|
| 1 | +set(PART_CUDA_KERNEL_FILES) |
| 2 | +function(op_library TARGET) |
| 3 | + # op_library is a function to create op library. The interface is same as |
| 4 | + # cc_library. But it handle split GPU/CPU code and link some common library |
| 5 | + # for ops. |
| 6 | + set(cc_srcs) |
| 7 | + set(cu_srcs) |
| 8 | + set(hip_cu_srcs) |
| 9 | + set(miopen_hip_cc_srcs) |
| 10 | + set(cu_cc_srcs) |
| 11 | + set(cudnn_cu_cc_srcs) |
| 12 | + set(CUDNN_FILE) |
| 13 | + set(mkldnn_cc_srcs) |
| 14 | + set(MKLDNN_FILE) |
| 15 | + set(op_common_deps operator op_registry math_function) |
| 16 | + set(options "") |
| 17 | + set(oneValueArgs "") |
| 18 | + set(multiValueArgs SRCS DEPS) |
| 19 | + set(pybind_flag 0) |
| 20 | + cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" |
| 21 | + "${multiValueArgs}" ${ARGN}) |
| 22 | + |
| 23 | + list(LENGTH op_library_SRCS op_library_SRCS_len) |
| 24 | + if (${op_library_SRCS_len} EQUAL 0) |
| 25 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc) |
| 26 | + list(APPEND cc_srcs ${TARGET}.cc) |
| 27 | + endif() |
| 28 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu.cc) |
| 29 | + list(APPEND cu_cc_srcs ${TARGET}.cu.cc) |
| 30 | + endif() |
| 31 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu) |
| 32 | + list(APPEND cu_srcs ${TARGET}.cu) |
| 33 | + endif() |
| 34 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu) |
| 35 | + set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu |
| 36 | + ${PART_CUDA_KERNEL_FILES} PARENT_SCOPE) |
| 37 | + list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu) |
| 38 | + endif() |
| 39 | + |
| 40 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu) |
| 41 | + list(APPEND hip_cu_srcs ${TARGET}.hip.cu) |
| 42 | + endif() |
| 43 | + string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}") |
| 44 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc) |
| 45 | + list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc) |
| 46 | + endif() |
| 47 | + if(WITH_AMD_GPU) |
| 48 | + string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}") |
| 49 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc) |
| 50 | + list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc) |
| 51 | + endif() |
| 52 | + endif() |
| 53 | + if(WITH_MKLDNN) |
| 54 | + string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}") |
| 55 | + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc) |
| 56 | + list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc) |
| 57 | + endif() |
| 58 | + endif() |
| 59 | + else() |
| 60 | + foreach(src ${op_library_SRCS}) |
| 61 | + if (${src} MATCHES ".*\\.hip.cu$") |
| 62 | + list(APPEND hip_cu_srcs ${src}) |
| 63 | + elseif (${src} MATCHES ".*\\.cu$") |
| 64 | + list(APPEND cu_srcs ${src}) |
| 65 | + elseif(${src} MATCHES ".*_cudnn_op.cu.cc$") |
| 66 | + list(APPEND cudnn_cu_cc_srcs ${src}) |
| 67 | + elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$") |
| 68 | + list(APPEND miopen_hip_cc_srcs ${src}) |
| 69 | + elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$") |
| 70 | + list(APPEND mkldnn_cc_srcs ${src}) |
| 71 | + elseif(${src} MATCHES ".*\\.cu.cc$") |
| 72 | + list(APPEND cu_cc_srcs ${src}) |
| 73 | + elseif(${src} MATCHES ".*\\.cc$") |
| 74 | + list(APPEND cc_srcs ${src}) |
| 75 | + else() |
| 76 | + message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu") |
| 77 | + endif() |
| 78 | + endforeach() |
| 79 | + endif() |
| 80 | + |
| 81 | + list(LENGTH cc_srcs cc_srcs_len) |
| 82 | + if (${cc_srcs_len} EQUAL 0) |
| 83 | + message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file") |
| 84 | + endif() |
| 85 | + if (WIN32) |
| 86 | + # remove windows unsupported op, because windows has no nccl, no warpctc such ops. |
| 87 | + foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op" |
| 88 | + "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op" |
| 89 | + "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op") |
| 90 | + if ("${TARGET}" STREQUAL "${windows_unsupport_op}") |
| 91 | + return() |
| 92 | + endif() |
| 93 | + endforeach() |
| 94 | + endif(WIN32) |
| 95 | + set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} CACHE INTERNAL "op libs") |
| 96 | + |
| 97 | + list(LENGTH op_library_DEPS op_library_DEPS_len) |
| 98 | + if (${op_library_DEPS_len} GREATER 0) |
| 99 | + set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE) |
| 100 | + endif() |
| 101 | + if (WITH_GPU) |
| 102 | + nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} |
| 103 | + ${op_common_deps}) |
| 104 | + elseif (WITH_AMD_GPU) |
| 105 | + hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS} |
| 106 | + ${op_common_deps}) |
| 107 | + else() |
| 108 | + cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS} |
| 109 | + ${op_common_deps}) |
| 110 | + endif() |
| 111 | + |
| 112 | + # Define operators that don't need pybind here. |
| 113 | + foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" |
| 114 | +"tensor_array_read_write_op" "tensorrt_engine_op") |
| 115 | + if ("${TARGET}" STREQUAL "${manual_pybind_op}") |
| 116 | + set(pybind_flag 1) |
| 117 | + endif() |
| 118 | + endforeach() |
| 119 | + |
| 120 | + # The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h. |
| 121 | + # Note that it's enough to just adding one operator to pybind in a *_op.cc file. |
| 122 | + # And for detail pybind information, please see generated paddle/pybind/pybind.h. |
| 123 | + file(READ ${TARGET}.cc TARGET_CONTENT) |
| 124 | + string(REGEX MATCH "REGISTER_OPERATOR\\(.*REGISTER_OPERATOR\\(" multi_register "${TARGET_CONTENT}") |
| 125 | + string(REGEX MATCH "REGISTER_OPERATOR\\([a-z0-9_]*," one_register "${multi_register}") |
| 126 | + if (one_register STREQUAL "") |
| 127 | + string(REPLACE "_op" "" TARGET "${TARGET}") |
| 128 | + else () |
| 129 | + string(REPLACE "REGISTER_OPERATOR(" "" TARGET "${one_register}") |
| 130 | + string(REPLACE "," "" TARGET "${TARGET}") |
| 131 | + endif() |
| 132 | + |
| 133 | + # pybind USE_NO_KERNEL_OP |
| 134 | + # HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel |
| 135 | + string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}") |
| 136 | + string(REPLACE "_op" "" TARGET "${TARGET}") |
| 137 | + if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "") |
| 138 | + file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n") |
| 139 | + set(pybind_flag 1) |
| 140 | + endif() |
| 141 | + |
| 142 | + # pybind USE_CPU_ONLY_OP |
| 143 | + list(LENGTH cu_srcs cu_srcs_len) |
| 144 | + list(LENGTH cu_cc_srcs cu_cc_srcs_len) |
| 145 | + list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len) |
| 146 | + list(LENGTH hip_cu_srcs hip_cu_srcs_len) |
| 147 | + list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len) |
| 148 | + 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 |
| 149 | + ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0) |
| 150 | + file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n") |
| 151 | + set(pybind_flag 1) |
| 152 | + endif() |
| 153 | + |
| 154 | + # pybind USE_OP_DEVICE_KERNEL for CUDNN |
| 155 | + list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len) |
| 156 | + if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0) |
| 157 | + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n") |
| 158 | + endif() |
| 159 | + |
| 160 | + # pybind USE_OP_DEVICE_KERNEL for MIOPEN |
| 161 | + if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0) |
| 162 | + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n") |
| 163 | + endif() |
| 164 | + |
| 165 | + # pybind USE_OP_DEVICE_KERNEL for MKLDNN |
| 166 | + if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0) |
| 167 | + # Append first implemented MKLDNN activation operator |
| 168 | + if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op") |
| 169 | + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n") |
| 170 | + else() |
| 171 | + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n") |
| 172 | + endif() |
| 173 | + endif() |
| 174 | + |
| 175 | + # pybind USE_OP |
| 176 | + if (${pybind_flag} EQUAL 0) |
| 177 | + # NOTE(*): activation use macro to regist the kernels, set use_op manually. |
| 178 | + if(${TARGET} STREQUAL "activation") |
| 179 | + file(APPEND ${pybind_file} "USE_OP(relu);\n") |
| 180 | + elseif(${TARGET} STREQUAL "fake_dequantize") |
| 181 | + file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n") |
| 182 | + elseif(${TARGET} STREQUAL "fake_quantize") |
| 183 | + file(APPEND ${pybind_file} "USE_OP(fake_quantize_abs_max);\n") |
| 184 | + elseif(${TARGET} STREQUAL "tensorrt_engine_op") |
| 185 | + message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference") |
| 186 | + elseif(${TARGET} STREQUAL "fc") |
| 187 | + # HACK: fc only have mkldnn and cpu, which would mismatch the cpu only condition |
| 188 | + file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n") |
| 189 | + else() |
| 190 | + file(APPEND ${pybind_file} "USE_OP(${TARGET});\n") |
| 191 | + endif() |
| 192 | + endif() |
| 193 | +endfunction() |
| 194 | + |
| 195 | + |
| 196 | +function(register_operators) |
| 197 | + set(options "") |
| 198 | + set(oneValueArgs "") |
| 199 | + set(multiValueArgs EXCLUDES) |
| 200 | + cmake_parse_arguments(register_operators "${options}" "${oneValueArgs}" |
| 201 | + "${multiValueArgs}" ${ARGN}) |
| 202 | + |
| 203 | + file(GLOB OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc") |
| 204 | + string(REPLACE "_mkldnn" "" OPS "${OPS}") |
| 205 | + string(REPLACE ".cc" "" OPS "${OPS}") |
| 206 | + list(REMOVE_DUPLICATES OPS) |
| 207 | + |
| 208 | + foreach(src ${OPS}) |
| 209 | + list(FIND register_operators_EXCLUDES ${src} _index) |
| 210 | + if (${_index} EQUAL -1) |
| 211 | + op_library(${src}) |
| 212 | + endif() |
| 213 | + endforeach() |
| 214 | +endfunction() |
0 commit comments