CMakeLists.txt 9.5 KB
Newer Older
W
Wu Yi 已提交
1
include(operators)
D
dzhwinter 已提交
2

3 4
add_subdirectory(generator)

W
Wilber 已提交
5
# solve "math constants not defined" problems caused by the order of inclusion
6 7 8
# of <cmath> and the definition of macro _USE_MATH_DEFINES
add_definitions(-D_USE_MATH_DEFINES)

W
Wu Yi 已提交
9 10 11
# clean cache and pybind_file content first when rebuild
unset(GLOB_OP_LIB CACHE)
unset(OP_LIBRARY CACHE)
12
set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h.tmp CACHE INTERNAL "pybind.h file")
13
set(pybind_file_prune ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h.prune CACHE INTERNAL "pybind.h file")
14
set(pybind_file_final ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h)
15
file(WRITE ${pybind_file} "#include \"paddle/phi/core/kernel_registry.h\" // Generated by the paddle/fluid/operators/CMakeLists.txt.  DO NOT EDIT!\n\n")
16

Q
qijun 已提交
17
add_subdirectory(math)
W
Wu Yi 已提交
18 19 20 21 22 23 24 25
add_subdirectory(controlflow)
add_subdirectory(detection)
add_subdirectory(elementwise)
add_subdirectory(fused)
add_subdirectory(metrics)
add_subdirectory(optimizers)
add_subdirectory(reduce_ops)
add_subdirectory(sequence_ops)
S
Steffy-zxf 已提交
26
add_subdirectory(string)
L
levi131 已提交
27
add_subdirectory(prim_ops)
Q
QI JUN 已提交
28

T
tangwei12 已提交
29

T
typhoonzero 已提交
30
if(WITH_DISTRIBUTE)
31
    add_subdirectory(collective)
W
Wu Yi 已提交
32
endif()
33

T
tangwei12 已提交
34 35 36 37
if (WITH_PSCORE)
    add_subdirectory(pscore)
endif()

38 39
add_subdirectory(amp)

W
wopeizl 已提交
40
add_subdirectory(reader)
T
typhoonzero 已提交
41

W
Wu Yi 已提交
42 43
if (NOT WIN32)
    add_subdirectory(nccl)
S
sneaxiy 已提交
44 45
endif()

46
if (WITH_GPU AND TENSORRT_FOUND)
W
Wu Yi 已提交
47
    add_subdirectory(tensorrt)
48
endif()
49

D
denglin-github 已提交
50 51 52 53
if (WITH_DLNNE)
    add_subdirectory(dlnne)
endif()

石晓伟 已提交
54 55 56 57
if (WITH_LITE)
    add_subdirectory(lite)
endif()

F
fwenguang 已提交
58 59 60 61
if (WITH_MLU)
    add_subdirectory(mlu)
endif()

62 63 64 65
if(WITH_CINN)
    add_subdirectory(cinn)
endif()

A
Allen Guo 已提交
66 67 68 69
if(WITH_IPU)
    add_subdirectory(ipu)
endif()

W
Wilber 已提交
70
SET(OP_HEADER_DEPS xxhash executor)
石晓伟 已提交
71

72
if (WITH_GPU)
73 74 75
    if (${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0)
        SET(OP_HEADER_DEPS ${OP_HEADER_DEPS} cub)
    endif()
76 77
endif()

78 79 80 81
if (WITH_POCKETFFT)
    SET(OP_HEADER_DEPS ${OP_HEADER_DEPS} pocketfft)
endif()

82

A
Aurelius84 已提交
83
SET(OP_MKL_DEPS "")
84
if (NOT WITH_MKL OR NOT WITH_AVX)
A
Aurelius84 已提交
85 86 87
    SET(OP_MKL_DEPS ${OP_MKL_DEPS} match_matrix_tensor_op)
    SET(OP_MKL_DEPS ${OP_MKL_DEPS} var_conv_2d_op)
endif()
88
if(WITH_COVERAGE OR WIN32 OR WITH_NV_JETSON)
A
Aurelius84 已提交
89
    SET(OP_MKL_DEPS ${OP_MKL_DEPS} pyramid_hash_op)
K
Kevin 已提交
90 91
endif()

92 93 94 95 96
if(WITH_UNITY_BUILD)
    # Load Unity Build rules for operators in paddle/fluid/operators.
    include(unity_build_rule.cmake)
endif()

97
set(OP_HEADER_DEPS ${OP_HEADER_DEPS} phi phi_utils backward_infermeta sparse_backward_infermeta static_prim_api get_expected_kernel_func)
98

99
register_operators(EXCLUDES py_func_op warpctc_op dgc_op generated_op1 generated_op2 generated_op3 generated_op4 load_combine_op lstm_op run_program_op eye_op quantize_linear_op
F
Feiyu Chan 已提交
100
        recurrent_op save_combine_op sparse_attention_op sync_batch_norm_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS})
武毅 已提交
101

102
op_library(generated_op UNITY SRCS generated_op1.cc generated_op2.cc generated_op3.cc generated_op4.cc DEPS ${OP_HEADER_DEPS})
103
op_library(run_program_op SRCS run_program_op.cc run_program_op.cu.cc run_program_op_npu.cc DEPS executor_cache ${OP_HEADER_DEPS})
104
target_link_libraries(run_program_op cuda_graph_with_memory_pool)
105
op_library(quantize_linear_op DEPS phi)
106
op_library(save_combine_op DEPS string_array phi)
S
Steffy-zxf 已提交
107
op_library(load_combine_op DEPS string_array)
108

109 110
if (WITH_GPU OR WITH_ROCM)
    if(WITH_ROCM)
0
0x45f 已提交
111
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc)
Q
qingqing01 已提交
112
    # warpctc_op needs cudnn 7 above
113
    elseif(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
0
0x45f 已提交
114
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc)
W
Wu Yi 已提交
115 116
    else()
        op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
W
Wu Yi 已提交
117
    endif()
118
    op_library(sync_batch_norm_op)
119
    if ((NOT WIN32) AND (NOT WITH_ROCM) AND (NOT PADDLE_WITH_ARM) AND (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 11.3) )
120 121
        op_library(sparse_attention_op)
    endif()
W
Wu Yi 已提交
122 123
else()
    op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
W
Wu Yi 已提交
124
endif()
T
tangwei12 已提交
125

126 127 128 129
if (WITH_ASCEND_CL)
  op_library(sync_batch_norm_op)
endif()

Q
qipengh 已提交
130 131 132 133
if (WITH_MLU)
  op_library(sync_batch_norm_op)
endif()

T
tangwei12 已提交
134 135 136
op_library(lstm_op DEPS ${OP_HEADER_DEPS}  lstm_compute)
op_library(eye_op DEPS ${OP_HEADER_DEPS})
op_library(recurrent_op DEPS ${OP_HEADER_DEPS})
137

138
set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
139

G
gongweibao 已提交
140
if (WITH_DGC)
141 142 143 144
    op_library(dgc_op DEPS dgc)
    set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dgc)
endif()

145
cc_library(common_infer_shape_functions SRCS common_infer_shape_functions.cc DEPS operator)
146
cc_library(ops_extra_info SRCS ops_extra_info.cc DEPS attribute cudnn_workspace_helper)
A
Aurelius84 已提交
147

148
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows_utils lapack_function
Y
yaoxuefeng 已提交
149
lod_tensor maxouting unpooling pooling lod_rank_table context_project
J
Jiabin Yang 已提交
150
sequence_pooling executor generator static_prim_api)
151
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc static_prim_api static_utils static_global_utils prim_utils)
X
xuezhong 已提交
152
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler sample_prob tree2col)
153
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions beam_search fc_functor matrix_inverse matrix_solve)
R
Ruibiao Chen 已提交
154
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} box_wrapper ps_gpu_wrapper)
155
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} common_infer_shape_functions)
156
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} eigen_function)
157
if (WITH_GPU OR WITH_ROCM)
158
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu bert_encoder_functor)
W
Wu Yi 已提交
159
endif()
160
if(WITH_XPU)
161
  cc_test(beam_search_decode_op_xpu_test SRCS beam_search_decode_op_xpu_test.cc DEPS lod_tensor)
162 163
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} xpulib)
endif()
164
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} layer)
H
Huihuang Zheng 已提交
165
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} tensor_formatter)
166
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} op_version_registry)
167 168 169
if (WITH_ASCEND)
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} ascend_wrapper)
endif()
170

171
if (WITH_ASCEND_CL)
172
  cc_test(assign_op_npu_test SRCS assign_op_npu_test.cc DEPS generated_static_op)
173 174 175
  set(COMMON_OP_DEPS ${COMMON_OP_DEPS} npu_op_runner)
endif()

W
Wu Yi 已提交
176 177 178 179 180 181
# FIXME(typhoonzero): operator deps may not needed.
# op_library(unsqueeze_op DEPS reshape_op)
# op_library(squeeze_op DEPS reshape_op)
# op_library(flatten_op DEPS reshape_op)
# op_library(unstack_op DEPS stack_op)
# op_library(tensor_array_to_tensor_op DEPS concat_op)
B
baiyf 已提交
182

W
Wu Yi 已提交
183 184
set(OPERATOR_DEPS ${OPERATOR_DEPS} ${COMMON_OP_DEPS})
set(GLOB_OPERATOR_DEPS ${OPERATOR_DEPS} CACHE INTERNAL "Global Op dependencies")
L
Luo Tao 已提交
185

186
cc_test(test_common_infer_shape_functions SRCS test_common_infer_shape_functions.cc DEPS common_infer_shape_functions ${COMMON_OP_DEPS} activation_op elementwise_add_op softmax_op softmax)
187
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
188
cc_test(assign_op_test SRCS assign_op_test.cc DEPS generated_static_op)
Y
Yu Yang 已提交
189
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor math_function)
Q
Qiao Longfei 已提交
190
cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
Y
Yu Yang 已提交
191
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
192
cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
Z
Zeng Jinle 已提交
193
if (WITH_GPU)
194
    nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator)
Z
Zeng Jinle 已提交
195
    nv_test(test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc test_leaky_relu_grad_grad_functor.cu DEPS tensor device_context eigen3)
196
    nv_test(feed_forward_test SRCS feed_forward_test.cu DEPS elementwise_add_op matmul_op tensor generator)
197 198 199
elseif(WITH_ROCM)
    hip_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator)
    hip_test(test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc test_leaky_relu_grad_grad_functor.cu DEPS tensor device_context eigen3)
Z
Zeng Jinle 已提交
200 201 202
else()
    cc_test(test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc DEPS tensor device_context eigen3)
endif()
203
cc_test(share_buffer_op_cpp_test SRCS share_buffer_op_test.cc DEPS lod_tensor device_context generated_static_op)
204

H
Huihuang Zheng 已提交
205
cc_library(tensor_formatter SRCS tensor_formatter.cc DEPS ${OP_HEADER_DEPS})
S
sneaxiy 已提交
206 207 208 209
if (WITH_PYTHON)
  cc_library(py_func_op SRCS py_func_op.cc DEPS op_registry python pybind)
endif()

210 211
if (WITH_ASCEND_CL)
  cc_test(range_op_npu_test SRCS range_op_npu_test.cc DEPS op_registry range_op scope device_context enforce executor)
212
  cc_test(expand_op_npu_test SRCS expand_op_npu_test.cc DEPS op_registry expand_op eigen_function scope device_context enforce executor compare_op)
213 214
endif()

W
Wu Yi 已提交
215
set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
216
add_subdirectory(benchmark)
217

218
cc_test_old(op_debug_string_test SRCS op_debug_string_test.cc DEPS elementwise_add_op ${COMMON_OP_DEPS})
219 220 221 222
if (WITH_ASCEND_CL)
    cc_test(transpose_op_npu_test SRCS transpose_op_npu_test.cc DEPS op_registry transpose_op scope device_context enforce executor)
endif()

223 224 225

if(WITH_MKLDNN)
include(mkldnn/inplace_op_tests.cmake)
226
include(mkldnn/caching_tests.cmake)
J
Jacek Czaja 已提交
227
include(mkldnn/nhwc_op_tests.cmake)
228
endif()
229 230 231 232 233

if(WITH_UNITY_BUILD)
    # Using Unity Build to compile operators, `register_operator` will cause
    # the unity library to lose some symbols.
    # The specified link dependency needs to be displayed here.
T
tangwei12 已提交
234
    target_link_libraries(paddle_operators_unity ${OP_HEADER_DEPS} ${COMMON_OP_DEPS})
235
endif()
236

B
Baibaifan 已提交
237 238 239
if (WITH_GPU OR WITH_ASCEND_CL)
cc_test(copy_cross_scope_test SRCS copy_cross_scope_test.cc DEPS op_registry copy_cross_scope_op scope device_context enforce executor)
endif()
240 241

copy_if_different(${pybind_file} ${pybind_file_final})
242 243 244 245

if (WITH_CUSTOM_DEVICE)
cc_library(custom_device_common_op_registry SRCS custom_device_common_op_registry.cc DEPS operator)
endif()
246 247 248 249

if(NOT "${OP_LIST}" STREQUAL "")
prune_pybind_h()
endif()