extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-17 21:02:03 +02:00
Code Issues Packages Projects Releases Wiki Activity

vulkan : add fp16 support for the conv_2d kernel (llama/14872)

Browse Source 
* add f16 to conv_2d testing
* weaken conv2d test error threshold
This commit is contained in:
Erik Scholz
2025-07-27 12:04:33 +02:00
committed by Georgi Gerganov
parent 5693b857d2
commit d96f4d8ea1
2 changed files with 20 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -484,6 +484,7 @@ struct vk_device_struct {
vk_pipeline pipeline_rwkv_wkv7_f32; vk_pipeline pipeline_rwkv_wkv7_f32;
vk_pipeline pipeline_opt_step_adamw_f32; vk_pipeline pipeline_opt_step_adamw_f32;
vk_pipeline pipeline_conv2d_f32; vk_pipeline pipeline_conv2d_f32;
vk_pipeline pipeline_conv2d_f16_f32;
vk_pipeline pipeline_conv2d_dw_whcn_f32; vk_pipeline pipeline_conv2d_dw_whcn_f32;
vk_pipeline pipeline_conv2d_dw_cwhn_f32; vk_pipeline pipeline_conv2d_dw_cwhn_f32;
@ -3074,12 +3075,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3, device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 }, sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
{ conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true); { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
ggml_vk_create_pipeline(
device, device->pipeline_conv2d_f16_f32, "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
{ conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
} else { } else {
ggml_vk_create_pipeline( ggml_vk_create_pipeline(
device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3, device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 }, sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
{ conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
false); false);
ggml_vk_create_pipeline(
device, device->pipeline_conv2d_f16_f32, "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
{ conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
false);
} }
ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@ -6958,9 +6968,13 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
} }
return nullptr; return nullptr;
case GGML_OP_CONV_2D: case GGML_OP_CONV_2D:
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) { ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
return ctx->device->pipeline_conv2d_f32; if (src0->type == GGML_TYPE_F32) {
return ctx->device->pipeline_conv2d_f32;
} else if (src0->type == GGML_TYPE_F16) {
return ctx->device->pipeline_conv2d_f16_f32;
}
} }
return nullptr; return nullptr;
case GGML_OP_CONV_2D_DW: case GGML_OP_CONV_2D_DW:
@ -8185,13 +8199,13 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0, static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_TENSOR_BINARY_OP_LOCALS GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT(nb00 == sizeof(float)); GGML_ASSERT(nb00 == sizeof(float) || nb00 == sizeof(ggml_fp16_t));
GGML_ASSERT(nb10 == sizeof(float)); GGML_ASSERT(nb10 == sizeof(float));
GGML_ASSERT(nb0 == sizeof(float)); GGML_ASSERT(nb0 == sizeof(float));
@ -10874,7 +10888,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
const vk_device& device = ggml_vk_get_device(ctx->device); const vk_device& device = ggml_vk_get_device(ctx->device);
bool is_Apple = ggml_vk_get_device(ctx->device)->vendor_id == VK_VENDOR_ID_APPLE; bool is_Apple = ggml_vk_get_device(ctx->device)->vendor_id == VK_VENDOR_ID_APPLE;
// Channel-contiguous format is not supported yet. // Channel-contiguous format is not supported yet.
return (op->src[0]->type == GGML_TYPE_F32 && return ((op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
op->src[1]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
op->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[0]) &&

1
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View File

@ -656,6 +656,7 @@ void process_shaders() {
string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}}); string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}})); string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}})); string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));