1 = 0, sumi2 = 0; + for (int ib = 0; ib < QK_K/64; ++ib) { + + q4bits = ggml_vld1q_u8_x2(q4); q4 += 32; + q8b = ggml_vld1q_s8_x4(q8); q8 += 64; + + q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[0], m4b)); + q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4)); + q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[1], m4b)); + q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4)); + + prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]); + prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]); + + int ls1 = ((x[ibl].scales_l[ib] & 0xf) | ((h << 4) & 0x30)) - 32; + int ls2 = ((x[ibl].scales_l[ib] >> 4) | ((h << 2) & 0x30)) - 32; + h >>= 4; + sumi1 += vaddvq_s32(prod_1) * ls1; + sumi2 += vaddvq_s32(prod_2) * ls2; + + } + + sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); + } + + *s = sumf; + +#elif defined __AVX2__ + + const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl); + const __m128i m4b = _mm_set1_epi8(0x0f); + + __m256 accum = _mm256_setzero_ps(); + for (int ibl = 0; ibl < nb; ++ibl) { + const uint8_t * qs = x[ibl].qs; + const int8_t * q8 = y[ibl].qs; + uint16_t sh = x[ibl].scales_h; + __m256i sumi1 = _mm256_setzero_si256(); + __m256i sumi2 = _mm256_setzero_si256(); + for (int ib = 0; ib < QK_K/32; ib += 2) { + const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs); qs += 16; + const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs); qs += 16; + const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; + const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b))); + const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b))); + const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1); + const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); + const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32; + const int16_t ls2 = ((x[ibl].scales_l[ib/2] >> 4) | ((sh << 2) & 0x30)) - 32; + sh >>= 4; + const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1)); + const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2)); + sumi1 = _mm256_add_epi32(p_1, sumi1); + sumi2 = _mm256_add_epi32(p_2, sumi2); + } + accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), + _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum); + } + + *s = hsum_float_8(accum); + +#else + float sumf = 0; + for (int ibl = 0; ibl < nb; ++ibl) { + const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + uint16_t h = x[ibl].scales_h; + const uint8_t * qs = x[ibl].qs; + const int8_t * q8 = y[ibl].qs; + for (int ib = 0; ib < QK_K/32; ib += 2) { + const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); + const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); + h >>= 4; + const float d1 = d4d8*(ls1 - 32); + const float d2 = d4d8*(ls2 - 32); + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < 16; ++j) { + sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; + sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; + } + sumf += d1 * (sumi1 + sumi2); + qs += 16; + q8 += 32; + sumi1 = sumi2 = 0; + for (int j = 0; j < 16; ++j) { + sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; + sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; + } + sumf += d2 * (sumi1 + sumi2); + qs += 16; + q8 += 32; + } + } + *s = sumf; +#endif +#endif +} + +// ================================ IQ2 quantization ============================================= + +typedef struct { + uint64_t * grid; + int * map; + uint16_t * neighbours; +} iq2_entry_t; + +static iq2_entry_t iq2_data[4] = { + {NULL, NULL, NULL}, + {NULL, NULL, NULL}, + {NULL, NULL, NULL}, + {NULL, NULL, NULL}, +}; + +static inline int iq2_data_index(enum ggml_type type) { + GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S); + return type == GGML_TYPE_IQ2_XXS ? 0 : + type == GGML_TYPE_IQ2_XS ? 1 : + type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 2 : 3; +} + +static inline int iq2_grid_size(enum ggml_type type) { + GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S); + return type == GGML_TYPE_IQ2_XXS ? 256 : + type == GGML_TYPE_IQ2_XS ? 512 : + type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? NGRID_IQ1S : 1024; +} + +static int iq2_compare_func(const void * left, const void * right) { + const int * l = (const int *)left; + const int * r = (const int *)right; + return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0; +} + +void iq2xs_init_impl(enum ggml_type type) { + const int gindex = iq2_data_index(type); + const int grid_size = iq2_grid_size(type); + if (iq2_data[gindex].grid) { + return; + } + static const uint16_t kgrid_2bit_256[256] = { + 0, 2, 5, 8, 10, 17, 20, 32, 34, 40, 42, 65, 68, 80, 88, 97, + 100, 128, 130, 138, 162, 257, 260, 272, 277, 320, 388, 408, 512, 514, 546, 642, + 1025, 1028, 1040, 1057, 1060, 1088, 1090, 1096, 1120, 1153, 1156, 1168, 1188, 1280, 1282, 1288, + 1312, 1350, 1385, 1408, 1425, 1545, 1552, 1600, 1668, 1700, 2048, 2053, 2056, 2068, 2088, 2113, + 2116, 2128, 2130, 2184, 2308, 2368, 2562, 2580, 4097, 4100, 4112, 4129, 4160, 4192, 4228, 4240, + 4245, 4352, 4360, 4384, 4432, 4442, 4480, 4644, 4677, 5120, 5128, 5152, 5157, 5193, 5248, 5400, + 5474, 5632, 5654, 6145, 6148, 6160, 6208, 6273, 6400, 6405, 6560, 6737, 8192, 8194, 8202, 8260, + 8289, 8320, 8322, 8489, 8520, 8704, 8706, 9217, 9220, 9232, 9280, 9302, 9472, 9537, 9572, 9872, + 10248, 10272, 10388, 10820, 16385, 16388, 16400, 16408, 16417, 16420, 16448, 16456, 16470, 16480, 16513, 16516, + 16528, 16640, 16672, 16737, 16768, 16773, 16897, 16912, 16968, 16982, 17000, 17408, 17416, 17440, 17536, 17561, + 17682, 17700, 17920, 18433, 18436, 18448, 18496, 18501, 18688, 18776, 18785, 18818, 19013, 19088, 20480, 20488, + 20497, 20505, 20512, 20608, 20616, 20740, 20802, 20900, 21137, 21648, 21650, 21770, 22017, 22100, 22528, 22545, + 22553, 22628, 22848, 23048, 24580, 24592, 24640, 24680, 24832, 24917, 25112, 25184, 25600, 25605, 25872, 25874, + 25988, 26690, 32768, 32770, 32778, 32833, 32898, 33028, 33048, 33088, 33297, 33793, 33796, 33808, 33813, 33856, + 33888, 34048, 34118, 34196, 34313, 34368, 34400, 34818, 35076, 35345, 36868, 36880, 36900, 36928, 37025, 37142, + 37248, 37445, 37888, 37922, 37956, 38225, 39041, 39200, 40962, 41040, 41093, 41225, 41472, 42008, 43088, 43268, + }; + static const uint16_t kgrid_2bit_512[512] = { + 0, 2, 5, 8, 10, 17, 20, 22, 25, 32, 34, 37, 40, 65, 68, 70, + 73, 80, 82, 85, 88, 97, 100, 128, 130, 133, 136, 145, 148, 153, 160, 257, + 260, 262, 265, 272, 274, 277, 280, 282, 289, 292, 320, 322, 325, 328, 337, 340, + 352, 360, 385, 388, 400, 512, 514, 517, 520, 529, 532, 544, 577, 580, 592, 597, + 640, 650, 1025, 1028, 1030, 1033, 1040, 1042, 1045, 1048, 1057, 1060, 1088, 1090, 1093, 1096, + 1105, 1108, 1110, 1120, 1153, 1156, 1168, 1280, 1282, 1285, 1288, 1297, 1300, 1312, 1345, 1348, + 1360, 1377, 1408, 1537, 1540, 1552, 1574, 1600, 1602, 1668, 2048, 2050, 2053, 2056, 2058, 2065, + 2068, 2080, 2085, 2113, 2116, 2128, 2136, 2176, 2208, 2218, 2305, 2308, 2320, 2368, 2433, 2441, + 2560, 2592, 2600, 2710, 2720, 4097, 4100, 4102, 4105, 4112, 4114, 4117, 4120, 4129, 4132, 4160, + 4162, 4165, 4168, 4177, 4180, 4192, 4202, 4225, 4228, 4240, 4352, 4354, 4357, 4360, 4369, 4372, + 4384, 4417, 4420, 4432, 4480, 4500, 4502, 4609, 4612, 4614, 4624, 4672, 4704, 5120, 5122, 5125, + 5128, 5137, 5140, 5152, 5185, 5188, 5193, 5200, 5220, 5248, 5377, 5380, 5392, 5440, 5632, 5652, + 5705, 6145, 6148, 6160, 6162, 6208, 6228, 6278, 6400, 6405, 6502, 6737, 6825, 8192, 8194, 8197, + 8200, 8202, 8209, 8212, 8224, 8257, 8260, 8272, 8320, 8352, 8449, 8452, 8464, 8512, 8520, 8549, + 8704, 8738, 8832, 8872, 9217, 9220, 9232, 9257, 9280, 9472, 9537, 9554, 9625, 9729, 9754, 9894, + 10240, 10248, 10250, 10272, 10325, 10376, 10402, 10600, 10640, 10760, 10784, 10882, 10888, 10890, 16385, 16388, + 16390, 16393, 16400, 16402, 16405, 16408, 16417, 16420, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16480, + 16485, 16513, 16516, 16528, 16640, 16642, 16645, 16648, 16657, 16660, 16672, 16705, 16708, 16720, 16768, 16773, + 16802, 16897, 16900, 16912, 16914, 16937, 16960, 17408, 17410, 17413, 17416, 17425, 17428, 17433, 17440, 17473, + 17476, 17488, 17536, 17556, 17665, 17668, 17680, 17700, 17728, 17818, 17920, 17930, 17988, 18000, 18433, 18436, + 18448, 18496, 18501, 18516, 18530, 18688, 18705, 18756, 18768, 18793, 18948, 20480, 20482, 20485, 20488, 20497, + 20500, 20512, 20520, 20545, 20548, 20560, 20608, 20737, 20740, 20752, 20757, 20800, 20802, 20992, 21060, 21162, + 21505, 21508, 21520, 21537, 21568, 21600, 21633, 21665, 21760, 21768, 21888, 21896, 22049, 22120, 22177, 22528, + 22548, 22593, 22608, 22681, 22810, 22848, 22850, 23173, 24577, 24580, 24592, 24640, 24660, 24674, 24710, 24745, + 24832, 25124, 25162, 25234, 25600, 25622, 25872, 25920, 25925, 26020, 26625, 26730, 26917, 27142, 27220, 27234, + 32768, 32770, 32773, 32776, 32785, 32788, 32800, 32810, 32833, 32836, 32848, 32896, 32898, 32936, 32938, 33025, + 33028, 33030, 33040, 33088, 33105, 33113, 33280, 33312, 33408, 33410, 33440, 33448, 33793, 33796, 33808, 33810, + 33813, 33856, 33888, 33929, 34048, 34116, 34213, 34328, 34410, 34816, 34824, 34853, 34906, 34944, 34946, 34984, + 35078, 35362, 35456, 35464, 35478, 35496, 36865, 36868, 36880, 36928, 36950, 36996, 37120, 37154, 37220, 37462, + 37513, 37888, 37893, 37956, 37968, 37976, 38185, 38288, 38290, 38465, 38993, 39078, 39241, 39445, 39520, 40960, + 40962, 40968, 40970, 40992, 41002, 41120, 41297, 41305, 41382, 41472, 41474, 41480, 41514, 41600, 41632, 42048, + 42133, 42597, 42648, 43018, 43040, 43042, 43048, 43168, 43176, 43268, 43396, 43398, 43560, 43562, 43665, 43690, + }; + static const uint16_t kgrid_1bit_2048[NGRID_IQ1S] = { + 0, 2, 5, 8, 10, 17, 21, 32, 34, 40, 42, 69, 81, 84, 86, 101, + 128, 130, 136, 138, 149, 160, 162, 168, 170, 260, 261, 273, 276, 278, 281, 282, + 293, 321, 326, 329, 338, 341, 346, 353, 356, 358, 360, 389, 401, 404, 406, 421, + 512, 514, 520, 522, 533, 544, 546, 552, 554, 581, 593, 601, 612, 617, 640, 642, + 648, 650, 657, 661, 665, 672, 674, 680, 682, 1041, 1044, 1046, 1061, 1089, 1097, 1109, + 1114, 1124, 1125, 1169, 1177, 1189, 1281, 1284, 1285, 1286, 1301, 1304, 1306, 1321, 1344, 1349, + 1354, 1360, 1361, 1364, 1365, 1366, 1369, 1376, 1378, 1381, 1384, 1386, 1409, 1425, 1429, 1432, + 1434, 1441, 1444, 1445, 1446, 1449, 1556, 1561, 1601, 1604, 1616, 1618, 1621, 1624, 1632, 1633, + 1638, 1641, 1669, 1681, 1684, 1689, 2048, 2050, 2056, 2058, 2069, 2080, 2082, 2088, 2090, 2117, + 2129, 2134, 2149, 2176, 2178, 2184, 2186, 2197, 2208, 2210, 2216, 2218, 2309, 2321, 2324, 2329, + 2340, 2341, 2369, 2384, 2385, 2389, 2401, 2404, 2409, 2449, 2452, 2454, 2457, 2469, 2560, 2562, + 2568, 2570, 2581, 2592, 2594, 2600, 2602, 2629, 2641, 2649, 2657, 2661, 2688, 2690, 2693, 2696, + 2698, 2709, 2720, 2722, 2728, 2730, 4112, 4113, 4116, 4121, 4132, 4133, 4161, 4164, 4176, 4181, + 4184, 4193, 4196, 4197, 4201, 4241, 4244, 4246, 4257, 4261, 4353, 4356, 4358, 4361, 4368, 4370, + 4373, 4376, 4385, 4388, 4393, 4421, 4426, 4432, 4433, 4434, 4436, 4437, 4438, 4441, 4448, 4453, + 4484, 4498, 4501, 4513, 4516, 4625, 4628, 4630, 4645, 4672, 4678, 4681, 4690, 4693, 4696, 4698, + 4708, 4710, 4741, 4753, 4756, 4758, 4773, 5121, 5126, 5129, 5140, 5141, 5144, 5145, 5153, 5158, + 5185, 5189, 5190, 5192, 5194, 5201, 5204, 5205, 5206, 5209, 5218, 5221, 5224, 5252, 5257, 5264, + 5268, 5269, 5272, 5273, 5274, 5281, 5284, 5285, 5289, 5378, 5381, 5386, 5393, 5396, 5397, 5398, + 5401, 5408, 5410, 5413, 5416, 5418, 5441, 5444, 5445, 5446, 5457, 5458, 5460, 5461, 5462, 5465, + 5466, 5473, 5476, 5477, 5478, 5481, 5504, 5506, 5508, 5509, 5512, 5514, 5520, 5521, 5524, 5525, + 5526, 5529, 5530, 5536, 5538, 5541, 5633, 5636, 5637, 5638, 5653, 5654, 5656, 5658, 5665, 5670, + 5696, 5698, 5700, 5701, 5704, 5706, 5713, 5717, 5718, 5720, 5721, 5729, 5732, 5733, 5736, 5737, + 5738, 5766, 5770, 5778, 5781, 5796, 5801, 6161, 6166, 6181, 6209, 6212, 6214, 6217, 6224, 6229, + 6232, 6234, 6240, 6241, 6244, 6246, 6249, 6277, 6289, 6292, 6309, 6416, 6418, 6421, 6426, 6433, + 6437, 6466, 6468, 6469, 6472, 6481, 6484, 6485, 6486, 6489, 6490, 6496, 6501, 6506, 6537, 6545, + 6546, 6549, 6552, 6561, 6566, 6569, 6665, 6678, 6692, 6694, 6724, 6726, 6729, 6736, 6738, 6741, + 6744, 6753, 6758, 6761, 6789, 6801, 6806, 6810, 8192, 8194, 8200, 8202, 8213, 8224, 8226, 8229, + 8232, 8234, 8261, 8273, 8281, 8289, 8293, 8320, 8322, 8328, 8330, 8341, 8352, 8354, 8357, 8360, + 8362, 8453, 8465, 8468, 8473, 8485, 8514, 8516, 8521, 8533, 8536, 8538, 8545, 8548, 8549, 8550, + 8581, 8592, 8598, 8601, 8613, 8705, 8712, 8714, 8721, 8725, 8736, 8738, 8744, 8746, 8773, 8785, + 8790, 8793, 8805, 8833, 8840, 8842, 8849, 8853, 8864, 8866, 8872, 8874, 9221, 9236, 9238, 9241, + 9253, 9284, 9285, 9286, 9289, 9298, 9301, 9304, 9306, 9318, 9349, 9361, 9364, 9369, 9377, 9381, + 9481, 9493, 9505, 9513, 9536, 9541, 9544, 9553, 9556, 9557, 9561, 9570, 9573, 9576, 9609, 9616, + 9620, 9621, 9624, 9626, 9633, 9636, 9638, 9641, 9733, 9744, 9746, 9753, 9765, 9793, 9801, 9813, + 9824, 9825, 9833, 9860, 9862, 9872, 9882, 10240, 10242, 10248, 10250, 10261, 10272, 10274, 10280, 10282, + 10309, 10321, 10324, 10341, 10368, 10370, 10376, 10378, 10400, 10402, 10408, 10410, 10505, 10513, 10516, 10521, + 10533, 10566, 10569, 10578, 10581, 10593, 10596, 10598, 10601, 10629, 10640, 10646, 10649, 10660, 10661, 10752, + 10754, 10760, 10762, 10784, 10786, 10792, 10794, 10821, 10833, 10838, 10841, 10853, 10880, 10882, 10888, 10890, + 10901, 10912, 10914, 10920, 10922, 16389, 16401, 16406, 16421, 16457, 16466, 16469, 16472, 16474, 16481, 16484, + 16486, 16532, 16537, 16545, 16550, 16640, 16641, 16644, 16646, 16649, 16658, 16661, 16662, 16664, 16666, 16673, + 16678, 16681, 16709, 16712, 16714, 16721, 16724, 16725, 16726, 16729, 16730, 16741, 16744, 16746, 16769, 16772, + 16774, 16784, 16786, 16789, 16800, 16801, 16802, 16901, 16913, 16916, 16918, 16933, 16961, 16978, 16981, 16986, + 16996, 17001, 17033, 17044, 17061, 17409, 17429, 17433, 17449, 17477, 17480, 17482, 17489, 17492, 17493, 17494, + 17505, 17506, 17509, 17512, 17514, 17537, 17542, 17545, 17552, 17554, 17557, 17568, 17569, 17577, 17665, 17666, + 17669, 17674, 17681, 17684, 17685, 17686, 17689, 17696, 17701, 17706, 17729, 17732, 17733, 17734, 17737, 17744, + 17745, 17748, 17749, 17750, 17752, 17753, 17761, 17764, 17765, 17766, 17769, 17794, 17796, 17797, 17800, 17809, + 17812, 17813, 17814, 17817, 17818, 17829, 17832, 17834, 17921, 17925, 17929, 17940, 17941, 17944, 17946, 17953, + 17956, 17961, 17984, 17986, 17989, 17992, 18000, 18001, 18002, 18005, 18006, 18009, 18018, 18021, 18024, 18049, + 18053, 18058, 18068, 18069, 18081, 18084, 18086, 18437, 18449, 18453, 18458, 18469, 18498, 18505, 18512, 18517, + 18520, 18529, 18532, 18534, 18537, 18565, 18577, 18580, 18582, 18585, 18597, 18689, 18693, 18694, 18698, 18704, + 18708, 18709, 18712, 18721, 18724, 18726, 18752, 18757, 18762, 18769, 18770, 18772, 18773, 18774, 18777, 18784, + 18786, 18789, 18790, 18794, 18822, 18825, 18834, 18837, 18838, 18840, 18849, 18852, 18854, 18857, 18966, 19012, + 19014, 19017, 19029, 19032, 19034, 19044, 19049, 19092, 19109, 20481, 20484, 20485, 20486, 20489, 20498, 20501, + 20506, 20513, 20516, 20521, 20544, 20549, 20552, 20561, 20564, 20565, 20566, 20569, 20581, 20584, 20614, 20617, + 20629, 20632, 20640, 20641, 20646, 20649, 20741, 20744, 20745, 20746, 20753, 20756, 20757, 20758, 20760, 20761, + 20768, 20773, 20774, 20776, 20778, 20801, 20804, 20805, 20806, 20809, 20816, 20817, 20818, 20820, 20821, 20822, + 20824, 20825, 20826, 20833, 20836, 20837, 20838, 20841, 20866, 20869, 20881, 20884, 20885, 20886, 20889, 20896, + 20901, 20906, 20993, 20998, 21010, 21013, 21018, 21025, 21028, 21058, 21061, 21066, 21073, 21076, 21077, 21078, + 21081, 21090, 21093, 21125, 21136, 21138, 21141, 21145, 21146, 21156, 21508, 21509, 21521, 21524, 21525, 21526, + 21528, 21529, 21537, 21541, 21544, 21546, 21569, 21572, 21573, 21574, 21577, 21578, 21584, 21585, 21588, 21589, + 21590, 21592, 21593, 21594, 21601, 21602, 21604, 21605, 21606, 21609, 21632, 21640, 21642, 21649, 21652, 21653, + 21654, 21657, 21665, 21668, 21669, 21674, 21761, 21762, 21764, 21765, 21766, 21769, 21776, 21777, 21778, 21780, + 21781, 21782, 21785, 21786, 21793, 21796, 21797, 21798, 21801, 21824, 21825, 21826, 21828, 21829, 21830, 21832, + 21833, 21840, 21841, 21842, 21844, 21845, 21846, 21848, 21849, 21850, 21856, 21857, 21860, 21861, 21862, 21864, + 21865, 21866, 21889, 21892, 21893, 21897, 21898, 21904, 21905, 21908, 21909, 21910, 21912, 21913, 21921, 21924, + 21925, 21926, 21929, 22016, 22017, 22018, 22020, 22022, 22024, 22025, 22033, 22036, 22037, 22040, 22041, 22048, + 22049, 22050, 22052, 22053, 22054, 22056, 22057, 22081, 22085, 22086, 22088, 22089, 22090, 22096, 22097, 22098, + 22100, 22101, 22102, 22104, 22105, 22106, 22113, 22116, 22117, 22121, 22146, 22149, 22150, 22152, 22153, 22154, + 22161, 22165, 22170, 22178, 22181, 22182, 22184, 22185, 22532, 22533, 22534, 22537, 22544, 22549, 22552, 22561, + 22570, 22597, 22600, 22602, 22609, 22612, 22613, 22614, 22616, 22617, 22624, 22626, 22628, 22629, 22658, 22665, + 22672, 22674, 22677, 22680, 22689, 22697, 22785, 22786, 22789, 22794, 22801, 22804, 22805, 22806, 22809, 22821, + 22849, 22852, 22853, 22854, 22857, 22864, 22865, 22866, 22868, 22869, 22870, 22872, 22873, 22874, 22881, 22884, + 22885, 22886, 22889, 22913, 22917, 22921, 22929, 22932, 22933, 22934, 22936, 22937, 22949, 23044, 23048, 23061, + 23066, 23072, 23077, 23078, 23081, 23109, 23112, 23113, 23121, 23125, 23126, 23128, 23129, 23138, 23141, 23144, + 23146, 23169, 23178, 23186, 23189, 23190, 23192, 23194, 23201, 24581, 24596, 24598, 24601, 24613, 24644, 24656, + 24661, 24662, 24664, 24666, 24673, 24676, 24678, 24681, 24705, 24726, 24741, 24833, 24836, 24838, 24841, 24850, + 24853, 24865, 24866, 24870, 24873, 24901, 24905, 24913, 24917, 24918, 24921, 24933, 24934, 24938, 24964, 24970, + 24978, 24981, 24993, 24998, 25001, 25105, 25110, 25113, 25152, 25153, 25158, 25173, 25174, 25176, 25184, 25221, + 25233, 25238, 25253, 25617, 25618, 25621, 25622, 25626, 25633, 25638, 25641, 25664, 25666, 25669, 25672, 25674, + 25681, 25684, 25685, 25686, 25689, 25690, 25696, 25698, 25701, 25732, 25733, 25737, 25744, 25746, 25748, 25749, + 25750, 25752, 25754, 25761, 25764, 25769, 25861, 25864, 25866, 25873, 25877, 25878, 25881, 25924, 25925, 25926, + 25929, 25936, 25937, 25940, 25941, 25942, 25945, 25953, 25956, 25957, 25958, 25961, 25990, 25993, 25994, 26001, + 26005, 26006, 26009, 26010, 26018, 26021, 26022, 26024, 26114, 26121, 26133, 26144, 26150, 26152, 26153, 26176, + 26181, 26184, 26186, 26193, 26196, 26197, 26198, 26200, 26202, 26208, 26213, 26216, 26240, 26242, 26245, 26250, + 26260, 26262, 26264, 26265, 26272, 26276, 26278, 26282, 26646, 26649, 26661, 26689, 26706, 26709, 26714, 26721, + 26729, 26757, 26769, 26776, 26790, 26881, 26884, 26896, 26901, 26913, 26916, 26918, 26921, 26944, 26945, 26949, + 26950, 26952, 26961, 26964, 26965, 26966, 26969, 26976, 26981, 26986, 27010, 27012, 27018, 27029, 27041, 27044, + 27045, 27049, 27153, 27158, 27160, 27201, 27204, 27209, 27216, 27221, 27224, 27226, 27236, 27237, 27241, 27270, + 27284, 27288, 27290, 27302, 32768, 32770, 32776, 32778, 32800, 32802, 32808, 32810, 32837, 32848, 32849, 32852, + 32854, 32857, 32869, 32896, 32898, 32904, 32906, 32917, 32928, 32930, 32936, 32938, 33029, 33041, 33044, 33046, + 33049, 33061, 33089, 33092, 33097, 33104, 33106, 33109, 33110, 33112, 33113, 33124, 33126, 33129, 33157, 33161, + 33172, 33174, 33177, 33189, 33280, 33282, 33288, 33290, 33301, 33312, 33314, 33320, 33322, 33361, 33364, 33369, + 33381, 33408, 33410, 33416, 33418, 33429, 33440, 33442, 33448, 33450, 33812, 33817, 33857, 33860, 33873, 33877, + 33882, 33889, 33892, 33897, 33940, 33945, 34049, 34057, 34066, 34069, 34074, 34086, 34089, 34112, 34113, 34117, + 34120, 34129, 34132, 34133, 34134, 34137, 34138, 34149, 34150, 34152, 34154, 34177, 34180, 34182, 34185, 34192, + 34194, 34197, 34200, 34214, 34321, 34326, 34329, 34341, 34369, 34372, 34377, 34378, 34384, 34389, 34393, 34394, + 34401, 34406, 34410, 34437, 34449, 34458, 34468, 34816, 34818, 34824, 34826, 34837, 34848, 34850, 34856, 34858, + 34881, 34885, 34897, 34900, 34905, 34917, 34921, 34944, 34946, 34952, 34954, 34965, 34976, 34978, 34984, 34986, + 35077, 35078, 35089, 35092, 35094, 35109, 35137, 35140, 35142, 35145, 35152, 35154, 35157, 35162, 35169, 35172, + 35205, 35222, 35225, 35237, 35328, 35330, 35336, 35338, 35349, 35360, 35362, 35368, 35370, 35397, 35409, 35412, + 35414, 35456, 35458, 35464, 35466, 35477, 35488, 35490, 35496, 35498, 36869, 36881, 36886, 36888, 36889, 36901, + 36929, 36934, 36937, 36949, 36952, 36954, 36969, 36970, 36997, 37009, 37012, 37014, 37017, 37029, 37121, 37124, + 37126, 37129, 37136, 37141, 37144, 37146, 37153, 37156, 37158, 37161, 37184, 37189, 37200, 37201, 37204, 37205, + 37206, 37209, 37218, 37221, 37252, 37254, 37266, 37269, 37272, 37281, 37284, 37286, 37289, 37381, 37393, 37396, + 37401, 37413, 37444, 37446, 37449, 37456, 37458, 37461, 37464, 37478, 37481, 37509, 37524, 37526, 37545, 37889, + 37892, 37894, 37904, 37909, 37912, 37926, 37952, 37962, 37969, 37972, 37973, 37974, 37976, 37977, 37984, 37985, + 37986, 37989, 38020, 38022, 38034, 38036, 38037, 38040, 38049, 38057, 38144, 38149, 38152, 38154, 38160, 38161, + 38164, 38165, 38166, 38169, 38177, 38181, 38185, 38186, 38209, 38212, 38213, 38214, 38217, 38224, 38225, 38226, + 38228, 38229, 38230, 38232, 38233, 38234, 38241, 38244, 38245, 38246, 38249, 38273, 38277, 38280, 38289, 38290, + 38292, 38293, 38294, 38297, 38298, 38304, 38306, 38309, 38312, 38314, 38401, 38404, 38416, 38421, 38425, 38432, + 38438, 38441, 38469, 38472, 38473, 38481, 38482, 38485, 38486, 38489, 38501, 38504, 38530, 38532, 38537, 38538, + 38546, 38548, 38549, 38564, 38566, 38569, 38917, 38934, 38937, 38949, 38977, 38982, 38992, 38994, 38997, 38998, + 39002, 39012, 39013, 39045, 39057, 39062, 39065, 39077, 39172, 39174, 39177, 39184, 39186, 39189, 39192, 39194, + 39200, 39201, 39204, 39206, 39232, 39234, 39237, 39240, 39242, 39249, 39252, 39253, 39254, 39257, 39266, 39269, + 39270, 39274, 39297, 39300, 39312, 39314, 39317, 39322, 39329, 39334, 39429, 39445, 39461, 39492, 39494, 39497, + 39504, 39509, 39512, 39521, 39557, 39569, 39572, 39573, 39574, 40960, 40962, 40968, 40970, 40981, 40992, 40994, + 41000, 41002, 41029, 41041, 41044, 41046, 41049, 41088, 41090, 41096, 41098, 41109, 41120, 41122, 41128, 41130, + 41221, 41225, 41233, 41236, 41238, 41241, 41242, 41286, 41289, 41297, 41301, 41304, 41306, 41313, 41316, 41349, + 41360, 41362, 41366, 41369, 41474, 41480, 41482, 41488, 41497, 41506, 41512, 41514, 41541, 41553, 41558, 41561, + 41573, 41600, 41602, 41608, 41610, 41621, 41632, 41634, 41640, 41642, 42009, 42021, 42049, 42052, 42064, 42068, + 42069, 42072, 42074, 42081, 42085, 42086, 42088, 42089, 42117, 42246, 42249, 42256, 42258, 42261, 42264, 42278, + 42281, 42306, 42309, 42321, 42324, 42325, 42326, 42329, 42341, 42346, 42369, 42372, 42373, 42374, 42377, 42386, + 42389, 42392, 42501, 42513, 42518, 42522, 42529, 42533, 42564, 42566, 42570, 42578, 42581, 42582, 42584, 42592, + 42594, 42630, 42640, 42645, 42646, 42649, 42657, 42660, 42662, 43008, 43010, 43016, 43018, 43040, 43042, 43048, + 43050, 43089, 43092, 43094, 43097, 43136, 43138, 43144, 43146, 43157, 43168, 43170, 43176, 43178, 43269, 43284, + 43289, 43297, 43301, 43329, 43344, 43349, 43354, 43361, 43366, 43369, 43408, 43414, 43520, 43522, 43528, 43530, + 43552, 43554, 43560, 43562, 43601, 43604, 43606, 43648, 43650, 43656, 43658, 43669, 43680, 43682, 43688, 43690, + }; + static const uint16_t kgrid_2bit_1024[1024] = { + 0, 2, 5, 8, 10, 17, 20, 22, 25, 32, 34, 37, 40, 65, 68, 70, + 73, 80, 82, 85, 88, 97, 100, 102, 105, 128, 130, 133, 136, 145, 148, 160, + 165, 170, 257, 260, 262, 265, 272, 274, 277, 280, 289, 292, 320, 322, 325, 328, + 337, 340, 342, 345, 352, 357, 360, 385, 388, 400, 402, 405, 417, 420, 512, 514, + 517, 520, 529, 532, 544, 554, 577, 580, 582, 585, 592, 597, 640, 645, 650, 660, + 674, 1025, 1028, 1030, 1033, 1040, 1042, 1045, 1048, 1057, 1060, 1062, 1065, 1088, 1090, 1093, + 1096, 1098, 1105, 1108, 1110, 1113, 1120, 1122, 1125, 1153, 1156, 1158, 1161, 1168, 1173, 1176, + 1185, 1188, 1280, 1282, 1285, 1288, 1290, 1297, 1300, 1302, 1305, 1312, 1317, 1320, 1345, 1348, + 1350, 1353, 1360, 1362, 1365, 1368, 1377, 1380, 1408, 1410, 1413, 1416, 1425, 1428, 1440, 1537, + 1540, 1542, 1545, 1552, 1557, 1600, 1605, 1608, 1617, 1620, 1632, 1665, 1668, 1680, 2048, 2050, + 2053, 2056, 2065, 2068, 2070, 2073, 2080, 2085, 2090, 2113, 2116, 2118, 2121, 2128, 2130, 2133, + 2136, 2145, 2148, 2176, 2181, 2196, 2218, 2305, 2308, 2320, 2322, 2325, 2328, 2337, 2368, 2373, + 2376, 2385, 2388, 2400, 2433, 2448, 2560, 2577, 2580, 2594, 2600, 2602, 2640, 2713, 4097, 4100, + 4102, 4105, 4112, 4114, 4117, 4120, 4129, 4132, 4134, 4160, 4162, 4165, 4168, 4177, 4180, 4182, + 4185, 4192, 4194, 4197, 4200, 4225, 4228, 4230, 4240, 4245, 4248, 4257, 4260, 4352, 4354, 4357, + 4360, 4362, 4369, 4372, 4374, 4377, 4384, 4386, 4389, 4392, 4417, 4420, 4422, 4425, 4432, 4434, + 4437, 4440, 4449, 4452, 4480, 4482, 4485, 4488, 4497, 4500, 4609, 4612, 4617, 4624, 4629, 4641, + 4644, 4672, 4677, 4689, 4692, 4737, 4740, 4752, 5120, 5122, 5125, 5128, 5137, 5140, 5142, 5145, + 5152, 5157, 5160, 5185, 5188, 5190, 5193, 5200, 5202, 5205, 5208, 5217, 5220, 5248, 5250, 5253, + 5256, 5265, 5268, 5280, 5377, 5380, 5382, 5385, 5392, 5394, 5397, 5400, 5409, 5412, 5440, 5442, + 5445, 5448, 5457, 5460, 5472, 5505, 5508, 5520, 5632, 5637, 5640, 5649, 5652, 5664, 5697, 5700, + 5712, 5760, 5802, 6145, 6148, 6150, 6153, 6160, 6165, 6168, 6177, 6208, 6210, 6213, 6216, 6225, + 6228, 6240, 6273, 6276, 6400, 6402, 6405, 6408, 6417, 6420, 6432, 6465, 6468, 6480, 6505, 6562, + 6660, 6672, 6720, 6742, 8192, 8194, 8197, 8200, 8209, 8212, 8214, 8217, 8224, 8229, 8234, 8257, + 8260, 8272, 8274, 8277, 8292, 8320, 8330, 8340, 8362, 8449, 8452, 8464, 8466, 8469, 8481, 8512, + 8514, 8517, 8529, 8532, 8544, 8577, 8580, 8592, 8704, 8714, 8738, 8744, 8746, 8772, 8784, 8840, + 8842, 8872, 9217, 9220, 9222, 9225, 9232, 9237, 9240, 9249, 9252, 9280, 9282, 9285, 9288, 9297, + 9300, 9312, 9345, 9348, 9360, 9472, 9477, 9480, 9489, 9492, 9504, 9537, 9540, 9552, 9574, 9600, + 9729, 9732, 9744, 9792, 9817, 10240, 10245, 10257, 10260, 10305, 10308, 10320, 10378, 10410, 10497, 10500, + 10512, 10645, 10762, 10786, 10852, 10888, 10890, 16385, 16388, 16390, 16393, 16400, 16402, 16405, 16408, 16410, + 16417, 16420, 16422, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16470, 16473, 16480, 16482, 16485, 16513, + 16516, 16528, 16533, 16536, 16545, 16548, 16640, 16642, 16645, 16648, 16657, 16660, 16662, 16665, 16672, 16674, + 16677, 16705, 16708, 16710, 16713, 16720, 16722, 16725, 16728, 16737, 16740, 16768, 16770, 16773, 16776, 16785, + 16788, 16800, 16897, 16900, 16912, 16914, 16917, 16920, 16932, 16960, 16965, 16968, 16977, 16980, 16992, 17025, + 17028, 17408, 17410, 17413, 17416, 17418, 17425, 17428, 17430, 17433, 17440, 17442, 17445, 17448, 17473, 17476, + 17478, 17481, 17488, 17490, 17493, 17496, 17505, 17508, 17536, 17538, 17541, 17544, 17553, 17556, 17568, 17665, + 17668, 17670, 17673, 17680, 17682, 17685, 17688, 17697, 17700, 17728, 17730, 17733, 17736, 17745, 17748, 17760, + 17770, 17793, 17796, 17808, 17920, 17922, 17925, 17928, 17937, 17940, 17952, 17985, 17988, 18000, 18048, 18085, + 18433, 18436, 18441, 18448, 18450, 18453, 18456, 18465, 18468, 18496, 18498, 18501, 18504, 18513, 18516, 18528, + 18564, 18576, 18688, 18690, 18693, 18696, 18705, 18708, 18720, 18753, 18756, 18768, 18816, 18838, 18945, 18948, + 18960, 19008, 20480, 20482, 20485, 20488, 20497, 20500, 20502, 20505, 20512, 20514, 20517, 20520, 20545, 20548, + 20550, 20553, 20560, 20562, 20565, 20568, 20577, 20580, 20608, 20610, 20613, 20616, 20625, 20628, 20737, 20740, + 20742, 20745, 20752, 20754, 20757, 20760, 20769, 20772, 20800, 20802, 20805, 20808, 20817, 20820, 20832, 20865, + 20868, 20880, 20992, 20997, 21000, 21009, 21012, 21024, 21057, 21060, 21072, 21097, 21120, 21505, 21508, 21510, + 21513, 21520, 21522, 21525, 21528, 21537, 21540, 21568, 21570, 21573, 21576, 21585, 21588, 21600, 21633, 21636, + 21648, 21760, 21762, 21765, 21768, 21777, 21780, 21792, 21825, 21828, 21840, 21888, 22017, 22020, 22032, 22054, + 22080, 22528, 22530, 22533, 22536, 22545, 22548, 22560, 22593, 22596, 22608, 22618, 22656, 22785, 22788, 22800, + 22848, 23040, 23065, 23173, 23208, 24577, 24580, 24582, 24592, 24594, 24597, 24600, 24609, 24612, 24640, 24645, + 24648, 24657, 24660, 24672, 24708, 24720, 24832, 24834, 24837, 24840, 24849, 24852, 24864, 24897, 24900, 24912, + 24960, 24985, 25092, 25104, 25152, 25174, 25249, 25600, 25605, 25608, 25617, 25620, 25632, 25665, 25668, 25680, + 25728, 25857, 25860, 25872, 25920, 25930, 25960, 26002, 26112, 26260, 26625, 26628, 26640, 26725, 26776, 26880, + 26922, 27202, 27297, 32768, 32770, 32773, 32776, 32785, 32788, 32793, 32800, 32805, 32833, 32836, 32848, 32850, + 32853, 32856, 32865, 32896, 32901, 32913, 32916, 33025, 33028, 33033, 33040, 33042, 33045, 33048, 33057, 33060, + 33088, 33090, 33093, 33096, 33105, 33108, 33153, 33156, 33168, 33193, 33280, 33285, 33290, 33297, 33300, 33345, + 33348, 33360, 33793, 33796, 33798, 33801, 33808, 33810, 33813, 33816, 33825, 33856, 33858, 33861, 33864, 33873, + 33876, 33888, 33921, 33924, 33936, 34048, 34050, 34053, 34056, 34065, 34068, 34080, 34113, 34116, 34128, 34176, + 34186, 34305, 34308, 34320, 34345, 34368, 34816, 34821, 34833, 34836, 34881, 34884, 34896, 34978, 35073, 35076, + 35136, 35173, 35362, 35416, 35418, 35458, 35490, 36865, 36868, 36873, 36880, 36882, 36885, 36888, 36900, 36928, + 36930, 36933, 36936, 36945, 36948, 36960, 36993, 36996, 37008, 37120, 37125, 37137, 37140, 37185, 37188, 37200, + 37210, 37377, 37380, 37392, 37440, 37542, 37888, 37890, 37893, 37896, 37905, 37908, 37920, 37953, 37956, 37968, + 38016, 38038, 38145, 38148, 38160, 38208, 38296, 38305, 38400, 38470, 38500, 38913, 38916, 38928, 38950, 38976, + 39081, 39168, 39241, 39250, 39568, 40960, 40965, 40970, 40980, 40994, 41002, 41025, 41028, 41040, 41122, 41130, + 41280, 41317, 41474, 41482, 41506, 41512, 41514, 41602, 41608, 41610, 41640, 41985, 41988, 42000, 42048, 42121, + 42148, 42240, 42265, 42577, 43018, 43048, 43170, 43348, 43398, 43528, 43530, 43552, 43554, 43560, 43656, 43690, + }; + + const int kmap_size = 43692; + //const int nwant = type == GGML_TYPE_IQ1_S ? 3 : 2; + const int nwant = type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 3 : type == GGML_TYPE_IQ2_S ? 1 : 2; + const uint16_t * kgrid = type == GGML_TYPE_IQ2_XXS ? kgrid_2bit_256 : + type == GGML_TYPE_IQ2_XS ? kgrid_2bit_512 : + type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? kgrid_1bit_2048 : kgrid_2bit_1024; + uint64_t * kgrid_q2xs; + int * kmap_q2xs; + uint16_t * kneighbors_q2xs; + + //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size); + uint64_t * the_grid = (uint64_t *)malloc(grid_size*sizeof(uint64_t)); + for (int k = 0; k < grid_size; ++k) { + int8_t * pos = (int8_t *)(the_grid + k); + for (int i = 0; i < 8; ++i) { + int l = (kgrid[k] >> 2*i) & 0x3; + pos[i] = 2*l + 1; + } + } + kgrid_q2xs = the_grid; + iq2_data[gindex].grid = the_grid; + kmap_q2xs = (int *)malloc(kmap_size*sizeof(int)); + iq2_data[gindex].map = kmap_q2xs; + for (int i = 0; i < kmap_size; ++i) kmap_q2xs[i] = -1; + uint64_t aux64; + uint8_t * aux8 = (uint8_t *)&aux64; + for (int i = 0; i < grid_size; ++i) { + aux64 = kgrid_q2xs[i]; + uint16_t index = 0; + for (int k=0; k<8; ++k) { + uint16_t q = (aux8[k] - 1)/2; + index |= (q << 2*k); + } + kmap_q2xs[index] = i; + } + int8_t pos[8]; + int * dist2 = (int *)malloc(2*grid_size*sizeof(int)); + int num_neighbors = 0, num_not_in_map = 0; + for (int i = 0; i < kmap_size; ++i) { + if (kmap_q2xs[i] >= 0) continue; + ++num_not_in_map; + for (int k = 0; k < 8; ++k) { + int l = (i >> 2*k) & 0x3; + pos[k] = 2*l + 1; + } + for (int j = 0; j < grid_size; ++j) { + const int8_t * pg = (const int8_t *)(kgrid_q2xs + j); + int d2 = 0; + for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]); + dist2[2*j+0] = d2; + dist2[2*j+1] = j; + } + qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func); + int n = 0; int d2 = dist2[0]; + int nhave = 1; + for (int j = 0; j < grid_size; ++j) { + if (dist2[2*j] > d2) { + if (nhave == nwant) break; + d2 = dist2[2*j]; + ++nhave; + } + ++n; + } + num_neighbors += n; + } + //printf("%s: %d neighbours in total\n", __func__, num_neighbors); + kneighbors_q2xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t)); + iq2_data[gindex].neighbours = kneighbors_q2xs; + int counter = 0; + for (int i = 0; i < kmap_size; ++i) { + if (kmap_q2xs[i] >= 0) continue; + for (int k = 0; k < 8; ++k) { + int l = (i >> 2*k) & 0x3; + pos[k] = 2*l + 1; + } + for (int j = 0; j < grid_size; ++j) { + const int8_t * pg = (const int8_t *)(kgrid_q2xs + j); + int d2 = 0; + for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]); + dist2[2*j+0] = d2; + dist2[2*j+1] = j; + } + qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func); + kmap_q2xs[i] = -(counter + 1); + int d2 = dist2[0]; + uint16_t * start = &kneighbors_q2xs[counter++]; + int n = 0, nhave = 1; + for (int j = 0; j < grid_size; ++j) { + if (dist2[2*j] > d2) { + if (nhave == nwant) break; + d2 = dist2[2*j]; + ++nhave; + } + kneighbors_q2xs[counter++] = dist2[2*j+1]; + ++n; + } + *start = n; + } + free(dist2); +} + +void iq2xs_free_impl(enum ggml_type type) { + GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S); + const int gindex = iq2_data_index(type); + if (iq2_data[gindex].grid) { + free(iq2_data[gindex].grid); iq2_data[gindex].grid = NULL; + free(iq2_data[gindex].map); iq2_data[gindex].map = NULL; + free(iq2_data[gindex].neighbours); iq2_data[gindex].neighbours = NULL; + } +} + +static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid, + const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) { + int num_neighbors = neighbours[0]; + GGML_ASSERT(num_neighbors > 0); + float best_d2 = FLT_MAX; + int grid_index = -1; + for (int j = 1; j <= num_neighbors; ++j) { + const int8_t * pg = (const int8_t *)(grid + neighbours[j]); + float d2 = 0; + for (int i = 0; i < 8; ++i) { + float q = pg[i]; + float diff = scale*q - xval[i]; + d2 += weight[i]*diff*diff; + } + if (d2 < best_d2) { + best_d2 = d2; grid_index = neighbours[j]; + } + } + GGML_ASSERT(grid_index >= 0); + const int8_t * pg = (const int8_t *)(grid + grid_index); + for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2; + return grid_index; +} + +static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) { + + const int gindex = iq2_data_index(GGML_TYPE_IQ2_XXS); + + const uint64_t * kgrid_q2xs = iq2_data[gindex].grid; + const int * kmap_q2xs = iq2_data[gindex].map; + const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; + + GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + const int kMaxQ = 3; + + const int64_t nbl = n/QK_K; + + block_iq2_xxs * y = vy; + + float scales[QK_K/32]; + float weight[32]; + float xval[32]; + int8_t L[32]; + int8_t Laux[32]; + float waux[32]; + uint8_t block_signs[4]; + uint32_t q2[2*(QK_K/32)]; + + for (int ibl = 0; ibl < nbl; ++ibl) { + + y[ibl].d = GGML_FP32_TO_FP16(0.f); + memset(q2, 0, QK_K/4); + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = sumx2/QK_K; + + for (int ib = 0; ib < QK_K/32; ++ib) { + const float * xb = xbl + 32*ib; + const float * qw = quant_weights + QK_K*ibl + 32*ib; + for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]); + for (int k = 0; k < 4; ++k) { + int nflip = 0; + uint8_t s = 0; + for (int i = 0; i < 8; ++i) { + if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i]; + else { + xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i); + } + } + if (nflip%2) { + int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin]; + for (int i = 1; i < 8; ++i) { + float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i]; + if (ax < min) { + min = ax; imin = i; + } + } + xval[8*k+imin] = -xval[8*k+imin]; + s ^= (1 << imin); + } + block_signs[k] = s & 127; + } + float max = xval[0]; + for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]); + if (!max) { + scales[ib] = 0; + memset(L, 0, 32); + continue; + } + float scale = make_qp_quants(32, kMaxQ+1, xval, (uint8_t*)L, weight); + float eff_max = scale*kMaxQ; + float best = 0; + for (int is = -6; is <= 6; ++is) { + float id = (2*kMaxQ-1+is*0.1f)/eff_max; + float this_scale = 1/id; + for (int k = 0; k < 4; ++k) { + for (int i = 0; i < 8; ++i) { + int l = nearest_int(0.5f*(id*xval[8*k+i]-1)); + Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l)); + } + uint16_t u = 0; + for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i); + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 32; ++i) { + float w = weight[i]; + float q = 2*Laux[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0 && sumqx*sumqx > best*sumq2) { + scale = sumqx/sumq2; best = scale*sumqx; + memcpy(L, Laux, 32); + } + } + if (scale > 0) { + float id = 1/scale; + for (int k = 0; k < 4; ++k) { + uint16_t u = 0; + for (int i = 0; i < 8; ++i) { + int l = nearest_int(0.5f*(id*xval[8*k+i]-1)); + l = MAX(0, MIN(kMaxQ-1, l)); + u |= (l << 2*i); + } + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k); + } + const int8_t * pg = (const int8_t *)(kgrid_q2xs + grid_index); + for (int i = 0; i < 8; ++i) L[8*k+i] = (pg[i] - 1)/2; + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 32; ++i) { + float w = weight[i]; + float q = 2*L[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0) scale = sumqx/sumq2; + } + if (scale < 0) { + // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale) + // and correspondingly flip quant signs. + scale = -scale; + for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127; + } + for (int k = 0; k < 4; ++k) { + uint16_t u = 0; + for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i); + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + printf("Oops: found point %u not on grid:", u); + for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]); + printf("\n"); + GGML_ASSERT(false); + } + q2[2*ib+0] |= (grid_index << 8*k); + q2[2*ib+1] |= (block_signs[k] << 7*k); + } + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + memset(y[ibl].qs, 0, QK_K/4); + continue; + } + + float d = max_scale/31; + y[ibl].d = GGML_FP32_TO_FP16(d); + float id = 1/d; + for (int ib = 0; ib < QK_K/32; ++ib) { + int l = nearest_int(0.5f*(id*scales[ib]-1)); + l = MAX(0, MIN(15, l)); + q2[2*ib+1] |= ((uint32_t)l << 28); + } + memcpy(y[ibl].qs, q2, QK_K/4); + } +} + +static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) { + + const int gindex = iq2_data_index(GGML_TYPE_IQ2_XS); + + const uint64_t * kgrid_q2xs = iq2_data[gindex].grid; + const int * kmap_q2xs = iq2_data[gindex].map; + const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; + + GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + const int kMaxQ = 3; + + const int64_t nbl = n/QK_K; + + block_iq2_xs * y = vy; + + float scales[QK_K/16]; + float weight[16]; + float xval[16]; + int8_t L[16]; + int8_t Laux[16]; + float waux[16]; + bool is_on_grid[2]; + bool is_on_grid_aux[2]; + uint8_t block_signs[2]; + uint16_t q2[2*(QK_K/16)]; + + for (int ibl = 0; ibl < nbl; ++ibl) { + + y[ibl].d = GGML_FP32_TO_FP16(0.f); + memset(q2, 0, QK_K/4); + memset(y[ibl].scales, 0, QK_K/32); + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = sumx2/QK_K; + + for (int ib = 0; ib < QK_K/16; ++ib) { + const float * xb = xbl + 16*ib; + const float * qw = quant_weights + QK_K*ibl + 16*ib; + for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]); + for (int k = 0; k < 2; ++k) { + int nflip = 0; + uint8_t s = 0; + for (int i = 0; i < 8; ++i) { + if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i]; + else { + xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i); + } + } + if (nflip%2) { + int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin]; + for (int i = 1; i < 8; ++i) { + float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i]; + if (ax < min) { + min = ax; imin = i; + } + } + xval[8*k+imin] = -xval[8*k+imin]; + s ^= (1 << imin); + } + block_signs[k] = s & 127; + } + float max = xval[0]; + for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]); + if (!max) { + scales[ib] = 0; + memset(L, 0, 16); + continue; + } + float best = 0; + float scale = max/(2*kMaxQ-1); + is_on_grid[0] = is_on_grid[1] = true; + for (int is = -9; is <= 9; ++is) { + float id = (2*kMaxQ-1+is*0.1f)/max; + float this_scale = 1/id; + for (int k = 0; k < 2; ++k) { + for (int i = 0; i < 8; ++i) { + int l = nearest_int(0.5f*(id*xval[8*k+i]-1)); + Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l)); + } + uint16_t u = 0; + for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i); + int grid_index = kmap_q2xs[u]; + is_on_grid_aux[k] = true; + if (grid_index < 0) { + is_on_grid_aux[k] = false; + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 16; ++i) { + float w = weight[i]; + float q = 2*Laux[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0 && sumqx*sumqx > best*sumq2) { + scale = sumqx/sumq2; best = scale*sumqx; + for (int i = 0; i < 16; ++i) L[i] = Laux[i]; + for (int k = 0; k < 2; ++k) is_on_grid[k] = is_on_grid_aux[k]; + } + } + int n_not_ongrid = 0; + for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid; + if (n_not_ongrid > 0 && scale > 0) { + float id = 1/scale; + for (int k = 0; k < 2; ++k) { + if (is_on_grid[k]) continue; + uint16_t u = 0; + for (int i = 0; i < 8; ++i) { + int l = nearest_int(0.5f*(id*xval[8*k+i]-1)); + l = MAX(0, MIN(kMaxQ-1, l)); + u |= (l << 2*i); + L[8*k + i] = l; + } + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 16; ++i) { + float w = weight[i]; + float q = 2*L[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0) scale = sumqx/sumq2; + } + if (scale < 0) { + scale = -scale; + for (int k = 0; k < 2; ++k) block_signs[k] = (~block_signs[k]) & 127; + } + for (int k = 0; k < 2; ++k) { + uint16_t u = 0; + for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i); + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + printf("Oops: found point %u not on grid:", u); + for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]); + printf("\n"); + GGML_ASSERT(false); + } + q2[2*ib+k] = grid_index | (block_signs[k] << 9); + } + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + memset(y[ibl].qs, 0, QK_K/4); + continue; + } + + float d = max_scale/31; + y[ibl].d = GGML_FP32_TO_FP16(d); + float id = 1/d; + for (int ib = 0; ib < QK_K/16; ++ib) { + int l = nearest_int(0.5f*(id*scales[ib]-1)); + l = MAX(0, MIN(15, l)); + if (ib%2 == 0) y[ibl].scales[ib/2] = l; + else y[ibl].scales[ib/2] |= (l << 4); + } + memcpy(y[ibl].qs, q2, QK_K/4); + + } +} + +size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq2_xxs_impl(src, qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += nblock*sizeof(block_iq2_xxs); + } + return nrow * nblock * sizeof(block_iq2_xxs); +} + +size_t quantize_iq2_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq2_xs_impl(src, qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += nblock*sizeof(block_iq2_xs); + } + return nrow * nblock * sizeof(block_iq2_xs); +} + +// +// ============================================= 3-bit using D4 lattice +// + +typedef struct { + uint32_t * grid; + int * map; + uint16_t * neighbours; +} iq3_entry_t; + +static iq3_entry_t iq3_data[2] = { + {NULL, NULL, NULL}, + {NULL, NULL, NULL}, +}; + +static inline int iq3_data_index(int grid_size) { + (void)grid_size; + GGML_ASSERT(grid_size == 256 || grid_size == 512); + return grid_size == 256 ? 0 : 1; +} + +static int iq3_compare_func(const void * left, const void * right) { + const int * l = (const int *)left; + const int * r = (const int *)right; + return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0; +} + +void iq3xs_init_impl(int grid_size) { + const int gindex = iq3_data_index(grid_size); + if (iq3_data[gindex].grid) { + return; + } + static const uint16_t kgrid_256[256] = { + 0, 2, 4, 9, 11, 15, 16, 18, 25, 34, 59, 61, 65, 67, 72, 74, + 81, 85, 88, 90, 97, 108, 120, 128, 130, 132, 137, 144, 146, 153, 155, 159, + 169, 175, 189, 193, 199, 200, 202, 213, 248, 267, 287, 292, 303, 315, 317, 321, + 327, 346, 362, 413, 436, 456, 460, 462, 483, 497, 513, 515, 520, 522, 529, 531, + 536, 538, 540, 551, 552, 576, 578, 585, 592, 594, 641, 643, 648, 650, 657, 664, + 698, 704, 706, 720, 729, 742, 758, 769, 773, 808, 848, 852, 870, 889, 901, 978, + 992, 1024, 1026, 1033, 1035, 1040, 1042, 1046, 1049, 1058, 1089, 1091, 1093, 1096, 1098, 1105, + 1112, 1139, 1143, 1144, 1152, 1154, 1161, 1167, 1168, 1170, 1183, 1184, 1197, 1217, 1224, 1228, + 1272, 1276, 1309, 1323, 1347, 1367, 1377, 1404, 1473, 1475, 1486, 1509, 1537, 1544, 1546, 1553, + 1555, 1576, 1589, 1594, 1600, 1602, 1616, 1625, 1636, 1638, 1665, 1667, 1672, 1685, 1706, 1722, + 1737, 1755, 1816, 1831, 1850, 1856, 1862, 1874, 1901, 1932, 1950, 1971, 2011, 2032, 2052, 2063, + 2077, 2079, 2091, 2095, 2172, 2192, 2207, 2208, 2224, 2230, 2247, 2277, 2308, 2345, 2356, 2389, + 2403, 2424, 2501, 2504, 2506, 2520, 2570, 2593, 2616, 2624, 2630, 2646, 2669, 2700, 2714, 2746, + 2754, 2795, 2824, 2835, 2839, 2874, 2882, 2905, 2984, 3028, 3042, 3092, 3108, 3110, 3124, 3153, + 3185, 3215, 3252, 3288, 3294, 3364, 3397, 3434, 3483, 3523, 3537, 3587, 3589, 3591, 3592, 3610, + 3626, 3670, 3680, 3722, 3749, 3754, 3776, 3789, 3803, 3824, 3857, 3873, 3904, 3906, 3924, 3992, + }; + static const uint16_t kgrid_512[512] = { + 0, 1, 2, 5, 7, 8, 9, 10, 12, 14, 16, 17, 21, 27, 32, 34, + 37, 39, 41, 43, 48, 50, 57, 60, 63, 64, 65, 66, 68, 72, 73, 77, + 80, 83, 87, 89, 93, 100, 113, 117, 122, 128, 129, 133, 135, 136, 139, 142, + 145, 149, 152, 156, 162, 165, 167, 169, 171, 184, 187, 195, 201, 205, 208, 210, + 217, 219, 222, 228, 232, 234, 247, 249, 253, 256, 267, 271, 273, 276, 282, 288, + 291, 297, 312, 322, 324, 336, 338, 342, 347, 353, 357, 359, 374, 379, 390, 393, + 395, 409, 426, 441, 448, 450, 452, 464, 466, 470, 475, 488, 492, 512, 513, 514, + 516, 520, 521, 523, 525, 527, 528, 530, 537, 540, 542, 556, 558, 561, 570, 576, + 577, 579, 582, 584, 588, 593, 600, 603, 609, 616, 618, 632, 638, 640, 650, 653, + 655, 656, 660, 666, 672, 675, 685, 688, 698, 705, 708, 711, 712, 715, 721, 727, + 728, 732, 737, 754, 760, 771, 773, 778, 780, 793, 795, 802, 806, 808, 812, 833, + 840, 843, 849, 856, 858, 873, 912, 916, 919, 932, 934, 961, 963, 968, 970, 977, + 989, 993, 1010, 1016, 1024, 1025, 1027, 1029, 1031, 1032, 1034, 1036, 1038, 1041, 1043, 1047, + 1048, 1050, 1057, 1059, 1061, 1064, 1066, 1079, 1080, 1083, 1085, 1088, 1090, 1096, 1099, 1103, + 1106, 1109, 1113, 1116, 1122, 1129, 1153, 1156, 1159, 1169, 1171, 1176, 1183, 1185, 1195, 1199, + 1209, 1212, 1216, 1218, 1221, 1225, 1234, 1236, 1241, 1243, 1250, 1256, 1270, 1281, 1287, 1296, + 1299, 1306, 1309, 1313, 1338, 1341, 1348, 1353, 1362, 1375, 1376, 1387, 1400, 1408, 1410, 1415, + 1425, 1453, 1457, 1477, 1481, 1494, 1496, 1507, 1512, 1538, 1545, 1547, 1549, 1551, 1554, 1561, + 1563, 1565, 1570, 1572, 1575, 1577, 1587, 1593, 1601, 1603, 1605, 1612, 1617, 1619, 1632, 1648, + 1658, 1662, 1664, 1674, 1680, 1690, 1692, 1704, 1729, 1736, 1740, 1745, 1747, 1751, 1752, 1761, + 1763, 1767, 1773, 1787, 1795, 1801, 1806, 1810, 1817, 1834, 1840, 1844, 1857, 1864, 1866, 1877, + 1882, 1892, 1902, 1915, 1934, 1953, 1985, 1987, 2000, 2002, 2013, 2048, 2052, 2058, 2064, 2068, + 2071, 2074, 2081, 2088, 2104, 2114, 2119, 2121, 2123, 2130, 2136, 2141, 2147, 2153, 2157, 2177, + 2179, 2184, 2189, 2193, 2203, 2208, 2223, 2226, 2232, 2244, 2249, 2251, 2256, 2258, 2265, 2269, + 2304, 2306, 2324, 2335, 2336, 2361, 2373, 2375, 2385, 2418, 2443, 2460, 2480, 2504, 2509, 2520, + 2531, 2537, 2562, 2568, 2572, 2578, 2592, 2596, 2599, 2602, 2614, 2620, 2625, 2627, 2629, 2634, + 2641, 2650, 2682, 2688, 2697, 2707, 2712, 2718, 2731, 2754, 2759, 2760, 2775, 2788, 2793, 2805, + 2811, 2817, 2820, 2832, 2842, 2854, 2890, 2902, 2921, 2923, 2978, 3010, 3012, 3026, 3081, 3083, + 3085, 3097, 3099, 3120, 3136, 3152, 3159, 3188, 3210, 3228, 3234, 3245, 3250, 3256, 3264, 3276, + 3281, 3296, 3349, 3363, 3378, 3392, 3395, 3420, 3440, 3461, 3488, 3529, 3531, 3584, 3588, 3591, + 3600, 3602, 3614, 3616, 3628, 3634, 3650, 3657, 3668, 3683, 3685, 3713, 3716, 3720, 3726, 3729, + 3736, 3753, 3778, 3802, 3805, 3819, 3841, 3845, 3851, 3856, 3880, 3922, 3938, 3970, 3993, 4032, + }; + + const int kmap_size = 4096; + const int nwant = grid_size == 256 ? 2 : 3; + const uint16_t * kgrid = grid_size == 256 ? kgrid_256 : kgrid_512; + uint32_t * kgrid_q3xs; + int * kmap_q3xs; + uint16_t * kneighbors_q3xs; + + //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size); + uint32_t * the_grid = (uint32_t *)malloc(grid_size*sizeof(uint32_t)); + for (int k = 0; k < grid_size; ++k) { + int8_t * pos = (int8_t *)(the_grid + k); + for (int i = 0; i < 4; ++i) { + int l = (kgrid[k] >> 3*i) & 0x7; + pos[i] = 2*l + 1; + } + } + kgrid_q3xs = the_grid; + iq3_data[gindex].grid = the_grid; + kmap_q3xs = (int *)malloc(kmap_size*sizeof(int)); + iq3_data[gindex].map = kmap_q3xs; + for (int i = 0; i < kmap_size; ++i) kmap_q3xs[i] = -1; + uint32_t aux32; + uint8_t * aux8 = (uint8_t *)&aux32; + for (int i = 0; i < grid_size; ++i) { + aux32 = kgrid_q3xs[i]; + uint16_t index = 0; + for (int k=0; k<4; ++k) { + uint16_t q = (aux8[k] - 1)/2; + index |= (q << 3*k); + } + kmap_q3xs[index] = i; + } + int8_t pos[4]; + int * dist2 = (int *)malloc(2*grid_size*sizeof(int)); + int num_neighbors = 0, num_not_in_map = 0; + for (int i = 0; i < kmap_size; ++i) { + if (kmap_q3xs[i] >= 0) continue; + ++num_not_in_map; + for (int k = 0; k < 4; ++k) { + int l = (i >> 3*k) & 0x7; + pos[k] = 2*l + 1; + } + for (int j = 0; j < grid_size; ++j) { + const int8_t * pg = (const int8_t *)(kgrid_q3xs + j); + int d2 = 0; + for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]); + dist2[2*j+0] = d2; + dist2[2*j+1] = j; + } + qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func); + int n = 0; int d2 = dist2[0]; + int nhave = 1; + for (int j = 0; j < grid_size; ++j) { + if (dist2[2*j] > d2) { + if (nhave == nwant) break; + d2 = dist2[2*j]; + ++nhave; + } + ++n; + } + num_neighbors += n; + } + //printf("%s: %d neighbours in total\n", __func__, num_neighbors); + kneighbors_q3xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t)); + iq3_data[gindex].neighbours = kneighbors_q3xs; + int counter = 0; + for (int i = 0; i < kmap_size; ++i) { + if (kmap_q3xs[i] >= 0) continue; + for (int k = 0; k < 4; ++k) { + int l = (i >> 3*k) & 0x7; + pos[k] = 2*l + 1; + } + for (int j = 0; j < grid_size; ++j) { + const int8_t * pg = (const int8_t *)(kgrid_q3xs + j); + int d2 = 0; + for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]); + dist2[2*j+0] = d2; + dist2[2*j+1] = j; + } + qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func); + kmap_q3xs[i] = -(counter + 1); + int d2 = dist2[0]; + uint16_t * start = &kneighbors_q3xs[counter++]; + int n = 0, nhave = 1; + for (int j = 0; j < grid_size; ++j) { + if (dist2[2*j] > d2) { + if (nhave == nwant) break; + d2 = dist2[2*j]; + ++nhave; + } + kneighbors_q3xs[counter++] = dist2[2*j+1]; + ++n; + } + *start = n; + } + free(dist2); +} + +void iq3xs_free_impl(int grid_size) { + GGML_ASSERT(grid_size == 256 || grid_size == 512); + const int gindex = iq3_data_index(grid_size); + if (iq3_data[gindex].grid) { + free(iq3_data[gindex].grid); iq3_data[gindex].grid = NULL; + free(iq3_data[gindex].map); iq3_data[gindex].map = NULL; + free(iq3_data[gindex].neighbours); iq3_data[gindex].neighbours = NULL; + } +} + +static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const uint32_t * restrict grid, + const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) { + int num_neighbors = neighbours[0]; + GGML_ASSERT(num_neighbors > 0); + float best_d2 = FLT_MAX; + int grid_index = -1; + for (int j = 1; j <= num_neighbors; ++j) { + const int8_t * pg = (const int8_t *)(grid + neighbours[j]); + float d2 = 0; + for (int i = 0; i < 4; ++i) { + float q = pg[i]; + float diff = scale*q - xval[i]; + d2 += weight[i]*diff*diff; + } + if (d2 < best_d2) { + best_d2 = d2; grid_index = neighbours[j]; + } + } + GGML_ASSERT(grid_index >= 0); + const int8_t * pg = (const int8_t *)(grid + grid_index); + for (int i = 0; i < 4; ++i) L[i] = (pg[i] - 1)/2; + return grid_index; +} + +static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, void * restrict vy, int64_t n, + const float * restrict quant_weights) { + + const int gindex = iq3_data_index(grid_size); + + const uint32_t * kgrid_q3xs = iq3_data[gindex].grid; + const int * kmap_q3xs = iq3_data[gindex].map; + const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours; + + //GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kgrid_q3xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kmap_q3xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + const int kMaxQ = 8; + + const int64_t nbl = n/QK_K; + + ggml_fp16_t * dh; + uint8_t * qs; + int block_size; + if (grid_size == 256) { + block_iq3_xxs * y = vy; + dh = &y->d; + qs = y->qs; + block_size = sizeof(block_iq3_xxs); + } else { + block_iq3_s * y = vy; + dh = &y->d; + qs = y->qs; + block_size = sizeof(block_iq3_s); + } + int quant_size = block_size - sizeof(ggml_fp16_t); + + float scales[QK_K/32]; + float weight[32]; + float xval[32]; + int8_t L[32]; + int8_t Laux[32]; + float waux[32]; + bool is_on_grid[8]; + bool is_on_grid_aux[8]; + uint8_t block_signs[8]; + uint8_t q3[3*(QK_K/8)+QK_K/32]; + uint32_t * scales_and_signs = (uint32_t *)(q3 + QK_K/4); + uint8_t * qh = q3 + 3*(QK_K/8); + + for (int ibl = 0; ibl < nbl; ++ibl) { + + dh[0] = GGML_FP32_TO_FP16(0.f); + memset(q3, 0, 3*QK_K/8+QK_K/32); + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = 2*sumx2/QK_K; + + for (int ib = 0; ib < QK_K/32; ++ib) { + const float * xb = xbl + 32*ib; + if (quant_weights) { + const float * qw = quant_weights + QK_K*ibl + 32*ib; + for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + } else { + for (int i = 0; i < 32; ++i) weight[i] = xb[i]*xb[i]; + } + for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]); + for (int k = 0; k < 4; ++k) { + int nflip = 0; + uint8_t s = 0; + for (int i = 0; i < 8; ++i) { + if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i]; + else { + xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i); + } + } + if (nflip%2) { + int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin]; + for (int i = 1; i < 8; ++i) { + float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i]; + if (ax < min) { + min = ax; imin = i; + } + } + xval[8*k+imin] = -xval[8*k+imin]; + s ^= (1 << imin); + } + block_signs[k] = s & 127; + } + float max = xval[0]; + for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]); + if (!max) { + scales[ib] = 0; + memset(L, 0, 32); + continue; + } + float best = 0; + float scale = max/(2*kMaxQ-1); + for (int is = -15; is <= 15; ++is) { + float id = (2*kMaxQ-1+is*0.2f)/max; + float this_scale = 1/id; + for (int k = 0; k < 8; ++k) { + for (int i = 0; i < 4; ++i) { + int l = nearest_int(0.5f*(id*xval[4*k+i]-1)); + Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l)); + } + uint16_t u = 0; + for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i); + int grid_index = kmap_q3xs[u]; + is_on_grid_aux[k] = true; + if (grid_index < 0) { + is_on_grid_aux[k] = false; + const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1; + grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 32; ++i) { + float w = weight[i]; + float q = 2*Laux[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0 && sumqx*sumqx > best*sumq2) { + scale = sumqx/sumq2; best = scale*sumqx; + for (int i = 0; i < 32; ++i) L[i] = Laux[i]; + for (int k = 0; k < 8; ++k) is_on_grid[k] = is_on_grid_aux[k]; + } + } + int n_not_ongrid = 0; + for (int k = 0; k < 8; ++k) if (!is_on_grid[k]) ++n_not_ongrid; + if (n_not_ongrid > 0 && scale > 0) { + float id = 1/scale; + for (int k = 0; k < 8; ++k) { + if (is_on_grid[k]) continue; + uint16_t u = 0; + for (int i = 0; i < 4; ++i) { + int l = nearest_int(0.5f*(id*xval[4*k+i]-1)); + l = MAX(0, MIN(kMaxQ-1, l)); + u |= (l << 3*i); + } + int grid_index = kmap_q3xs[u]; + if (grid_index < 0) { + const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1; + grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k); + } + const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index); + for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2; + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 32; ++i) { + float w = weight[i]; + float q = 2*L[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0) scale = sumqx/sumq2; + } + if (scale < 0) { + // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale) + // and correspondingly flip quant signs. + scale = -scale; + for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127; + } + for (int k = 0; k < 8; ++k) { + uint16_t u = 0; + for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i); + int grid_index = kmap_q3xs[u]; + if (grid_index < 0) { + printf("Oops: found point %u not on grid:", u); + for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]); + printf("\n"); + GGML_ASSERT(false); + } + if (grid_size == 256) { + q3[8*ib+k] = grid_index; + } else { + q3[8*ib+k] = grid_index & 255; + qh[ib] |= ((grid_index >> 8) << k); + } + + } + scales_and_signs[ib] = block_signs[0] | (block_signs[1] << 7) | (block_signs[2] << 14) | (block_signs[3] << 21); + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + memset(qs, 0, quant_size); + dh += block_size/sizeof(ggml_fp16_t); + qs += block_size; + continue; + } + + float d = max_scale/31; + dh[0] = GGML_FP32_TO_FP16(d * 1.0125f); // small improvement via this fudge factor + float id = 1/d; + for (int ib = 0; ib < QK_K/32; ++ib) { + int l = nearest_int(0.5f*(id*scales[ib]-1)); + l = MAX(0, MIN(15, l)); + scales_and_signs[ib] |= ((uint32_t)l << 28); + } + memcpy(qs, q3, quant_size); + + dh += block_size/sizeof(ggml_fp16_t); + qs += block_size; + + } +} + +size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq3_xxs_impl(256, src, qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += nblock*sizeof(block_iq3_xxs); + } + return nrow * nblock * sizeof(block_iq3_xxs); +} + +void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int64_t k) { + assert(k % QK_K == 0); + block_iq3_xxs * restrict y = vy; + quantize_row_iq3_xxs_reference(x, y, k); +} + +void quantize_row_iq3_xxs_reference(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) { + assert(k % QK_K == 0); + quantize_row_iq3_xxs_impl(256, x, y, k, NULL); +} + +static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, void * restrict vy, int n, + const float * restrict quant_weights, + float * scales, + float * weight, + float * xval, + int8_t * L, + int8_t * Laux, + float * waux, + bool * is_on_grid, + bool * is_on_grid_aux, + uint8_t * block_signs) { + + const int gindex = iq3_data_index(512); + + const uint32_t * kgrid_q3xs = iq3_data[gindex].grid; + const int * kmap_q3xs = iq3_data[gindex].map; + const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours; + + //GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kgrid_q3xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kmap_q3xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + const int kMaxQ = 8; + + const int64_t nbl = n/QK_K; + + block_iq3_s * y = vy; + + const int bs4 = block_size/4; + const int bs8 = block_size/8; + + for (int ibl = 0; ibl < nbl; ++ibl) { + + memset(&y[ibl], 0, sizeof(block_iq3_s)); + y[ibl].d = GGML_FP32_TO_FP16(0.f); + + uint8_t * qs = y[ibl].qs; + uint8_t * qh = y[ibl].qh; + uint8_t * signs = y[ibl].signs; + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = 2*sumx2/QK_K; + + for (int ib = 0; ib < QK_K/block_size; ++ib) { + const float * xb = xbl + block_size*ib; + if (quant_weights) { + const float * qw = quant_weights + QK_K*ibl + block_size*ib; + for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + } else { + for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i]; + } + for (int i = 0; i < block_size; ++i) waux[i] = sqrtf(weight[i]); + for (int k = 0; k < bs8; ++k) { + uint8_t s = 0; + for (int i = 0; i < 8; ++i) { + if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i]; + else { + xval[8*k + i] = -xb[8*k + i]; s |= (1 << i); + } + } + block_signs[k] = s; + } + float max = xval[0]; + for (int i = 1; i < block_size; ++i) max = MAX(max, xval[i]); + if (!max) { + scales[ib] = 0; + continue; + } + float best = 0; + float scale = max/(2*kMaxQ-1); + for (int k = 0; k < bs4; ++k) is_on_grid[k] = false; + for (int is = -9; is <= 9; ++is) { + float id = (2*kMaxQ-1+is*0.2f)/max; + float this_scale = 1/id; + for (int k = 0; k < bs4; ++k) { + for (int i = 0; i < 4; ++i) { + int l = nearest_int(0.5f*(id*xval[4*k+i]-1)); + Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l)); + } + uint16_t u = 0; + for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i); + int grid_index = kmap_q3xs[u]; + is_on_grid_aux[k] = true; + if (grid_index < 0) { + is_on_grid_aux[k] = false; + const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1; + grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < block_size; ++i) { + float w = weight[i]; + float q = 2*Laux[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0 && sumqx*sumqx > best*sumq2) { + scale = sumqx/sumq2; best = scale*sumqx; + for (int i = 0; i < block_size; ++i) L[i] = Laux[i]; + for (int k = 0; k < bs4; ++k) is_on_grid[k] = is_on_grid_aux[k]; + } + } + int n_not_ongrid = 0; + for (int k = 0; k < bs4; ++k) if (!is_on_grid[k]) ++n_not_ongrid; + if (n_not_ongrid > 0 && scale > 0) { + float id = 1/scale; + for (int k = 0; k < bs4; ++k) { + //if (is_on_grid[k]) continue; + uint16_t u = 0; + for (int i = 0; i < 4; ++i) { + int l = nearest_int(0.5f*(id*xval[4*k+i]-1)); + l = MAX(0, MIN(kMaxQ-1, l)); + u |= (l << 3*i); + } + int grid_index = kmap_q3xs[u]; + if (grid_index < 0) { + const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1; + grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k); + } + const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index); + for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2; + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < block_size; ++i) { + float w = weight[i]; + float q = 2*L[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0) scale = sumqx/sumq2; + } + if (scale < 0) { + // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale) + // and correspondingly flip quant signs. + scale = -scale; + for (int k = 0; k < bs8; ++k) block_signs[k] = ~block_signs[k]; + } + for (int k = 0; k < bs4; ++k) { + uint16_t u = 0; + for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i); + int grid_index = kmap_q3xs[u]; + if (grid_index < 0) { + printf("Oops: found point %u not on grid:", u); + for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]); + printf("\n"); + GGML_ASSERT(false); + } + qs[k] = grid_index & 255; + qh[(ib*bs4+k)/8] |= ((grid_index >> 8) << ((ib*bs4+k)%8)); + } + qs += bs4; + for (int k = 0; k < bs8; ++k) signs[k] = block_signs[k]; + signs += bs8; + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + continue; + } + + float d = max_scale/31; + y[ibl].d = GGML_FP32_TO_FP16(d * 1.033f); + float id = 1/d; + for (int ib = 0; ib < QK_K/block_size; ib += 2) { + int l1 = nearest_int(0.5f*(id*scales[ib+0]-1)); + l1 = MAX(0, MIN(15, l1)); + int l2 = nearest_int(0.5f*(id*scales[ib+1]-1)); + l2 = MAX(0, MIN(15, l2)); + y[ibl].scales[ib/2] = l1 | (l2 << 4); + } + + } +} + +#define IQ3S_BLOCK_SIZE 32 +size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + int64_t nblock = n_per_row/QK_K; + float scales[QK_K/IQ3S_BLOCK_SIZE]; + float weight[IQ3S_BLOCK_SIZE]; + float xval[IQ3S_BLOCK_SIZE]; + int8_t L[IQ3S_BLOCK_SIZE]; + int8_t Laux[IQ3S_BLOCK_SIZE]; + float waux[IQ3S_BLOCK_SIZE]; + bool is_on_grid[IQ3S_BLOCK_SIZE/4]; + bool is_on_grid_aux[IQ3S_BLOCK_SIZE/4]; + uint8_t block_signs[IQ3S_BLOCK_SIZE/8]; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq3_s_impl(IQ3S_BLOCK_SIZE, src, qrow, n_per_row, quant_weights, + scales, weight, xval, L, Laux, waux, is_on_grid, is_on_grid_aux, block_signs); + src += n_per_row; + qrow += nblock*sizeof(block_iq3_s); + } + return nrow * nblock * sizeof(block_iq3_s); +} + +void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int64_t k) { + assert(k % QK_K == 0); + block_iq3_s * restrict y = vy; + quantize_row_iq3_s_reference(x, y, k); +} + +void quantize_row_iq3_s_reference(const float * restrict x, block_iq3_s * restrict y, int64_t k) { + assert(k % QK_K == 0); + quantize_iq3_s(x, y, 1, k, NULL); +} + + +// =================================== 1.5 bpw =================================================== + +static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid, + const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) { + int num_neighbors = neighbours[0]; + GGML_ASSERT(num_neighbors > 0); + float best_score = 0; + int grid_index = -1; + for (int j = 1; j <= num_neighbors; ++j) { + const int8_t * pg = (const int8_t *)(grid + neighbours[j]); + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 8; ++i) { + float q = (pg[i] - 3)/2; + float w = weight[i]; + sumqx += w*q*xval[i]; + sumq2 += w*q*q; + } + if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) { + *scale = sumqx/sumq2; best_score = *scale * sumqx; + grid_index = neighbours[j]; + } + } + if (grid_index < 0) { + for (int i = 0; i < ngrid; ++i) { + const int8_t * grid_i = (const int8_t *)(grid + i); + float sumqx = 0, sumq2 = 0; + for (int j = 0; j < 8; ++j) { + float w = weight[j]; + float q = (grid_i[j] - 3)/2; + sumqx += w*q*xval[j]; + sumq2 += w*q*q; + } + if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) { + *scale = sumqx/sumq2; best_score = *scale*sumqx; + grid_index = i; + } + } + } + if (grid_index < 0) { + printf("Oops, did not find grid point\n"); + printf("Have %d neighbours\n", num_neighbors); + for (int j = 1; j <= num_neighbors; ++j) { + const int8_t * pg = (const int8_t *)(grid + neighbours[j]); + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 8; ++i) { + float q = (pg[i] - 3)/2; + float w = weight[i]; + sumqx += w*q*xval[i]; + sumq2 += w*q*q; + } + printf(" neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2); + } + } + GGML_ASSERT(grid_index >= 0); + //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! + *scale *= 1.05f; // This is a fudge factor. Don't ask me why it improves the result. + //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! + const int8_t * pg = (const int8_t *)(grid + grid_index); + for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2; + return grid_index; +} + +static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const uint64_t * restrict grid, + const float * restrict xval, const float * restrict weight, float scale, const float * restrict xg, int8_t * restrict L, int ngrid) { + int num_neighbors = neighbours[0]; + GGML_ASSERT(num_neighbors > 0); + float best_score = FLT_MAX; + int grid_index = -1; + for (int j = 1; j <= num_neighbors; ++j) { + const int8_t * pg = (const int8_t *)(grid + neighbours[j]); + float d2 = 0; + for (int i = 0; i < 8; ++i) { + float q = xg[(pg[i] - 1)/2]; + float w = weight[i]; + float diff = scale*q - xval[i]; + d2 += w*diff*diff; + } + if (d2 < best_score) { + best_score = d2; + grid_index = neighbours[j]; + } + } + if (grid_index < 0) { + for (int i = 0; i < ngrid; ++i) { + const int8_t * grid_i = (const int8_t *)(grid + i); + float d2 = 0; + for (int j = 0; j < 8; ++j) { + float w = weight[j]; + float q = xg[(grid_i[j] - 1)/2]; + float diff = scale*q - xval[i]; + d2 += w*diff*diff; + } + if (d2 < best_score) { + best_score = d2; + grid_index = i; + } + } + } + if (grid_index < 0) { + printf("Oops, did not find grid point\n"); + printf("Have %d neighbours\n", num_neighbors); + for (int j = 1; j <= num_neighbors; ++j) { + const int8_t * pg = (const int8_t *)(grid + neighbours[j]); + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 8; ++i) { + float q = xg[(pg[i] - 1)/2]; + float w = weight[i]; + sumqx += w*q*xval[i]; + sumq2 += w*q*q; + } + printf(" neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2); + } + } + GGML_ASSERT(grid_index >= 0); + const int8_t * pg = (const int8_t *)(grid + grid_index); + for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2; + return grid_index; +} + +static int iq1_sort_helper(const void * left, const void * right) { + const float * l = left; + const float * r = right; + return *l < *r ? -1 : *l > *r ? 1 : 0; +} + +#define IQ1S_BLOCK_SIZE 32 +#define IQ1M_BLOCK_SIZE 16 +static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights, + float * scales, + float * weight, + float * sumx, + float * sumw, + float * pairs, + int8_t * L, + uint16_t * index, + int8_t * shifts) { + + const int gindex = iq2_data_index(GGML_TYPE_IQ1_S); + + const uint64_t * kgrid_q2xs = iq2_data[gindex].grid; + const int * kmap_q2xs = iq2_data[gindex].map; + const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; + + GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + block_iq1_s * y = vy; + + const int64_t nbl = n/QK_K; + + const int block_size = IQ1S_BLOCK_SIZE; + + const float x_p[3] = {-1 + IQ1S_DELTA, IQ1S_DELTA, 1 + IQ1S_DELTA}; + const float x_m[3] = {-1 - IQ1S_DELTA, -IQ1S_DELTA, 1 - IQ1S_DELTA}; + + + int * idx = (int *)(pairs + 1); + + for (int ibl = 0; ibl < nbl; ++ibl) { + + y[ibl].d = GGML_FP32_TO_FP16(0.f); + memset(y[ibl].qs, 0, QK_K/8); + memset(y[ibl].qh, 0, QK_K/16); + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = 2*sumx2/QK_K; + + for (int ib = 0; ib < QK_K/block_size; ++ib) { + const float * xb = xbl + block_size*ib; + const float * qw = quant_weights + QK_K*ibl + block_size*ib; + for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + float max = fabsf(xb[0]); + for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i])); + if (!max) { + scales[ib] = 0; + memset(L, 1, block_size); + continue; + } + // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem. + // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two + // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights + // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and + // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale + // for each possible and score for each split. + for (int j = 0; j < block_size; ++j) { + pairs[2*j] = xb[j]; + idx[2*j] = j; + } + qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper); + { + sumx[0] = sumw[0] = 0; + for (int j = 0; j < block_size; ++j) { + int i = idx[2*j]; + sumx[j+1] = sumx[j] + weight[i]*xb[i]; + sumw[j+1] = sumw[j] + weight[i]; + } + } + float best_score = 0, scale = max; + int besti1 = -1, besti2 = -1, best_shift = 0; + for (int i1 = 0; i1 <= block_size; ++i1) { + for (int i2 = i1; i2 <= block_size; ++i2) { + float sumqx = (sumx[i1] - sumx[0])*x_p[0] + (sumx[i2] - sumx[i1])*x_p[1] + (sumx[block_size] - sumx[i2])*x_p[2]; + float sumq2 = (sumw[i1] - sumw[0])*x_p[0]*x_p[0] + (sumw[i2] - sumw[i1])*x_p[1]*x_p[1] + (sumw[block_size] - sumw[i2])*x_p[2]*x_p[2]; + if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) { + scale = sumqx/sumq2; best_score = scale*sumqx; + besti1 = i1; besti2 = i2; best_shift = 1; + } + sumqx = (sumx[i1] - sumx[0])*x_m[0] + (sumx[i2] - sumx[i1])*x_m[1] + (sumx[block_size] - sumx[i2])*x_m[2]; + sumq2 = (sumw[i1] - sumw[0])*x_m[0]*x_m[0] + (sumw[i2] - sumw[i1])*x_m[1]*x_m[1] + (sumw[block_size] - sumw[i2])*x_m[2]*x_m[2]; + if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) { + scale = sumqx/sumq2; best_score = scale*sumqx; + besti1 = i1; besti2 = i2; best_shift = -1; + } + } + } + GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_shift != 0); + for (int j = 0; j < besti1; ++j) L[idx[2*j]] = 0; + for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1; + for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2; + if (scale < 0) { + for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j]; + scale = -scale; best_shift = -best_shift; + } + bool all_on_grid = true; + const float * xx = best_shift == 1 ? x_p : x_m; + for (int k = 0; k < block_size/8; ++k) { + uint16_t u = 0; + for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j); + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + all_on_grid = false; + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S); + GGML_ASSERT(grid_index >= 0); + } + index[k] = grid_index; + } + if (!all_on_grid) { + float sumqx = 0, sumq2 = 0; + for (int k = 0; k < block_size/8; ++k) { + const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]); + for (int j = 0; j < 8; ++j) { + float w = weight[8*k + j]; + float q = xx[(pg[j] - 1)/2]; + sumqx += w*q*xb[8*k+j]; + sumq2 += w*q*q; + } + } + if (sumqx > 0 && sumq2 > 0) scale = sumqx/sumq2; + } + uint16_t h = 0; + for (int k = 0; k < block_size/8; ++k) { + y[ibl].qs[(block_size/8)*ib + k] = index[k] & 255; + h |= (index[k] >> 8) << 3*k; + } + y[ibl].qh[ib] = h; + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + shifts[ib] = best_shift; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + continue; + } + + float d = max_scale/15; + y[ibl].d = GGML_FP32_TO_FP16(d*1.125f); // 1.125f is another fudge factor. Don't ask me why it is needed. + float id = 1/d; + for (int ib = 0; ib < QK_K/block_size; ++ib) { + int l = nearest_int(0.5f*(id*scales[ib]-1)); + l = MAX(0, MIN(7, l)); + if (shifts[ib] == -1) l |= 8; + y[ibl].qh[ib] |= (l << 12); + } + } +} + +size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + float scales[QK_K/IQ1S_BLOCK_SIZE]; + float weight[IQ1S_BLOCK_SIZE]; + int8_t L[IQ1S_BLOCK_SIZE]; + float sumx[IQ1S_BLOCK_SIZE+1]; + float sumw[IQ1S_BLOCK_SIZE+1]; + float pairs[2*IQ1S_BLOCK_SIZE]; + uint16_t index[IQ1S_BLOCK_SIZE/8]; + int8_t shifts[QK_K/IQ1S_BLOCK_SIZE]; + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq1_s_impl(src, qrow, n_per_row, quant_weights, scales, weight, sumx, sumw, pairs, L, index, shifts); + src += n_per_row; + qrow += nblock*sizeof(block_iq1_s); + } + return nrow * nblock * sizeof(block_iq1_s); +} + +static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights, + float * scales, + float * weight, + float * pairs, + int8_t * L, + uint16_t * index, + int8_t * shifts) { + + const int gindex = iq2_data_index(GGML_TYPE_IQ1_M); + + const uint64_t * kgrid_q2xs = iq2_data[gindex].grid; + const int * kmap_q2xs = iq2_data[gindex].map; + const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; + + //GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + block_iq1_m * y = vy; + + const int64_t nbl = n/QK_K; + + const int block_size = IQ1M_BLOCK_SIZE; + + const float x_p[3] = {-1 + IQ1M_DELTA, IQ1M_DELTA, 1 + IQ1M_DELTA}; + const float x_m[3] = {-1 - IQ1M_DELTA, -IQ1M_DELTA, 1 - IQ1M_DELTA}; + const uint8_t masks[4] = {0x00, 0x80, 0x08, 0x88}; + + int * idx = (int *)(pairs + 1); + + float sumqx[4], sumq2[4]; + + iq1m_scale_t s; + const float * xx; + + for (int ibl = 0; ibl < nbl; ++ibl) { + +#if QK_K == 64 + y[ibl].d = GGML_FP32_TO_FP16(0.f); +#endif + memset(y[ibl].qs, 0, QK_K/8); + memset(y[ibl].qh, 0, QK_K/16); + memset(y[ibl].scales, 0, QK_K/32); + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = 2*sumx2/QK_K; + + for (int ib = 0; ib < QK_K/block_size; ++ib) { + const float * xb = xbl + block_size*ib; + if (quant_weights) { + const float * qw = quant_weights + QK_K*ibl + block_size*ib; + for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + } else { + for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i]; + } + float max = fabsf(xb[0]); + for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i])); + if (!max) { + scales[ib] = 0; + memset(L, 1, block_size); + continue; + } + // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem. + // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two + // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights + // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and + // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale + // for each possible and score for each split. + for (int j = 0; j < block_size; ++j) { + pairs[2*j] = xb[j]; + idx[2*j] = j; + } + qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper); + float best_score = 0, scale = max; + int besti1 = -1, besti2 = -1, best_k = -1; + // 0: +, + + // 1: +, - + // 2: -, + + // 3: -, - + for (int i1 = 0; i1 <= block_size; ++i1) { + for (int i2 = i1; i2 <= block_size; ++i2) { + memset(sumqx, 0, 4*sizeof(float)); + memset(sumq2, 0, 4*sizeof(float)); + for (int j = 0; j < i1; ++j) { + int i = idx[2*j]; + if (i < block_size/2) { + sumqx[0] += weight[i]*x_p[0]*xb[i]; + sumqx[1] += weight[i]*x_p[0]*xb[i]; + sumqx[2] += weight[i]*x_m[0]*xb[i]; + sumqx[3] += weight[i]*x_m[0]*xb[i]; + sumq2[0] += weight[i]*x_p[0]*x_p[0]; + sumq2[1] += weight[i]*x_p[0]*x_p[0]; + sumq2[2] += weight[i]*x_m[0]*x_m[0]; + sumq2[3] += weight[i]*x_m[0]*x_m[0]; + } else { + sumqx[0] += weight[i]*x_p[0]*xb[i]; + sumqx[2] += weight[i]*x_p[0]*xb[i]; + sumqx[1] += weight[i]*x_m[0]*xb[i]; + sumqx[3] += weight[i]*x_m[0]*xb[i]; + sumq2[0] += weight[i]*x_p[0]*x_p[0]; + sumq2[2] += weight[i]*x_p[0]*x_p[0]; + sumq2[1] += weight[i]*x_m[0]*x_m[0]; + sumq2[3] += weight[i]*x_m[0]*x_m[0]; + } + } + for (int j = i1; j < i2; ++j) { + int i = idx[2*j]; + if (i < block_size/2) { + sumqx[0] += weight[i]*x_p[1]*xb[i]; + sumqx[1] += weight[i]*x_p[1]*xb[i]; + sumqx[2] += weight[i]*x_m[1]*xb[i]; + sumqx[3] += weight[i]*x_m[1]*xb[i]; + sumq2[0] += weight[i]*x_p[1]*x_p[1]; + sumq2[1] += weight[i]*x_p[1]*x_p[1]; + sumq2[2] += weight[i]*x_m[1]*x_m[1]; + sumq2[3] += weight[i]*x_m[1]*x_m[1]; + } else { + sumqx[0] += weight[i]*x_p[1]*xb[i]; + sumqx[2] += weight[i]*x_p[1]*xb[i]; + sumqx[1] += weight[i]*x_m[1]*xb[i]; + sumqx[3] += weight[i]*x_m[1]*xb[i]; + sumq2[0] += weight[i]*x_p[1]*x_p[1]; + sumq2[2] += weight[i]*x_p[1]*x_p[1]; + sumq2[1] += weight[i]*x_m[1]*x_m[1]; + sumq2[3] += weight[i]*x_m[1]*x_m[1]; + } + } + for (int j = i2; j < block_size; ++j) { + int i = idx[2*j]; + if (i < block_size/2) { + sumqx[0] += weight[i]*x_p[2]*xb[i]; + sumqx[1] += weight[i]*x_p[2]*xb[i]; + sumqx[2] += weight[i]*x_m[2]*xb[i]; + sumqx[3] += weight[i]*x_m[2]*xb[i]; + sumq2[0] += weight[i]*x_p[2]*x_p[2]; + sumq2[1] += weight[i]*x_p[2]*x_p[2]; + sumq2[2] += weight[i]*x_m[2]*x_m[2]; + sumq2[3] += weight[i]*x_m[2]*x_m[2]; + } else { + sumqx[0] += weight[i]*x_p[2]*xb[i]; + sumqx[2] += weight[i]*x_p[2]*xb[i]; + sumqx[1] += weight[i]*x_m[2]*xb[i]; + sumqx[3] += weight[i]*x_m[2]*xb[i]; + sumq2[0] += weight[i]*x_p[2]*x_p[2]; + sumq2[2] += weight[i]*x_p[2]*x_p[2]; + sumq2[1] += weight[i]*x_m[2]*x_m[2]; + sumq2[3] += weight[i]*x_m[2]*x_m[2]; + } + } + for (int k = 0; k < 4; ++k) { + if (sumq2[k] > 0 && sumqx[k]*sumqx[k] > best_score*sumq2[k]) { + scale = sumqx[k]/sumq2[k]; best_score = scale*sumqx[k]; + besti1 = i1; besti2 = i2; best_k = k; + } + } + } + } + GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_k >= 0); + for (int j = 0; j < besti1; ++j) L[idx[2*j]] = 0; + for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1; + for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2; + if (scale < 0) { + for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j]; + scale = -scale; + best_k = best_k == 0 ? 3 : best_k == 1 ? 2 : best_k == 2 ? 1 : 0; + } + bool all_on_grid = true; + for (int k = 0; k < block_size/8; ++k) { + if (k == 0) xx = best_k < 2 ? x_p : x_m; + else xx = best_k%2 == 0 ? x_p : x_m; + uint16_t u = 0; + for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j); + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + all_on_grid = false; + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S); + GGML_ASSERT(grid_index >= 0); + } + index[k] = grid_index; + } + if (!all_on_grid) { + float sumqx_f = 0, sumq2_f = 0; + for (int k = 0; k < block_size/8; ++k) { + if (k == 0) xx = best_k < 2 ? x_p : x_m; + else xx = best_k%2 == 0 ? x_p : x_m; + const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]); + for (int j = 0; j < 8; ++j) { + float w = weight[8*k + j]; + float q = xx[(pg[j] - 1)/2]; + sumqx_f += w*q*xb[8*k+j]; + sumq2_f += w*q*q; + } + } + if (sumqx_f > 0 && sumq2_f > 0) scale = sumqx_f/sumq2_f; + } + y[ibl].qs[2*ib + 0] = index[0] & 255; + y[ibl].qs[2*ib + 1] = index[1] & 255; + y[ibl].qh[ib] = (index[0] >> 8) | ((index[1] >> 8) << 4); + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + shifts[ib] = best_k; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + continue; + } + + uint16_t * sc = (uint16_t *)y[ibl].scales; +#if QK_K == 64 + float d = max_scale/31; +#else + float d = max_scale/15; +#endif + float id = 1/d; + float sumqx_f = 0, sumq2_f = 0; + for (int ib = 0; ib < QK_K/block_size; ++ib) { + int l = nearest_int(0.5f*(id*scales[ib+0]-1)); +#if QK_K == 64 + l = MAX(0, MIN(15, l)); + sc[ib/4] |= (l << 4*(ib%4)); +#else + l = MAX(0, MIN(7, l)); + sc[ib/4] |= (l << 3*(ib%4)); +#endif + y[ibl].qh[ib] |= masks[shifts[ib]]; + const float * xb = xbl + block_size*ib; + if (quant_weights) { + const float * qw = quant_weights + QK_K*ibl + block_size*ib; + for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + } else { + for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i]; + } + for (int k = 0; k < block_size/8; ++k) { + if (k == 0) xx = shifts[ib] < 2 ? x_p : x_m; + else xx = shifts[ib]%2 == 0 ? x_p : x_m; + const int8_t * pg = (const int8_t *)(kgrid_q2xs + y[ibl].qs[2*ib+k] + ((y[ibl].qh[ib] << (8 - 4*k)) & 0x700)); + for (int j = 0; j < 8; ++j) { + float w = weight[8*k + j]; + float q = xx[(pg[j] - 1)/2]*(2*l+1); + sumqx_f += w*q*xb[8*k+j]; + sumq2_f += w*q*q; + } + } + } + if (sumq2_f > 0) d = sumqx_f/sumq2_f; + s.f16 = GGML_FP32_TO_FP16(d*1.1125f); // 1.1125f is another fudge factor. Don't ask me why it is needed. +#if QK_K == 64 + y[ibl].d = s.f16; +#else + sc[0] |= ((s.u16 & 0x000f) << 12); + sc[1] |= ((s.u16 & 0x00f0) << 8); + sc[2] |= ((s.u16 & 0x0f00) << 4); + sc[3] |= ((s.u16 & 0xf000) << 0); +#endif + } +} + +size_t quantize_iq1_m(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + float scales[QK_K/IQ1M_BLOCK_SIZE]; + float weight[IQ1M_BLOCK_SIZE]; + int8_t L[IQ1M_BLOCK_SIZE]; + float pairs[2*IQ1M_BLOCK_SIZE]; + uint16_t index[IQ1M_BLOCK_SIZE/8]; + int8_t shifts[QK_K/IQ1M_BLOCK_SIZE]; + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq1_m_impl(src, qrow, n_per_row, quant_weights, scales, weight, pairs, L, index, shifts); + src += n_per_row; + qrow += nblock*sizeof(block_iq1_m); + } + return nrow * nblock * sizeof(block_iq1_m); +} + +// ============================ 4-bit non-linear quants + +static inline int best_index_int8(int n, const int8_t * val, float x) { + if (x <= val[0]) return 0; + if (x >= val[n-1]) return n-1; + int ml = 0, mu = n-1; + while (mu-ml > 1) { + int mav = (ml+mu)/2; + if (x < val[mav]) mu = mav; else ml = mav; + } + return x - val[mu-1] < val[mu] - x ? mu-1 : mu; +} + +static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * restrict x, + ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l, + float * scales, float * weight, uint8_t * L, + const int8_t * values, + const float * quant_weights, + const int ntry) { + + float sigma2 = 0; + for (int j = 0; j < super_block_size; ++j) sigma2 += x[j]*x[j]; + sigma2 *= 2.f/super_block_size; + + memset(q4, 0, super_block_size/2); + dh[0] = GGML_FP32_TO_FP16(0.f); + + float max_scale = 0, amax_scale = 0; + for (int ib = 0; ib < super_block_size/block_size; ++ib) { + const float * xb = x + ib*block_size; + uint8_t * Lb = L + ib*block_size; + if (quant_weights) { + const float * qw = quant_weights + ib*block_size; + for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]); + } else { + for (int j = 0; j < block_size; ++j) weight[j] = xb[j]*xb[j]; + } + float amax = 0, max = 0; + for (int j = 0; j < block_size; ++j) { + float ax = fabsf(xb[j]); + if (ax > amax) { + amax = ax; max = xb[j]; + } + } + if (!amax) { + scales[ib] = 0; + continue; + } + float d = ntry > 0 ? -max/values[0] : max/values[0]; + float id = 1/d; + float sumqx = 0, sumq2 = 0; + for (int j = 0; j < block_size; ++j) { + float al = id*xb[j]; + int l = best_index_int8(16, values, al); + Lb[j] = l; + float q = values[l]; + float w = weight[j]; + sumqx += w*q*xb[j]; + sumq2 += w*q*q; + } + d = sumqx/sumq2; + float best = d*sumqx; + for (int itry = -ntry; itry <= ntry; ++itry) { + id = (itry + values[0])/max; + sumqx = sumq2 = 0; + for (int j = 0; j < block_size; ++j) { + float al = id*xb[j]; + int l = best_index_int8(16, values, al); + float q = values[l]; + float w = weight[j]; + sumqx += w*q*xb[j]; + sumq2 += w*q*q; + } + if (sumq2 > 0 && sumqx*sumqx > best*sumq2) { + d = sumqx/sumq2; best = d * sumqx; + } + } + scales[ib] = d; + float abs_d = fabsf(d); + if (abs_d > amax_scale) { + amax_scale = abs_d; max_scale = d; + } + } + + if (super_block_size/block_size > 1) { + int nb = super_block_size/block_size; + memset(scales_h, 0, ((nb+7)/8)*sizeof(uint16_t)); + float d = -max_scale/32; + dh[0] = GGML_FP32_TO_FP16(d); + float id = d ? 1/d : 0.f; + for (int ib = 0; ib < super_block_size/block_size; ++ib) { + int l = nearest_int(id*scales[ib]); + l = MAX(-32, MIN(31, l)); + float dl = d * l; + float idl = dl ? 1/dl : 0.f; + uint8_t * Lb = L + ib*block_size; + const float * xb = x + ib*block_size; + for (int j = 0; j < block_size; ++j) { + Lb[j] = best_index_int8(16, values, idl*xb[j]); + } + l += 32; + uint8_t l_l = l & 0xf; + uint8_t l_h = l >> 4; + if (ib%2 == 0) scales_l[ib/2] = l_l; + else scales_l[ib/2] |= (l_l << 4); + scales_h[ib/8] |= (l_h << 2*(ib%8)); + } + } else { + dh[0] = GGML_FP32_TO_FP16(scales[0]); + if (ntry > 0) { + float id = scales[0] ? 1/scales[0] : 0; + for (int j = 0; j < super_block_size; ++j) { + L[j] = best_index_int8(16, values, id*x[j]); + } + } + } + + for (int i = 0; i < super_block_size/32; ++i) { + for (int j = 0; j < 16; ++j) { + q4[16*i + j] = L[32*i + j] | (L[32*i + 16 + j] << 4); + } + } +} + +size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK4_NL == 0); + int64_t nblock = n_per_row/QK4_NL; + char * qrow = (char *)dst; + uint8_t L[QK4_NL]; + float weight[QK4_NL]; + uint16_t unused_h; + uint8_t * unused_l = NULL; + float scale; + for (int64_t row = 0; row < nrow; ++row) { + block_iq4_nl * iq4 = (block_iq4_nl *)qrow; + for (int ibl = 0; ibl < nblock; ++ibl) { + const float * qw = quant_weights ? quant_weights + QK4_NL*ibl : NULL; + quantize_row_iq4_nl_impl(QK4_NL, 32, src + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l, + &scale, weight, L, kvalues_iq4nl, qw, 7); + } + src += n_per_row; + qrow += nblock*sizeof(block_iq4_nl); + } + return nrow * nblock * sizeof(block_iq4_nl); +} + +void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int64_t k) { + GGML_ASSERT(k%QK4_NL == 0); + int64_t nblock = k/QK4_NL; + uint8_t L[QK4_NL]; + float weight[QK4_NL]; + uint16_t unused_h; + uint8_t * unused_l = NULL; + float scale; + block_iq4_nl * iq4 = (block_iq4_nl *)vy; + for (int ibl = 0; ibl < nblock; ++ibl) { + quantize_row_iq4_nl_impl(QK4_NL, 32, x + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l, + &scale, weight, L, kvalues_iq4nl, NULL, -1); + } +} + +void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int64_t k) { + assert(k % QK4_NL == 0); + quantize_row_iq4_nl(x, y, k); +} + +size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +#if QK_K == 64 + return quantize_iq4_nl(src, dst, nrow, n_per_row, quant_weights); +#else + GGML_ASSERT(n_per_row%QK_K == 0); + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + uint8_t L[QK_K]; + float weight[32]; + float scales[QK_K/32]; + for (int64_t row = 0; row < nrow; ++row) { + block_iq4_xs * iq4 = (block_iq4_xs *)qrow; + for (int ibl = 0; ibl < nblock; ++ibl) { + const float * qw = quant_weights ? quant_weights + QK_K*ibl : NULL; + quantize_row_iq4_nl_impl(QK_K, 32, src + QK_K*ibl, &iq4[ibl].d, iq4[ibl].qs, &iq4[ibl].scales_h, iq4[ibl].scales_l, + scales, weight, L, kvalues_iq4nl, qw, 7); + } + src += n_per_row; + qrow += nblock*sizeof(block_iq4_xs); + } + return nrow * nblock * sizeof(block_iq4_xs); +#endif +} + +void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) { + assert(k % QK_K == 0); + block_iq4_xs * restrict y = vy; + quantize_row_iq4_xs_reference(x, y, k); +} + +void quantize_row_iq4_xs_reference(const float * restrict x, block_iq4_xs * restrict y, int64_t k) { + assert(k % QK_K == 0); + quantize_iq4_xs(x, y, 1, k, NULL); +} + +// =============================== 2.5625 bpw + +static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) { + + const int gindex = iq2_data_index(GGML_TYPE_IQ2_S); + + const uint64_t * kgrid_q2xs = iq2_data[gindex].grid; + const int * kmap_q2xs = iq2_data[gindex].map; + const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; + + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(n%QK_K == 0); + + const int kMaxQ = 3; + + const int64_t nbl = n/QK_K; + + block_iq2_s * y = vy; + + float scales[QK_K/16]; + float weight[16]; + float xval[16]; + int8_t L[16]; + int8_t Laux[16]; + float waux[16]; + bool is_on_grid[2]; + bool is_on_grid_aux[2]; + uint8_t block_signs[2]; + + for (int ibl = 0; ibl < nbl; ++ibl) { + + memset(&y[ibl], 0, sizeof(block_iq2_s)); + y[ibl].d = GGML_FP32_TO_FP16(0.f); + + float max_scale = 0; + + const float * xbl = x + QK_K*ibl; + float sumx2 = 0; + for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i]; + float sigma2 = 2*sumx2/QK_K; + + for (int ib = 0; ib < QK_K/16; ++ib) { + const float * xb = xbl + 16*ib; + if (quant_weights) { + const float * qw = quant_weights + QK_K*ibl + 16*ib; + for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]); + } else { + for (int i = 0; i < 16; ++i) weight[i] = 0.25f*sigma2 + xb[i]*xb[i]; + } + for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]); + for (int k = 0; k < 2; ++k) { + uint8_t s = 0; + for (int i = 0; i < 8; ++i) { + if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i]; + else { + xval[8*k + i] = -xb[8*k + i]; s |= (1 << i); + } + } + block_signs[k] = s; + } + float max = xval[0]; + for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]); + if (!max) { + scales[ib] = 0; + continue; + } + float best = 0; + float scale = max/(2*kMaxQ-1); + is_on_grid[0] = is_on_grid[1] = true; + for (int is = -9; is <= 9; ++is) { + float id = (2*kMaxQ-1+is*0.1f)/max; + float this_scale = 1/id; + for (int k = 0; k < 2; ++k) { + for (int i = 0; i < 8; ++i) { + int l = nearest_int(0.5f*(id*xval[8*k+i]-1)); + Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l)); + } + uint16_t u = 0; + for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i); + int grid_index = kmap_q2xs[u]; + is_on_grid_aux[k] = true; + if (grid_index < 0) { + is_on_grid_aux[k] = false; + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 16; ++i) { + float w = weight[i]; + float q = 2*Laux[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0 && sumqx*sumqx > best*sumq2) { + scale = sumqx/sumq2; best = scale*sumqx; + for (int i = 0; i < 16; ++i) L[i] = Laux[i]; + for (int k = 0; k < 2; ++k) is_on_grid[k] = is_on_grid_aux[k]; + } + } + int n_not_ongrid = 0; + for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid; + if (n_not_ongrid > 0 && scale > 0) { + float id = 1/scale; + for (int k = 0; k < 2; ++k) { + if (is_on_grid[k]) continue; + uint16_t u = 0; + for (int i = 0; i < 8; ++i) { + int l = nearest_int(0.5f*(id*xval[8*k+i]-1)); + l = MAX(0, MIN(kMaxQ-1, l)); + u |= (l << 2*i); + L[8*k + i] = l; + } + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1; + grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k); + } + } + float sumqx = 0, sumq2 = 0; + for (int i = 0; i < 16; ++i) { + float w = weight[i]; + float q = 2*L[i] + 1; + sumqx += w*xval[i]*q; + sumq2 += w*q*q; + } + if (sumq2 > 0) scale = sumqx/sumq2; + } + if (scale < 0) { + scale = -scale; + for (int k = 0; k < 2; ++k) block_signs[k] = ~block_signs[k]; + } + for (int k = 0; k < 2; ++k) { + uint16_t u = 0; + for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i); + int grid_index = kmap_q2xs[u]; + if (grid_index < 0) { + printf("Oops: found point %u not on grid:", u); + for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]); + printf("\n"); + GGML_ASSERT(false); + } + const int i8 = 2*ib + k; + y[ibl].qs[i8] = grid_index & 255; + y[ibl].qh[i8/4] |= ((grid_index >> 8) << 2*(i8%4)); + y[ibl].qs[QK_K/8 + i8] = block_signs[k]; + } + GGML_ASSERT(scale >= 0); + scales[ib] = scale; + max_scale = MAX(max_scale, scale); + } + + if (!max_scale) { + continue; + } + + float d = max_scale/31; + y[ibl].d = GGML_FP32_TO_FP16(d * 0.9875f); + float id = 1/d; + for (int ib = 0; ib < QK_K/16; ++ib) { + int l = nearest_int(0.5f*(id*scales[ib]-1)); + l = MAX(0, MIN(15, l)); + if (ib%2 == 0) y[ibl].scales[ib/2] = l; + else y[ibl].scales[ib/2] |= (l << 4); + } + } +} + +size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { + GGML_ASSERT(n_per_row%QK_K == 0); + int64_t nblock = n_per_row/QK_K; + char * qrow = (char *)dst; + for (int64_t row = 0; row < nrow; ++row) { + quantize_row_iq2_s_impl(src, qrow, n_per_row, quant_weights); + src += n_per_row; + qrow += nblock*sizeof(block_iq2_s); + } + return nrow * nblock * sizeof(block_iq2_s); +} + +void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restrict y, int64_t k) { + assert(k % QK_K == 0); + quantize_iq2_s(x, y, 1, k, NULL); +} + +void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int64_t k) { + assert(k % QK_K == 0); + block_iq2_s * restrict y = vy; + quantize_row_iq2_s_reference(x, y, k); +} diff --git a/bindings/ruby/ext/ggml-quants.h b/bindings/ruby/ext/ggml-quants.h index 70c12c27..4d436a8f 100644 --- a/bindings/ruby/ext/ggml-quants.h +++ b/bindings/ruby/ext/ggml-quants.h @@ -1,224 +1,133 @@ #pragma once -#include "ggml-impl.h" +#define GGML_COMMON_DECL_C +#include "ggml-common.h" + +#include "ggml.h" // GGML internal header -#include -#include - -#define QK4_0 32 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qs[QK4_0 / 2]; // nibbles / quants -} block_q4_0; -static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding"); - -#define QK4_1 32 -typedef struct { - ggml_fp16_t d; // delta - ggml_fp16_t m; // min - uint8_t qs[QK4_1 / 2]; // nibbles / quants -} block_q4_1; -static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding"); - -#define QK5_0 32 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qh[4]; // 5-th bit of quants - uint8_t qs[QK5_0 / 2]; // nibbles / quants -} block_q5_0; -static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); - -#define QK5_1 32 -typedef struct { - ggml_fp16_t d; // delta - ggml_fp16_t m; // min - uint8_t qh[4]; // 5-th bit of quants - uint8_t qs[QK5_1 / 2]; // nibbles / quants -} block_q5_1; -static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); - -#define QK8_0 32 -typedef struct { - ggml_fp16_t d; // delta - int8_t qs[QK8_0]; // quants -} block_q8_0; -static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding"); - -#define QK8_1 32 -typedef struct { - float d; // delta - float s; // d * sum(qs[i]) - int8_t qs[QK8_1]; // quants -} block_q8_1; -static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); - -// -// Super-block quantization structures -// - -// Super-block size -#ifdef GGML_QKK_64 -#define QK_K 64 -#define K_SCALE_SIZE 4 -#else -#define QK_K 256 -#define K_SCALE_SIZE 12 +#ifdef __cplusplus +extern "C" { #endif -// 2-bit quantization -// weight is represented as x = a * q + b -// 16 blocks of 16 elements each -// Effectively 2.5625 bits per weight -typedef struct { - uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits - uint8_t qs[QK_K/4]; // quants - ggml_fp16_t d; // super-block scale for quantized scales - ggml_fp16_t dmin; // super-block scale for quantized mins -} block_q2_K; -static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding"); - -// 3-bit quantization -// weight is represented as x = a * q -// 16 blocks of 16 elements each -// Effectively 3.4375 bits per weight -#ifdef GGML_QKK_64 -typedef struct { - uint8_t hmask[QK_K/8]; // quants - high bit - uint8_t qs[QK_K/4]; // quants - low 2 bits - uint8_t scales[2]; - ggml_fp16_t d; // super-block scale -} block_q3_K; -static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding"); -#else -typedef struct { - uint8_t hmask[QK_K/8]; // quants - high bit - uint8_t qs[QK_K/4]; // quants - low 2 bits - uint8_t scales[12]; // scales, quantized with 6 bits - ggml_fp16_t d; // super-block scale -} block_q3_K; -static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding"); -#endif - -// 4-bit quantization -// 8 blocks of 32 elements each -// weight is represented as x = a * q + b -// Effectively 4.5 bits per weight -#ifdef GGML_QKK_64 -typedef struct { - ggml_fp16_t d[2]; // super-block scales/mins - uint8_t scales[2]; // 4-bit block scales/mins - uint8_t qs[QK_K/2]; // 4--bit quants -} block_q4_K; -static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding"); -#else -typedef struct { - ggml_fp16_t d; // super-block scale for quantized scales - ggml_fp16_t dmin; // super-block scale for quantized mins - uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits - uint8_t qs[QK_K/2]; // 4--bit quants -} block_q4_K; -static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding"); -#endif - -// 5-bit quantization -// 8 blocks of 32 elements each -// weight is represented as x = a * q + b -// Effectively 5.5 bits per weight -#ifdef GGML_QKK_64 -typedef struct { - ggml_fp16_t d; // super-block scale - int8_t scales[QK_K/16]; // 8-bit block scales - uint8_t qh[QK_K/8]; // quants, high bit - uint8_t qs[QK_K/2]; // quants, low 4 bits -} block_q5_K; -static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding"); -#else -typedef struct { - ggml_fp16_t d; // super-block scale for quantized scales - ggml_fp16_t dmin; // super-block scale for quantized mins - uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits - uint8_t qh[QK_K/8]; // quants, high bit - uint8_t qs[QK_K/2]; // quants, low 4 bits -} block_q5_K; -static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding"); -#endif - -// 6-bit quantization -// weight is represented as x = a * q -// 16 blocks of 16 elements each -// Effectively 6.5625 bits per weight -typedef struct { - uint8_t ql[QK_K/2]; // quants, lower 4 bits - uint8_t qh[QK_K/4]; // quants, upper 2 bits - int8_t scales[QK_K/16]; // scales, quantized with 8 bits - ggml_fp16_t d; // super-block scale -} block_q6_K; -static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding"); - -// This is only used for intermediate quantization and dot products -typedef struct { - float d; // delta - int8_t qs[QK_K]; // quants - int16_t bsums[QK_K/16]; // sum of quants in groups of 16 -} block_q8_K; -static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding"); - - // Quantization -void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k); -void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k); -void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k); -void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k); -void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k); -void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k); +void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k); +void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k); -void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k); -void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k); -void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k); -void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k); -void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k); -void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k); +void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k); +void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k); +void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k); +void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k); -void quantize_row_q4_0(const float * restrict x, void * restrict y, int k); -void quantize_row_q4_1(const float * restrict x, void * restrict y, int k); -void quantize_row_q5_0(const float * restrict x, void * restrict y, int k); -void quantize_row_q5_1(const float * restrict x, void * restrict y, int k); -void quantize_row_q8_0(const float * restrict x, void * restrict y, int k); -void quantize_row_q8_1(const float * restrict x, void * restrict y, int k); +void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k); +void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k); +void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k); +void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_s_reference (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k); -void quantize_row_q2_K(const float * restrict x, void * restrict y, int k); -void quantize_row_q3_K(const float * restrict x, void * restrict y, int k); -void quantize_row_q4_K(const float * restrict x, void * restrict y, int k); -void quantize_row_q5_K(const float * restrict x, void * restrict y, int k); -void quantize_row_q6_K(const float * restrict x, void * restrict y, int k); -void quantize_row_q8_K(const float * restrict x, void * restrict y, int k); +void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); + +void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); + +void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); // Dequantization -void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k); -void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k); -void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k); -void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k); -void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k); -//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k); +void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); -void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k); -void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k); -void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k); -void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k); -void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k); -void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k); +void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); + +void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq1_m (const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq4_xs (const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); // Dot product -void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); + +void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); + +void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); + +// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") +size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq2_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq1_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq1_m (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_iq3_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); + +size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); + +void iq2xs_init_impl(enum ggml_type type); +void iq2xs_free_impl(enum ggml_type type); +void iq3xs_init_impl(int grid_size); +void iq3xs_free_impl(int grid_size); + +#ifdef __cplusplus +} +#endif -void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); -void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); diff --git a/bindings/ruby/ext/ggml-sycl.h b/bindings/ruby/ext/ggml-sycl.h new file mode 100644 index 00000000..a9f776fc --- /dev/null +++ b/bindings/ruby/ext/ggml-sycl.h @@ -0,0 +1,49 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define GGML_SYCL_MAX_DEVICES 48 +#define GGML_SYCL_NAME "SYCL" + +// backend API +GGML_API ggml_backend_t ggml_backend_sycl_init(int device); + +// devide buffer +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device); + +// split tensor buffer that splits matrices by rows across multiple devices +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); + +// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void); + +GGML_API void ggml_backend_sycl_print_sycl_devices(void); +GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len); +GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size); +GGML_API GGML_CALL int ggml_backend_sycl_get_device_count(); +GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); +GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id); + +// TODO: these are temporary +// ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670 +GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index); +GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id); +GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode(); + +// SYCL doesn't support registering host memory, keep here for reference +// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size); +// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer); +#ifdef __cplusplus +} +#endif diff --git a/bindings/ruby/ext/ggml-vulkan.h b/bindings/ruby/ext/ggml-vulkan.h new file mode 100644 index 00000000..af661c2d --- /dev/null +++ b/bindings/ruby/ext/ggml-vulkan.h @@ -0,0 +1,29 @@ +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define GGML_VK_NAME "Vulkan" +#define GGML_VK_MAX_DEVICES 16 + +GGML_API void ggml_vk_instance_init(void); + +// backend API +GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num); + +GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend); +GGML_API GGML_CALL int ggml_backend_vk_get_device_count(void); +GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size); +GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total); + +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num); +// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void); + +#ifdef __cplusplus +} +#endif diff --git a/bindings/ruby/whispercpp.gemspec b/bindings/ruby/whispercpp.gemspec new file mode 100644 index 00000000..508a6a94 --- /dev/null +++ b/bindings/ruby/whispercpp.gemspec @@ -0,0 +1,28 @@ +Gem::Specification.new do |s| + s.name = "whispercpp" + s.authors = ["Georgi Gerganov", "Todd A. Fisher"] + s.version = '1.3.0' + s.date = '2024-05-14' + s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby} + s.email = 'todd.fisher@gmail.com' + s.extra_rdoc_files = ['LICENSE', 'README.md'] + + s.files = ["LICENSE", "README.md", "Rakefile", "ext/extconf.rb", "ext/ggml.c", "ext/ruby_whisper.cpp", "ext/whisper.cpp", "ext/dr_wav.h", "ext/ggml.h", "ext/ruby_whisper.h", "ext/whisper.h"] + + #### Load-time details + s.require_paths = ['lib','ext'] + s.summary = %q{Ruby whisper.cpp bindings} + s.test_files = ["tests/test_whisper.rb"] + + s.extensions << 'ext/extconf.rb' + + + #### Documentation and testing. + s.homepage = 'https://github.com/ggerganov/whisper.cpp' + s.rdoc_options = ['--main', '../../README.md'] + + + s.platform = Gem::Platform::RUBY + + s.licenses = ['MIT'] +end