CMakeLists.txt 13.8 KB
Newer Older
L
Luo Tao 已提交
1
file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
2
string(REPLACE "_mkldnn" "" GENERAL_OPS "${GENERAL_OPS}")
L
Luo Tao 已提交
3
string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
4
list(REMOVE_DUPLICATES GENERAL_OPS)
5
set(DEPS_OPS "")
6 7
set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
Y
Yu Yang 已提交
8 9 10 11
function(op_library TARGET)
    # op_library is a function to create op library. The interface is same as
    # cc_library. But it handle split GPU/CPU code and link some common library
    # for ops.
L
Luo Tao 已提交
12
    set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
Y
Yu Yang 已提交
13 14
    set(cc_srcs)
    set(cu_srcs)
S
sabreshao 已提交
15 16
    set(hip_cu_srcs)
    set(miopen_hip_cc_srcs)
17
    set(cu_cc_srcs)
C
chengduoZH 已提交
18
    set(cudnn_cu_cc_srcs)
C
chengduoZH 已提交
19
    set(CUDNN_FILE)
20 21
    set(mkldnn_cc_srcs)
    set(MKLDNN_FILE)
Y
update  
Yancey1989 已提交
22
    set(op_common_deps operator op_registry math_function)
Y
Yu Yang 已提交
23 24 25
    set(options "")
    set(oneValueArgs "")
    set(multiValueArgs SRCS DEPS)
L
Luo Tao 已提交
26
    set(pybind_flag 0)
Y
Yu Yang 已提交
27 28 29
    cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
            "${multiValueArgs}" ${ARGN})

30 31 32 33
    list(LENGTH op_library_SRCS op_library_SRCS_len)
    if (${op_library_SRCS_len} EQUAL 0)
        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
            list(APPEND cc_srcs ${TARGET}.cc)
Y
Yu Yang 已提交
34
        endif()
35 36 37
        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc)
            list(APPEND cu_cc_srcs ${TARGET}.cu.cc)
        endif()
38
        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
39
            list(APPEND cu_srcs ${TARGET}.cu)
40
        endif()
S
sabreshao 已提交
41 42 43
        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu)
            list(APPEND hip_cu_srcs ${TARGET}.hip.cu)
        endif()
C
chengduoZH 已提交
44 45 46 47
        string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}")
        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
            list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
        endif()
S
sabreshao 已提交
48 49 50 51 52 53
        if(WITH_AMD_GPU)
            string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}")
            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc)
                list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc)
            endif()
        endif()
54 55 56 57 58 59
        if(WITH_MKLDNN)
            string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
                list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
            endif()
        endif()
60 61
    else()
        foreach(src ${op_library_SRCS})
S
sabreshao 已提交
62 63 64
            if (${src} MATCHES ".*\\.hip.cu$")
                list(APPEND hip_cu_srcs ${src})
            elseif (${src} MATCHES ".*\\.cu$")
65
                list(APPEND cu_srcs ${src})
C
chengduoZH 已提交
66 67
            elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
                list(APPEND cudnn_cu_cc_srcs ${src})
S
sabreshao 已提交
68 69
            elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$")
                list(APPEND miopen_hip_cc_srcs ${src})
70 71
            elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
                list(APPEND mkldnn_cc_srcs ${src})
72 73
            elseif(${src} MATCHES ".*\\.cu.cc$")
                list(APPEND cu_cc_srcs ${src})
74 75 76 77 78 79 80
            elseif(${src} MATCHES ".*\\.cc$")
                list(APPEND cc_srcs ${src})
            else()
                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
            endif()
        endforeach()
    endif()
Y
Yu Yang 已提交
81 82 83 84 85 86

    list(LENGTH cc_srcs cc_srcs_len)
    if (${cc_srcs_len} EQUAL 0)
        message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
    endif()

87 88 89 90
    list(LENGTH op_library_DEPS op_library_DEPS_len)
    if (${op_library_DEPS_len} GREATER 0)
        set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
    endif()
Y
Yu Yang 已提交
91
    if (WITH_GPU)
92
        nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
Y
Yu Yang 已提交
93
                ${op_common_deps})
94
    elseif (WITH_AMD_GPU)
S
sabreshao 已提交
95 96
        hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
                ${op_common_deps})
Y
Yu Yang 已提交
97
    else()
98 99
        cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
            ${op_common_deps})
Y
Yu Yang 已提交
100
    endif()
L
Luo Tao 已提交
101

102
    # Define operators that don't need pybind here.
103
    foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
104 105 106 107
        if ("${TARGET}" STREQUAL "${manual_pybind_op}")
            set(pybind_flag 1)
        endif()
    endforeach()
Q
qijun 已提交
108

S
sabreshao 已提交
109
    # The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h.
110 111
    # Note that it's enough to just adding one operator to pybind in a *_op.cc file.
    # And for detail pybind information, please see generated paddle/pybind/pybind.h.
112
    file(READ ${TARGET}.cc TARGET_CONTENT)
Y
Yang Yang 已提交
113 114
    string(REGEX MATCH "REGISTER_OPERATOR\\(.*REGISTER_OPERATOR\\(" multi_register "${TARGET_CONTENT}")
    string(REGEX MATCH "REGISTER_OPERATOR\\([a-z0-9_]*," one_register "${multi_register}")
115 116 117
    if (one_register STREQUAL "")
        string(REPLACE "_op" "" TARGET "${TARGET}")
    else ()
Y
Yang Yang 已提交
118
        string(REPLACE "REGISTER_OPERATOR(" "" TARGET "${one_register}")
119
        string(REPLACE "," "" TARGET "${TARGET}")
Y
Yu Yang 已提交
120 121
    endif()

L
Luo Tao 已提交
122
    # pybind USE_NO_KERNEL_OP
T
typhoonzero 已提交
123
    # HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel
T
typhoonzero 已提交
124
    string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}")
L
Luo Tao 已提交
125 126 127 128 129 130 131 132
    string(REPLACE "_op" "" TARGET "${TARGET}")
    if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
        set(pybind_flag 1)
    endif()

    # pybind USE_CPU_ONLY_OP
    list(LENGTH cu_srcs cu_srcs_len)
133
    list(LENGTH cu_cc_srcs cu_cc_srcs_len)
134
    list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
S
sabreshao 已提交
135 136 137 138
    list(LENGTH hip_cu_srcs hip_cu_srcs_len)
    list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len)
    if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND
        ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0)
L
Luo Tao 已提交
139 140 141 142
        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
        set(pybind_flag 1)
    endif()

C
chengduoZH 已提交
143 144
    # pybind USE_OP_DEVICE_KERNEL for CUDNN
    list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
C
chengduoZH 已提交
145
    if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
C
chengduoZH 已提交
146 147 148
        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
    endif()

S
sabreshao 已提交
149 150 151 152 153
    # pybind USE_OP_DEVICE_KERNEL for MIOPEN
    if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0)
        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n")
    endif()

154 155
    # pybind USE_OP_DEVICE_KERNEL for MKLDNN
    if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
156 157 158 159
      # Append first implemented MKLDNN activation operator
      if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
      else()
160
        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
161
      endif()
162 163
    endif()

L
Luo Tao 已提交
164 165
    # pybind USE_OP
    if (${pybind_flag} EQUAL 0)
D
dzhwinter 已提交
166 167 168
      # NOTE(*): activation use macro to regist the kernels, set use_op manually.
      if(${TARGET} STREQUAL "activation")
        file(APPEND ${pybind_file} "USE_OP(relu);\n")
169 170
      elseif(${TARGET} STREQUAL "fake_dequantize")
        file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n")
171 172
      elseif(${TARGET} STREQUAL "tensorrt_engine_op")
          message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference")
D
dzhwinter 已提交
173
      else()
L
Luo Tao 已提交
174
        file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
D
dzhwinter 已提交
175
      endif()
L
Luo Tao 已提交
176
    endif()
Y
Yu Yang 已提交
177 178
endfunction()

Q
qijun 已提交
179
add_subdirectory(math)
D
Dong Zhihong 已提交
180
add_subdirectory(nccl)
181

L
Luo Tao 已提交
182 183
if(WITH_GPU)
    op_library(nccl_op DEPS nccl_common)
184
    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
L
Luo Tao 已提交
185 186 187
else()
    set(DEPS_OPS ${DEPS_OPS} nccl_op)
endif()
Q
QI JUN 已提交
188

G
gongweibao 已提交
189
set(DISTRIBUTE_DEPS "")
T
typhoonzero 已提交
190
if(WITH_DISTRIBUTE)
191
    add_subdirectory(distributed)
Y
Yancey1989 已提交
192
    
G
gongweibao 已提交
193 194
    set(DISTRIBUTE_DEPS "")
    if(WITH_GRPC)
X
Xin Pan 已提交
195
        set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
G
gongweibao 已提交
196
    else()
X
Xin Pan 已提交
197
        set(DISTRIBUTE_DEPS sendrecvop_brpc brpc leveldb snappystream snappy protobuf ssl crypto zlib node)
G
gongweibao 已提交
198 199 200 201 202 203 204 205 206 207 208 209
        if(WITH_BRPC_RDMA)
            find_library(IBVERBS_LIBRARY NAMES ibverbs)
            ADD_LIBRARY(ibverbs SHARED IMPORTED GLOBAL)
            SET_PROPERTY(TARGET ibverbs PROPERTY IMPORTED_LOCATION ${IBVERBS_LIBRARY})


            find_library(RDMACM_LIBRARY NAMES rdmacm)
            ADD_LIBRARY(rdmacm SHARED IMPORTED GLOBAL)
            SET_PROPERTY(TARGET rdmacm PROPERTY IMPORTED_LOCATION ${RDMACM_LIBRARY})

            set(DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} ibverbs rdmacm)
        endif()
G
gongweibao 已提交
210 211
    endif()

212
    set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
T
merge  
tangwei12 已提交
213
    foreach(dist_op "prefetch_op" "checkpoint_notify_op" "listen_and_serv_op" "send_op" "recv_op" "send_barrier_op" "fetch_barrier_op")
Y
Yancey1989 已提交
214 215 216 217
        op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS})
        set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    endforeach()
    
Y
yuyang18 已提交
218 219 220
    #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
    #        listen_and_serv_op sum_op executor SERIAL)
W
Wu Yi 已提交
221
    if(WITH_GPU)
X
Xin Pan 已提交
222
        set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
G
gongweibao 已提交
223
        cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op ${DISTRIBUTE_DEPS} executor SERIAL)
G
gongweibao 已提交
224 225 226 227 228
        if(WITH_GRPC)
            op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc)
        else()
            op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_brpc)
        endif()
W
Wu Yi 已提交
229 230 231 232
        set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
    else()
        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
    endif()
233
else()
T
tangwei12 已提交
234
    set(DEPS_OPS ${DEPS_OPS}  checkpoint_notify_op prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op)
T
typhoonzero 已提交
235 236
endif()

237 238
op_library(cross_entropy_op DEPS cross_entropy)
op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
239 240
op_library(softmax_op DEPS softmax)
op_library(sequence_softmax_op DEPS softmax)
241
if (WITH_GPU AND TENSORRT_FOUND)
242
    op_library(tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter)
243
    nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc
244
      DEPS tensorrt_engine_op
245
      analysis)
246 247
else()
    set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
248
endif()
Q
QI JUN 已提交
249 250
op_library(sum_op DEPS selected_rows_functor)
op_library(sgd_op DEPS selected_rows_functor)
Y
Yan Chunwei 已提交
251
op_library(print_op DEPS lod_tensor)
Q
QI JUN 已提交
252
op_library(adagrad_op DEPS selected_rows_functor)
W
wanghaox 已提交
253
op_library(maxout_op DEPS maxouting)
S
sweetsky0901 已提交
254
op_library(unpool_op DEPS unpooling)
255
op_library(pool_op DEPS pooling)
C
chengduoZH 已提交
256
op_library(pool_with_index_op DEPS pooling)
257 258 259 260
op_library(lod_rank_table_op DEPS lod_rank_table)
op_library(lod_tensor_to_array_op DEPS lod_rank_table_op)
op_library(array_to_lod_tensor_op DEPS lod_rank_table_op)
op_library(max_sequence_len_op DEPS lod_rank_table)
C
chengduoZH 已提交
261
op_library(sequence_conv_op DEPS context_project)
262
op_library(sequence_pool_op DEPS sequence_pooling)
D
dangqingqing 已提交
263
op_library(lstm_op DEPS sequence2batch lstm_compute)
W
weixing02 已提交
264
op_library(hierarchical_sigmoid_op DEPS matrix_bit_code)
265
op_library(lstmp_op DEPS sequence2batch lstm_compute)
G
guosheng 已提交
266
op_library(gru_op DEPS sequence2batch gru_compute)
267
op_library(recurrent_op DEPS executor)
C
chengduoZH 已提交
268
op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
C
chengduoZH 已提交
269
op_library(cos_sim_op DEPS cos_sim_functor)
Y
Yang Yang 已提交
270
op_library(parallel_do_op DEPS executor)
271
op_library(unsqueeze_op DEPS reshape_op)
Y
yuyang18 已提交
272
op_library(squeeze_op DEPS reshape_op)
273 274

if (WITH_GPU)
275
    op_library(conv_op DEPS vol2col depthwise_conv im2col)
276
else()
277
    op_library(conv_op DEPS vol2col im2col)
278
endif()
279
op_library(conv_transpose_op DEPS vol2col im2col)
280

武毅 已提交
281 282 283
# FIXME(typhoonzero): save/load depends lodtensor serialization functions
op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
284 285
op_library(save_combine_op DEPS lod_tensor)
op_library(load_combine_op DEPS lod_tensor)
286
op_library(concat_op DEPS concat)
武毅 已提交
287

T
Thuan Nguyen 已提交
288 289 290 291 292
# FIXME(thuan): Move CSP operators to paddle/fluid/framework/operators/concurrency
add_subdirectory(concurrency)
op_library(channel_send_op DEPS concurrency)
op_library(channel_recv_op DEPS concurrency)

293
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
294 295 296 297 298 299 300

# The fully connected layer is deleted when the WITH_MKLDNN flag is OFF
# Because the fully connected layer has only one MKLDNN's operator
if(NOT WITH_MKLDNN)
    list(REMOVE_ITEM GENERAL_OPS fc_op)
endif(NOT WITH_MKLDNN)

L
Luo Tao 已提交
301
foreach(src ${GENERAL_OPS})
302
    op_library(${src})
L
Luo Tao 已提交
303
endforeach()
304

305
file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
306

Y
FIX CI  
Yu Yang 已提交
307 308 309 310
add_subdirectory(reader)
foreach(src ${READER_LIBRARY})
    set(OP_LIBRARY ${src} ${OP_LIBRARY})
endforeach()
311

B
baiyf 已提交
312 313 314 315 316
add_subdirectory(detection)
foreach(src ${DETECTION_LIBRARY})
    set(OP_LIBRARY ${src} ${OP_LIBRARY})
endforeach()

L
Luo Tao 已提交
317
set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
G
gongweibao 已提交
318
set(GLOB_DISTRIBUTE_DEPS ${DISTRIBUTE_DEPS} CACHE INTERNAL "distributed dependency")
L
Luo Tao 已提交
319

320 321
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
Q
Qiao Longfei 已提交
322
cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
Y
Yan Chunwei 已提交
323
cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_search_op)
Y
Yi Wang 已提交
324
cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
Y
Yu Yang 已提交
325
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
326
cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
Q
QI JUN 已提交
327
nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
G
gongweibao 已提交
328
nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)