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
56863b9a
Commit
56863b9a
authored
Aug 16, 2023
by
Jing Zhang
Browse files
add fp8 support
parents
54df59bf
d4c84256
Changes
250
Hide whitespace changes
Inline
Side-by-side
Showing
20 changed files
with
188 additions
and
116 deletions
+188
-116
example/21_gemm_layernorm/CMakeLists.txt
example/21_gemm_layernorm/CMakeLists.txt
+2
-0
example/22_cgemm/CMakeLists.txt
example/22_cgemm/CMakeLists.txt
+15
-10
example/24_batched_gemm/CMakeLists.txt
example/24_batched_gemm/CMakeLists.txt
+16
-12
example/25_gemm_bias_e_permute/CMakeLists.txt
example/25_gemm_bias_e_permute/CMakeLists.txt
+4
-2
example/26_contraction/CMakeLists.txt
example/26_contraction/CMakeLists.txt
+8
-5
example/27_layernorm/CMakeLists.txt
example/27_layernorm/CMakeLists.txt
+4
-2
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
+3
-1
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
+5
-3
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
+25
-15
example/31_batched_gemm_gemm/CMakeLists.txt
example/31_batched_gemm_gemm/CMakeLists.txt
+12
-5
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
+22
-14
example/35_splitK_gemm/CMakeLists.txt
example/35_splitK_gemm/CMakeLists.txt
+16
-11
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
+3
-1
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
+3
-1
example/39_permute/CMakeLists.txt
example/39_permute/CMakeLists.txt
+9
-7
example/40_conv2d_fwd_quantization/CMakeLists.txt
example/40_conv2d_fwd_quantization/CMakeLists.txt
+14
-16
example/41_grouped_conv_conv_fwd/CMakeLists.txt
example/41_grouped_conv_conv_fwd/CMakeLists.txt
+12
-4
example/42_groupnorm/CMakeLists.txt
example/42_groupnorm/CMakeLists.txt
+5
-3
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
+6
-2
example/44_elementwise_permute/CMakeLists.txt
example/44_elementwise_permute/CMakeLists.txt
+4
-2
No files found.
example/21_gemm_layernorm/CMakeLists.txt
View file @
56863b9a
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
()
example/22_cgemm/CMakeLists.txt
View file @
56863b9a
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
)
...
...
example/24_batched_gemm/CMakeLists.txt
View file @
56863b9a
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
)
...
...
example/25_gemm_bias_e_permute/CMakeLists.txt
View file @
56863b9a
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
()
example/26_contraction/CMakeLists.txt
View file @
56863b9a
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
()
example/27_layernorm/CMakeLists.txt
View file @
56863b9a
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
()
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
View file @
56863b9a
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
()
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @
56863b9a
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
()
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View file @
56863b9a
...
...
@@ -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
()
example/31_batched_gemm_gemm/CMakeLists.txt
View file @
56863b9a
...
...
@@ -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
()
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @
56863b9a
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
()
example/35_splitK_gemm/CMakeLists.txt
View file @
56863b9a
...
...
@@ -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
)
...
...
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
View file @
56863b9a
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
()
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View file @
56863b9a
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
()
\ No newline at end of file
endforeach
()
endif
()
example/39_permute/CMakeLists.txt
View file @
56863b9a
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
()
example/40_conv2d_fwd_quantization/CMakeLists.txt
View file @
56863b9a
...
...
@@ -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
per
channel
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
_
per
layer_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
()
\ No newline at end of file
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @
56863b9a
...
...
@@ -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
()
example/42_groupnorm/CMakeLists.txt
View file @
56863b9a
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
()
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
View file @
56863b9a
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
()
example/44_elementwise_permute/CMakeLists.txt
View file @
56863b9a
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
()
Prev
1
2
3
4
5
6
7
…
13
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
