replica
/
llama.cpp
mirror of https://github.com/ggml-org/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

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

* 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 GitHub
parent f1a4e72de5
commit 89d1029559
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 49 additions and 20 deletions
Show Stats 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_opt_step_adamw_f32;
vk_pipeline pipeline_conv2d_f32;
vk_pipeline pipeline_conv2d_f16_f32;
vk_pipeline pipeline_conv2d_dw_whcn_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,
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);
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 {
ggml_vk_create_pipeline(
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 },
{ 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_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);
@ -6958,9 +6968,13 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
}
return nullptr;
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)) {
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;
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,
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(dst->type == GGML_TYPE_F32);
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(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);
bool is_Apple = ggml_vk_get_device(ctx->device)->vendor_id == VK_VENDOR_ID_APPLE;
// 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->type == GGML_TYPE_F32 &&
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("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_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));

44
tests/test-backend-ops.cpp
View File

@ -3734,6 +3734,7 @@ struct test_im2col : public test_case {
struct test_conv_2d : public test_case {
const std::array<int64_t, 4> ne_input;
const std::array<int64_t, 4> ne_kernel;
const ggml_type type_kernel;
const int stride0;
const int stride1;
const int padding0;
@ -3751,7 +3752,11 @@ struct test_conv_2d : public test_case {
// IM2COL -> MUL_MM graph will be built.
std::string vars() override {
return VARS_TO_STR9(ne_input, ne_kernel, stride0, stride1, padding0, padding1, dilation0, dilation1, cwhn);
return VARS_TO_STR10(ne_input, ne_kernel, type_kernel, stride0, stride1, padding0, padding1, dilation0, dilation1, cwhn);
}
double max_nmse_err() override {
return 5e-4;
}
uint64_t op_flops(ggml_tensor * t) override {
@ -3782,10 +3787,11 @@ struct test_conv_2d : public test_case {
}
test_conv_2d(std::array<int64_t, 4> ne_input = { 64, 64, 16, 1 },
std::array<int64_t, 4> ne_kernel = { 3, 3, 1, 16 }, int stride0 = 1, int stride1 = 1, int padding0 = 0,
int padding1 = 0, int dilation0 = 1, int dilation1 = 1, bool cwhn = false) :
std::array<int64_t, 4> ne_kernel = { 3, 3, 1, 16 }, ggml_type type_kernel = GGML_TYPE_F32, int stride0 = 1,
int stride1 = 1, int padding0 = 0, int padding1 = 0, int dilation0 = 1, int dilation1 = 1, bool cwhn = false) :
ne_input(ne_input),
ne_kernel(ne_kernel),
type_kernel(type_kernel),
stride0(stride0),
stride1(stride1),
padding0(padding0),
@ -3798,7 +3804,7 @@ struct test_conv_2d : public test_case {
ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
ggml_set_name(input, "input");
ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data());
ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data());
ggml_set_name(kernel, "kernel");
if (cwhn) {
@ -5165,10 +5171,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
{ 16, 3, 256, 128, 8 }
};
for (auto act_case : cases) {
test_cases.emplace_back(new test_conv_2d(
{ act_case[iwh_idx], act_case[iwh_idx], act_case[Cin_idx], act_case[B_idx] },
{ act_case[kwh_idx], act_case[kwh_idx], act_case[Cin_idx], act_case[Cout_idx] }, 1, 1, 0, 0, 1, 1, false));
for (auto kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
for (auto act_case : cases) {
test_cases.emplace_back(new test_conv_2d(
{ act_case[iwh_idx], act_case[iwh_idx], act_case[Cin_idx], act_case[B_idx] },
{ act_case[kwh_idx], act_case[kwh_idx], act_case[Cin_idx], act_case[Cout_idx] },
kernel_type, 1, 1, 0, 0, 1, 1, false));
}
}
#endif
@ -5194,8 +5203,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
for (uint32_t W : { 1, 141 }) {
if (calc_conv_output_size(W, KW, s0, p0, d0) > 0 &&
calc_conv_output_size(H, KH, s1, p1, d1) > 0) {
test_cases.emplace_back(new test_conv_2d(
{ W, H, Cin, 2 }, { KW, KH, Cin, Cout }, s0, s1, p0, p1, d0, d1, false));
for (auto kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_conv_2d(
{ W, H, Cin, 2 }, { KW, KH, Cin, Cout }, kernel_type, s0, s1, p0, p1, d0, d1, false));
}
}
}
}
@ -5840,11 +5851,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
{ 16, 3, 512, 128, 8 },
};
for (auto act_case : cases) {
// Direct CONV_2D
test_cases.emplace_back(new test_conv_2d(
{ act_case[iwh_idx], act_case[iwh_idx], act_case[Cin_idx], act_case[B_idx] },
{ act_case[kwh_idx], act_case[kwh_idx], act_case[Cin_idx], act_case[Cout_idx] }, 1, 1, 0, 0, 1, 1, false));
for (auto kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
for (auto act_case : cases) {
// Direct CONV_2D
test_cases.emplace_back(new test_conv_2d(
{ act_case[iwh_idx], act_case[iwh_idx], act_case[Cin_idx], act_case[B_idx] },
{ act_case[kwh_idx], act_case[kwh_idx], act_case[Cin_idx], act_case[Cout_idx] },
kernel_type, 1, 1, 0, 0, 1, 1, false));
}
}
test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 1, 1, 1}));