sync : ggml (HBM + Metal + style) (#1264)

This commit is contained in:
Georgi Gerganov
2023-09-08 17:58:31 +03:00
committed by GitHub
parent 3e9edc6845
commit b39809668a
3 changed files with 42 additions and 39 deletions

30
ggml.c
View File

@@ -106,6 +106,9 @@ typedef void * thread_ret_t;
#include <sys/stat.h>
#include <unistd.h>
#endif
#ifdef GGML_USE_CPU_HBM
#include <hbwmalloc.h>
#endif
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
@@ -195,8 +198,14 @@ typedef void * thread_ret_t;
#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr)
#else
inline static void * ggml_aligned_malloc(size_t size) {
if (size == 0) {
GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
return NULL;
}
void * aligned_memory = NULL;
#ifdef GGML_USE_METAL
#ifdef GGML_USE_CPU_HBM
int result = hbw_posix_memalign(&aligned_memory, 16, size);
#elif GGML_USE_METAL
int result = posix_memalign(&aligned_memory, sysconf(_SC_PAGESIZE), size);
#else
int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
@@ -218,8 +227,12 @@ inline static void * ggml_aligned_malloc(size_t size) {
return aligned_memory;
}
#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
#ifdef GGML_USE_CPU_HBM
#define GGML_ALIGNED_FREE(ptr) if(NULL != ptr) hbw_free(ptr)
#else
#define GGML_ALIGNED_FREE(ptr) free(ptr)
#endif
#endif
#define UNUSED GGML_UNUSED
#define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
@@ -4571,6 +4584,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
return NULL;
}
// allow to call ggml_init with 0 size
if (params.mem_size == 0) {
params.mem_size = GGML_MEM_ALIGN;
}
const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN);
*ctx = (struct ggml_context) {
@@ -4773,7 +4791,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
size_t obj_alloc_size = 0;
if (view_src == NULL && ctx->no_alloc == false) {
if (view_src == NULL && !ctx->no_alloc) {
if (ctx->scratch.data != NULL) {
// allocate tensor data in the scratch buffer
if (ctx->scratch.offs + data_size > ctx->scratch.size) {
@@ -5474,7 +5492,7 @@ static struct ggml_tensor * ggml_mul_impl(
}
if (inplace) {
GGML_ASSERT(is_node == false);
GGML_ASSERT(!is_node);
}
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
@@ -5517,7 +5535,7 @@ static struct ggml_tensor * ggml_div_impl(
}
if (inplace) {
GGML_ASSERT(is_node == false);
GGML_ASSERT(!is_node);
}
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
@@ -19961,7 +19979,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
struct ggml_tensor * data = NULL;
if (params.no_alloc == false) {
if (!params.no_alloc) {
data = ggml_new_tensor_1d(ctx_data, GGML_TYPE_I8, ctx->size);
ok = ok && data != NULL;
@@ -20002,7 +20020,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
}
// point the data member to the appropriate location in the binary blob using the tensor infos
if (params.no_alloc == false) {
if (!params.no_alloc) {
//cur->data = (char *) data->data + ctx->infos[i].offset - ctx->offset; // offset from start of file
cur->data = (char *) data->data + ctx->infos[i].offset; // offset from data
}