Skip to content

GitLab

Commit 11279540 authored by Astha Rai's avatar Astha Rai
Browse files

Merge branch 'transpose_5d' of github.com:ROCmSoftwarePlatform/composable_kernel into transpose_5d

parents 14daa201 33e78b9a
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 796 additions and 481 deletions
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View file @ 11279540
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_convnd_fwd_reduce_xdl) 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_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8) add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
endif()
add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp) add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
if(result EQUAL 0) add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
endif() add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp) add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
if(result EQUAL 0)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16) add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
endif() add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
if(result EQUAL 0) if(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32) add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
endif() add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
if(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp) set(target 1)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4) endif()
endif(USE_BITINT_EXTENSION_INT4) endforeach()
set(target 1)
endif()
endforeach()
\ No newline at end of file
example/12_reduce/README.md
View file @ 11279540
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
## Run ```example_reduce_blockwise``` ## Run ```example_reduce_blockwise```
```bash ```bash
# -D <xxx> : input 3d/4d/5d tensor lengths # -D <xxx> : input 3D/4D/5D tensor lengths
# -R <xxx> : reduce dimension ids # -R <xxx> : reduce dimension ids
# -v <x> : verification (0=no, 1=yes) # -v <x> : verification (0=no, 1=yes)
#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4) #arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)
...@@ -22,7 +22,7 @@ Perf: 0.238063 ms, 264.285 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSr ...@@ -22,7 +22,7 @@ Perf: 0.238063 ms, 264.285 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSr
## Run ```example_reduce_multiblock_atomic_add``` ## Run ```example_reduce_multiblock_atomic_add```
```bash ```bash
# -D <xxx> : input 3d/4d/5d tensor lengths # -D <xxx> : input 3D/4D/5D tensor lengths
# -R <xxx> : reduce dimension ids # -R <xxx> : reduce dimension ids
# -v <x> : verification (0=no, 1=yes) # -v <x> : verification (0=no, 1=yes)
#arg1: data type (0: fp32, 1: fp64) #arg1: data type (0: fp32, 1: fp64)
......
example/15_grouped_gemm/CMakeLists.txt
View file @ 11279540
add_custom_target(example_grouped_gemm_xdl) add_custom_target(example_grouped_gemm_xdl)
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_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
endif()
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp) add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp16)
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) add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_dl_fp16)
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) add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_splitk_fp16)
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) add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp16)
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) add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_bias_fp16)
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) add_example_executable(example_grouped_gemm_xdl_bf16 grouped_gemm_xdl_bf16.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bf16)
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) add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
endif()
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_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp) add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
if(result EQUAL 0) add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
endif()
endif() endif()
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View file @ 11279540
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)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
if(result EQUAL 0) add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16) add_example_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) add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
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)
add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp) add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
if(result EQUAL 0)
add_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)
endif() add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
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)
if(result EQUAL 0) add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
endif() add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp) add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8) add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
endif() add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp) add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
if(result EQUAL 0) add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
endif() add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp) add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32) add_example_dependencies(example_gemm_reduce_xdl
endif() example_gemm_reduce_xdl_mean_meansquare
example_gemm_reduce_xdl_max
add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp) example_gemm_add_add_mean_meansquare_xdl)
if(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16) if(USE_BITINT_EXTENSION_INT4)
endif() add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp) add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
if(result EQUAL 0) endif()
add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16) set(target 1)
endif() endif()
add_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_mean_meansquare
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(result EQUAL 0)
add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
endif()
endif()
set(target 1)
endif()
endforeach() endforeach()
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 11279540
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)
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_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)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp) add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16) add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8 grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp)
endif() add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8)
if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942") set(target 1)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8 grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp) endif()
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8) 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() endif()
endif()
set(target 1)
endif()
endforeach() endforeach()
add_custom_target(example_grouped_conv_bwd_weight_dl) 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)
if(result EQUAL 0) 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()
example/20_grouped_conv_bwd_weight/common.hpp
View file @ 11279540
...@@ -46,25 +46,21 @@ struct CommonLayoutSetting ...@@ -46,25 +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>
{ {
}; };
...@@ -84,10 +80,10 @@ struct ExecutionConfig final ...@@ -84,10 +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 @ 11279540
...@@ -76,4 +76,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe ...@@ -76,4 +76,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#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 @ 11279540
// 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 @ 11279540
...@@ -78,4 +78,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe ...@@ -78,4 +78,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#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 @ 11279540
...@@ -77,4 +77,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe ...@@ -77,4 +77,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#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
View file @ 11279540
...@@ -83,4 +83,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe ...@@ -83,4 +83,23 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
#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/run_grouped_conv_bwd_weight_example.inc
View file @ 11279540
...@@ -5,7 +5,7 @@ template <ck::index_t NDimSpatial> ...@@ -5,7 +5,7 @@ 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 =
...@@ -143,23 +143,3 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, ...@@ -143,23 +143,3 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
return true; return true;
} }
bool run_grouped_conv_bwd_weight_example(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return false;
}
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);
}
return false;
}
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
View file @ 11279540
...@@ -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 @ 11279540
add_custom_target(example_cgemm_xdl) add_custom_target(example_cgemm_xdl)
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp) add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
if(result EQUAL 0) 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) add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
if(result EQUAL 0) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
endif()
add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp) add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
if(result EQUAL 0) add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
endif()
add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp) add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
if(result EQUAL 0) 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 @ 11279540
add_custom_target(example_batched_gemm_xdl) add_custom_target(example_batched_gemm_xdl)
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)
if(result EQUAL 0) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
endif()
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)
if(result EQUAL 0) 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) add_example_executable(example_batched_gemm_xdl_bf16 batched_gemm_xdl_bf16.cpp)
if(result EQUAL 0) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16)
endif()
add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp) add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
if(result EQUAL 0) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
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)
if(result EQUAL 0) add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
endif()
endif() endif()
example/26_contraction/CMakeLists.txt
View file @ 11279540
add_custom_target(example_contraction)
add_custom_target(example_contraction_scale)
add_custom_target(example_contraction_bilinear)
# FP32
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_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_fp32)
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_dependencies(example_contraction_scale example_contraction_scale_xdl_fp32)
add_example_executable(example_contraction_bilinear_xdl_fp32_compute_bf16 contraction_bilinear_xdl_fp32_compute_bf16.cpp)
add_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_fp32_compute_bf16)
add_example_executable(example_contraction_scale_xdl_fp32_compute_bf16 contraction_scale_xdl_fp32_compute_bf16.cpp)
add_dependencies(example_contraction_scale example_contraction_scale_xdl_fp32_compute_bf16)
add_example_executable(example_contraction_bilinear_xdl_fp32_compute_fp16 contraction_bilinear_xdl_fp32_compute_fp16.cpp)
add_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_fp32_compute_fp16)
add_example_executable(example_contraction_scale_xdl_fp32_compute_fp16 contraction_scale_xdl_fp32_compute_fp16.cpp)
add_dependencies(example_contraction_scale example_contraction_scale_xdl_fp32_compute_fp16)
# FP64
add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp) add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
add_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_fp64)
add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp) add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
add_dependencies(example_contraction_scale example_contraction_scale_xdl_fp64)
add_example_executable(example_contraction_bilinear_xdl_fp64_compute_fp32 contraction_bilinear_xdl_fp64_compute_fp32.cpp)
add_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_fp64_compute_fp32)
add_example_executable(example_contraction_scale_xdl_fp64_compute_fp32 contraction_scale_xdl_fp64_compute_fp32.cpp)
add_dependencies(example_contraction_scale example_contraction_scale_xdl_fp64_compute_fp32)
# FP16
add_example_executable(example_contraction_bilinear_xdl_fp16_compute_fp32 contraction_bilinear_xdl_fp16_compute_fp32.cpp)
add_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_fp16_compute_fp32)
add_example_executable(example_contraction_scale_xdl_fp16_compute_fp32 contraction_scale_xdl_fp16_compute_fp32.cpp)
add_dependencies(example_contraction_scale example_contraction_scale_xdl_fp16_compute_fp32)
# BF16
add_example_executable(example_contraction_bilinear_xdl_bf16_compute_fp32 contraction_bilinear_xdl_bf16_compute_fp32.cpp)
add_dependencies(example_contraction_bilinear example_contraction_bilinear_xdl_bf16_compute_fp32)
add_example_executable(example_contraction_scale_xdl_bf16_compute_fp32 contraction_scale_xdl_bf16_compute_fp32.cpp)
add_dependencies(example_contraction_scale example_contraction_scale_xdl_bf16_compute_fp32)
add_dependencies(example_contraction example_contraction_scale)
add_dependencies(example_contraction example_contraction_bilinear)
example/26_contraction/common_instances.hpp 0 → 100644
View file @ 11279540
This diff is collapsed.
example/26_contraction/contraction_bilinear_xdl_bf16_compute_fp32.cpp 0 → 100644
View file @ 11279540
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = BF16;
using DDataType = BF16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = BF16;
using ComputeDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
using DeviceOpInstanceKKNN = DeviceOpInstanceKK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNNN = DeviceOpInstanceKN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKNN = DeviceOpInstanceMK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNNN = DeviceOpInstanceMN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKNN;
#include "run_contraction_bilinear_example.inc"
int main(int argc, char* argv[]) { return run_contraction_bilinear_example(argc, argv); }
example/26_contraction/contraction_bilinear_xdl_fp16_compute_fp32.cpp 0 → 100644
View file @ 11279540
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "common_instances.hpp"
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
using ComputeDataType = F32;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
using DeviceOpInstanceKKNN = DeviceOpInstanceKK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceKNNN = DeviceOpInstanceKN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMKNN = DeviceOpInstanceMK_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstanceMNNN = DeviceOpInstanceMN_Generic<NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
ComputeDataType,
AElementOp,
BElementOp,
CDEElementOp>;
using DeviceOpInstance = DeviceOpInstanceKKNN;
#include "run_contraction_bilinear_example.inc"
int main(int argc, char* argv[]) { return run_contraction_bilinear_example(argc, argv); }
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