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
a8629a98
".github/vscode:/vscode.git/clone" did not exist on "451a1fe3eccf26e6b8bf275f803dc7ec26e102c0"
Unverified
Commit
a8629a98
authored
Sep 26, 2023
by
zjing14
Committed by
GitHub
Sep 26, 2023
Browse files
Merge branch 'develop' into gemm_v2r3_kpad_fix
parents
8dc713ea
94bfa502
Changes
334
Show whitespace changes
Inline
Side-by-side
Showing
20 changed files
with
112 additions
and
130 deletions
+112
-130
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
example/41_grouped_conv_conv_fwd/CMakeLists.txt
example/41_grouped_conv_conv_fwd/CMakeLists.txt
+4
-12
example/42_groupnorm/CMakeLists.txt
example/42_groupnorm/CMakeLists.txt
+3
-5
example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
+10
-6
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
+2
-6
example/44_elementwise_permute/CMakeLists.txt
example/44_elementwise_permute/CMakeLists.txt
+2
-4
example/46_gemm_add_multiply/CMakeLists.txt
example/46_gemm_add_multiply/CMakeLists.txt
+2
-6
example/48_pool3d_fwd/CMakeLists.txt
example/48_pool3d_fwd/CMakeLists.txt
+1
-3
example/49_maxpool2d_bwd/CMakeLists.txt
example/49_maxpool2d_bwd/CMakeLists.txt
+3
-9
example/50_put_element/CMakeLists.txt
example/50_put_element/CMakeLists.txt
+1
-3
example/52_image_to_column/CMakeLists.txt
example/52_image_to_column/CMakeLists.txt
+10
-0
No files found.
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @
a8629a98
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 @
a8629a98
...
@@ -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
)
endif
()
add_example_executable
(
example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp
)
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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4
)
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
)
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
)
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 @
a8629a98
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
)
endif
()
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp
)
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
)
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 @
a8629a98
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 @
a8629a98
...
@@ -3,26 +3,29 @@ set(target 0)
...
@@ -3,26 +3,29 @@ 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
)
if
(
result EQUAL 0
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_int4
)
add_dependencies
(
example_splitK_gemm_xdl example_splitK_gemm_xdl_int4
)
endif
()
endif
()
endif
()
set
(
target 1
)
set
(
target 1
)
endif
()
endif
()
endforeach
()
endforeach
()
example/35_splitK_gemm/splitK_gemm_xdl_bf
p
16.cpp
→
example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
View file @
a8629a98
File moved
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
View file @
a8629a98
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 @
a8629a98
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
)
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_fp16
)
endif
()
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_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 @
a8629a98
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 @
a8629a98
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
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @
a8629a98
...
@@ -3,15 +3,9 @@ list(APPEND gpu_list2 gfx908 gfx90a)
...
@@ -3,15 +3,9 @@ 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_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp
)
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
)
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
)
add_example_executable
(
example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp
)
endif
()
if
(
USE_BITINT_EXTENSION_INT4
)
if
(
USE_BITINT_EXTENSION_INT4
)
add_example_executable
(
example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp
)
add_example_executable
(
example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_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_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp
)
add_example_executable
(
example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp
)
endif
()
endif
()
endif
()
example/42_groupnorm/CMakeLists.txt
View file @
a8629a98
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_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp
)
add_example_executable
(
example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp
)
add_example_executable
(
example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp
)
add_example_executable
(
example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp
)
add_example_executable
(
example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp
)
endif
()
example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
View file @
a8629a98
...
@@ -14,18 +14,22 @@ using ComputeDataType = float;
...
@@ -14,18 +14,22 @@ using ComputeDataType = float;
struct
YElementOp
struct
YElementOp
{
{
template
<
typename
T
>
template
<
typename
Y
,
typename
X
>
__host__
__device__
void
operator
()(
T
&
y
,
const
T
&
x
)
const
__host__
__device__
void
operator
()(
Y
&
y
,
const
X
&
x
)
const
{
{
static_assert
(
ck
::
is_same
<
T
,
float
>::
value
||
ck
::
is_same
<
T
,
double
>::
value
||
static_assert
(
ck
::
is_same
<
X
,
float
>::
value
||
ck
::
is_same
<
X
,
double
>::
value
||
ck
::
is_same
<
T
,
ck
::
half_t
>::
value
,
ck
::
is_same
<
X
,
ck
::
half_t
>::
value
,
"Data type is not supported by this operation!"
);
"Data type is not supported by this operation!"
);
T
a
;
static_assert
(
ck
::
is_same
<
Y
,
float
>::
value
||
ck
::
is_same
<
Y
,
double
>::
value
||
ck
::
is_same
<
Y
,
ck
::
half_t
>::
value
,
"Data type is not supported by this operation!"
);
X
a
;
ck
::
tensor_operation
::
element_wise
::
Sigmoid
{}(
a
,
x
);
ck
::
tensor_operation
::
element_wise
::
Sigmoid
{}(
a
,
x
);
y
=
x
*
a
;
y
=
ck
::
type_convert
<
Y
>
(
x
*
a
)
;
};
};
};
};
...
...
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
View file @
a8629a98
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
)
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
)
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
()
example/44_elementwise_permute/CMakeLists.txt
View file @
a8629a98
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 elementwise_permute_4D_fp16.cpp
)
add_example_executable
(
example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp
)
add_example_executable
(
example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp
)
endif
()
example/46_gemm_add_multiply/CMakeLists.txt
View file @
a8629a98
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp
)
if
(
DL_KERNELS
)
add_example_executable
(
example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp
)
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
()
example/48_pool3d_fwd/CMakeLists.txt
View file @
a8629a98
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp
)
add_example_executable
(
example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp
)
endif
()
example/49_maxpool2d_bwd/CMakeLists.txt
View file @
a8629a98
if
(
DTYPES MATCHES
"bf16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp
)
add_example_executable
(
example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp
)
add_example_executable
(
example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp
)
endif
()
add_example_executable
(
example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp
)
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
()
example/50_put_element/CMakeLists.txt
View file @
a8629a98
if
(
DTYPES MATCHES
"fp16"
OR NOT DEFINED DTYPES
)
add_example_executable
(
example_put_element_fp16 put_element_fp16.cpp
)
add_example_executable
(
example_put_element_fp16 put_element_fp16.cpp
)
endif
()
example/52_image_to_column/CMakeLists.txt
0 → 100644
View file @
a8629a98
list
(
APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942
)
set
(
target 0
)
foreach
(
gpu IN LISTS GPU_TARGETS
)
if
(
gpu IN_LIST gpu_list AND target EQUAL 0
)
add_custom_target
(
example_image_to_column
)
add_example_executable
(
example_image_to_column_f32 image_to_column_f32.cpp
)
add_dependencies
(
example_image_to_column example_image_to_column_f32
)
set
(
target 1
)
endif
()
endforeach
()
Prev
1
2
3
4
5
6
7
…
17
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
