CMakeLists.txt 7.2 KB
Newer Older
1
2
3
4
5
6
# ckProfiler
set(PROFILER_SOURCES
    profiler.cpp
    profile_gemm.cpp
    profile_gemm_splitk.cpp
    profile_gemm_bias_add_reduce.cpp
ltqin's avatar
ltqin committed
7
    profile_gemm_add_multiply.cpp
8
    profile_gemm_multiply_add.cpp
9
10
11
12
13
14
15
16
17
18
    profile_gemm_reduce.cpp
    profile_batched_gemm.cpp
    profile_batched_gemm_reduce.cpp
    profile_conv_fwd.cpp
    profile_conv_fwd_bias_relu.cpp
    profile_conv_fwd_bias_relu_add.cpp
    profile_conv_bwd_data.cpp
    profile_grouped_conv_fwd.cpp
    profile_grouped_conv_bwd_weight.cpp
    profile_reduce.cpp
19
    profile_groupnorm_bwd_data.cpp
rocking's avatar
rocking committed
20
    profile_groupnorm_fwd.cpp
21
    profile_layernorm_bwd_data.cpp
22
23
    profile_layernorm_bwd_gamma_beta.cpp
    profile_groupnorm_bwd_gamma_beta.cpp
rocking's avatar
rocking committed
24
    profile_layernorm_fwd.cpp
rocking's avatar
rocking committed
25
    profile_max_pool3d_fwd.cpp
26
27
    profile_avg_pool3d_bwd.cpp
    profile_max_pool3d_bwd.cpp
28
29
30
    profile_softmax.cpp
    profile_batchnorm_fwd.cpp
    profile_batchnorm_bwd.cpp
31
    profile_batchnorm_infer.cpp
32
    profile_grouped_conv_bwd_data.cpp
33
    profile_conv_tensor_rearrange.cpp
arai713's avatar
arai713 committed
34
    profile_transpose.cpp
35
    profile_permute_scale.cpp
36
)
37

38
39
40
if(DL_KERNELS)
  list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
endif()
41

42
43
44
45
46
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
  list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
47
  list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) 
48
  list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
49
50
  list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
51
52
53
54
  list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
  list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
55
  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
56
57
  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
endif()
58

59
60
61
62
63
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
  list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
  list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
endif()

64
65
66
67
68
set(PROFILER_EXECUTABLE ckProfiler)

add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)

69
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
70
71
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
ltqin's avatar
ltqin committed
72
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
73
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
rocking's avatar
rocking committed
90
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
91
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
92
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
93
94
95
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
rocking's avatar
rocking committed
96
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
97
98
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
99
100
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
101
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
102
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
arai713's avatar
arai713 committed
103
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
104
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
105
106
107
108
109
110
111
112

if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
endif()



113
114
115
if(DL_KERNELS)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
endif()
116

117
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
118
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
119
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
120
121
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
122
123
124
125
126
127
128
129
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
130
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
131
132
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
endif()
133

134
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)