Skip to content

GitLab

Commit 930b2872 authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

best performing kernel for GEMV codex problem with M=1 with inverted B matrix

parents a1e17d18 a4f72a31
Hide whitespace changes
Inline Side-by-side
Showing with 182 additions and 138 deletions
example/01_gemm/gemm_xdl_f8.cpp → example/01_gemm/gemm_xdl_fp8.cpp
View file @ 930b2872
......@@ -7,9 +7,9 @@
using ADataType = ck::f8_t;
using BDataType = ck::f8_t;
using CDataType = ck::f8_t;
using CDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::f8_t;
using CShuffleDataType = float;
using ALayout = Row;
using BLayout = Col;
......@@ -27,7 +27,7 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| 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_MWaveMPerXdl| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 16>;
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
......
example/01_gemm/gemm_xdl_fp8_bf8.cpp 0 → 100755
View file @ 930b2872
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::bf8_t;
using CDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = float;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto LoopSched = ck::make_default_loop_scheduler();
static constexpr auto PipelineVer = ck::PipelineVersion::v1;
using ComputeTypeA = ck::f8_t;
using ComputeTypeB = ck::bf8_t;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| 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|
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| 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_MWaveMPerXdl| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, LoopSched, PipelineVer, ComputeTypeA, ComputeTypeB>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp,
ComputeTypeA,
ComputeTypeB>;
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/02_gemm_bilinear/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102)
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
......@@ -19,4 +18,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()
example/03_gemm_bias_relu/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
......@@ -7,4 +6,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View file @ 930b2872
......@@ -3,24 +3,24 @@ set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_gemm_add_add_fastgelu_xdl)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
endif()
set(target 1)
......
example/09_convnd_fwd/CMakeLists.txt
View file @ 930b2872
......@@ -2,34 +2,16 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
endif()
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
endif()
set(target 1)
endif()
endforeach()
if(DL_KERNELS)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
endif()
endif()
add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View file @ 930b2872
......@@ -2,27 +2,27 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_convnd_fwd_reduce_xdl)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
add_custom_target(example_convnd_fwd_reduce_xdl)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
endif()
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
endif()
add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
endif()
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
endif()
if(USE_BITINT_EXTENSION_INT4)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
set(target 1)
endif(USE_BITINT_EXTENSION_INT4)
set(target 1)
endif()
endforeach()
\ No newline at end of file
example/13_pool2d_fwd/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
endif()
add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
example/14_gemm_quantization/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
# dlops
if(DL_KERNELS)
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
endif()
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
# xdlops
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
......@@ -14,4 +10,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()
\ No newline at end of file
example/15_grouped_gemm/CMakeLists.txt
View file @ 930b2872
add_custom_target(example_grouped_gemm_xdl)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
add_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp16
example_grouped_gemm_multiple_d_dl_fp16
example_grouped_gemm_xdl_splitk_fp16
example_grouped_gemm_xdl_fixed_nk_fp16
example_grouped_gemm_xdl_fixed_nk_bias_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bfp16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp16)
endif()
add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_dl_fp16)
endif()
add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_splitk_fp16)
endif()
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp16)
endif()
add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_bias_fp16)
endif()
add_example_executable(example_grouped_gemm_xdl_bf16 grouped_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bf16)
endif()
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
endif()
if(DTYPES MATCHES "f8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
if(result EQUAL 0)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
endif()
endif()
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
View file @ 930b2872
......@@ -296,13 +296,11 @@ int main(int argc, char* argv[])
problem_size.group_count = 16;
problem_size.Ms = {
167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ns.push_back(768);
problem_size.Ks.push_back(4608);
problem_size.Ms.push_back(256 + 256 * i);
problem_size.Ns.push_back(128 + 128 * i);
problem_size.Ks.push_back(128 + 64 * i);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
......
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
View file @ 930b2872
......@@ -297,13 +297,11 @@ int main(int argc, char* argv[])
problem_size.group_count = 16;
problem_size.Ms = {
167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ns.push_back(768);
problem_size.Ks.push_back(4608);
problem_size.Ms.push_back(256 + 256 * i);
problem_size.Ns.push_back(128 + 128 * i);
problem_size.Ks.push_back(128 + 64 * i);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
......
example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
View file @ 930b2872
......@@ -66,13 +66,11 @@ int main(int argc, char* argv[])
problem_size.group_count = 16;
problem_size.Ms = {
167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ns.push_back(768);
problem_size.Ks.push_back(4608);
problem_size.Ms.push_back(256 + 256 * i);
problem_size.Ns.push_back(128 + 128 * i);
problem_size.Ks.push_back(128 + 64 * i);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
......
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ 930b2872
......@@ -6,30 +6,43 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
endif()
add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
endif()
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
endif()
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
endif()
add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
endif()
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
endif()
......@@ -40,7 +53,9 @@ foreach(gpu IN LISTS GPU_TARGETS)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
endif()
set(target 1)
endif()
......
example/17_convnd_bwd_data/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
if(result EQUAL 0)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
endif()
set(target 1)
endif()
endforeach()
if(DL_KERNELS)
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
endif()
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
endif()
example/18_batched_gemm_reduce/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
......@@ -7,4 +6,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 930b2872
......@@ -3,22 +3,26 @@ set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_weight)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
endif()
if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8 grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8)
endif()
endif()
set(target 1)
endif()
endforeach()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if(DL_KERNELS)
add_custom_target(example_grouped_conv_bwd_weight_dl)
add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
endif()
endif()
\ No newline at end of file
add_custom_target(example_grouped_conv_bwd_weight_dl)
add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
endif()
example/20_grouped_conv_bwd_weight/common.hpp
View file @ 930b2872
......@@ -23,6 +23,12 @@
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#ifdef CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......
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