ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-17 17:19:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/develop' into ck_tile/refactor_moe_sorting

Browse Source
This commit is contained in:
carlushuang
2025-02-08 19:22:18 +08:00
parent a0a6bc721c a8c5bd9b9a
commit 8b3e32a96d
271 changed files with 13275 additions and 1532 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -12,7 +12,13 @@
#include "ck_tile/host.hpp"
#include "gemm_basic.hpp"
template <typename ALayout, typename BLayout, typename CLayout>
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)
{
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
@@ -20,16 +26,12 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr int kBlockPerCu = 1;
// This part comes from the Codegen
constexpr ck_tile::index_t M_Tile = 128;
constexpr ck_tile::index_t N_Tile = 128;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t K_Tile = 64;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
@@ -37,40 +39,33 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
constexpr ck_tile::index_t K_Warp_Tile = 16;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
@@ -110,12 +105,32 @@ int run_gemm_example(int argc, char* argv[])
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
std::string data_type = arg_parser.get_str("prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
if(data_type == "fp16")
{
return run_gemm_example_with_layouts<ck_tile::half_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
return run_gemm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
return run_gemm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else
{

53
example/ck_tile/03_gemm/gemm_basic.hpp
View File

@@ -18,7 +18,7 @@
#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE
#endif
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrMem
#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrMem
#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Interwave
@@ -43,6 +43,33 @@ struct GemmBasicTypeConfig<ck_tile::half_t>
// ToDo: Add more bias config to support different categories of GEMM.
};
template <>
struct GemmBasicTypeConfig<ck_tile::bf16_t>
{
using ADataType = ck_tile::bf16_t;
using BDataType = ck_tile::bf16_t;
using AccDataType = float;
using CDataType = ck_tile::bf16_t;
};
template <>
struct GemmBasicTypeConfig<ck_tile::fp8_t>
{
using ADataType = ck_tile::fp8_t;
using BDataType = ck_tile::fp8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <>
struct GemmBasicTypeConfig<ck_tile::bf8_t>
{
using ADataType = ck_tile::bf8_t;
using BDataType = ck_tile::bf8_t;
using AccDataType = float;
using CDataType = ck_tile::half_t;
};
template <typename T>
struct DataTypeTraits;
@@ -64,13 +91,23 @@ struct DataTypeTraits<ck_tile::half_t>
static constexpr const char* name = "fp16";
};
using Types = GemmBasicTypeConfig<ck_tile::half_t>;
template <>
struct DataTypeTraits<ck_tile::bf16_t>
{
static constexpr const char* name = "bf16";
};
// Specific type aliases for easy access
using ADataType = Types::ADataType;
using BDataType = Types::BDataType;
using AccDataType = Types::AccDataType;
using CDataType = Types::CDataType;
template <>
struct DataTypeTraits<ck_tile::fp8_t>
{
static constexpr const char* name = "fp8";
};
template <>
struct DataTypeTraits<ck_tile::bf8_t>
{
static constexpr const char* name = "bf8";
};
auto create_args(int argc, char* argv[])
{
@@ -79,7 +116,7 @@ auto create_args(int argc, char* argv[])
.insert("n", "4096", "n dimension")
.insert("k", "2048", "k dimension")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "R", "B tensor data layout - Row by default")
.insert("b_layout", "C", "B tensor data layout - Column by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")

29
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -9,6 +9,7 @@ static constexpr inline auto is_row_major(Layout layout_)
ck_tile::tensor_layout::gemm::RowMajor>>{};
}
template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
@@ -29,7 +30,8 @@ auto calculate_rtol_atol(const ck_tile::index_t K,
return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
}
template <typename ALayout, typename BLayout, typename CLayout>
template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType,
typename ALayout, typename BLayout, typename CLayout>
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
ck_tile::DeviceMem& b_k_n_dev_buf,
ck_tile::DeviceMem& c_m_n_dev_buf,
@@ -55,7 +57,8 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
args.stride_B = stride_B;
args.stride_C = stride_C;
float ave_time = gemm_calc<ALayout, BLayout, CLayout>(
float ave_time = gemm_calc<ADataType, BDataType, AccDataType, CDataType,
ALayout, BLayout, CLayout>(
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
std::size_t flop = std::size_t(2) * M * N * K;
@@ -66,13 +69,19 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
std::cout << "Run Gemm kernel with M =" << M << " N =" << N << " K =" << K
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
<< " A_Layout =" << ALayout::name
<< " B_Layout =" << BLayout::name
<< " C_Layout =" << CLayout::name
<< " A Type = " << DataTypeTraits<ADataType>::name
<< " B Type = " << DataTypeTraits<BDataType>::name
<< " C Type = " << DataTypeTraits<CDataType>::name
<< " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< std::endl;
return ave_time;
}
template <typename ALayout, typename BLayout, typename CLayout>
template <typename PrecType, typename ALayout, typename BLayout, typename CLayout>
int run_gemm_example_with_layouts(int argc,
char* argv[],
const ALayout a_layout = ALayout{},
@@ -83,6 +92,11 @@ int run_gemm_example_with_layouts(int argc,
if(!result)
return -1;
using ADataType = typename GemmBasicTypeConfig<PrecType>::ADataType;
using BDataType = typename GemmBasicTypeConfig<PrecType>::BDataType;
using CDataType = typename GemmBasicTypeConfig<PrecType>::CDataType;
using AccDataType = typename GemmBasicTypeConfig<PrecType>::AccDataType;
ck_tile::index_t M = arg_parser.get_int("m");
ck_tile::index_t N = arg_parser.get_int("n");
ck_tile::index_t K = arg_parser.get_int("k");
@@ -119,7 +133,8 @@ int run_gemm_example_with_layouts(int argc,
c_m_n_dev_buf.SetZero();
c_m_n_dev_result.SetZero();
invoke_gemm<ALayout, BLayout, CLayout>(a_m_k_dev_buf,
invoke_gemm<ADataType, BDataType, AccDataType, CDataType,
ALayout, BLayout, CLayout>(a_m_k_dev_buf,
b_k_n_dev_buf,
c_m_n_dev_buf,
M,
@@ -145,7 +160,8 @@ int run_gemm_example_with_layouts(int argc,
a_m_k, b_k_n, c_m_n_host_ref);
const float max_accumulated_value =
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>
(K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_host_ref,
"Error: Incorrect results!",
@@ -202,7 +218,8 @@ int run_gemm_example_with_layouts(int argc,
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
const float max_accumulated_value =
*std::max_element(c_m_n_gpu_ref.mData.begin(), c_m_n_gpu_ref.mData.end());
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>
(K, kbatch, max_accumulated_value);
pass = ck_tile::check_err(c_m_n_dev_result,
c_m_n_gpu_ref,
"Error: Incorrect results!",

7
example/ck_tile/03_gemm/script/benchmark_basic.sh
View File

@@ -1,12 +1,13 @@
#!/bin/sh
EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
VALID=0
VALID=1
for b_matrix_layout in "R" "C"; do
for b_matrix_layout in "C"; do
for m in "64" "512" "1024" "2048"; do
for n in "512" "1024" "2048"; do
for k in "64" "512" "1024" "2048"; do
$EXE -prec=fp16 -b=1 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
$EXE -prec=fp16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
done
done
done

0
example/ck_tile/03_gemm/script/benchmark_basic_bf16.sh Normal file
View File

0
example/ck_tile/03_gemm/script/benchmark_basic_bf8.sh Normal file
View File

14
example/ck_tile/03_gemm/script/benchmark_basic_fp8.sh Normal file
View File

@@ -0,0 +1,14 @@
#!/bin/sh
EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
VALID=1
for b_matrix_layout in "C"; do
for m in "64" "512" "1024" "2048"; do
for n in "512" "1024" "2048"; do
for k in "64" "512" "1024" "2048"; do
$EXE -prec=fp8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
done
done
done
done

10
example/ck_tile/03_gemm/script/benchmark_mem_pipeline.sh
View File

@@ -1,12 +1,12 @@
#!/bin/sh
EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
VALID=0
VALID=1
for b_matrix_layout in "R" "C"; do
for m in "64" "512" "1024" "2048"; do
for b_matrix_layout in "C"; do
for m in "512" "1024" "2048" "4096"; do
for n in "512" "1024" "2048"; do
for k in "64" "512" "1024" "2048"; do
$EXE -prec=fp16 -b=1 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
for k in "512" "1024" "2048"; do
$EXE -prec=fp16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
done
done
done

13
example/ck_tile/03_gemm/script/benchmark_mem_pipeline_bf16.sh Normal file
View File

@@ -0,0 +1,13 @@
#!/bin/sh
EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
VALID=1
for b_matrix_layout in "C"; do
for m in "512" "1024" "2048" "4096"; do
for n in "512" "1024" "2048"; do
for k in "512" "1024" "2048"; do
$EXE -prec=bf16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
done
done
done
done

13
example/ck_tile/03_gemm/script/benchmark_mem_pipeline_bf8.sh Normal file
View File

@@ -0,0 +1,13 @@
#!/bin/sh
EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
VALID=1
for b_matrix_layout in "C"; do
for m in "512" "1024" "2048" "4096"; do
for n in "512" "1024" "2048"; do
for k in "512" "1024" "2048"; do
$EXE -prec=bf8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
done
done
done
done

13
example/ck_tile/03_gemm/script/benchmark_mem_pipeline_fp8.sh Normal file
View File

@@ -0,0 +1,13 @@
#!/bin/sh
EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
VALID=1
for b_matrix_layout in "C"; do
for m in "512" "1024" "2048" "4096"; do
for n in "512" "1024" "2048"; do
for k in "512" "1024" "2048"; do
$EXE -prec=fp8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
done
done
done
done

31
example/ck_tile/03_gemm/script/smoke_test_basic.sh
View File

@@ -7,22 +7,20 @@ export CK_REPEAT=1
COMMON_ARGS='-v=2 -warmup=0 -repeat=1'
run_fp16_tests() {
for batch in 1 2; do
for m in 128 1024; do
for n in 128 2048; do
for k in 32 64; do
run_tests() {
for m in 128 1024; do
for n in 128 2048; do
for k in 64 128; do
$EXE -b=$batch -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -e=1e-5 -prec=fp16 $COMMON_ARGS
if [ $? -eq 0 ]; then
echo "Success: Test with batch=$batch, m=$m, n=$n, k=$k executed successfully."
else
echo "Error: Test with batch=$batch, m=$m, n=$n, k=$k failed to execute properly."
# Optionally, exit or break if you need to halt further execution
# exit 1
fi
$EXE -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -prec=$1 $COMMON_ARGS
if [ $? -eq 0 ]; then
echo "Success: Test with m=$m, n=$n, k=$k executed successfully."
else
echo "Error: Test with m=$m, n=$n, k=$k failed to execute properly."
# Optionally, exit or break if you need to halt further execution
# exit 1
fi
done
done
done
done
@@ -30,6 +28,9 @@ run_fp16_tests() {
set -x
run_fp16_tests
run_tests "fp16"
run_tests "bf16"
run_tests "fp8"
run_tests "bf8"
set +x

31
example/ck_tile/03_gemm/script/smoke_test_mem_pipeline.sh
View File

@@ -7,22 +7,20 @@ export CK_REPEAT=1
COMMON_ARGS='-v=2 -warmup=0 -repeat=1'
run_fp16_tests() {
for batch in 1 2; do
for m in 128 1024; do
for n in 128 2048; do
for k in 32 64; do
run_tests() {
for m in 512 1024; do
for n in 512 2048; do
for k in 512 1024; do
$EXE -b=$batch -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -e=1e-5 -prec=fp16 $COMMON_ARGS
if [ $? -eq 0 ]; then
echo "Success: Test with batch=$batch, m=$m, n=$n, k=$k executed successfully."
else
echo "Error: Test with batch=$batch, m=$m, n=$n, k=$k failed to execute properly."
# Optionally, exit or break if you need to halt further execution
# exit 1
fi
$EXE -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -prec=$1 $COMMON_ARGS
if [ $? -eq 0 ]; then
echo "Success: Test with batch=$batch, m=$m, n=$n, k=$k executed successfully."
else
echo "Error: Test with batch=$batch, m=$m, n=$n, k=$k failed to execute properly."
# Optionally, exit or break if you need to halt further execution
# exit 1
fi
done
done
done
done
@@ -30,6 +28,9 @@ run_fp16_tests() {
set -x
run_fp16_tests
run_tests "fp16"
run_tests "bf16"
run_tests "fp8"
run_tests "bf8"
set +x

124
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
@@ -12,7 +12,13 @@
#include "ck_tile/host.hpp"
#include "gemm_basic.hpp"
template <typename ALayout, typename BLayout, typename CLayout>
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)
{
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
@@ -33,7 +39,7 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
// Compute friendly for Intrawave scheduler
constexpr ck_tile::index_t M_Tile = 256;
constexpr ck_tile::index_t N_Tile = 256;
constexpr ck_tile::index_t K_Tile = 32;
constexpr ck_tile::index_t K_Tile = 64;
constexpr ck_tile::index_t M_Warp = 2;
constexpr ck_tile::index_t N_Warp = 2;
@@ -50,7 +56,9 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr bool TransposeC = false;
constexpr int kBlockPerCu = 1;
constexpr int kBlockPerCu = 1;
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
constexpr ck_tile::index_t TileParitionerM01 = 4;
// ===============================================
@@ -58,10 +66,8 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using TilePartitioner = ck_tile::
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits = ck_tile::
@@ -95,6 +101,19 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
using GemmPipeline =
GEMM_PIPELINE<UniversalGemmProblem, ck_tile::UniversalGemmPipelineAgBgCrPolicy>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
UniversalGemmProblem::TransposeC>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
@@ -230,24 +249,101 @@ int run_gemm_example(int argc, char* argv[])
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
std::string data_type = arg_parser.get_str("prec");
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
if(data_type == "fp16")
{
return run_gemm_example_with_layouts<ck_tile::half_t>(argc, argv, Row{}, Row{}, Row{});
}
else if(data_type == "bf16")
{
return run_gemm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Row{}, Row{}, Row{});
}
else if(data_type == "fp8")
{
return run_gemm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Row{}, Row{}, Row{});
}
else if(data_type == "bf8")
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Row{}, Row{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
if(data_type == "fp16")
{
return run_gemm_example_with_layouts<ck_tile::half_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
return run_gemm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
return run_gemm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Row{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else if(a_layout == "C" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
if(data_type == "fp16")
{
return run_gemm_example_with_layouts<ck_tile::half_t>(argc, argv, Col{}, Col{}, Row{});
}
else if(data_type == "bf16")
{
return run_gemm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Col{}, Col{}, Row{});
}
else if(data_type == "fp8")
{
return run_gemm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Col{}, Col{}, Row{});
}
else if(data_type == "bf8")
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Col{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else if(a_layout == "C" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
if(data_type == "fp16")
{
return run_gemm_example_with_layouts<ck_tile::half_t>(argc, argv, Col{}, Row{}, Row{});
}
else if(data_type == "bf16")
{
return run_gemm_example_with_layouts<ck_tile::bf16_t>(argc, argv, Col{}, Row{}, Row{});
}
else if(data_type == "fp8")
{
return run_gemm_example_with_layouts<ck_tile::fp8_t>(argc, argv, Col{}, Row{}, Row{});
}
else if(data_type == "bf8")
{
return run_gemm_example_with_layouts<ck_tile::bf8_t>(argc, argv, Col{}, Row{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data_type!");
}
}
else
{

44
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View File

@@ -19,12 +19,9 @@ template <typename ALayout, typename BLayout, typename CLayout>
float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s)
{
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr int kBlockPerCu = 1;
@@ -41,38 +38,31 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;

46
example/ck_tile/17_grouped_gemm/grouped_gemm.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
@@ -20,12 +20,9 @@ namespace {
struct GroupedGemmKernelParam
{
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const bool kTilePermute = false;
static const ck_tile::index_t kOutputRank = 2;
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const int kBlockPerCu = 1;
static const ck_tile::index_t M_Tile = 128;
@@ -54,24 +51,6 @@ using CodegenGemmShape =
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
template <typename CLayout>
using GemmEpilogue = std::conditional_t<
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
GroupedGemmKernelParam::kPadM,
GroupedGemmKernelParam::kPadN,
GroupedGemmKernelParam::kTilePermute,
GroupedGemmKernelParam::kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<AccDataType,
CDataType,
GroupedGemmKernelParam::kPadM,
GroupedGemmKernelParam::kPadN>>>;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmTraits = ck_tile::TileGemmTraits<GroupedGemmKernelParam::kPadM,
GroupedGemmKernelParam::kPadN,
@@ -92,10 +71,25 @@ template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>>;
template <typename ALayout, typename BLayout, typename CLayout>
using GemmEpilogue = ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<
AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem<ALayout, BLayout, CLayout>::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GroupedGemmKernelParam::M_Warp,
GroupedGemmKernelParam::N_Warp,
GroupedGemmKernelParam::M_Warp_Tile,
GroupedGemmKernelParam::N_Warp_Tile,
GroupedGemmKernelParam::K_Warp_Tile,
CodegenPipelineProblem<ALayout, BLayout, CLayout>::TransposeC>>;
template <typename ALayout, typename BLayout, typename CLayout>
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,
CodegenGemmPipeline<ALayout, BLayout, CLayout>,
GemmEpilogue<CLayout>>;
GemmEpilogue<ALayout, BLayout, CLayout>>;
}; // namespace
std::size_t get_workspace_size(const std::vector<grouped_gemm_kargs>& gemm_descs)