已合并
【PR】:[fix] [autofuse] remove hccl related code which is unused. #1224
邢智雄创建于 7月7日
【PR】:[fix] [autofuse] remove hccl related code which is unused. #1224
已合并
共 18 个文件变更+16-101
| @@ -58,7 +58,7 @@ repos: | |||
| 58 | - id: codespell | 58 | - id: codespell |
| 59 | args: [ | 59 | args: [ |
| 60 | "-L", | 60 | "-L", |
| 61 | - "CANN,cann,NNAL,nnal,ASCEND,ascend,EnQue,CopyIn,ArchType,AND,ND,tbe,copyin,alog,Vistor", | 61 | + "CANN,cann,NNAL,nnal,ASCEND,ascend,EnQue,CopyIn,ArchType,AND,ND,tbe,copyin,alog,Vistor,BUILDIN,buildin", |
| 62 | "--skip", | 62 | "--skip", |
| 63 | "*.py,*.cpp,*.hpp,*.c,*.h", | 63 | "*.py,*.cpp,*.hpp,*.c,*.h", |
| 64 | ] | 64 | ] |
| @@ -815,7 +815,6 @@ static const char* GetActivityKindString(msptiActivityKind kind) { | |||
| 815 | {MSPTI_ACTIVITY_KIND_MARKER, "MARKER"}, | 815 | {MSPTI_ACTIVITY_KIND_MARKER, "MARKER"}, |
| 816 | {MSPTI_ACTIVITY_KIND_KERNEL, "KERNEL"}, | 816 | {MSPTI_ACTIVITY_KIND_KERNEL, "KERNEL"}, |
| 817 | {MSPTI_ACTIVITY_KIND_API, "API"}, | 817 | {MSPTI_ACTIVITY_KIND_API, "API"}, |
| 818 | - {MSPTI_ACTIVITY_KIND_HCCL, "HCCL"}, | ||
| 819 | {MSPTI_ACTIVITY_KIND_MEMORY, "MEMORY"}, | 818 | {MSPTI_ACTIVITY_KIND_MEMORY, "MEMORY"}, |
| 820 | {MSPTI_ACTIVITY_KIND_MEMSET, "MEMSET"}, | 819 | {MSPTI_ACTIVITY_KIND_MEMSET, "MEMSET"}, |
| 821 | {MSPTI_ACTIVITY_KIND_MEMCPY, "MEMCPY"}, | 820 | {MSPTI_ACTIVITY_KIND_MEMCPY, "MEMCPY"}, |
| @@ -182,7 +182,6 @@ const std::string ATTR_NAME_IS_ROOT_GRAPH = "_is_root_graph"; | |||
| 182 | const std::string ATTR_NAME_NODE_CONNECT_INPUT = "_is_connected_to_data"; | 182 | const std::string ATTR_NAME_NODE_CONNECT_INPUT = "_is_connected_to_data"; |
| 183 | const std::string ATTR_NAME_NODE_CONNECT_OUTPUT = "_is_connected_to_netoutput"; | 183 | const std::string ATTR_NAME_NODE_CONNECT_OUTPUT = "_is_connected_to_netoutput"; |
| 184 | 184 | ||
| 185 | -// Need Map rank id when hccl task init for NPU | ||
| 186 | const std::string ATTR_NAME_NEED_MAP_RANK_ID = "_need_map_rank_id"; | 185 | const std::string ATTR_NAME_NEED_MAP_RANK_ID = "_need_map_rank_id"; |
| 187 | 186 | ||
| 188 | // To be deleted | 187 | // To be deleted |
| @@ -866,7 +865,7 @@ const std::string AVG_POOL_GRAD_OUTPUT_SHAPE = "avg_pool_grad_output_shape"; | |||
| 866 | // Pad | 865 | // Pad |
| 867 | const std::string ATTR_PAD_FORMAT = "attr_pad_format"; | 866 | const std::string ATTR_PAD_FORMAT = "attr_pad_format"; |
| 868 | 867 | ||
| 869 | -// Varible | 868 | +// Variable |
| 870 | const std::string VAR_ATTR_FORMAT = "_var_format"; | 869 | const std::string VAR_ATTR_FORMAT = "_var_format"; |
| 871 | const std::string VAR_ATTR_NAME = "var_name"; | 870 | const std::string VAR_ATTR_NAME = "var_name"; |
| 872 | const std::string VAR_ATTR_FRACTALZ_FORMAT = "FZ"; | 871 | const std::string VAR_ATTR_FRACTALZ_FORMAT = "FZ"; |
| @@ -1160,12 +1159,6 @@ const std::string ATTR_NAME_UNREGST_ATTRLIST = "_unregst_attrlist"; | |||
| 1160 | // used for Horovod | 1159 | // used for Horovod |
| 1161 | const std::string ATTR_INTER_EVENT_IDENTIFY = "event_id"; | 1160 | const std::string ATTR_INTER_EVENT_IDENTIFY = "event_id"; |
| 1162 | const std::string ATTR_HOROVOD_ATTR_REDUCE_TYPE = "reduce_op"; | 1161 | const std::string ATTR_HOROVOD_ATTR_REDUCE_TYPE = "reduce_op"; |
| 1163 | -// used for allreduce tailing optimization | ||
| 1164 | -const std::string ATTR_NAME_HCCL_FUSED_GROUP = "_hccl_fused_group"; | ||
| 1165 | -const std::string ATTR_NAME_HCCL_FUSED_FLAG = "_hccl_fused_node"; | ||
| 1166 | -// used for recording the number of tasks to be issued for each operator | ||
| 1167 | -const std::string ATTR_NAME_HCCL_TASK_NUM = "_hccl_task_num"; | ||
| 1168 | -const std::string ATTR_NAME_HCCL_ATTACHED_TASK_NUM = "_hccl_attached_task_num"; | ||
| 1169 | // used for recording task num of RTS nodes such as MemcpyAsync | 1162 | // used for recording task num of RTS nodes such as MemcpyAsync |
| 1170 | const std::string ATTR_NAME_NODE_SQE_NUM = "_node_sqe_num"; | 1163 | const std::string ATTR_NAME_NODE_SQE_NUM = "_node_sqe_num"; |
| 1171 | // used for parallel group | 1164 | // used for parallel group |
| @@ -1173,8 +1166,6 @@ const std::string ATTR_NAME_PARALLEL_GROUP = "_parallel_group"; | |||
| 1173 | 1166 | ||
| 1174 | const std::string ATTR_NAME_IS_SUPPORT_ADDR_REFRESH = "_is_support_addr_refresh"; | 1167 | const std::string ATTR_NAME_IS_SUPPORT_ADDR_REFRESH = "_is_support_addr_refresh"; |
| 1175 | 1168 | ||
| 1176 | -const std::string ATTR_NAME_HCCL_GROUP_ID_LIST = "_hccl_group_id_list"; | ||
| 1177 | - | ||
| 1178 | // dynamic shape attr | 1169 | // dynamic shape attr |
| 1179 | const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR = "_alloc_fixed_addr"; | 1170 | const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR = "_alloc_fixed_addr"; |
| 1180 | const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR_INDEX = "_alloc_fixed_addr_index"; | 1171 | const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR_INDEX = "_alloc_fixed_addr_index"; |
| @@ -475,11 +475,6 @@ static graphStatus HiddenInputParser(const OpDescPtr &op_desc, const std::string | |||
| 475 | arg_descs.emplace_back(arg); | 475 | arg_descs.emplace_back(arg); |
| 476 | return ge::GRAPH_SUCCESS; | 476 | return ge::GRAPH_SUCCESS; |
| 477 | } | 477 | } |
| 478 | - if (sscanf_s(pattern_str.c_str(), "hi.hcclsk%d*", &arg.ir_idx) == kDigitFormatCnt) { | ||
| 479 | - *reinterpret_cast<uint32_t *>(arg.reserved) = static_cast<uint32_t>(HiddenInputsType::HCCLSUPERKERNEL); | ||
| 480 | - arg_descs.emplace_back(arg); | ||
| 481 | - return ge::GRAPH_SUCCESS; | ||
| 482 | - } | ||
| 483 | GELOGE(ge::GRAPH_FAILED, "Hidden input type [%s] is unsupported.", pattern_str.c_str()); | 478 | GELOGE(ge::GRAPH_FAILED, "Hidden input type [%s] is unsupported.", pattern_str.c_str()); |
| 484 | return ge::GRAPH_FAILED; | 479 | return ge::GRAPH_FAILED; |
| 485 | } | 480 | } |
| @@ -491,10 +486,6 @@ static void HiddenInputSerializer(std::stringstream &ss, const std::string &patt | |||
| 491 | if (*reinterpret_cast<const uint32_t *>(arg_desc.reserved) == static_cast<uint32_t>(HiddenInputsType::TILEFWK)) { | 486 | if (*reinterpret_cast<const uint32_t *>(arg_desc.reserved) == static_cast<uint32_t>(HiddenInputsType::TILEFWK)) { |
| 492 | ss << pattern << ".tilefwk" << arg_desc.ir_idx << "*"; | 487 | ss << pattern << ".tilefwk" << arg_desc.ir_idx << "*"; |
| 493 | } | 488 | } |
| 494 | - if (*reinterpret_cast<const uint32_t *>(arg_desc.reserved) == | ||
| 495 | - static_cast<uint32_t>(HiddenInputsType::HCCLSUPERKERNEL)) { | ||
| 496 | - ss << pattern << ".hcclsk" << arg_desc.ir_idx << "*"; | ||
| 497 | - } | ||
| 498 | return; | 489 | return; |
| 499 | } | 490 | } |
| 500 | 491 | ||
| @@ -68,9 +68,9 @@ const std::map<domi::FrameworkType, std::string> kFmkTypeToString = { | |||
| 68 | }; | 68 | }; |
| 69 | 69 | ||
| 70 | const std::map<domi::ImplyType, std::string> kImplyTypeToString = { | 70 | const std::map<domi::ImplyType, std::string> kImplyTypeToString = { |
| 71 | - {domi::ImplyType::BUILDIN, "buildin"}, {domi::ImplyType::TVM, "tvm"}, {domi::ImplyType::CUSTOM, "custom"}, | 71 | + {domi::ImplyType::BUILDIN, "buildin"}, {domi::ImplyType::TVM, "tvm"}, {domi::ImplyType::CUSTOM, "custom"}, |
| 72 | - {domi::ImplyType::AI_CPU, "ai_cpu"}, {domi::ImplyType::CCE, "cce"}, {domi::ImplyType::GELOCAL, "gelocal"}, | 72 | + {domi::ImplyType::AI_CPU, "ai_cpu"}, {domi::ImplyType::CCE, "cce"}, {domi::ImplyType::GELOCAL, "gelocal"}, |
| 73 | - {domi::ImplyType::HCCL, "hccl"}, {domi::ImplyType::INVALID, "invalid"}}; | 73 | + {domi::ImplyType::INVALID, "invalid"}}; |
| 74 | } // namespace | 74 | } // namespace |
| 75 | 75 | ||
| 76 | std::string TypeUtils::DataTypeToSerialString(const DataType data_type) { | 76 | std::string TypeUtils::DataTypeToSerialString(const DataType data_type) { |
| @@ -201,11 +201,7 @@ GE_ERRORNO_RUNTIME(GE_RTI_CALL_CCE_DESTORY_HANDLE_FAILED, 39, "destory handle fa | |||
| 201 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_CREATE_EVENT_FAILED, 40, "call rutime create event failed"); | 201 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_CREATE_EVENT_FAILED, 40, "call rutime create event failed"); |
| 202 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_EVENT_RECORD_FAILED, 41, "call rutime event record failed"); | 202 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_EVENT_RECORD_FAILED, 41, "call rutime event record failed"); |
| 203 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_STREAM_WAIT_EVENT_FAILED, 42, "call rutime stream wait event failed"); | 203 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_STREAM_WAIT_EVENT_FAILED, 42, "call rutime stream wait event failed"); |
| 204 | -GE_ERRORNO_RUNTIME(GE_RTI_CALL_HCCL_BROADCAST_FAILED, 43, "call hccl hcom broadcast failed"); | ||
| 205 | -GE_ERRORNO_RUNTIME(GE_RTI_CALL_HCCL_ALL_GATHER_FAILED, 44, "call hccl hcom all gather failed"); | ||
| 206 | -GE_ERRORNO_RUNTIME(GE_RTI_CALL_HCCL_ALL_REDUCE_FAILED, 45, "call hccl hcom all reduce failed"); | ||
| 207 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_DESTORY_EVENT_FAILED, 46, "destory rt event failed"); | 204 | GE_ERRORNO_RUNTIME(GE_RTI_CALL_RUNTIME_DESTORY_EVENT_FAILED, 46, "destory rt event failed"); |
| 208 | -GE_ERRORNO_RUNTIME(GE_RTI_CALL_HCCL_REDUCE_SCATTER_FAILED, 47, "call hccl hcom reduce scatter failed"); | ||
| 209 | 205 | ||
| 210 | // Executor module error code definition | 206 | // Executor module error code definition |
| 211 | GE_ERRORNO_EXECUTOR(GE_EXEC_NOT_INIT, 1, "GE Executor is not yet initialized."); | 207 | GE_ERRORNO_EXECUTOR(GE_EXEC_NOT_INIT, 1, "GE Executor is not yet initialized."); |
| @@ -84,7 +84,6 @@ const std::string kTaskTypeInvalid = "TASK_TYPE_INVALID"; | |||
| 84 | const std::string kTaskTypeFftsPlus = "FFTS_PLUS"; | 84 | const std::string kTaskTypeFftsPlus = "FFTS_PLUS"; |
| 85 | const std::string kEngineNameVectorCore = "VectorEngine"; | 85 | const std::string kEngineNameVectorCore = "VectorEngine"; |
| 86 | 86 | ||
| 87 | -const std::string kEngineNameHccl = "ops_kernel_info_hccl"; | ||
| 88 | const std::string kEngineNameRts = "DNN_VM_RTS_OP_STORE"; | 87 | const std::string kEngineNameRts = "DNN_VM_RTS_OP_STORE"; |
| 89 | const std::string kEngineNameHostCpu = "DNN_VM_HOST_CPU_OP_STORE"; | 88 | const std::string kEngineNameHostCpu = "DNN_VM_HOST_CPU_OP_STORE"; |
| 90 | const std::string kEngineNameGeLocal = "DNN_VM_GE_LOCAL_OP_STORE"; | 89 | const std::string kEngineNameGeLocal = "DNN_VM_GE_LOCAL_OP_STORE"; |
| @@ -47,7 +47,6 @@ typedef enum tagHiAiNpuModuleId { | |||
| 47 | HIAI_TOOLCHIAN = 15, | 47 | HIAI_TOOLCHIAN = 15, |
| 48 | HIAI_ALG = 16, | 48 | HIAI_ALG = 16, |
| 49 | HIAI_PROFILING = 17, | 49 | HIAI_PROFILING = 17, |
| 50 | - HIAI_HCCL = 18, | ||
| 51 | HIAI_SIMULATION = 19, | 50 | HIAI_SIMULATION = 19, |
| 52 | HIAI_BIOS = 20, | 51 | HIAI_BIOS = 20, |
| 53 | HIAI_SEC = 21, | 52 | HIAI_SEC = 21, |
| @@ -19,47 +19,12 @@ | |||
| 19 | 19 | ||
| 20 | 20 | ||
| 21 | namespace af { | 21 | namespace af { |
| 22 | -struct HcclDumpInfo { | ||
| 23 | - uint32_t task_id; | ||
| 24 | - uint32_t stream_id; | ||
| 25 | - uint32_t sub_task_type; | ||
| 26 | - void *input_addr; | ||
| 27 | - uint64_t input_size; | ||
| 28 | - void *output_addr; | ||
| 29 | - uint64_t output_size; | ||
| 30 | -}; | ||
| 31 | - | ||
| 32 | struct DvppInfo { | 22 | struct DvppInfo { |
| 33 | OpDescPtr op_desc; | 23 | OpDescPtr op_desc; |
| 34 | std::vector<void *> io_addrs; | 24 | std::vector<void *> io_addrs; |
| 35 | uint32_t sqe[16]; | 25 | uint32_t sqe[16]; |
| 36 | }; | 26 | }; |
| 37 | 27 | ||
| 38 | -// when need to eliminate GETaskKernelHcclInfo, so not need DAVINCI_TRAIN/DAVINCI_CLOUD | ||
| 39 | -struct GETaskKernelHcclInfo { | ||
| 40 | - std::string input_name; | ||
| 41 | - std::string hccl_type; | ||
| 42 | - void *inputDataAddr; | ||
| 43 | - void *outputDataAddr; | ||
| 44 | - void *workSpaceAddr; | ||
| 45 | - int64_t count; | ||
| 46 | - int32_t dataType; | ||
| 47 | - int32_t opType; | ||
| 48 | - int64_t rootId; | ||
| 49 | - uint64_t workSpaceMemSize; | ||
| 50 | - std::vector<int64_t> dims; | ||
| 51 | - std::vector<aclrtStream> hcclStreamList; | ||
| 52 | - std::vector<HcclDumpInfo> hccl_dump_info; | ||
| 53 | - std::vector<void *> global_workspace_addr; | ||
| 54 | - uint32_t hcclQosCfg; | ||
| 55 | - std::vector<void *> inputDataAddrs; | ||
| 56 | - std::vector<void *> outputDataAddrs; | ||
| 57 | - std::vector<void *> workSpaceAddrs; | ||
| 58 | - std::vector<uint64_t> workSpaceMemSizes; | ||
| 59 | - std::vector<int32_t> inputZeroCopyFlags; | ||
| 60 | - std::vector<int32_t> outputZeroCopyFlags; | ||
| 61 | -}; | ||
| 62 | - | ||
| 63 | struct GETaskInfo { | 28 | struct GETaskInfo { |
| 64 | uint32_t id; | 29 | uint32_t id; |
| 65 | uint16_t type; | 30 | uint16_t type; |
| @@ -69,7 +34,6 @@ struct GETaskInfo { | |||
| 69 | void *privateDef; | 34 | void *privateDef; |
| 70 | uint32_t privateDefLen; | 35 | uint32_t privateDefLen; |
| 71 | void *opsKernelStorePtr; | 36 | void *opsKernelStorePtr; |
| 72 | - std::vector<GETaskKernelHcclInfo> kernelHcclInfo; | ||
| 73 | DvppInfo dvpp_info; | 37 | DvppInfo dvpp_info; |
| 74 | bool needRefresh{false}; | 38 | bool needRefresh{false}; |
| 75 | std::vector<void *> rt_attached_streams; | 39 | std::vector<void *> rt_attached_streams; |
| @@ -59,8 +59,7 @@ enum class ModelTaskType : uint32_t { | |||
| 59 | MODEL_TASK_FUSION_START, | 59 | MODEL_TASK_FUSION_START, |
| 60 | MODEL_TASK_FUSION_END, | 60 | MODEL_TASK_FUSION_END, |
| 61 | MODEL_TASK_KERNEL_EX, | 61 | MODEL_TASK_KERNEL_EX, |
| 62 | - MODEL_TASK_HCCL, | 62 | + MODEL_TASK_STREAM_SWITCH = 7, |
| 63 | - MODEL_TASK_STREAM_SWITCH, | ||
| 64 | MODEL_TASK_STREAM_ACTIVE, | 63 | MODEL_TASK_STREAM_ACTIVE, |
| 65 | MODEL_TASK_LABEL_SET, | 64 | MODEL_TASK_LABEL_SET, |
| 66 | MODEL_TASK_LABEL_SWITCH, | 65 | MODEL_TASK_LABEL_SWITCH, |
| @@ -57,8 +57,6 @@ const char_t *const OPTION_EXEC_PROFILING_BPPONIT_OPTIONS = "ge.exec.profilingBp | |||
| 57 | // profiling flag | 57 | // profiling flag |
| 58 | const char_t *const OPTION_EXEC_PROFILING_MODE = "ge.exec.profilingMode"; | 58 | const char_t *const OPTION_EXEC_PROFILING_MODE = "ge.exec.profilingMode"; |
| 59 | const char_t *const OPTION_EXEC_PROFILING_OPTIONS = "ge.exec.profilingOptions"; | 59 | const char_t *const OPTION_EXEC_PROFILING_OPTIONS = "ge.exec.profilingOptions"; |
| 60 | -// Hccl flag, if ge.exec.hcclFlag =1, it means load plugin for opskernel, else:ge.exec.hcclFlag =0 | ||
| 61 | -const char_t *const OPTION_EXEC_HCCL_FLAG = "ge.exec.hcclFlag"; | ||
| 62 | const char_t *const OPTION_EXEC_ATOMIC_FLAG = "ge.exec.enable_atomic"; | 60 | const char_t *const OPTION_EXEC_ATOMIC_FLAG = "ge.exec.enable_atomic"; |
| 63 | const char_t *const OPTION_EXEC_DISABLE_REUSED_MEMORY = "ge.exec.disableReuseMemory"; | 61 | const char_t *const OPTION_EXEC_DISABLE_REUSED_MEMORY = "ge.exec.disableReuseMemory"; |
| 64 | const char_t *const OPTION_EXEC_ENABLE_TAILING_OPTIMIZATION = "ge.exec.isTailingOptimization"; | 62 | const char_t *const OPTION_EXEC_ENABLE_TAILING_OPTIMIZATION = "ge.exec.isTailingOptimization"; |
| @@ -151,8 +149,6 @@ const char_t *const GRAPH_PARALLEL_OPTION_PATH = "ge.graphParallelOptionPath"; | |||
| 151 | const std::string DISTRIBUTED_CLUSTER_BUILD = "ge.distributed_cluster_build"; | 149 | const std::string DISTRIBUTED_CLUSTER_BUILD = "ge.distributed_cluster_build"; |
| 152 | const std::string MODEL_RELATION_CONFIG = "ge.offline_model_relation"; | 150 | const std::string MODEL_RELATION_CONFIG = "ge.offline_model_relation"; |
| 153 | const std::string CLUSTER_CONFIG = "ge.cluster_config"; | 151 | const std::string CLUSTER_CONFIG = "ge.cluster_config"; |
| 154 | -const std::string OPTION_HCCL_COMPILER_OFFLINE = "ge.offline_hccl_compile"; | ||
| 155 | - | ||
| 156 | // option for screen log | 152 | // option for screen log |
| 157 | constexpr const char_t *OPTION_SCREEN_PRINT_MODE = "ge.screen_print_mode"; | 153 | constexpr const char_t *OPTION_SCREEN_PRINT_MODE = "ge.screen_print_mode"; |
| 158 | 154 | ||
| @@ -165,13 +161,11 @@ const char_t *const OPTION_TUNINGPATH = "ge.tuningPath"; | |||
| 165 | const char_t *const OPTION_AOE_CONFIG_FILE = "ge.aoe_config_file"; | 161 | const char_t *const OPTION_AOE_CONFIG_FILE = "ge.aoe_config_file"; |
| 166 | const char_t *const OPTION_SESSION_DEVICE_ID = "ge.session_device_id"; | 162 | const char_t *const OPTION_SESSION_DEVICE_ID = "ge.session_device_id"; |
| 167 | const char_t *const OPTION_DISTRIBUTE_CONFIG = "distribute_config"; | 163 | const char_t *const OPTION_DISTRIBUTE_CONFIG = "distribute_config"; |
| 168 | -const char_t *const OPTION_EXEC_HCCL_EXECUTE_TIMEOUT = "ge.exec.hcclExecuteTimeOut"; | ||
| 169 | const char_t *const OPTION_EXEC_PLACEMENT = "ge.exec.placement"; | 164 | const char_t *const OPTION_EXEC_PLACEMENT = "ge.exec.placement"; |
| 170 | const char_t *const OPTION_IS_VAR_INIT_GRAPH = "ge.exec.isVarInitGraph"; | 165 | const char_t *const OPTION_IS_VAR_INIT_GRAPH = "ge.exec.isVarInitGraph"; |
| 171 | const char_t *const OPTION_EXEC_OVERFLOW = "ge.exec.overflow"; | 166 | const char_t *const OPTION_EXEC_OVERFLOW = "ge.exec.overflow"; |
| 172 | const char_t *const OPTION_DATAFLOW_DEPLOY_INFO_PATH = "ge.experiment.data_flow_deploy_info_path"; | 167 | const char_t *const OPTION_DATAFLOW_DEPLOY_INFO_PATH = "ge.experiment.data_flow_deploy_info_path"; |
| 173 | const char_t *const OPTION_MOMORY_POOL_THRESHOLD = "ge.experiment.memory_pool_threshold"; | 168 | const char_t *const OPTION_MOMORY_POOL_THRESHOLD = "ge.experiment.memory_pool_threshold"; |
| 174 | -const char_t *const OPTION_HCCL_ALGORITHM = "HCCL_algorithm"; | ||
| 175 | const char_t *const OPTION_ES_CLUSTER_CONFIG = "ge.esClusterConfig"; | 169 | const char_t *const OPTION_ES_CLUSTER_CONFIG = "ge.esClusterConfig"; |
| 176 | const char_t *const OPTION_EXECUTE_TIMES = "execute_times"; | 170 | const char_t *const OPTION_EXECUTE_TIMES = "execute_times"; |
| 177 | const char_t *const OPTION_ES_MAX_REMOTEOP_NUM_PER_STREAM = "es_max_remoteop_num_per_stream"; | 171 | const char_t *const OPTION_ES_MAX_REMOTEOP_NUM_PER_STREAM = "es_max_remoteop_num_per_stream"; |
| @@ -339,8 +333,8 @@ const std::string DDK_VERSION_FLAG = "ge.DDK_version"; | |||
| 339 | const std::string GE_FE_FLAG = "ge.feFlag"; | 333 | const std::string GE_FE_FLAG = "ge.feFlag"; |
| 340 | 334 | ||
| 341 | // Configure stream max parallel num only by Session constructor options param, | 335 | // Configure stream max parallel num only by Session constructor options param, |
| 342 | -// its value should be stream:int, such as "DNN_V100:2,DNN_HCCL:3", | 336 | +// its value should be stream:int, such as "DNN_V100:2", |
| 343 | -// default value is "1", such as "DNN_V100:1,DNN_HCCL:1" | 337 | +// default value is "1", such as "DNN_V100:1" |
| 344 | // this option is to obtain stream max parallel num | 338 | // this option is to obtain stream max parallel num |
| 345 | const std::string STREAM_MAX_PARALLEL_NUM = "ge.streamMaxParallelNum"; | 339 | const std::string STREAM_MAX_PARALLEL_NUM = "ge.streamMaxParallelNum"; |
| 346 | 340 | ||
| @@ -542,7 +542,6 @@ enum class ImplyType : unsigned int { | |||
| 542 | AI_CPU, | 542 | AI_CPU, |
| 543 | CCE, | 543 | CCE, |
| 544 | GELOCAL, | 544 | GELOCAL, |
| 545 | - HCCL, | ||
| 546 | INVALID = 0xFFFFFFFF, | 545 | INVALID = 0xFFFFFFFF, |
| 547 | }; | 546 | }; |
| 548 | using char_t = ge::char_t; | 547 | using char_t = ge::char_t; |
| @@ -103,8 +103,6 @@ enum class ScalarType : uint32_t { | |||
| 103 | INVALID_DTYPE = static_cast<uint32_t>(-1), | 103 | INVALID_DTYPE = static_cast<uint32_t>(-1), |
| 104 | }; | 104 | }; |
| 105 | 105 | ||
| 106 | -enum class HcclServerType : uint32_t { AICPU = 0, AICORE = 1, CCU = 2, MAX }; | ||
| 107 | - | ||
| 108 | union ScalarNum { | 106 | union ScalarNum { |
| 109 | uint64_t value_u64; | 107 | uint64_t value_u64; |
| 110 | int64_t value_i64; | 108 | int64_t value_i64; |
| @@ -458,9 +456,6 @@ class OpMC2Def { | |||
| 458 | OpMC2Def(const OpMC2Def &mc2_def); | 456 | OpMC2Def(const OpMC2Def &mc2_def); |
| 459 | ~OpMC2Def(); | 457 | ~OpMC2Def(); |
| 460 | OpMC2Def &operator=(const OpMC2Def &mc2_def); | 458 | OpMC2Def &operator=(const OpMC2Def &mc2_def); |
| 461 | - OpMC2Def &HcclGroup(const char *value); | ||
| 462 | - OpMC2Def &HcclGroup(std::vector<const char *> value); | ||
| 463 | - void HcclServerType(enum HcclServerType type, const char *soc = nullptr); | ||
| 464 | 459 | ||
| 465 | private: | 460 | private: |
| 466 | friend class AclnnFallBackGenerator; | 461 | friend class AclnnFallBackGenerator; |
| @@ -473,8 +468,6 @@ class OpMC2Def { | |||
| 473 | friend class OpDefImpl; | 468 | friend class OpDefImpl; |
| 474 | friend class OpMC2DefImpl; | 469 | friend class OpMC2DefImpl; |
| 475 | 470 | ||
| 476 | - std::vector<ge::AscendString> &GetHcclGroups(void) const; | ||
| 477 | - ops::HcclServerType GetHcclServerType(const ge::AscendString &soc_version = "") const; | ||
| 478 | std::unique_ptr<OpMC2DefImpl> impl_; | 471 | std::unique_ptr<OpMC2DefImpl> impl_; |
| 479 | }; | 472 | }; |
| 480 | 473 | ||
| @@ -1166,12 +1166,6 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAM | |||
| 1166 | // Used for support Horovod | 1166 | // Used for support Horovod |
| 1167 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_INTER_EVENT_IDENTIFY; | 1167 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_INTER_EVENT_IDENTIFY; |
| 1168 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_HOROVOD_ATTR_REDUCE_TYPE; | 1168 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_HOROVOD_ATTR_REDUCE_TYPE; |
| 1169 | -// for gradient group | ||
| 1170 | -GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_HCCL_FUSED_GROUP; | ||
| 1171 | -GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_HCCL_FUSED_FLAG; | ||
| 1172 | -// used for recording the number of tasks to be issued for each operator | ||
| 1173 | -GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_HCCL_TASK_NUM; | ||
| 1174 | -GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_HCCL_ATTACHED_TASK_NUM; | ||
| 1175 | // used for recording task num of RTS nodes such as MemcpyAsync | 1169 | // used for recording task num of RTS nodes such as MemcpyAsync |
| 1176 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_NODE_SQE_NUM; | 1170 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_NODE_SQE_NUM; |
| 1177 | // for parallel group | 1171 | // for parallel group |
| @@ -1179,8 +1173,6 @@ GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAM | |||
| 1179 | 1173 | ||
| 1180 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_IS_SUPPORT_ADDR_REFRESH; | 1174 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_IS_SUPPORT_ADDR_REFRESH; |
| 1181 | 1175 | ||
| 1182 | -GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_NAME_HCCL_GROUP_ID_LIST; | ||
| 1183 | - | ||
| 1184 | // dynamic shape attrs | 1176 | // dynamic shape attrs |
| 1185 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR; | 1177 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR; |
| 1186 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR_INDEX; | 1178 | GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY extern const std::string ATTR_DYNAMIC_SHAPE_FIXED_ADDR_INDEX; |
| @@ -81,7 +81,7 @@ class SlogStub { | |||
| 81 | std::map<int, std::string> level_str = {{DLOG_DEBUG, "[DEBUG]"}, {DLOG_INFO, "[INFO]"}, {DLOG_WARN, "[WARNING]"}, | 81 | std::map<int, std::string> level_str = {{DLOG_DEBUG, "[DEBUG]"}, {DLOG_INFO, "[INFO]"}, {DLOG_WARN, "[WARNING]"}, |
| 82 | {DLOG_ERROR, "[ERROR]"}, {DLOG_DEBUG, "[TRACE]"}, {DLOG_NULL, "[NULL]"}, | 82 | {DLOG_ERROR, "[ERROR]"}, {DLOG_DEBUG, "[TRACE]"}, {DLOG_NULL, "[NULL]"}, |
| 83 | {DLOG_EVENT, "[EVENT]"}}; | 83 | {DLOG_EVENT, "[EVENT]"}}; |
| 84 | - std::map<int, std::string> module_id_str = {{GE, "GE"}, {FE, "FE"}, {HCCL, "HCCL"}, {RUNTIME, "RUNTIME"}}; | 84 | + std::map<int, std::string> module_id_str = {{GE, "GE"}, {FE, "FE"}, {RUNTIME, "RUNTIME"}}; |
| 85 | }; | 85 | }; |
| 86 | } // namespace ge | 86 | } // namespace ge |
| 87 | 87 | ||
| @@ -5224,8 +5224,8 @@ inline Tensor HcomAllToAllVC(const Tensor &send_data, const Tensor &send_count_m | |||
| 5224 | return out; | 5224 | return out; |
| 5225 | } | 5225 | } |
| 5226 | inline Tensor HcomCollRemoteLookup(const Tensor &table_id, const Tensor &keys, int64_t tag, int64_t max_num, | 5226 | inline Tensor HcomCollRemoteLookup(const Tensor &table_id, const Tensor &keys, int64_t tag, int64_t max_num, |
| 5227 | - int64_t embedding_dim, int64_t insert_option = 0, | 5227 | + int64_t embedding_dim, int64_t insert_option = 0, const char *group = "world_group", |
| 5228 | - const char *group = "hccl_world_group", int64_t flags = 0) { | 5228 | + int64_t flags = 0) { |
| 5229 | auto out = EsHcomCollRemoteLookup(table_id.GetEsbTensor(), keys.GetEsbTensor(), tag, max_num, embedding_dim, | 5229 | auto out = EsHcomCollRemoteLookup(table_id.GetEsbTensor(), keys.GetEsbTensor(), tag, max_num, embedding_dim, |
| 5230 | insert_option, group, flags); | 5230 | insert_option, group, flags); |
| 5231 | return out; | 5231 | return out; |
| @@ -5240,7 +5240,7 @@ struct HcomCollRemoteLookupPairedOutput { | |||
| 5240 | inline HcomCollRemoteLookupPairedOutput HcomCollRemoteLookupPaired(const Tensor &table_id, const Tensor &keys, | 5240 | inline HcomCollRemoteLookupPairedOutput HcomCollRemoteLookupPaired(const Tensor &table_id, const Tensor &keys, |
| 5241 | int64_t tag, int64_t max_num, int64_t embedding_dim, | 5241 | int64_t tag, int64_t max_num, int64_t embedding_dim, |
| 5242 | int64_t insert_option = 0, | 5242 | int64_t insert_option = 0, |
| 5243 | - const char *group = "hccl_world_group", | 5243 | + const char *group = "world_group", |
| 5244 | int64_t flags = 0) { | 5244 | int64_t flags = 0) { |
| 5245 | auto out = EsHcomCollRemoteLookupPaired(table_id.GetEsbTensor(), keys.GetEsbTensor(), tag, max_num, embedding_dim, | 5245 | auto out = EsHcomCollRemoteLookupPaired(table_id.GetEsbTensor(), keys.GetEsbTensor(), tag, max_num, embedding_dim, |
| 5246 | insert_option, group, flags); | 5246 | insert_option, group, flags); |
| @@ -5256,14 +5256,14 @@ struct HcomCollRemoteLookupUniquedAndPairedOutput { | |||
| 5256 | inline HcomCollRemoteLookupUniquedAndPairedOutput HcomCollRemoteLookupUniquedAndPaired( | 5256 | inline HcomCollRemoteLookupUniquedAndPairedOutput HcomCollRemoteLookupUniquedAndPaired( |
| 5257 | const Tensor &table_id, const Tensor &keys, const Tensor &key_num_input, const Tensor &unique_indices, | 5257 | const Tensor &table_id, const Tensor &keys, const Tensor &key_num_input, const Tensor &unique_indices, |
| 5258 | const Tensor &key_count, int64_t tag, int64_t max_num, int64_t embedding_dim, int64_t flags, | 5258 | const Tensor &key_count, int64_t tag, int64_t max_num, int64_t embedding_dim, int64_t flags, |
| 5259 | - int64_t insert_option = 0, const char *group = "hccl_world_group") { | 5259 | + int64_t insert_option = 0, const char *group = "world_group") { |
| 5260 | auto out = EsHcomCollRemoteLookupUniquedAndPaired( | 5260 | auto out = EsHcomCollRemoteLookupUniquedAndPaired( |
| 5261 | table_id.GetEsbTensor(), keys.GetEsbTensor(), key_num_input.GetEsbTensor(), unique_indices.GetEsbTensor(), | 5261 | table_id.GetEsbTensor(), keys.GetEsbTensor(), key_num_input.GetEsbTensor(), unique_indices.GetEsbTensor(), |
| 5262 | key_count.GetEsbTensor(), tag, max_num, embedding_dim, flags, insert_option, group); | 5262 | key_count.GetEsbTensor(), tag, max_num, embedding_dim, flags, insert_option, group); |
| 5263 | return {out.values, out.indices, out.num_uniqued, out.ps_segments, out.ps_segments_num}; | 5263 | return {out.values, out.indices, out.num_uniqued, out.ps_segments, out.ps_segments_num}; |
| 5264 | } | 5264 | } |
| 5265 | inline Tensor HcomCollRemoteUpdate(const Tensor &table_id, const Tensor &keys, const Tensor &values, int64_t tag, | 5265 | inline Tensor HcomCollRemoteUpdate(const Tensor &table_id, const Tensor &keys, const Tensor &values, int64_t tag, |
| 5266 | - int64_t max_num, int64_t embedding_dim, const char *group = "hccl_world_group") { | 5266 | + int64_t max_num, int64_t embedding_dim, const char *group = "world_group") { |
| 5267 | auto out = EsHcomCollRemoteUpdate(table_id.GetEsbTensor(), keys.GetEsbTensor(), values.GetEsbTensor(), tag, max_num, | 5267 | auto out = EsHcomCollRemoteUpdate(table_id.GetEsbTensor(), keys.GetEsbTensor(), values.GetEsbTensor(), tag, max_num, |
| 5268 | embedding_dim, group); | 5268 | embedding_dim, group); |
| 5269 | return out; | 5269 | return out; |
| @@ -5271,7 +5271,7 @@ inline Tensor HcomCollRemoteUpdate(const Tensor &table_id, const Tensor &keys, c | |||
| 5271 | inline Tensor HcomCollRemoteUpdatePaired(const Tensor &table_id, const Tensor &keys, const Tensor &values, | 5271 | inline Tensor HcomCollRemoteUpdatePaired(const Tensor &table_id, const Tensor &keys, const Tensor &values, |
| 5272 | const Tensor &indices, const Tensor &num_uniqued, const Tensor &ps_segments, | 5272 | const Tensor &indices, const Tensor &num_uniqued, const Tensor &ps_segments, |
| 5273 | const Tensor &ps_segments_num, const Tensor &global_step, int64_t tag, | 5273 | const Tensor &ps_segments_num, const Tensor &global_step, int64_t tag, |
| 5274 | - int64_t max_num, int64_t embedding_dim, const char *group = "hccl_world_group", | 5274 | + int64_t max_num, int64_t embedding_dim, const char *group = "world_group", |
| 5275 | int64_t padding_key = 0, int64_t flags = 0) { | 5275 | int64_t padding_key = 0, int64_t flags = 0) { |
| 5276 | auto out = EsHcomCollRemoteUpdatePaired( | 5276 | auto out = EsHcomCollRemoteUpdatePaired( |
| 5277 | table_id.GetEsbTensor(), keys.GetEsbTensor(), values.GetEsbTensor(), indices.GetEsbTensor(), | 5277 | table_id.GetEsbTensor(), keys.GetEsbTensor(), values.GetEsbTensor(), indices.GetEsbTensor(), |
| @@ -795,7 +795,6 @@ static const char* GetActivityKindString(msptiActivityKind kind) { | |||
| 795 | {MSPTI_ACTIVITY_KIND_MARKER, "MARKER"}, | 795 | {MSPTI_ACTIVITY_KIND_MARKER, "MARKER"}, |
| 796 | {MSPTI_ACTIVITY_KIND_KERNEL, "KERNEL"}, | 796 | {MSPTI_ACTIVITY_KIND_KERNEL, "KERNEL"}, |
| 797 | {MSPTI_ACTIVITY_KIND_API, "API"}, | 797 | {MSPTI_ACTIVITY_KIND_API, "API"}, |
| 798 | - {MSPTI_ACTIVITY_KIND_HCCL, "HCCL"}, | ||
| 799 | {MSPTI_ACTIVITY_KIND_MEMORY, "MEMORY"}, | 798 | {MSPTI_ACTIVITY_KIND_MEMORY, "MEMORY"}, |
| 800 | {MSPTI_ACTIVITY_KIND_MEMSET, "MEMSET"}, | 799 | {MSPTI_ACTIVITY_KIND_MEMSET, "MEMSET"}, |
| 801 | {MSPTI_ACTIVITY_KIND_MEMCPY, "MEMCPY"}, | 800 | {MSPTI_ACTIVITY_KIND_MEMCPY, "MEMCPY"}, |