vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Vulkan: VK_KHR_cooperative_matrix support to speed up prompt processing (#10597)

Browse source 
* Vulkan: Implement VK_KHR_cooperative_matrix support in the matrix matrix multiplication shader

* Improve performance with better q4_k and q5_k dequant and store unrolling

* Add Vulkan MUL_MAT and MUL_MAT_ID accumulator precision selection

* Rework mulmat shader selection and compilation logic, avoid compiling shaders that won't get used by device

* Vulkan: Implement accumulator switch for specific mul mat mat shaders

* Vulkan: Unroll more loops for more mul mat mat performance

* Vulkan: Add VK_AMD_shader_core_properties2 support to read Compute Unit count for split_k logic

* Disable coopmat support on AMD proprietary driver

* Remove redundant checks

* Add environment variable GGML_VK_DISABLE_COOPMAT to disable VK_KHR_cooperative_matrix support

* Fix rebase typo

* Fix coopmat2 MUL_MAT_ID pipeline selection
This commit is contained in:
0cc4m 2024-12-07 10:24:15 +01:00 • committed by GitHub
parent 86a1934978
commit 3df784b305
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 739 additions and 386 deletions
Download patch file Download diff file Expand all files Collapse all files

73
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View file

@ -60,6 +60,7 @@ const std::vector<std::string> type_names = {
"iq4_nl"
};
namespace {
void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
#ifdef _WIN32
HANDLE stdout_read, stdout_write;
@ -198,8 +199,8 @@ static uint32_t compile_count = 0;
static std::mutex compile_count_mutex;
static std::condition_variable compile_count_cond;
void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat2 = false, bool f16acc = false) {
std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat ? "_coopmat" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
std::string out_fname = join_paths(output_dir, name + ".spv");
std::string in_path = join_paths(input_dir, in_fname);
@ -258,7 +259,7 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
}
static std::vector<std::future<void>> compiles;
void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat2 = false, bool f16acc = false) {
void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
{
// wait until fewer than N compiles are in progress.
// 16 is an arbitrary limit, the goal is to avoid "failed to create pipe" errors.
@ -269,10 +270,10 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
}
compile_count++;
}
compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat2, f16acc));
compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat, coopmat2, f16acc));
}
void matmul_shaders(bool fp16, bool matmul_id, bool coopmat2, bool f16acc) {
void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool f16acc) {
std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
@ -291,14 +292,20 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat2, bool f16acc) {
base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
if (coopmat) {
base_dict["COOPMAT"] = "1";
}
base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
// Shaders with f16 B_TYPE
string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
for (const auto& tname : type_names) {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
@ -307,12 +314,12 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat2, bool f16acc) {
// For aligned matmul loads
std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : "2";
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
if (tname != "f16" && tname != "f32") {
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
}
}
}
@ -322,25 +329,24 @@ void process_shaders() {
std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
// matmul
for (const auto& fp16 : {false, true}) {
for (const auto& matmul_id : {false, true}) {
for (const auto& coopmat2 : {false, true}) {
for (const auto& f16acc : {false, true}) {
#if !defined(VK_NV_cooperative_matrix2)
if (coopmat2) {
continue;
}
for (const auto& matmul_id : {false, true}) {
// No coopmats
// fp32
matmul_shaders(false, matmul_id, false, false, false);
// fp16, fp32acc and fp16acc
matmul_shaders(true, matmul_id, false, false, false);
matmul_shaders(true, matmul_id, false, false, true);
// Coopmat, fp32acc and fp16acc
matmul_shaders(true, matmul_id, true, false, false);
matmul_shaders(true, matmul_id, true, false, true);
#if defined(VK_NV_cooperative_matrix2)
// Coopmat2, fp32acc and fp16acc
matmul_shaders(true, matmul_id, false, true, false);
matmul_shaders(true, matmul_id, false, true, true);
#endif
if (coopmat2 && !fp16) {
continue;
}
if (!coopmat2 && f16acc) {
continue;
}
matmul_shaders(fp16, matmul_id, coopmat2, f16acc);
}
}
}
}
#if defined(VK_NV_cooperative_matrix2)
@ -355,11 +361,11 @@ void process_shaders() {
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, true, f16acc);
merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc);
} else {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, true, f16acc);
merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
}
}
}
@ -524,6 +530,7 @@ void write_output_files() {
fclose(hdr);
fclose(src);
}
}
int main(int argc, char** argv) {
std::map<std::string, std::string> args;