Skip to content

GitLab

Commit d1e9d6cc authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into refactor_threadwise_copy_using_sfcurve

parents 7fcff6f5 992f71e3
Hide whitespace changes
Inline Side-by-side
Showing with 1915 additions and 271 deletions
.gitignore 0 → 100644
View file @ d1e9d6cc
# Compiled Object files
*.slo
*.lo
*.o
*.obj
# Precompiled Headers
*.gch
*.pch
*.ipch
# Compiled Dynamic libraries
*.so
*.dylib
*.dll
# Fortran module files
*.mod
# Compiled Static libraries
*.lai
*.la
*.a
*.lib
# Executables
*.exe
*.out
*.app
# vim tags
tags
.tags
.*.swp
# Editors
.vscode
# build-in-source directory
build*
# emacs temporary/backup files
.\#*
\#*\#
*~
# GDB temporary files
.gdb_history
\ No newline at end of file
CMakeLists.txt
View file @ d1e9d6cc
...@@ -240,9 +240,8 @@ file(GLOB_RECURSE DEVICE_OPS_SOURCE "device_operation/*.cpp") ...@@ -240,9 +240,8 @@ file(GLOB_RECURSE DEVICE_OPS_SOURCE "device_operation/*.cpp")
set(CK_HEADERS ${COMPOSABLE_KERNEL_HEADERS} ${DEVICE_OPS_HEADERS}) set(CK_HEADERS ${COMPOSABLE_KERNEL_HEADERS} ${DEVICE_OPS_HEADERS})
set(CK_SOURCE ${DEVICE_OPS_SOURCE}) set(CK_SOURCE ${DEVICE_OPS_SOURCE})
add_library(composable_kernel add_library(composable_kernel ${CK_SOURCE})
${CK_SOURCE}
)
target_include_directories(composable_kernel PUBLIC target_include_directories(composable_kernel PUBLIC
$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/composable_kernel/include> $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/composable_kernel/include>
......
Dockerfile
View file @ d1e9d6cc
FROM ubuntu:18.04 FROM ubuntu:18.04
ARG ROCMVERSION=4.5 ARG ROCMVERSION=4.3.1
ARG OSDB_BKC_VERSION ARG OSDB_BKC_VERSION
RUN set -xe RUN set -xe
......
Jenkinsfile
View file @ d1e9d6cc
...@@ -17,7 +17,7 @@ def cmake_build(Map conf=[:]){ ...@@ -17,7 +17,7 @@ def cmake_build(Map conf=[:]){
def compiler = conf.get("compiler","/opt/rocm/bin/hipcc") def compiler = conf.get("compiler","/opt/rocm/bin/hipcc")
def config_targets = conf.get("config_targets","check") def config_targets = conf.get("config_targets","check")
def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined " + conf.get("extradebugflags", "") def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined " + conf.get("extradebugflags", "")
def build_envs = "CTEST_PARALLEL_LEVEL=4 MIOPEN_CONV_PRECISE_ROCBLAS_TIMING=0 " + conf.get("build_env","") def build_envs = "CTEST_PARALLEL_LEVEL=4 " + conf.get("build_env","")
def prefixpath = conf.get("prefixpath","/opt/rocm") def prefixpath = conf.get("prefixpath","/opt/rocm")
def setup_args = conf.get("setup_args","") def setup_args = conf.get("setup_args","")
...@@ -60,7 +60,7 @@ def cmake_build(Map conf=[:]){ ...@@ -60,7 +60,7 @@ def cmake_build(Map conf=[:]){
cd build cd build
""" """
def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ") def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args} .. ")
def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(nproc) ${config_targets}") def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j\$(( \$(nproc) / 4 )) ${config_targets}")
def execute_cmd = conf.get("execute_cmd", "") def execute_cmd = conf.get("execute_cmd", "")
def cmd = conf.get("cmd", """ def cmd = conf.get("cmd", """
...@@ -177,15 +177,27 @@ pipeline { ...@@ -177,15 +177,27 @@ pipeline {
// buildHipClangJobAndReboot(build_cmd: build_cmd, no_reboot:true, prefixpath: '/opt/rocm', build_type: 'debug') // buildHipClangJobAndReboot(build_cmd: build_cmd, no_reboot:true, prefixpath: '/opt/rocm', build_type: 'debug')
// } // }
// } // }
stage('Build Profiler: gfx908') stage('Build Profiler: Release, gfx908')
{ {
agent { label rocmnode("gfx908")} agent { label rocmnode("nogpu")}
environment{ environment{
setup_args = """ -D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " -DBUILD_DEV=On """ setup_args = """ -D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " -DBUILD_DEV=On """
build_cmd = "make -j\$(nproc) -k ckProfiler"
} }
steps{ steps{
buildHipClangJobAndReboot(setup_args:setup_args, build_cmd:build_cmd, no_reboot:true, build_type: 'Release') buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
}
}
stage('Build Profiler: Debug, gfx908')
{
agent { label rocmnode("nogpu")}
environment{
setup_args = """ -D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " -DBUILD_DEV=On """
}
steps{
// until we stabilize debug build due to compiler crashes
catchError(buildResult: 'SUCCESS', stageResult: 'FAILURE') {
buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Debug')
}
} }
} }
stage('Clang Format') { stage('Clang Format') {
...@@ -207,6 +219,24 @@ pipeline { ...@@ -207,6 +219,24 @@ pipeline {
} }
} }
} }
stage("Tests")
{
parallel
{
stage("Run Tests: gfx908")
{
agent{ label rocmnode("gfx908")}
environment{
setup_args = """ -D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " -DBUILD_DEV=On """
}
steps{
buildHipClangJobAndReboot(setup_args:setup_args, config_targets: "check", no_reboot:true, build_type: 'Release')
}
}
}
}
// enable after the cmake file supports packaging // enable after the cmake file supports packaging
// stage("Packages") { // stage("Packages") {
// when { // when {
...@@ -222,4 +252,4 @@ pipeline { ...@@ -222,4 +252,4 @@ pipeline {
// } // }
// } // }
} }
} }
\ No newline at end of file
composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp
View file @ d1e9d6cc
...@@ -17,7 +17,7 @@ template <index_t BlockSize, ...@@ -17,7 +17,7 @@ template <index_t BlockSize,
index_t NPerXDL, index_t NPerXDL,
index_t MRepeat, index_t MRepeat,
index_t NRepeat, index_t NRepeat,
index_t K1> index_t KPack>
struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
{ {
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
...@@ -29,10 +29,15 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -29,10 +29,15 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1); static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1); static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
static constexpr index_t KPerBlock =
BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
static constexpr index_t K0 = BK0NK1BlockDesc{}.GetLength(I0); static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{}; static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL); static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL); static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
...@@ -66,7 +71,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -66,7 +71,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex(); const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
return make_tuple(xdlops_a_idx[I0], 0, waveId_m, xdlops_a_idx[I1], 0); return make_tuple(0, waveId_m, xdlops_a_idx[I1], Number<KPack>{} * xdlops_a_idx[I0]);
} }
__device__ static auto CalculateBThreadOriginDataIndex() __device__ static auto CalculateBThreadOriginDataIndex()
...@@ -77,7 +82,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -77,7 +82,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex(); const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
return make_tuple(xdlops_b_idx[I0], 0, waveId_n, xdlops_b_idx[I1], 0); return make_tuple(0, waveId_n, xdlops_b_idx[I1], Number<KPack>{} * xdlops_b_idx[I0]);
} }
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i> template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
...@@ -115,12 +120,6 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -115,12 +120,6 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
BK0NK1BlockDesc::IsKnownAtCompileTime(), BK0NK1BlockDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time"); "wrong! Desc should be known at compile-time");
static_assert(AK0MK1BlockDesc{}.GetLength(I0) == BK0NK1BlockDesc{}.GetLength(I0),
"wrong! K0 dimension not consistent");
static_assert(AK0MK1BlockDesc{}.GetLength(I2) == BK0NK1BlockDesc{}.GetLength(I2),
"wrong! K1 dimension not consistent");
static_assert(BlockSize == MWaves * NWaves * WaveSize, static_assert(BlockSize == MWaves * NWaves * WaveSize,
"BlockSize != MWaves * NWaves * WaveSize\n"); "BlockSize != MWaves * NWaves * WaveSize\n");
...@@ -219,32 +218,32 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -219,32 +218,32 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
c_grid_desc_g_m0_n0_m1_n1_m2_n2); c_grid_desc_g_m0_n0_m1_n1_m2_n2);
} }
__host__ __device__ static constexpr auto MakeABlockDescriptor_K0_M0_M1_M2_K1() __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
{ {
return transform_tensor_descriptor( return transform_tensor_descriptor(
AK0MK1BlockDesc{}, AK0MK1BlockDesc{},
make_tuple(make_pass_through_transform(Number<K0>{}), make_tuple(
make_unmerge_transform( make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{})), make_unmerge_transform(
make_pass_through_transform(Number<K1>{})), make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{})); make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
} }
__host__ __device__ static constexpr auto MakeBBlockDescriptor_K0_N0_N1_N2_K1() __host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
{ {
return transform_tensor_descriptor( return transform_tensor_descriptor(
BK0NK1BlockDesc{}, BK0NK1BlockDesc{},
make_tuple(make_pass_through_transform(Number<K0>{}), make_tuple(
make_unmerge_transform( make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{})), make_unmerge_transform(
make_pass_through_transform(Number<K1>{})), make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{})); make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
} }
static constexpr auto a_block_desc_k0_m0_m1_m2_k1 = MakeABlockDescriptor_K0_M0_M1_M2_K1(); static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
static constexpr auto b_block_desc_k0_n0_n1_n2_k1 = MakeBBlockDescriptor_K0_N0_N1_N2_K1(); static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer> template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf, __device__ void Run(const ABlockBuffer& a_block_buf,
...@@ -258,31 +257,31 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -258,31 +257,31 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A // read A
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1, a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(I0, m0, I0, I0, I0), make_tuple(m0, I0, I0, I0),
a_block_buf, a_block_buf,
a_thread_desc_, a_thread_desc_,
make_tuple(I0, I0, I0, I0, I0), make_tuple(I0, I0, I0, I0),
a_thread_buf); a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B // read B
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1, b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(I0, n0, I0, I0, I0), make_tuple(n0, I0, I0, I0),
b_block_buf, b_block_buf,
b_thread_desc_, b_thread_desc_,
make_tuple(I0, I0, I0, I0, I0), make_tuple(I0, I0, I0, I0),
b_thread_buf); b_thread_buf);
static_for<0, K0, xdlops_gemm.K0PerXdlops>{}([&](auto k0) { static_for<0, KPerBlock, KPack * xdlops_gemm.K0PerXdlops>{}([&](auto k) {
vector_type<FloatAB, K1> a_thread_vec; vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, K1> b_thread_vec; vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, K1, 1>{}([&](auto i) { static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
[Number<a_thread_desc_.CalculateOffset(make_tuple(k0, 0, 0, 0, i))>{}]; [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
[Number<b_thread_desc_.CalculateOffset(make_tuple(k0, 0, 0, 0, i))>{}]; [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
}); });
using mfma_input_type = using mfma_input_type =
...@@ -301,13 +300,13 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -301,13 +300,13 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
} }
private: private:
// A[K0, M0, M1, M2, K1] // A[M0, M1, M2, KPerBlock]
static constexpr auto a_thread_desc_ = static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{})); make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerBlock>{}));
// B[K0, N0, N1, N2, K1] // B[N0, N1, N2, KPerBlock]
static constexpr auto b_thread_desc_ = static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{})); make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerBlock>{}));
// C[M, N, NumRegXdlops] // C[M, N, NumRegXdlops]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed( static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
...@@ -315,23 +314,23 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -315,23 +314,23 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB, FloatAB,
decltype(a_block_desc_k0_m0_m1_m2_k1), decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_), decltype(a_thread_desc_),
Sequence<K0, 1, 1, 1, K1>, Sequence<1, 1, 1, KPerBlock>,
Sequence<0, 1, 2, 3, 4>, Sequence<0, 1, 2, 3>,
4, 3,
K1, A_K1,
K1>; A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB, FloatAB,
decltype(b_block_desc_k0_n0_n1_n2_k1), decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_), decltype(b_thread_desc_),
Sequence<K0, 1, 1, 1, K1>, Sequence<1, 1, 1, KPerBlock>,
Sequence<0, 1, 2, 3, 4>, Sequence<0, 1, 2, 3>,
4, 3,
K1, B_K1,
K1>; B_K1>;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()}; AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()}; BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r4r2.hpp 0 → 100644
View file @ d1e9d6cc
#ifndef CK_GRIDWISE_GEMM_XDLOPS_V2R4R2_HPP
#define CK_GRIDWISE_GEMM_XDLOPS_V2R4R2_HPP
#include "common_header.hpp"
#include "multi_index_transform_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_xdlops.hpp"
#include "blockwise_tensor_slice_transfer_v4r1.hpp"
#include "blockwise_tensor_slice_transfer_v6r1.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
namespace ck {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename CBlockClusterAdaptor,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_xdlops_v2r4r2(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_c_grid,
p_shared_block,
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
c_element_op,
c_block_cluster_adaptor);
}
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CMNGridDesc,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t K1Value,
index_t MRepeat,
index_t NRepeat,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_K1,
bool BThreadTransferSrcResetCoordinateAfterRun,
bool BBlockLdsExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
constexpr auto max_lds_align = K1;
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_k0_m_k1_block_desc = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
}
}();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_k0_n_k1_block_desc = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
}
}();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size =
math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size =
math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
constexpr auto c_block_size =
GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
return math::max((a_block_space_size + b_block_space_size) * sizeof(FloatAB),
c_block_size * sizeof(FloatC));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CMNGridDesc& c_m_n_grid_desc,
index_t M01,
index_t N01)
{
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
"wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerXDL)) == 0,
"Invalid tuning param!");
const auto M = a_b_k0_m_k1_grid_desc.GetLength(I2);
const auto N = b_b_k0_n_k1_grid_desc.GetLength(I2);
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
return false;
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false;
// check M01, N01
constexpr auto M1 = Number<MPerBlock>{};
constexpr auto N1 = Number<NPerBlock>{};
const auto M0 = M / M1;
const auto N0 = N / N1;
if(!(M0 % M01 == 0 && N0 % N01 == 0))
return false;
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static constexpr index_t
CalculateGridSize(const CMNGridDesc& c_m_n_grid_desc, index_t KBatch)
{
const auto M = c_m_n_grid_desc.GetLength(I0);
const auto N = c_m_n_grid_desc.GetLength(I1);
const index_t grid_size = (M / MPerBlock) * (N / NPerBlock) * KBatch;
return grid_size;
}
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{
const bool has_main_k0_block_loop = K0 > K0PerBlock;
return has_main_k0_block_loop;
}
__host__ __device__ static constexpr auto
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
{
const auto M = c_m_n_grid_desc.GetLength(I0);
const auto N = c_m_n_grid_desc.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
return transform_tensor_descriptor(
c_m_n_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
const CMNGridDesc& c_m_n_grid_desc, index_t M01, index_t N01, index_t KBatch)
{
const auto M = c_m_n_grid_desc.GetLength(I0);
const auto N = c_m_n_grid_desc.GetLength(I1);
constexpr auto M1 = Number<MPerBlock>{};
constexpr auto N1 = Number<NPerBlock>{};
const auto M0 = M / M1;
const auto N0 = N / N1;
const auto M00 = M0 / M01;
const auto N00 = N0 / N01;
const auto kbatch_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_pass_through_transform(KBatch),
make_unmerge_transform(make_tuple(M00, M01)),
make_unmerge_transform(make_tuple(N00, N01))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
const auto c_blockid_to_kbatch_m00_m01_n00_n01_block_cluster_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(KBatch, M00, N00, M01, N01))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto c_blockid_to_kbatch_m0_n0_block_cluster_adaptor =
chain_tensor_adaptors(kbatch_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
c_blockid_to_kbatch_m00_m01_n00_n01_block_cluster_adaptor);
return c_blockid_to_kbatch_m0_n0_block_cluster_adaptor;
}
__host__ __device__ static constexpr auto
GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
return make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMRepeatPerShuffle * MWaves * MPerXDL>{},
I1,
Number<CShuffleNRepeatPerShuffle * NWaves * NPerXDL>{}));
}
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
template <bool HasMainKBlockLoop>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
FloatAB* __restrict__ p_shared_block,
const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
// divide block work by [M, N]
const auto block_work_idx =
c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
const index_t k_batch_id = block_work_idx[I0];
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = K1;
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_k0_m_k1_block_desc = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
}
}();
constexpr auto a_b_k0_m_k1_block_desc = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
Number<MPerBlock + 1>{} * K1,
K1,
I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
max_lds_align);
}
}();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_k0_n_k1_block_desc = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
}
}();
constexpr auto b_b_k0_n_k1_block_desc = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
Number<NPerBlock + 1>{} * K1,
K1,
I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
max_lds_align);
}
}();
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<1, K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(a_b_k0_m_k1_grid_desc),
decltype(a_b_k0_m_k1_block_desc),
ABlockTransferSrcAccessOrder,
Sequence<0, 2, 1, 3>,
ABlockTransferSrcVectorDim,
3,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true>(
a_b_k0_m_k1_grid_desc,
make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
a_element_op,
a_b_k0_m_k1_block_desc,
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<1, K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(b_b_k0_n_k1_grid_desc),
decltype(b_b_k0_n_k1_block_desc),
BBlockTransferSrcAccessOrder,
Sequence<0, 2, 1, 3>,
BBlockTransferSrcVectorDim,
3,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true>(
b_b_k0_n_k1_grid_desc,
make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
b_element_op,
b_b_k0_n_k1_block_desc,
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[K0PerBlock, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
K1>{};
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size =
math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
FloatAB* p_a_block = p_shared_block;
FloatAB* p_b_block = p_shared_block + a_block_space_size;
constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
// preload data into LDS
{
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
}
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainKBlockLoop)
{
index_t k0_block_data_begin = 0;
do
{
a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
block_sync_lds();
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
block_sync_lds();
a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
k0_block_data_begin += K0PerBlock;
} while(k0_block_data_begin < (K0 - K0PerBlock));
}
// tail
{
block_sync_lds();
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
// output: register to global memory
{
constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatC*>(p_shared_block),
c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
static_assert(M1 == MWaves, "");
static_assert(N1 == NWaves, "");
static_assert(M2 * M3 * M4 == MPerXDL, "");
static_assert(N2 == NPerXDL, "");
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0), // freeze mblock
make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
M1,
M2,
M3,
M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
make_freeze_transform(I0), // freeze nblock
make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
N1,
N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
FloatC,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum_t::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
auto c_block_copy_lds_to_global = BlockwiseTensorSliceTransfer_v6r1<
BlockSize, // index_t BlockSize,
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMRepeatPerShuffle * MWaves * MPerXDL,
1,
CShuffleNRepeatPerShuffle * NWaves * NPerXDL>, // BlockSliceLengths,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
FloatC, // typename SrcData,
FloatC, // typename DstData,
decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun
{c_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
c_element_op};
constexpr auto mxdlperwave_forward_step =
make_multi_index(0, CShuffleMRepeatPerShuffle * MWaves * MPerXDL, 0, 0);
constexpr auto nxdlperwave_forward_step =
make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWaves * NPerXDL);
constexpr auto nxdlperwave_backward_step =
make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWaves * NPerXDL);
static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
constexpr auto mxdlperwave = mxdlperwave_iter;
static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
constexpr bool nxdlperwave_forward_sweep =
(mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
constexpr index_t nxdlperwave_value =
nxdlperwave_forward_sweep
? nxdlperwave_iter
: (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
// make sure it's safe to do ds_write
block_sync_lds();
// VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(
c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_block_buf);
// make sure it's safe to do ds_read
block_sync_lds();
// LDS to global
c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
c_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
// move on nxdlperwave dimension
if constexpr(nxdlperwave_forward_sweep &&
(nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
{
c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock,
nxdlperwave_forward_step);
}
else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
{
c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock,
nxdlperwave_backward_step);
}
});
// move on mxdlperwave dimension
if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
{
c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
}
});
}
}
}; // namespace ck
} // namespace ck
#endif
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v3r1.hpp
View file @ d1e9d6cc
...@@ -16,8 +16,8 @@ namespace ck { ...@@ -16,8 +16,8 @@ namespace ck {
template <typename GridwiseGemm, template <typename GridwiseGemm,
typename FloatAB, typename FloatAB,
typename FloatC, typename FloatC,
typename AGridDesc_K0_M_K1, typename AGridDesc_AK0_M_AK1,
typename BGridDesc_K0_N_K1, typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl, typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
...@@ -32,8 +32,8 @@ __global__ void ...@@ -32,8 +32,8 @@ __global__ void
const FloatAB* __restrict__ p_a_grid, const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid, FloatC* __restrict__ p_c_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1, const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1, const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl, c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
const AElementwiseOperation a_element_op, const AElementwiseOperation a_element_op,
...@@ -48,8 +48,8 @@ __global__ void ...@@ -48,8 +48,8 @@ __global__ void
p_b_grid, p_b_grid,
p_c_grid, p_c_grid,
p_shared, p_shared,
a_grid_desc_k0_m_k1, a_grid_desc_ak0_m_ak1,
b_grid_desc_k0_n_k1, b_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl, c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
a_element_op, a_element_op,
b_element_op, b_element_op,
...@@ -63,21 +63,22 @@ template < ...@@ -63,21 +63,22 @@ template <
typename FloatAcc, typename FloatAcc,
typename FloatC, typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation, InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
typename AGridDesc_K0_M_K1, typename AGridDesc_AK0_M_AK1,
typename BGridDesc_K0_N_K1, typename BGridDesc_BK0_N_BK1,
typename CGridDesc_M_N, typename CGridDesc_M_N,
typename AElementwiseOperation, typename AElementwiseOperation,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
index_t MPerBlock, index_t MPerBlock,
index_t NPerBlock, index_t NPerBlock,
index_t K0PerBlock, index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl, index_t MPerXdl,
index_t NPerXdl, index_t NPerXdl,
index_t K1Value,
index_t MXdlPerWave, index_t MXdlPerWave,
index_t NXdlPerWave, index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder, typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcVectorDim,
...@@ -85,7 +86,7 @@ template < ...@@ -85,7 +86,7 @@ template <
index_t ABlockTransferDstScalarPerVector_K1, index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun, bool AThreadTransferSrcResetCoordinateAfterRun,
bool ABlockLdsExtraM, bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1, typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder, typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcVectorDim,
...@@ -110,50 +111,53 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -110,50 +111,53 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
static constexpr auto I7 = Number<7>{}; static constexpr auto I7 = Number<7>{};
// K1 should be Number<...> // K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{}; static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1() __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{ {
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = AK1;
// A matrix in LDS memory, dst of blockwise copy // A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_k0_m_k1 = [&]() { constexpr auto a_block_desc_ak0_m_ak1 = [&]() {
if constexpr(ABlockLdsExtraM) if constexpr(ABlockLdsExtraM)
{ {
return make_naive_tensor_descriptor( return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), make_tuple(AK0, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1)); make_tuple(Number<MPerBlock + 1>{} * AK1, AK1, I1));
} }
else else
{ {
return make_naive_tensor_descriptor_aligned( return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align); make_tuple(AK0, Number<MPerBlock>{}, AK1), max_lds_align);
} }
}(); }();
return a_block_desc_k0_m_k1; return a_block_desc_ak0_m_ak1;
} }
__host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1() __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{ {
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = BK1;
// B matrix in LDS memory, dst of blockwise copy // B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_k0_n_k1 = [&]() { constexpr auto b_block_desc_bk0_n_bk1 = [&]() {
if constexpr(BBlockLdsExtraN) if constexpr(BBlockLdsExtraN)
{ {
return make_naive_tensor_descriptor( return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), make_tuple(BK0, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1)); make_tuple(Number<NPerBlock + 1>{} * BK1, BK1, I1));
} }
else else
{ {
return make_naive_tensor_descriptor_aligned( return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align); make_tuple(BK0, Number<NPerBlock>{}, BK1), max_lds_align);
} }
}(); }();
return b_block_desc_k0_n_k1; return b_block_desc_bk0_n_bk1;
} }
__host__ __device__ static constexpr auto __host__ __device__ static constexpr auto
...@@ -178,17 +182,15 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -178,17 +182,15 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte() __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{ {
// LDS allocation for A and B: be careful of alignment // LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1(); constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1(); constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
constexpr auto max_lds_align = K1;
constexpr auto a_block_space_size_aligned = constexpr auto a_block_space_size_aligned =
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align); math::integer_least_multiple(a_block_desc_ak0_m_ak1.GetElementSpaceSize(), AK1);
constexpr auto b_block_space_size_aligned = constexpr auto b_block_space_size_aligned =
math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align); math::integer_least_multiple(b_block_desc_bk0_n_bk1.GetElementSpaceSize(), BK1);
// LDS allocation for C shuffle in LDS // LDS allocation for C shuffle in LDS
constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl = constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
...@@ -205,29 +207,28 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -205,29 +207,28 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01} // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
__host__ __device__ static constexpr bool __host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1, CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1, const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const CGridDesc_M_N& c_grid_desc_m_n, const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01, index_t M01,
index_t N01) index_t N01)
{ {
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value, // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
"wrong! K1 need to be known at compile-time"); // is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
// "wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) && static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0, (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!"); "Invalid tuning param!");
const auto M = a_grid_desc_k0_m_k1.GetLength(I1); const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
const auto N = b_grid_desc_k0_n_k1.GetLength(I1); const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0); const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) && if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
return false; return false;
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0)) if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
return false; return false;
// check NumPrefetch // check NumPrefetch
...@@ -239,7 +240,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -239,7 +240,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
{ {
// 2-stage prefetch currently only support even number of K0 loop // 2-stage prefetch currently only support even number of K0 loop
// TODO: add support for odd number of K0 loop // TODO: add support for odd number of K0 loop
if(!((K0 / K0PerBlock) % 2 == 0)) if(!((K / KPerBlock) % 2 == 0))
{ {
return false; return false;
} }
...@@ -277,7 +278,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -277,7 +278,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// TODO move this function into GEMM-pipeline class // TODO move this function into GEMM-pipeline class
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0) __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{ {
const bool has_main_k0_block_loop = (K0 / (NumPrefetch * K0PerBlock)) > 1; const bool has_main_k0_block_loop = ((K0 * AK1) / (NumPrefetch * KPerBlock)) > 1;
return has_main_k0_block_loop; return has_main_k0_block_loop;
} }
...@@ -357,8 +358,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -357,8 +358,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid, FloatC* __restrict__ p_c_grid,
void* __restrict__ p_shared, void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1, const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1, const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl& const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl, c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
const AElementwiseOperation& a_element_op, const AElementwiseOperation& a_element_op,
...@@ -367,16 +368,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -367,16 +368,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
const Block2CTileMap& block_2_ctile_map) const Block2CTileMap& block_2_ctile_map)
{ {
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>( const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize()); p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>( const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize()); p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>( auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_c_grid, p_c_grid,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize()); .GetElementSpaceSize());
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
// divide block work by [M, N] // divide block work by [M, N]
const auto block_work_idx = const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id())); block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
...@@ -389,13 +388,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -389,13 +388,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock); __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
// lds max alignment // lds max alignment
constexpr auto max_lds_align = K1; constexpr auto max_lds_align = math::lcm(AK1, BK1);
// A matrix in LDS memory, dst of blockwise copy // A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1(); constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy // B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1(); constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy // A matrix blockwise copy
auto a_blockwise_copy = auto a_blockwise_copy =
...@@ -403,13 +402,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -403,13 +402,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
AElementwiseOperation, AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set, InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, MPerBlock, K1>, Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder, ABlockTransferThreadClusterArrangeOrder,
FloatAB, FloatAB,
FloatAB, FloatAB,
decltype(a_grid_desc_k0_m_k1), decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_k0_m_k1), decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder, ABlockTransferSrcAccessOrder,
Sequence<1, 0, 2>, Sequence<1, 0, 2>,
ABlockTransferSrcVectorDim, ABlockTransferSrcVectorDim,
...@@ -421,10 +420,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -421,10 +420,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
AThreadTransferSrcResetCoordinateAfterRun, AThreadTransferSrcResetCoordinateAfterRun,
true, true,
NumPrefetch>( NumPrefetch>(
a_grid_desc_k0_m_k1, a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0), make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op, a_element_op,
a_block_desc_k0_m_k1, a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0), make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{}); ck::tensor_operation::element_wise::PassThrough{});
...@@ -434,13 +433,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -434,13 +433,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
BElementwiseOperation, BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set, InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, NPerBlock, K1>, Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_K0_N_K1, BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder, BBlockTransferThreadClusterArrangeOrder,
FloatAB, FloatAB,
FloatAB, FloatAB,
decltype(b_grid_desc_k0_n_k1), decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_k0_n_k1), decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder, BBlockTransferSrcAccessOrder,
Sequence<1, 0, 2>, Sequence<1, 0, 2>,
BBlockTransferSrcVectorDim, BBlockTransferSrcVectorDim,
...@@ -452,10 +451,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -452,10 +451,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
BThreadTransferSrcResetCoordinateAfterRun, BThreadTransferSrcResetCoordinateAfterRun,
true, true,
NumPrefetch>( NumPrefetch>(
b_grid_desc_k0_n_k1, b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0), make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op, b_element_op,
b_block_desc_k0_n_k1, b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0), make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{}); ck::tensor_operation::element_wise::PassThrough{});
...@@ -466,45 +465,47 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -466,45 +465,47 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in // c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register // register
// sanity check // sanity check
constexpr index_t k_pack = math::max(
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize, BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB, FloatAB,
FloatAcc, FloatAcc,
decltype(a_block_desc_k0_m_k1), decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_k0_n_k1), decltype(b_block_desc_bk0_n_bk1),
MPerXdl, MPerXdl,
NPerXdl, NPerXdl,
MXdlPerWave, MXdlPerWave,
NXdlPerWave, NXdlPerWave,
K1>{}; k_pack>{};
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer(); auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment // LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align); a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>( auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize()); static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>( auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned, static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_k0_n_k1.GetElementSpaceSize()); b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0); constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0); constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
// gridwise GEMM pipeline // gridwise GEMM pipeline
const auto gridwise_gemm_pipeline = const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_k0_m_k1)>, GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_ak0_m_ak1)>,
remove_cvref_t<decltype(a_block_desc_k0_m_k1)>, remove_cvref_t<decltype(a_block_desc_ak0_m_ak1)>,
remove_cvref_t<decltype(a_blockwise_copy)>, remove_cvref_t<decltype(a_blockwise_copy)>,
remove_cvref_t<decltype(a_grid_buf)>, remove_cvref_t<decltype(a_grid_buf)>,
remove_cvref_t<decltype(a_block_buf)>, remove_cvref_t<decltype(a_block_buf)>,
remove_cvref_t<decltype(a_block_slice_copy_step)>, remove_cvref_t<decltype(a_block_slice_copy_step)>,
remove_cvref_t<decltype(b_grid_desc_k0_n_k1)>, remove_cvref_t<decltype(b_grid_desc_bk0_n_bk1)>,
remove_cvref_t<decltype(b_block_desc_k0_n_k1)>, remove_cvref_t<decltype(b_block_desc_bk0_n_bk1)>,
remove_cvref_t<decltype(b_blockwise_copy)>, remove_cvref_t<decltype(b_blockwise_copy)>,
remove_cvref_t<decltype(b_grid_buf)>, remove_cvref_t<decltype(b_grid_buf)>,
remove_cvref_t<decltype(b_block_buf)>, remove_cvref_t<decltype(b_block_buf)>,
...@@ -514,23 +515,25 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1 ...@@ -514,23 +515,25 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
NumPrefetch, NumPrefetch,
HasMainK0BlockLoop>{}; HasMainK0BlockLoop>{};
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock); const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
gridwise_gemm_pipeline.Run(a_grid_desc_k0_m_k1, gridwise_gemm_pipeline.Run(a_grid_desc_ak0_m_ak1,
a_block_desc_k0_m_k1, a_block_desc_ak0_m_ak1,
a_blockwise_copy, a_blockwise_copy,
a_grid_buf, a_grid_buf,
a_block_buf, a_block_buf,
a_block_slice_copy_step, a_block_slice_copy_step,
b_grid_desc_k0_n_k1, b_grid_desc_bk0_n_bk1,
b_block_desc_k0_n_k1, b_block_desc_bk0_n_bk1,
b_blockwise_copy, b_blockwise_copy,
b_grid_buf, b_grid_buf,
b_block_buf, b_block_buf,
b_block_slice_copy_step, b_block_slice_copy_step,
blockwise_gemm, blockwise_gemm,
c_thread_buf, c_thread_buf,
K0BlockMainLoop); num_k_block_main_loop);
// shuffle C and write out // shuffle C and write out
{ {
......
composable_kernel/include/utility/common_header.hpp
View file @ d1e9d6cc
...@@ -35,6 +35,7 @@ ...@@ -35,6 +35,7 @@
#include "transpose_vectors.hpp" #include "transpose_vectors.hpp"
#include "inner_product.hpp" #include "inner_product.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "debug.hpp"
// TODO: remove this // TODO: remove this
#if CK_USE_AMD_INLINE_ASM #if CK_USE_AMD_INLINE_ASM
......
composable_kernel/include/utility/debug.hpp 0 → 100644
View file @ d1e9d6cc
#ifndef UTILITY_DEBUG_HPP
#define UTILITY_DEBUG_HPP
namespace ck {
namespace debug {
namespace detail {
template <typename T, typename Enable = void>
struct PrintAsType;
template <typename T>
struct PrintAsType<T, typename std::enable_if<std::is_floating_point<T>::value>::value>
{
using type = float;
};
template <>
struct PrintAsType<ck::half_t, void>
{
using type = float;
};
template <typename T>
struct PrintAsType<T, typename std::enable_if<std::is_integral<T>::value>::value>
{
using type = int;
};
} // namespace detail
// Print at runtime the data in shared memory in 128 bytes per row format given shared mem pointer
// and the number of elements. Can optionally specify strides between elements and how many bytes'
// worth of data per row.
//
// Usage example:
//
// debug::print_shared(a_block_buf.p_data_, index_t(a_block_desc_k0_m_k1.GetElementSpaceSize()));
//
template <typename T, index_t element_stride = 1, index_t row_bytes = 128>
__device__ void print_shared(T const* p_shared, index_t num_elements)
{
using PrintType = typename detail::PrintAsType<T>::type;
constexpr index_t row_elements = row_bytes / sizeof(T);
static_assert((element_stride >= 1 && element_stride <= row_elements),
"element_stride should between [1, row_elements]");
index_t wgid = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
index_t tid =
(threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x) + threadIdx.x;
__syncthreads();
if(tid == 0)
{
printf("\nWorkgroup id %d, bytes per row %d, element stride %d\n\n",
wgid,
row_bytes,
element_stride);
for(index_t i = 0; i < num_elements; i += row_elements)
{
printf("elem %5d: ", i);
for(index_t j = 0; j < row_elements; j += element_stride)
{
printf("%.0f ", static_cast<PrintType>(p_shared[i + j]));
}
printf("\n");
}
printf("\n");
}
__syncthreads();
}
} // namespace debug
} // namespace ck
#endif // UTILITY_DEBUG_HPP
device_operation/CMakeLists.txt
View file @ d1e9d6cc
...@@ -31,7 +31,11 @@ set(DEVICE_GEMM_INSTANCE_SOURCE ...@@ -31,7 +31,11 @@ set(DEVICE_GEMM_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp;
) ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp;
)
# device_gemm_bias_2d_instance # device_gemm_bias_2d_instance
set(DEVICE_GEMM_BIAS_2D_INSTANCE_SOURCE set(DEVICE_GEMM_BIAS_2D_INSTANCE_SOURCE
......
device_operation/include/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ d1e9d6cc
...@@ -432,10 +432,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W ...@@ -432,10 +432,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
OutElementwiseOperation, OutElementwiseOperation,
MPerBlock, MPerBlock,
NPerBlock, NPerBlock,
K0PerBlock, K0PerBlock * K1,
K1, // AK1
K1, // BK1
MPerXdl, MPerXdl,
NPerXdl, NPerXdl,
K1,
MXdlPerWave, MXdlPerWave,
NXdlPerWave, NXdlPerWave,
ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterLengths_K0_M_K1,
......
device_operation/include/device_gemm_xdl_c_shuffle.hpp
View file @ d1e9d6cc
...@@ -29,8 +29,9 @@ template < ...@@ -29,8 +29,9 @@ template <
ck::index_t BlockSize, ck::index_t BlockSize,
ck::index_t MPerBlock, ck::index_t MPerBlock,
ck::index_t NPerBlock, ck::index_t NPerBlock,
ck::index_t K0PerBlock, ck::index_t KPerBlock,
ck::index_t K1, ck::index_t AK1,
ck::index_t BK1,
ck::index_t MPerXDL, ck::index_t MPerXDL,
ck::index_t NPerXDL, ck::index_t NPerXDL,
ck::index_t MXdlPerWave, ck::index_t MXdlPerWave,
...@@ -61,13 +62,11 @@ struct DeviceGemmXdl_C_Shuffle ...@@ -61,13 +62,11 @@ struct DeviceGemmXdl_C_Shuffle
static constexpr auto I1 = Number<1>{}; static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{}; static constexpr auto I2 = Number<2>{};
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA) static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{ {
assert(K % K1 == 0); assert(K % AK1 == 0);
const index_t K0 = K / K1; const index_t K0 = K / AK1;
const auto a_grid_desc_m_k = [&]() { const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value) if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
...@@ -80,21 +79,20 @@ struct DeviceGemmXdl_C_Shuffle ...@@ -80,21 +79,20 @@ struct DeviceGemmXdl_C_Shuffle
} }
}(); }();
const auto a_grid_desc_k0_m_k1 = const auto a_grid_desc_k0_m_k1 = transform_tensor_descriptor(
transform_tensor_descriptor(a_grid_desc_m_k, a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), make_tuple(make_unmerge_transform(make_tuple(K0, AK1)), make_pass_through_transform(M)),
make_pass_through_transform(M)), make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}), make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_k0_m_k1; return a_grid_desc_k0_m_k1;
} }
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB) static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
{ {
assert(K % K1 == 0); assert(K % BK1 == 0);
const index_t K0 = K / K1; const index_t K0 = K / BK1;
const auto b_grid_desc_k_n = [&]() { const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value) if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
...@@ -107,12 +105,11 @@ struct DeviceGemmXdl_C_Shuffle ...@@ -107,12 +105,11 @@ struct DeviceGemmXdl_C_Shuffle
} }
}(); }();
const auto b_grid_desc_k0_n_k1 = const auto b_grid_desc_k0_n_k1 = transform_tensor_descriptor(
transform_tensor_descriptor(b_grid_desc_k_n, b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)), make_tuple(make_unmerge_transform(make_tuple(K0, BK1)), make_pass_through_transform(N)),
make_pass_through_transform(N)), make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}), make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_k0_n_k1; return b_grid_desc_k0_n_k1;
} }
...@@ -148,10 +145,11 @@ struct DeviceGemmXdl_C_Shuffle ...@@ -148,10 +145,11 @@ struct DeviceGemmXdl_C_Shuffle
CElementwiseOperation, CElementwiseOperation,
MPerBlock, MPerBlock,
NPerBlock, NPerBlock,
K0PerBlock, KPerBlock,
AK1,
BK1,
MPerXDL, MPerXDL,
NPerXDL, NPerXDL,
K1,
MXdlPerWave, MXdlPerWave,
NXdlPerWave, NXdlPerWave,
ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterLengths_K0_M_K1,
...@@ -461,7 +459,9 @@ struct DeviceGemmXdl_C_Shuffle ...@@ -461,7 +459,9 @@ struct DeviceGemmXdl_C_Shuffle
<< BlockSize << ", " << BlockSize << ", "
<< MPerBlock << ", " << MPerBlock << ", "
<< NPerBlock << ", " << NPerBlock << ", "
<< K0PerBlock << KPerBlock << ", "
<< AK1 << ", "
<< BK1
<< ">"; << ">";
// clang-format on // clang-format on
......
device_operation/include/device_gemm_xdl_splitk_c_shuffle.hpp 0 → 100644
View file @ d1e9d6cc
#ifndef DEVICE_GEMM_XDL_SPLITK_C_SHUFFLE_HPP
#define DEVICE_GEMM_XDL_SPLITK_C_SHUFFLE_HPP
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r4r2.hpp"
#include "gemm_specialization.hpp"
#ifndef CK_RUN_KERNEL_AND_TIME
#define CK_RUN_KERNEL_AND_TIME 1
#endif
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization_t GemmSpecialization,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t MXdlPerWave,
ck::index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool ABlockLdsAddExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CBlockTransferScalarPerVector_NWaveNPerXDL>
struct DeviceGemmXdlSplitKCShuffle
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto K1Number = Number<K1>{};
static auto
MakeAGridDescriptor_KBatch_K0_M_K1(index_t M, index_t K, index_t StrideA, int KBatch, int KPad)
{
assert(KPad % (K1 * KBatch) == 0);
const index_t K0 = KPad / (K1 * KBatch);
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(M)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
static auto
MakeBGridDescriptor_KBatch_K0_N_K1(index_t K, index_t N, index_t StrideB, int KBatch, int KPad)
{
assert(KPad % (K1 * KBatch) == 0);
const index_t K0 = KPad / (K1 * KBatch);
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
static auto GetKPad(index_t K, index_t KBatch)
{
const index_t K0 = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
const index_t KPad = KBatch * K0 * K1;
return KPad;
}
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_KBatch_K0_M_K1(1, 1, 1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_KBatch_K0_N_K1(1, 1, 1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CDataType,
InMemoryDataOperationEnum_t::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
MPerBlock,
NPerBlock,
K0PerBlock,
MPerXDL,
NPerXDL,
K1,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
// GridwiseGemm
using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CDataType,
InMemoryDataOperationEnum_t::AtomicAdd,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
MPerBlock,
NPerBlock,
K0PerBlock,
MPerXDL,
NPerXDL,
K1,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
using Block2CTileMap =
decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t k_batch)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
a_grid_desc_kbatch_k0_m_k1_{},
b_grid_desc_kbatch_k0_n_k1_{},
c_grid_desc_m_n_{},
c_grid_desc_mblock_mperblock_nblock_nperblock_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op},
k_batch_{k_batch}
{
int KPad = DeviceGemmXdlSplitKCShuffle::GetKPad(K, k_batch_);
a_grid_desc_kbatch_k0_m_k1_ =
DeviceGemmXdlSplitKCShuffle::MakeAGridDescriptor_KBatch_K0_M_K1(
M, K, StrideA, k_batch_, KPad);
b_grid_desc_kbatch_k0_n_k1_ =
DeviceGemmXdlSplitKCShuffle::MakeBGridDescriptor_KBatch_K0_N_K1(
K, N, StrideB, k_batch_, KPad);
c_grid_desc_m_n_ = DeviceGemmXdlSplitKCShuffle::MakeCGridDescriptor_M_N(M, N, StrideC);
if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
b_grid_desc_kbatch_k0_n_k1_,
c_grid_desc_m_n_,
M01_,
N01_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
index_t k_batch_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceGemmXdlSplitKCShuffle::Argument;
void ShowInfo(const Argument& arg)
{
std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
<< arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
<< arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
float Run(const Argument& arg, int nrepeat = 1)
{
const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_, kbatch);
const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
float ave_time = 0;
const auto Run = [&](const auto& kernel) {
if(nrepeat > 0)
{
ShowInfo(arg);
ave_time =
launch_and_time_kernel(kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
if(kbatch > 1 || nrepeat <= 0)
{
hipGetErrorString(hipMemset(
arg.p_c_grid_,
0,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
sizeof(CDataType)));
launch_kernel(kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
};
if(has_main_k0_block_loop)
{
if(kbatch == 1)
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
true>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
GridwiseGemmAtomicAdd,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
true>;
Run(kernel);
}
}
else
{
if(kbatch == 1)
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
false>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r4r2<
GridwiseGemmAtomicAdd,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::
CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdlSplitKCShuffle::Block2CTileMap>,
false>;
Run(kernel);
}
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
arg.b_grid_desc_kbatch_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t KBatch)
{
return Argument{p_a,
p_b,
p_c,
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op,
KBatch};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op,
KBatch);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGemmXdlSplitKCShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
device_operation/src/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ d1e9d6cc
...@@ -23,23 +23,23 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough; ...@@ -23,23 +23,23 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n] // Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple< using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off // clang-format off
//#####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num| //#####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Prefetch| //#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Prefetch|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | //#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2> DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8, 2>
// clang-format on // clang-format on
>; >;
......
device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
View file @ d1e9d6cc
...@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>; ...@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n] // Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
std::tuple< // clang-format off
// clang-format off //#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| //#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8> DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
// clang-format on DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
>; DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances) std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
......
device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
View file @ d1e9d6cc
...@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>; ...@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n] // Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
std::tuple< // clang-format off
// clang-format off //#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| //#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8> DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
// clang-format on DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
>; DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances) std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
......
device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
View file @ d1e9d6cc
...@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>; ...@@ -21,23 +21,30 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n] // Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
std::tuple< // clang-format off
// clang-format off //#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| //#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8> DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
// clang-format on DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
>; DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances) std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
......
device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ d1e9d6cc
...@@ -21,28 +21,27 @@ using S = ck::Sequence<Is...>; ...@@ -21,28 +21,27 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n] // Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
std::tuple< // clang-format off
// clang-format off //#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| //#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>, DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
DeviceGemmXdl_C_Shuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8> // clang-format on
// clang-format on >;
>;
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances) std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
......
device_operation/src/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
View file @ d1e9d6cc
...@@ -20,7 +20,8 @@ using S = ck::Sequence<Is...>; ...@@ -20,7 +20,8 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n] // Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances = using device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances =
...@@ -54,8 +55,8 @@ using device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = ...@@ -54,8 +55,8 @@ using device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances =
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| //###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| //###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1>, DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1> DeviceGemmXdl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1>
// clang-format on // clang-format on
>; >;
......
device_operation/src/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp 0 → 100644
View file @ d1e9d6cc
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl_splitk_c_shuffle.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment