Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
gaoqiong
composable_kernel
Commits
8ed29492
"...composable_kernel_rocm.git" did not exist on "2fd6c6d4e7d7c458bd44f4df49271b95ce745b20"
Commit
8ed29492
authored
Sep 21, 2023
by
Rostyslav Geyyer
Browse files
Merge branch 'develop' into lwpck-929
parents
560919ab
bba085d2
Changes
128
Hide whitespace changes
Inline
Side-by-side
Showing
20 changed files
with
113 additions
and
128 deletions
+113
-128
example/18_batched_gemm_reduce/CMakeLists.txt
example/18_batched_gemm_reduce/CMakeLists.txt
+0
-2
example/20_grouped_conv_bwd_weight/CMakeLists.txt
example/20_grouped_conv_bwd_weight/CMakeLists.txt
+9
-11
example/21_gemm_layernorm/CMakeLists.txt
example/21_gemm_layernorm/CMakeLists.txt
+1
-2
example/22_cgemm/CMakeLists.txt
example/22_cgemm/CMakeLists.txt
+8
-8
example/24_batched_gemm/CMakeLists.txt
example/24_batched_gemm/CMakeLists.txt
+12
-10
example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
+0
-0
example/25_gemm_bias_e_permute/CMakeLists.txt
example/25_gemm_bias_e_permute/CMakeLists.txt
+2
-4
example/26_contraction/CMakeLists.txt
example/26_contraction/CMakeLists.txt
+4
-8
example/27_layernorm/CMakeLists.txt
example/27_layernorm/CMakeLists.txt
+2
-4
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
+1
-3
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
+3
-5
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
+16
-16
example/31_batched_gemm_gemm/CMakeLists.txt
example/31_batched_gemm_gemm/CMakeLists.txt
+5
-13
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
+21
-14
example/35_splitK_gemm/CMakeLists.txt
example/35_splitK_gemm/CMakeLists.txt
+13
-10
example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
+0
-0
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
+1
-3
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
+6
-5
example/39_permute/CMakeLists.txt
example/39_permute/CMakeLists.txt
+9
-6
example/40_conv2d_fwd_quantization/CMakeLists.txt
example/40_conv2d_fwd_quantization/CMakeLists.txt
+0
-4
No files found.
example/18_batched_gemm_reduce/CMakeLists.txt
View file @
8ed29492
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 @
8ed29492
...
@@ -3,22 +3,20 @@ set(target 0)
...
@@ -3,22 +3,20 @@ 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_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
)
if
(
result EQUAL 0
)
add_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
()
endif
()
if
(
DTYPES MATCHES
"bf16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp
)
add_example_executable
(
example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16
)
add_dependencies
(
example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16
)
endif
()
endif
()
set
(
target 1
)
set
(
target 1
)
endif
()
endif
()
endforeach
()
endforeach
()
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_custom_target
(
example_grouped_conv_bwd_weight_dl
)
if
(
DL_KERNELS
)
add_example_executable
(
example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp
)
add_custom_target
(
example_grouped_conv_bwd_weight_dl
)
if
(
result EQUAL 0
)
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
)
add_dependencies
(
example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16
)
endif
()
endif
()
endif
()
\ No newline at end of file
example/21_gemm_layernorm/CMakeLists.txt
View file @
8ed29492
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/22_cgemm/CMakeLists.txt
View file @
8ed29492
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_cgemm_xdl example_cgemm_xdl_bf16
)
add_dependencies
(
example_cgemm_xdl example_cgemm_xdl_bf16
)
endif
()
endif
()
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
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_dependencies
(
example_cgemm_xdl example_cgemm_xdl_fp16
)
add_dependencies
(
example_cgemm_xdl example_cgemm_xdl_fp16
)
endif
()
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_cgemm_xdl example_cgemm_xdl_fp32
)
endif
()
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_cgemm_xdl example_cgemm_xdl_int8
)
add_dependencies
(
example_cgemm_xdl example_cgemm_xdl_int8
)
endif
()
endif
()
if
(
USE_BITINT_EXTENSION_INT4
)
if
(
USE_BITINT_EXTENSION_INT4
)
...
...
example/24_batched_gemm/CMakeLists.txt
View file @
8ed29492
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_fp32
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_fp32
)
endif
()
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_fp16
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_fp16
)
endif
()
endif
()
if
(
DTYPES MATCHES
"bf16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_batched_gemm_xdl_bf16 batched_gemm_xdl_bf16.cpp
)
add_example_executable
(
example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_bf
p
16
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_bf16
)
endif
()
endif
()
if
(
DTYPES MATCHES
"int8"
OR NOT DEFINED DTYPES
)
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_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_int8
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_int8
)
endif
()
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_batched_gemm_xdl example_batched_gemm_xdl_int4
)
endif
()
endif
()
endif
()
example/24_batched_gemm/batched_gemm_xdl_bf
p
16.cpp
→
example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
View file @
8ed29492
File moved
example/25_gemm_bias_e_permute/CMakeLists.txt
View file @
8ed29492
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 @
8ed29492
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 @
8ed29492
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/28_grouped_gemm_bias_e_permute/CMakeLists.txt
View file @
8ed29492
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
)
add_example_executable
(
example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp
)
endif
()
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @
8ed29492
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
)
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"
)
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
)
add_example_executable
(
example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp
)
endif
()
endif
()
endif
()
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View file @
8ed29492
...
@@ -5,27 +5,31 @@ set(target 0)
...
@@ -5,27 +5,31 @@ set(target 0)
foreach
(
gpu IN LISTS GPU_TARGETS
)
foreach
(
gpu IN LISTS GPU_TARGETS
)
if
(
gpu IN_LIST gpu_list1 AND target EQUAL 0
)
if
(
gpu IN_LIST gpu_list1 AND target EQUAL 0
)
add_custom_target
(
example_grouped_conv_fwd_multiple_d
)
add_custom_target
(
example_grouped_conv_fwd_multiple_d
)
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_example_executable
(
example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
)
if
(
result EQUAL 0
)
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_fp16
)
add_example_executable
(
example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp
)
endif
()
add_example_executable
(
example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16
)
add_dependencies
(
example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16
)
endif
()
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_example_executable
(
example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
)
if
(
result EQUAL 0
)
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_fp32
)
endif
()
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_example_executable
(
example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
)
if
(
result EQUAL 0
)
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_bf16
)
endif
()
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_example_executable
(
example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8
)
add_dependencies
(
example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8
)
endif
()
endif
()
if
(
USE_BITINT_EXTENSION_INT4
)
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_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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4
)
endif
()
endif
()
# USE_BITINT_EXTENSION_INT4
endif
()
# USE_BITINT_EXTENSION_INT4
set
(
target 1
)
set
(
target 1
)
...
@@ -35,12 +39,8 @@ endforeach()
...
@@ -35,12 +39,8 @@ endforeach()
set
(
target 0
)
set
(
target 0
)
foreach
(
gpu IN LISTS GPU_TARGETS
)
foreach
(
gpu IN LISTS GPU_TARGETS
)
if
(
gpu IN_LIST gpu_list2 AND target EQUAL 0
)
if
(
gpu IN_LIST gpu_list2 AND target EQUAL 0
)
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
)
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
)
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
)
set
(
target 1
)
endif
()
endif
()
endforeach
()
endforeach
()
example/31_batched_gemm_gemm/CMakeLists.txt
View file @
8ed29492
list
(
APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942
)
list
(
APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942
)
list
(
APPEND gpu_list2 gfx908 gfx90a
)
set
(
target 0
)
set
(
target 0
)
foreach
(
gpu IN LISTS GPU_TARGETS
)
foreach
(
gpu IN LISTS GPU_TARGETS
)
if
(
gpu IN_LIST gpu_list1 AND target EQUAL 0
)
if
(
gpu IN_LIST gpu_list1 AND target EQUAL 0
)
if
(
DTYPES MATCHES
"fp32"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp
)
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
)
endif
()
add_example_executable
(
example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp
)
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
)
if
(
USE_BITINT_EXTENSION_INT4
)
add_example_executable
(
example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp
)
add_example_executable
(
example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp
)
endif
(
USE_BITINT_EXTENSION_INT4
)
endif
(
USE_BITINT_EXTENSION_INT4
)
...
@@ -20,7 +14,5 @@ foreach(gpu IN LISTS GPU_TARGETS)
...
@@ -20,7 +14,5 @@ foreach(gpu IN LISTS GPU_TARGETS)
endforeach
()
endforeach
()
if
(
NOT GPU_TARGETS MATCHES
"gfx94"
AND NOT GPU_TARGETS MATCHES
"gfx1"
)
if
(
NOT GPU_TARGETS MATCHES
"gfx94"
AND NOT GPU_TARGETS MATCHES
"gfx1"
)
if
(
DTYPES MATCHES
"int8"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp
)
add_example_executable
(
example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp
)
endif
()
endif
()
endif
()
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @
8ed29492
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_custom_target
(
example_gemm_scale_softmax_gemm
)
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
)
if
(
result EQUAL 0
)
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_fp16
)
endif
()
add_example_executable
(
example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
)
if
(
result EQUAL 0
)
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_fp16
)
endif
()
add_example_executable
(
example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16
)
add_dependencies
(
example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16
)
endif
()
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
)
if
(
result EQUAL 0
)
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_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16
)
endif
()
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
(
result EQUAL 0
)
add_dependencies
(
example_gemm_scale_softmax_gemm example_grouped_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
()
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
)
if
(
result EQUAL 0
)
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_xdl_bf16
)
endif
()
add_example_executable
(
example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16
)
add_dependencies
(
example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16
)
endif
()
endif
()
example/35_splitK_gemm/CMakeLists.txt
View file @
8ed29492
...
@@ -3,25 +3,28 @@ set(target 0)
...
@@ -3,25 +3,28 @@ 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_splitK_gemm_xdl
)
add_custom_target
(
example_splitK_gemm_xdl
)
if
(
DTYPES MATCHES
"fp32"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp
)
add_example_executable
(
example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32
)
endif
()
endif
()
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp
)
add_example_executable
(
example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16
)
endif
()
endif
()
if
(
DTYPES MATCHES
"bf16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp
)
add_example_executable
(
example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_bf
p
16
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16
)
endif
()
endif
()
if
(
DTYPES MATCHES
"int8"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp
)
add_example_executable
(
example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_int8
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_int8
)
endif
()
endif
()
if
(
USE_BITINT_EXTENSION_INT4
)
if
(
USE_BITINT_EXTENSION_INT4
)
add_example_executable
(
example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp
)
add_example_executable
(
example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_int4
)
if
(
result EQUAL 0
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_int4
)
endif
()
endif
()
endif
()
set
(
target 1
)
set
(
target 1
)
endif
()
endif
()
...
...
example/35_splitK_gemm/splitK_gemm_xdl_bf
p
16.cpp
→
example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
View file @
8ed29492
File moved
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
View file @
8ed29492
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
)
add_example_executable
(
example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
)
endif
()
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View file @
8ed29492
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_custom_target
(
example_grouped_conv_bwd_data
)
add_custom_target
(
example_grouped_conv_bwd_data
)
add_example_executable
(
example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp
)
add_example_executable
(
example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16
)
endif
()
add_example_executable
(
example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp
)
add_example_executable
(
example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16
)
add_dependencies
(
example_grouped_conv_bwd_data example_grouped_conv_bwd_data_
bias_relu_
fp16
)
add_dependencies
(
example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16
)
endif
(
)
set
(
target 1
)
set
(
target 1
)
endif
()
endif
()
endforeach
()
endforeach
()
endif
()
example/39_permute/CMakeLists.txt
View file @
8ed29492
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_custom_target
(
example_permute
)
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_permute example_permute_1xHxW_fp16
)
add_dependencies
(
example_permute example_permute_1xHxW_fp16
)
endif
()
add_example_executable
(
example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_permute example_permute_NxHxW_fp16
)
add_dependencies
(
example_permute example_permute_NxHxW_fp16
)
endif
()
add_example_executable
(
example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp
)
if
(
result EQUAL 0
)
add_dependencies
(
example_permute example_permute_HxWx4_fp16
)
add_dependencies
(
example_permute example_permute_HxWx4_fp16
)
endif
()
endif
()
example/40_conv2d_fwd_quantization/CMakeLists.txt
View file @
8ed29492
if
(
DTYPES MATCHES
"int8"
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
)
...
@@ -11,7 +10,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
...
@@ -11,7 +10,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
endif
()
endif
()
endforeach
()
endforeach
()
if
(
DL_KERNELS
)
# Conv perlayer quantization
# Conv perlayer quantization
add_example_executable
(
example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp
)
add_example_executable
(
example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp
)
# Conv perchannel quantization
# Conv perchannel quantization
...
@@ -24,5 +22,3 @@ endforeach()
...
@@ -24,5 +22,3 @@ endforeach()
add_example_executable
(
example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
)
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
# 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
)
add_example_executable
(
example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
)
endif
()
endif
()
\ No newline at end of file
Prev
1
2
3
4
5
6
7
Next
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment
