ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Allow building CK for specific data types and split off last remaining DL instances. (#830)

Browse Source 
* properly split conv_nd_bwd_data instances

* split conv2d_fwd instance data types

* split the gemm, conv2d_fwd and batched_gemm_softamx_gemm

* split the tests by data types where possible

* filter examples by DTYPES

* split few remaining examples by DTYPES

* filter most instances by DTYPES

* add new lines at end of headers, fix grouped_gemm profiler

* fix syntax

* split the ckprofiler instances by DTYPES

* split the conv2d and quantization DL and XDL instances

* fix the splitting of conv2d DL instances

* split softmax and pool_fwd tests for fp16 and fp32 types

* fix syntax

* fix the dl_int8 quantization instances isolation

[ROCm/composable_kernel commit: 08eb176929]
This commit is contained in:
Illia Silin
2023-08-07 14:56:10 -07:00
committed by GitHub
parent 33496a08cc
commit 25efecc13e
117 changed files with 976 additions and 590 deletions
Download Patch File Download Diff File Expand all files Collapse all files

78
example/01_gemm/CMakeLists.txt
View File

@@ -1,31 +1,44 @@
add_custom_target(example_gemm_dl)
if(DL_KERNELS)
add_custom_target(example_gemm_dl)
add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_fp32)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_fp16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_int8)
endif()
add_dependencies(example_gemm_dl example_gemm_dl_fp32)
add_dependencies(example_gemm_dl example_gemm_dl_fp16)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_int4)
endif(USE_BITINT_EXTENSION_INT4)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
add_dependencies(example_gemm_dl example_gemm_dl_int4)
endif(USE_BITINT_EXTENSION_INT4)
add_custom_target(example_gemm_xdl)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
endif()
add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
@@ -37,22 +50,17 @@ if(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_xdl example_gemm_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
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_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
endif()
add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
add_example_executable(example_gemm_xdl_f8 gemm_xdl_f8.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
add_example_executable(example_gemm_xdl_f8 gemm_xdl_f8.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
endif()
endif()

2
example/02_gemm_bilinear/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
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)
@@ -15,3 +16,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()

2
example/03_gemm_bias_relu/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
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)
@@ -6,3 +7,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()

30
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View File

@@ -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_gemm_add_add_fastgelu_xdl)
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_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
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_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_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_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)
add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
if(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
endif(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
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_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
endif()
set(target 1)
endif()
endforeach()

34
example/09_convnd_fwd/CMakeLists.txt
View File

@@ -2,16 +2,34 @@ 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_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
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
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
endif()
set(target 1)
endif()
endforeach()
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)
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()

24
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View File

@@ -3,14 +3,22 @@ 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)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
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_fp16 convnd_fwd_max_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
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)
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)
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)
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)
add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
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)

9
example/13_pool2d_fwd/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
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()

4
example/14_gemm_quantization/CMakeLists.txt
View File

@@ -1,6 +1,8 @@
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
# dlops
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
if(DL_KERNELS)
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
endif()
# xdlops
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)

38
example/15_grouped_gemm/CMakeLists.txt
View File

@@ -1,21 +1,25 @@
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_fp16 grouped_gemm_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.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_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp32
example_grouped_gemm_xdl_fp16
example_grouped_gemm_xdl_bfp16
example_grouped_gemm_xdl_int8
example_grouped_gemm_multiple_d_dl_fp16
example_grouped_gemm_xdl_splitk_fp16)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
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_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp16
example_grouped_gemm_multiple_d_dl_fp16
example_grouped_gemm_xdl_splitk_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_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
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)

54
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View File

@@ -6,33 +6,33 @@ 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)
add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
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)
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_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
add_dependencies(example_gemm_reduce_xdl_max
example_gemm_max_xdl_bf16
example_gemm_max_xdl_fp16
example_gemm_max_xdl_fp32
example_gemm_max_xdl_int8)
add_dependencies(example_gemm_reduce_xdl_mean_meansquare
example_gemm_mean_meansquare_xdl_fp16
example_gemm_mean_meansquare_xdl_fp32
example_gemm_mean_meansquare_xdl_bf16
example_gemm_add_addsquare_xdl_int8)
add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
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_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
add_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()
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_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()
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_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()
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_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_dependencies(example_gemm_reduce_xdl
example_gemm_reduce_xdl_mean_meansquare
example_gemm_reduce_xdl_max

8
example/17_convnd_bwd_data/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
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)
@@ -7,5 +8,8 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
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(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()
endif()

2
example/18_batched_gemm_reduce/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
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)
@@ -6,3 +7,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()

26
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View File

@@ -3,18 +3,22 @@ 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)
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_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
example_grouped_conv_bwd_weight_xdl_bf16)
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_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_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
endif()
set(target 1)
endif()
endforeach()
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)
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()

2
example/21_gemm_layernorm/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
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)
@@ -9,3 +10,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
endif()

25
example/22_cgemm/CMakeLists.txt
View File

@@ -1,16 +1,21 @@
add_custom_target(example_cgemm_xdl)
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
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_int8 cgemm_xdl_int8.cpp)
add_dependencies(example_cgemm_xdl
example_cgemm_xdl_bf16
example_cgemm_xdl_fp16
example_cgemm_xdl_fp32
example_cgemm_xdl_int8)
add_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_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)

28
example/24_batched_gemm/CMakeLists.txt
View File

@@ -1,16 +1,20 @@
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_fp16 batched_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
add_dependencies(example_batched_gemm_xdl
example_batched_gemm_xdl_fp32
example_batched_gemm_xdl_fp16
example_batched_gemm_xdl_bfp16
example_batched_gemm_xdl_int8)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
add_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_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bfp16)
endif()
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)
add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)

6
example/25_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1,2 +1,4 @@
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)
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_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
endif()

13
example/26_contraction/CMakeLists.txt
View File

@@ -1,5 +1,8 @@
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_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
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_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
endif()
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()

6
example/27_layernorm/CMakeLists.txt
View File

@@ -1,2 +1,4 @@
add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
endif()

4
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1 +1,3 @@
add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
endif()

8
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1,5 +1,7 @@
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
endif()
endif()

40
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View File

@@ -5,23 +5,29 @@ set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_custom_target(example_grouped_conv_fwd_multiple_d)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif() # USE_BITINT_EXTENSION_INT4
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
set(target 1)
endif()
endforeach()
@@ -29,8 +35,12 @@ endforeach()
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
endif()
set(target 1)
endif()
endforeach()

17
example/31_batched_gemm_gemm/CMakeLists.txt
View File

@@ -3,10 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4)
@@ -15,5 +20,7 @@ foreach(gpu IN LISTS GPU_TARGETS)
endforeach()
if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
endif()
endif()

36
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View File

@@ -1,16 +1,24 @@
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
endif()
add_custom_target(example_gemm_scale_softmax_gemm)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
endif()

27
example/35_splitK_gemm/CMakeLists.txt
View File

@@ -3,17 +3,22 @@ set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_splitK_gemm_xdl)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
add_dependencies(example_splitK_gemm_xdl
example_splitK_gemm_xdl_fp32
example_splitK_gemm_xdl_fp16
example_splitK_gemm_xdl_bfp16
example_splitK_gemm_xdl_int8)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bfp16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)

4
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
View File

@@ -1 +1,3 @@
add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
endif()

4
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View File

@@ -1,3 +1,4 @@
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)
@@ -10,4 +11,5 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
set(target 1)
endif()
endforeach()
endforeach()
endif()

16
example/39_permute/CMakeLists.txt
View File

@@ -1,9 +1,11 @@
add_custom_target(example_permute)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_custom_target(example_permute)
add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
add_dependencies(example_permute example_permute_1xHxW_fp16)
add_dependencies(example_permute example_permute_NxHxW_fp16)
add_dependencies(example_permute example_permute_HxWx4_fp16)
add_dependencies(example_permute example_permute_1xHxW_fp16)
add_dependencies(example_permute example_permute_NxHxW_fp16)
add_dependencies(example_permute example_permute_HxWx4_fp16)
endif()

30
example/40_conv2d_fwd_quantization/CMakeLists.txt
View File

@@ -10,21 +10,19 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
# Conv perlayer quantization
add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
# Conv perchannel quantization
add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
# Conv + bias + relu perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
# Conv + bias + relu perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
# Conv + bias + tanh perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
# Conv + bias + tanh perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
if(DL_KERNELS)
# Conv perlayer quantization
add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
# Conv perchannel quantization
add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
# Conv + bias + relu perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
# Conv + bias + relu perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
# Conv + bias + tanh perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
# Conv + bias + tanh perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
endif()
endif()

16
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View File

@@ -3,9 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4)
@@ -14,5 +20,7 @@ foreach(gpu IN LISTS GPU_TARGETS)
endforeach()
if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
endif()
endif()

8
example/42_groupnorm/CMakeLists.txt
View File

@@ -1,3 +1,5 @@
add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
endif()

8
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1,2 +1,6 @@
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
endif()

6
example/44_elementwise_permute/CMakeLists.txt
View File

@@ -1,2 +1,4 @@
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
endif()

8
example/46_gemm_add_multiply/CMakeLists.txt
View File

@@ -1,2 +1,6 @@
add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if(DL_KERNELS)
add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
endif()
add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
endif()

5
example/48_pool3d_fwd/CMakeLists.txt
View File

@@ -1,2 +1,3 @@
add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
endif()

12
example/49_maxpool2d_bwd/CMakeLists.txt
View File

@@ -1,3 +1,9 @@
add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
endif()

4
example/50_put_element/CMakeLists.txt
View File

@@ -1 +1,3 @@
add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
endif()