Skip to content

GitLab

Unverified Commit 9f8ab221 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into add_int8_wmma_example_instance

parents 755ace59 b4fc4d0b
Hide whitespace changes
Inline Side-by-side
Showing with 449 additions and 198 deletions
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ 9f8ab221
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_gemm_reduce_xdl) add_custom_target(example_gemm_reduce_xdl)
add_custom_target(example_gemm_reduce_xdl_max) add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare) add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
add_custom_target(example_gemm_add_add_mean_meansquare_xdl) 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_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_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16) add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16) add_example_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
endif() add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES) add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
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)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8) add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
endif() add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES) add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
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)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32) add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
endif() add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES) add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
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)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16) add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
endif() add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
add_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_mean_meansquare add_example_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_max example_gemm_reduce_xdl_mean_meansquare
example_gemm_add_add_mean_meansquare_xdl) example_gemm_reduce_xdl_max
example_gemm_add_add_mean_meansquare_xdl)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp) if(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4) add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
endif() add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
set(target 1) endif()
endif() set(target 1)
endif()
endforeach() endforeach()
example/17_convnd_bwd_data/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) 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) 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) set(target 1)
endif() endif()
endforeach() endforeach()
if(DL_KERNELS)
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp) 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) if(result EQUAL 0)
endif() target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
endif() endif()
example/18_batched_gemm_reduce/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
...@@ -7,4 +6,3 @@ foreach(gpu IN LISTS GPU_TARGETS) ...@@ -7,4 +6,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1) set(target 1)
endif() endif()
endforeach() endforeach()
endif()
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 9f8ab221
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0) if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_weight) 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) add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
endif() add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES) add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16) if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
endif() add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8 grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp)
set(target 1) add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8)
endif() endif()
set(target 1)
endif()
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_wmma_fp16 grouped_conv_bwd_weight_wmma_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_wmma_fp16)
set(target 1)
endif()
endforeach() endforeach()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) add_custom_target(example_grouped_conv_bwd_weight_dl)
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_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp) add_example_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
endif()
endif()
\ No newline at end of file
example/20_grouped_conv_bwd_weight/common.hpp
View file @ 9f8ab221
...@@ -23,6 +23,12 @@ ...@@ -23,6 +23,12 @@
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; 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> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -40,25 +46,21 @@ struct CommonLayoutSetting ...@@ -40,25 +46,21 @@ struct CommonLayoutSetting
using OutputLayout = OutputLay; using OutputLayout = OutputLay;
}; };
template <ck::index_t NDimSpatial>
struct CommonLayoutSettingSelector;
namespace ctl = ck::tensor_layout::convolution; namespace ctl = ck::tensor_layout::convolution;
template <ck::index_t NDimSpatial>
template <> struct CommonLayoutSettingSelector
struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK> : CommonLayoutSetting<ck::tuple_element_t<NDimSpatial - 1,
{ ck::Tuple<ck::tensor_layout::convolution::GNWC,
}; ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
template <> ck::tuple_element_t<NDimSpatial - 1,
struct CommonLayoutSettingSelector<2> final ck::Tuple<ck::tensor_layout::convolution::GKXC,
: CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK> ck::tensor_layout::convolution::GKYXC,
{ ck::tensor_layout::convolution::GKZYXC>>,
}; ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
template <> ck::tensor_layout::convolution::GNHWK,
struct CommonLayoutSettingSelector<3> final ck::tensor_layout::convolution::GNDHWK>>>
: CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
{ {
}; };
...@@ -78,10 +80,10 @@ struct ExecutionConfig final ...@@ -78,10 +80,10 @@ struct ExecutionConfig final
bool time_kernel = false; bool time_kernel = false;
}; };
#define DefaultConvParam \ #define DefaultConvParam \
ck::utils::conv::ConvParam \ ck::utils::conv::ConvParam \
{ \ { \
2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \ 3, 4, 1, 128, 256, {3, 3, 3}, {14, 14, 14}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, { 1, 1, 1 } \
} }
inline void print_help_msg() inline void print_help_msg()
......
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
View file @ 9f8ab221
...@@ -65,6 +65,34 @@ using DeviceConvBwdWeightInstance = ck::tensor_operation::device::DeviceGroupedC ...@@ -65,6 +65,34 @@ using DeviceConvBwdWeightInstance = ck::tensor_operation::device::DeviceGroupedC
5, // CThreadTransferSrcDstVectorDim 5, // CThreadTransferSrcDstVectorDim
4>; // CThreadTransferDstScalarPerVector 4>; // CThreadTransferDstScalarPerVector
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
#include "run_grouped_conv_bwd_weight_example.inc" #include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); } int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp 0 → 100644
View file @ 9f8ab221
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Wmma_CShuffle<
NDimSpatial,
ck::tensor_layout::convolution::GNDHWC,
ck::tensor_layout::convolution::GKZYXC,
ck::tensor_layout::convolution::GNDHWK,
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
16, // MPerWMMA
16, // NPerWMMA
4, // MRepeat
2, // NRepeat
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<0, 2, 1>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1>, // ABlockTransferSrcAccessOrder
1, // ABlockTransferSrcVectorDim
1, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
true, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<0, 2, 1>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1>, // BBlockTransferSrcAccessOrder
1, // BBlockTransferSrcVectorDim
1, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
true, // BBlockLdsExtraN
4,
2,
S<1, 32, 1, 8>,
1>;
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
View file @ 9f8ab221
...@@ -67,6 +67,34 @@ using DeviceConvBwdWeightInstance = ...@@ -67,6 +67,34 @@ using DeviceConvBwdWeightInstance =
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl 128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
#include "run_grouped_conv_bwd_weight_example.inc" #include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); } int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
View file @ 9f8ab221
...@@ -66,6 +66,34 @@ using DeviceConvBwdWeightInstance = ...@@ -66,6 +66,34 @@ using DeviceConvBwdWeightInstance =
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl 128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
#include "run_grouped_conv_bwd_weight_example.inc" #include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); } int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp 0 → 100644
View file @ 9f8ab221
// 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_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
using AccDataType = F32;
using ComputeTypeA = BF8;
using ComputeTypeB = F8;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
NDimSpatial,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWC,
ck::tensor_layout::convolution::GNHWC,
ck::tensor_layout::convolution::GNDHWC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GKXC,
ck::tensor_layout::convolution::GKYXC,
ck::tensor_layout::convolution::GKZYXC>>,
ck::tuple_element_t<NDimSpatial - 1,
ck::Tuple<ck::tensor_layout::convolution::GNWK,
ck::tensor_layout::convolution::GNHWK,
ck::tensor_layout::convolution::GNDHWK>>,
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
1, // ABlockTransferSrcScalarPerVector
1, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
1, // BBlockTransferSrcScalarPerVector
1, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
2, // CBlockTransferScalarPerVector_NWaveNPerXdl
ComputeTypeA, // ComputeTypeA
ComputeTypeB>; // ComputeTypeB
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ComputeTypeA,
ComputeTypeB>;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return 1;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
default: break;
}
return 1;
}
example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ 9f8ab221
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
const ck::utils::conv::ConvParam& conv_param) const ck::utils::conv::ConvParam& conv_param)
{ {
// Dl op doesn't support split_k > 1 // Dl and WMMA ops don't support split_k > 1
constexpr ck::index_t split_k = 1; constexpr ck::index_t split_k = 1;
const auto in_g_n_c_wis_desc = const auto in_g_n_c_wis_desc =
...@@ -46,8 +37,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, ...@@ -46,8 +37,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break; break;
default: default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0}); in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 0.2});
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5}); out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.1, 0.1});
} }
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize()); DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
...@@ -113,18 +104,7 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, ...@@ -113,18 +104,7 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
return true; return true;
} }
float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel}); invoker.Run(argument, StreamConfig{nullptr, false});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl
<< "DeviceOp: " << conv.GetTypeString() << std::endl;
if(config.do_verification) if(config.do_verification)
{ {
...@@ -148,25 +128,18 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, ...@@ -148,25 +128,18 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData); return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData);
} }
return true; float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
}
bool run_grouped_conv_bwd_weight_example(int argc, char* argv[]) std::size_t flop = conv_param.GetFlops();
{ std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param)) float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
{
return false;
}
switch(conv_param.num_dim_spatial_) float gb_per_sec = num_btype / 1.E6 / avg_time;
{
case 1: return run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return run_grouped_conv_bwd_weight<3>(config, conv_param);
}
return false; std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl
<< "DeviceOp: " << conv.GetTypeString() << std::endl;
return true;
} }
example/21_gemm_layernorm/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942) list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0) set(target 0)
foreach(gpu IN LISTS GPU_TARGETS) foreach(gpu IN LISTS GPU_TARGETS)
...@@ -10,4 +9,4 @@ foreach(gpu IN LISTS GPU_TARGETS) ...@@ -10,4 +9,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1) set(target 1)
endif() endif()
endforeach() endforeach()
endif()
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
View file @ 9f8ab221
...@@ -114,12 +114,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n, ...@@ -114,12 +114,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
BetaDataType, BetaDataType,
HDataType, HDataType,
AccDataType, AccDataType,
AccDataType,
HElementOp, HElementOp,
2, 2,
1>; 1>;
Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N}); Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N}); Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> save_mean({M});
Tensor<AccDataType> save_inv_std({M});
auto ref_gemm = ReferenceGemm{}; auto ref_gemm = ReferenceGemm{};
auto ref_gemm_invoker = ref_gemm.MakeInvoker(); auto ref_gemm_invoker = ref_gemm.MakeInvoker();
...@@ -145,7 +148,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n, ...@@ -145,7 +148,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
auto ref_layernorm_invoker = ref_layernorm.MakeInvoker(); auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
auto ref_layernorm_argument = ref_layernorm.MakeArgument( auto ref_layernorm_argument = ref_layernorm.MakeArgument(
e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon); e_m_n, gamma_n, beta_n, h_m_n, save_mean, save_inv_std, h_element_op, {M, N}, {1}, epsilon);
ref_layernorm_invoker.Run(ref_layernorm_argument); ref_layernorm_invoker.Run(ref_layernorm_argument);
} }
......
example/22_cgemm/CMakeLists.txt
View file @ 9f8ab221
add_custom_target(example_cgemm_xdl) add_custom_target(example_cgemm_xdl)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES) add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
endif() add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp) add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES) add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp) add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
endif() endif()
example/24_batched_gemm/CMakeLists.txt
View file @ 9f8ab221
add_custom_target(example_batched_gemm_xdl) add_custom_target(example_batched_gemm_xdl)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp) add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
endif() add_example_executable(example_batched_gemm_xdl_bf16 batched_gemm_xdl_bf16.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16)
add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bfp16) add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
endif() add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp) add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
endif() endif()
example/25_gemm_bias_e_permute/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp) add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
endif()
example/26_contraction/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES) add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp) add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp) add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
endif() add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
endif()
example/27_layernorm/CMakeLists.txt
View file @ 9f8ab221
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp) add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
endif()
example/27_layernorm/layernorm_fp16.cpp
View file @ 9f8ab221
...@@ -3,12 +3,15 @@ ...@@ -3,12 +3,15 @@
#include "common.hpp" #include "common.hpp"
using XDataType = ck::half_t; using XDataType = ck::half_t;
using GammaDataType = ck::half_t; using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t; using BetaDataType = ck::half_t;
using YDataType = ck::half_t; using YDataType = ck::half_t;
using ComputeDataType = float; using SaveMeanInvStdDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 2; constexpr int Rank = 2;
constexpr int NumReduceDim = 1; constexpr int NumReduceDim = 1;
...@@ -19,6 +22,7 @@ using DeviceInstance = ...@@ -19,6 +22,7 @@ using DeviceInstance =
BetaDataType, BetaDataType,
ComputeDataType, ComputeDataType,
YDataType, YDataType,
SaveMeanInvStdDataType,
PassThrough, PassThrough,
Rank, Rank,
NumReduceDim, NumReduceDim,
...@@ -33,7 +37,8 @@ using DeviceInstance = ...@@ -33,7 +37,8 @@ using DeviceInstance =
8, // GammaScalarPerVector 8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K) 1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector 8, // BetaScalarPerVector
8>; // OutScalarPerVector 8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm_example.inc" #include "run_layernorm_example.inc"
int main() { return run_groupnorm_example<DeviceInstance>(); } int main() { return run_groupnorm_example<DeviceInstance>(); }
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