ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-21 21:39:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '99857e10e64889e631f58a3cc38fca3a3f248bc4' into develop

Browse Source
This commit is contained in:
assistant-librarian[bot]
2025-05-28 06:07:44 +00:00
parent 8685301f50
commit a1a3ed5826
21 changed files with 478 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -214,4 +214,15 @@ int run_gemm_example(int argc, char* argv[])
}
}
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
int main(int argc, char* argv[])
{
try
{
return !run_gemm_example(argc, argv);
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

7
example/ck_tile/03_gemm/gemm_utils.hpp
View File

@@ -220,4 +220,11 @@ auto create_args(int argc, char* argv[])
}
// host API
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s);

2
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -178,7 +178,7 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
float ave_time =
gemm_calc<ADataType, BDataType, AccDataType, CDataType, ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte =

147
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -11,6 +11,7 @@
#include "ck_tile/host.hpp"
#include "gemm_utils.hpp"
#include "run_gemm_example.inc"
template <typename Pipeline, ck_tile::TailNumber TN>
void try_run(ck_tile::TailNumber tn)
@@ -74,64 +75,102 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
float ave_time{0};
const auto Run = [&](const auto has_hot_loop_,
const auto tail_number_,
const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = GEMM_PIPELINE_SCHEDULER;
constexpr auto memory_operation = memory_operation_.value;
const auto Run =
[&](const auto has_hot_loop_, const auto tail_number_, const auto memory_operation_) {
constexpr bool has_hot_loop_v = has_hot_loop_.value;
constexpr auto tail_number_v = tail_number_.value;
constexpr auto scheduler = GEMM_PIPELINE_SCHEDULER;
constexpr auto memory_operation = memory_operation_.value;
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
GemmUniversalTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
BDataType,
AccDataType,
GemmShape,
GemmUniversalTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
AccDataType,
CDataType,
CLayout,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
UniversalGemmProblem::TransposeC,
memory_operation>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
AccDataType,
CDataType,
CLayout,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
UniversalGemmProblem::TransposeC,
memory_operation>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(!Kernel::IsSupportedArgument(kargs))
{
throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
}
if(s.log_level_ > 0)
{
std::cout << "Launching kernel with args:"
<< " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z
<< "}" << std::endl;
}
if(s.flush_cache_)
{
std::cout << "Flushing cache..." << std::endl;
static constexpr ck_tile::index_t APackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
static constexpr ck_tile::index_t BPackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
ave_time = ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
kargs.a_ptr, kargs.b_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck_tile::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(args.k_batch > 1)
hipGetErrorString(hipMemsetAsync(
args.c_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
};
ave_time = ck_tile::launch_kernel_preprocess(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
const auto RunSplitk = [&](const auto has_hot_loop_, const auto tail_number_) {
if(args.k_batch == 1)
@@ -243,8 +282,6 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
return ave_time;
}
#include "run_gemm_example.inc"
template <typename APrecType, typename BPrecType = APrecType, typename CPrecType = APrecType>
int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int argc, char* argv[])
{
@@ -345,7 +382,7 @@ int main(int argc, char* argv[])
{
try
{
run_gemm_example(argc, argv);
return !run_gemm_example(argc, argv);
}
catch(const std::runtime_error& e)
{

30
example/ck_tile/13_moe_sorting/moe_sorting.cpp
View File

@@ -334,16 +334,26 @@ bool test_moe_sorting(ck_tile::ArgParser args)
int main(int argc, char** argv)
{
auto [result, args] = create_args(argc, argv);
if(!result)
return -1;
std::string index_prec = args.get_str("pr_i");
std::string weight_prec = args.get_str("pr_w");
bool r = true;
if(weight_prec.compare("fp32") == 0 && index_prec.compare("int32") == 0)
try
{
r &= test_moe_sorting<float, ck_tile::index_t>(args);
auto [result, args] = create_args(argc, argv);
if(!result)
return -1;
std::string index_prec = args.get_str("pr_i");
std::string weight_prec = args.get_str("pr_w");
bool r = true;
if(weight_prec == "fp32" && index_prec == "int32")
{
r &= test_moe_sorting<float, ck_tile::index_t>(args);
}
return r ? 0 : -1;
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
return r ? 0 : -1;
}

13
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View File

@@ -320,4 +320,15 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
int main(int argc, char* argv[])
{
try
{
return !run_batched_gemm_example(argc, argv);
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

13
example/ck_tile/17_grouped_gemm/grouped_gemm.cpp
View File

@@ -319,4 +319,15 @@ float grouped_gemm(const std::vector<grouped_gemm_kargs>& gemm_descs,
#include "run_grouped_gemm_example.inc"
constexpr bool Persistent = false;
int main(int argc, char* argv[]) { return !run_grouped_gemm_example<Persistent>(argc, argv); }
int main(int argc, char* argv[])
{
try
{
return !run_grouped_gemm_example<Persistent>(argc, argv);
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

58
example/ck_tile/18_flatmm/flatmm_basic.cpp
View File

@@ -11,6 +11,7 @@
#include "ck_tile/host.hpp"
#include "flatmm_basic.hpp"
#include "run_flatmm_example.inc"
template <typename ADataType,
typename BDataType,
@@ -115,9 +116,47 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
<< std::endl;
}
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
float ave_time{0};
if(s.flush_cache_)
{
std::cout << "Flushing cache..." << std::endl;
static constexpr ck_tile::index_t APackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
static constexpr ck_tile::index_t BPackedSize =
std::is_same_v<BDataType, ck_tile::pk_int4_t> ? 2 : 1;
ck_tile::HostTensor<ADataType> a_m(ck_tile::host_tensor_descriptor(
args.M, args.K, args.stride_A, is_row_major(ALayout{})));
ck_tile::HostTensor<BDataType> b_n(ck_tile::host_tensor_descriptor(
args.K, args.N, args.stride_B, is_row_major(BLayout{})));
auto size_a_buffer = a_m.get_element_space_size_in_bytes() / APackedSize;
auto size_b_buffer = b_n.get_element_space_size_in_bytes() / BPackedSize;
ck_tile::RotatingMemWrapper<ADataType, BDataType> rotating_mem(
kargs.a_ptr, kargs.b_shuffle_ptr, s.rotating_count_, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck_tile::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(args.k_batch > 1)
hipGetErrorString(hipMemsetAsync(
args.c_ptr, 0, args.M * args.N * sizeof(CDataType), s.stream_id_));
};
ave_time = ck_tile::launch_kernel_preprocess(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
if(args.k_batch == 1)
@@ -132,8 +171,6 @@ float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_con
}
}
#include "run_flatmm_example.inc"
int run_flatmm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
@@ -177,4 +214,15 @@ int run_flatmm_example(int argc, char* argv[])
return -1;
}
int main(int argc, char* argv[]) { return !run_flatmm_example(argc, argv); }
int main(int argc, char* argv[])
{
try
{
return !run_flatmm_example(argc, argv);
}
catch(const std::runtime_error& e)
{
std::cerr << "Runtime error: " << e.what() << '\n';
return EXIT_FAILURE;
}
}

7
example/ck_tile/18_flatmm/flatmm_basic.hpp
View File

@@ -133,4 +133,11 @@ auto create_args(int argc, char* argv[])
}
// host API
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
float flatmm_calc(const ck_tile::FlatmmHostArgs& args, const ck_tile::stream_config& s);

2
example/ck_tile/18_flatmm/run_flatmm_example.inc
View File

@@ -122,7 +122,7 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf,
float ave_time =
flatmm_calc<ADataType, BDataType, AccDataType, CDataType, ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat, true, true, 50});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_byte =