/** * Copyright 2020 Huawei Technologies Co., Ltd * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #include "graph/preprocess/graph_preprocess.h" #include #include #include #include "common/formats/format_transfers/format_transfer_fractal_nz.h" #include "common/formats/format_transfers/format_transfer_nchw_nc1hwc0.h" #include "common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.h" #include "common/formats/format_transfers/format_transfer_transpose.h" #include "common/formats/utils/formats_trans_utils.h" #include "common/util/error_manager/error_manager.h" #include "common/helper/model_helper.h" #include "common/math/math_util.h" #include "common/op/ge_op_utils.h" #include "graph/common/ge_call_wrapper.h" #include "graph/common/local_context.h" #include "graph/common/transop_util.h" #include "graph/ge_context.h" #include "graph/shape_refiner.h" #include "graph/manager/graph_var_manager.h" #include "graph/manager/util/rt_context_util.h" #include "graph/optimize/graph_optimize.h" #include "graph/passes/addn_pass.h" #include "graph/passes/aicpu_constant_folding_pass.h" #include "graph/passes/assert_pass.h" #include "ge/ge_api_types.h" #include "graph/passes/common_subexpression_elimination_pass.h" #include "graph/passes/cond_pass.h" #include "graph/passes/cond_remove_pass.h" #include "graph/passes/constant_folding_pass.h" #include "graph/passes/dimension_adjust_pass.h" #include "graph/passes/dimension_compute_pass.h" #include "graph/passes/dropout_pass.h" #include "graph/passes/enter_pass.h" #include "graph/passes/for_pass.h" #include "graph/passes/guarantee_const_pass.h" #include "graph/passes/hccl_memcpy_pass.h" #include "graph/passes/hccl_group_pass.h" #include "graph/passes/identity_pass.h" #include "graph/passes/infershape_pass.h" #include "graph/passes/merge_pass.h" #include "graph/passes/net_output_pass.h" #include "graph/passes/no_use_reshape_remove_pass.h" #include "graph/passes/parallel_concat_start_op_pass.h" #include "graph/passes/placeholder_with_default_pass.h" #include "graph/passes/prevent_gradient_pass.h" #include "graph/passes/print_op_pass.h" #include "graph/passes/prune_pass.h" #include "graph/passes/replace_transshape_pass.h" #include "graph/passes/replace_with_empty_const_pass.h" #include "graph/passes/resource_pair_add_control_pass.h" #include "graph/passes/resource_pair_remove_control_pass.h" #include "graph/passes/save_pass.h" #include "graph/passes/shape_operate_op_remove_pass.h" #include "graph/passes/snapshot_pass.h" #include "graph/passes/stop_gradient_pass.h" #include "graph/passes/switch_dead_branch_elimination.h" #include "graph/passes/unused_const_pass.h" #include "graph/passes/var_is_initialized_op_pass.h" #include "graph/passes/variable_prepare_op_pass.h" #include "graph/preprocess/insert_op/util_insert_aipp_op.h" #include "graph/utils/type_utils.h" #include "inc/pass_manager.h" #include "init/gelib.h" #include "multi_batch_copy_graph.h" #include "graph/passes/data_pass.h" #include "graph/passes/mark_agnostic_pass.h" namespace ge { namespace { static std::map output_type_str_to_datatype = { {"FP32", ge::DT_FLOAT}, {"FP16", ge::DT_FLOAT16}, {"INT8", ge::DT_INT8}, {"INT16", ge::DT_INT16}, {"UINT16", ge::DT_UINT16}, {"UINT8", ge::DT_UINT8}, {"INT32", ge::DT_INT32}, {"INT64", ge::DT_INT64}, {"UINT32", ge::DT_UINT32}, {"UINT64", ge::DT_UINT64}, {"DOUBLE", ge::DT_DOUBLE}}; const char *const kMbatchSwitchnName = "mbatch-switch-name"; // the size of user defined output datatype or format string after split by ":". const size_t kUserDefinedElementCount = 2; const int kDataOutIndex = 0; const int64_t kInvalidDynaimcDimsType = -1; OpDescPtr CreateTensorShape(const GeTensorDesc &data_tensor) { GeTensorPtr tensor = MakeShared(); if (tensor == nullptr) { GELOGE(INTERNAL_ERROR, "Create shared ptr for GeTensor failed"); return nullptr; } tensor->MutableTensorDesc().SetDataType(DT_INT32); tensor->MutableTensorDesc().SetFormat(FORMAT_ND); auto dst_ge_shape = data_tensor.GetShape(); auto dim_cnt = static_cast(dst_ge_shape.GetDimNum()); if (dim_cnt == 0) { // if the dim_cnt is 0, the tensor is a scalar tensor->MutableTensorDesc().SetShape(GeShape()); int32_t dst_shape = 1; if (tensor->SetData(reinterpret_cast(&dst_shape), sizeof(int32_t)) != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "tensor set data failed"); return nullptr; } } else { tensor->MutableTensorDesc().SetShape(GeShape(std::vector({dim_cnt}))); unique_ptr dst_shape(new (std::nothrow) int32_t[dim_cnt]()); if (dst_shape == nullptr) { GELOGE(INTERNAL_ERROR, "Create unique ptr failed"); return nullptr; } for (int64_t i = 0; i < dim_cnt; ++i) { dst_shape[i] = dst_ge_shape.GetDim(static_cast(i)); } GE_IF_BOOL_EXEC( tensor->SetData(reinterpret_cast(dst_shape.get()), dim_cnt * sizeof(int32_t)) != GRAPH_SUCCESS, GELOGE(INTERNAL_ERROR, "tensor set data failed"); return nullptr;) } GELOGD("Create shape input dim [%s]", dst_ge_shape.ToString().c_str()); return OpDescUtils::CreateConstOp(tensor); } void AddTransNodeAttr(const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output, OpDescPtr &op_desc) { // For format transfer node, the IR definition has src/dst format attrs if (node_type == TRANSDATA) { GE_IF_BOOL_EXEC( !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_SRC_FORMAT, TypeUtils::FormatToSerialString(input.GetFormat())), GELOGW("SetStr FORMAT_TRANSFER_SRC_FORMAT failed");) GE_IF_BOOL_EXEC( !AttrUtils::SetStr(op_desc, FORMAT_TRANSFER_DST_FORMAT, TypeUtils::FormatToSerialString(output.GetFormat())), GELOGW("SetStr FORMAT_TRANSFER_DST_FORMAT failed");) } // For TransposeD node, the IR definition has perm attrs if (node_type == TRANSPOSED) { Format src_format = input.GetFormat(); Format dst_format = output.GetFormat(); std::vector perm_arg; GE_CHK_BOOL_EXEC_WARN(formats::GetPermByForamt(src_format, dst_format, perm_arg) == SUCCESS, return, "Get perm by foramt failed."); GE_CHK_BOOL_EXEC_WARN(AttrUtils::SetListInt(op_desc, PERMUTE_ATTR_PERM, perm_arg), return, "SetStr PERMUTE_ATTR_PERM failed") } // For cast node, the IR definition has src/dst attrs if (node_type == CAST) { GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_SRCT, static_cast(input.GetDataType())), GELOGW("SetInt CAST_ATTR_SRCT failed");) GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DSTT, static_cast(output.GetDataType())), GELOGW("SetInt CAST_ATTR_DSTT failed");) GE_IF_BOOL_EXEC(!AttrUtils::SetInt(op_desc, CAST_ATTR_DST_TYPE, static_cast(output.GetDataType())), GELOGW("SetInt CAST_ATTR_DST_TYPE failed");) GE_IF_BOOL_EXEC(!AttrUtils::SetBool(op_desc, CAST_ATTR_TRUNCATE, false), GELOGW("SetBool CAST_ATTR_TRUNCATE failed");) } } NodePtr CreateTransNode(const std::string &name, const std::string &node_type, const GeTensorDesc &input, const GeTensorDesc &output, NodePtr &node) { if (node == nullptr) { GELOGE(PARAM_INVALID, "node is null."); return nullptr; } auto graph = node->GetOwnerComputeGraph(); if (graph == nullptr) { GELOGE(PARAM_INVALID, "Owner graph is null, node name:%s.", node->GetName().c_str()); return nullptr; } auto index = TransOpUtil::GetTransOpDataIndex(node_type); if (index < 0) { ErrorManager::GetInstance().ATCReportErrMessage( "E19025", {"situation", "reason"}, {"The trans node type[" + node_type + "]", "it must be " + TransOpUtil::TransopMapToString()}); GELOGE(INTERNAL_ERROR, "The trans node type %s does not exists", node_type.c_str()); return nullptr; } OpDescPtr op_desc = MakeShared(name, node_type); if (op_desc == nullptr) { GELOGE(INTERNAL_ERROR, "Create shared ptr for OpDesc failed"); return nullptr; } // for data dump GE_IF_BOOL_EXEC( !AttrUtils::SetListStr(op_desc, ATTR_NAME_DATA_DUMP_ORIGIN_OP_NAMES, std::move(std::vector())), GELOGW("CreateTransNode: SetListStr failed");) // Default single input and single output auto ret = op_desc->AddInputDesc(input); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to add input desc when create node %s type %s", name.c_str(), node_type.c_str()); return nullptr; } ret = op_desc->AddOutputDesc(output); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to add output desc when create node %s type %s", name.c_str(), node_type.c_str()); return nullptr; } AddTransNodeAttr(node_type, input, output, op_desc); NodePtr shape_node = nullptr; if (node_type == RESHAPE) { auto shape_desc = CreateTensorShape(output); if (shape_desc == nullptr) { GELOGE(INTERNAL_ERROR, "Failed to add shape for reshape %s, can not create the shape input", node->GetName().c_str()); return nullptr; } ret = op_desc->AddInputDesc(shape_desc->GetOutputDesc(0)); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to add the first input for reshape %s", name.c_str()); return nullptr; } shape_node = graph->AddNode(shape_desc); if (shape_node == nullptr) { GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the shape to graph", name.c_str()); return nullptr; } } auto trans_node = graph->AddNode(op_desc); if (trans_node == nullptr) { GELOGE(INTERNAL_ERROR, "Failed to add trans node %s to graph", name.c_str()); return nullptr; } if (node_type == RESHAPE) { if (GraphUtils::AddEdge(shape_node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(1)) != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to add shape node for reshape %s, can not add the edge", name.c_str()); return nullptr; } } return trans_node; } Status RecoverOneTransNodeForVar(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node, NodePtr &trans_node) { GE_CHECK_NOTNULL(node); trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.output, trans_node_info.input, node); if (trans_node == nullptr) { return INTERNAL_ERROR; } auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {}, {0}); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to replace out anchors when recover trans node for %s type %s", node->GetName().c_str(), node->GetType().c_str()); return INTERNAL_ERROR; } ret = GraphUtils::AddEdge(node->GetOutDataAnchor(0), trans_node->GetInDataAnchor(0)); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to connect node %s to trans node %s", node->GetName().c_str(), trans_node->GetName().c_str()); return INTERNAL_ERROR; } ret = GraphUtils::MoveOutCtrlEdges(node, trans_node); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to move out control edges from %s to %s when recover trans node.", node->GetName().c_str(), trans_node->GetName().c_str()); return INTERNAL_ERROR; } return SUCCESS; } Status RecoverOneTransNodeForVarRef(const std::string &name, const TransNodeInfo &trans_node_info, NodePtr node, NodePtr &trans_node) { GE_CHECK_NOTNULL(node); trans_node = CreateTransNode(name, trans_node_info.node_type, trans_node_info.input, trans_node_info.output, node); if (trans_node == nullptr) { return INTERNAL_ERROR; } auto ret = GraphUtils::ReplaceNodeDataAnchors(trans_node, node, {0}, {}); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to replace int anchors when recover trans node for %s type %s", node->GetName().c_str(), node->GetType().c_str()); return INTERNAL_ERROR; } ret = GraphUtils::AddEdge(trans_node->GetOutDataAnchor(0), node->GetInDataAnchor(0)); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to connect trans node %s to node %s", trans_node->GetName().c_str(), node->GetName().c_str()); return INTERNAL_ERROR; } ret = GraphUtils::MoveInCtrlEdges(node, trans_node); if (ret != GRAPH_SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to move int control edges from %s to %s when recover trans node.", node->GetName().c_str(), trans_node->GetName().c_str()); return INTERNAL_ERROR; } return SUCCESS; } Status UpdateVarFormats(const NodePtr &var, const GeTensorDesc &tensor_desc) { GE_IF_BOOL_EXEC(var == nullptr, GELOGW("node : var is nullptr"); return INTERNAL_ERROR); GE_CHECK_NOTNULL(var->GetOpDesc()); if (var->GetOpDesc()->GetOutputsSize() > 0) { auto output_desc = var->GetOpDesc()->GetOutputDesc(0); output_desc.SetFormat(tensor_desc.GetFormat()); output_desc.SetDataType(tensor_desc.GetDataType()); output_desc.SetShape(tensor_desc.GetShape()); output_desc.SetOriginFormat(tensor_desc.GetOriginFormat()); output_desc.SetOriginDataType(tensor_desc.GetOriginDataType()); output_desc.SetOriginShape(tensor_desc.GetOriginShape()); GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateOutputDesc(0, output_desc) != GRAPH_SUCCESS, GELOGE(INTERNAL_ERROR, "UpdateOutputDesc failed"); return INTERNAL_ERROR;); } if (var->GetOpDesc()->GetInputsSize() > 0) { auto desc = var->GetOpDesc()->GetInputDesc(0); desc.SetFormat(tensor_desc.GetFormat()); desc.SetDataType(tensor_desc.GetDataType()); desc.SetShape(tensor_desc.GetShape()); desc.SetOriginFormat(tensor_desc.GetOriginFormat()); desc.SetOriginDataType(tensor_desc.GetOriginDataType()); desc.SetOriginShape(tensor_desc.GetOriginShape()); GE_IF_BOOL_EXEC(var->GetOpDesc()->UpdateInputDesc(0, desc) != GRAPH_SUCCESS, GELOGE(INTERNAL_ERROR, "UpdateInputDesc failed"); return INTERNAL_ERROR;) } return SUCCESS; } Status RecoverTransRoadForVar(const NodePtr &var, const VarTransRoad &road) { GE_CHECK_NOTNULL(var); int index = 0; NodePtr last_node = var; for (auto iter = road.rbegin(); iter != road.rend(); ++iter) { auto trans_name = var->GetName() + "_trans_" + std::to_string(index++); auto ret = RecoverOneTransNodeForVar(trans_name, *iter, last_node, last_node); if (ret != SUCCESS) { ErrorManager::GetInstance().ATCReportErrMessage( "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type}); GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(), index, iter->node_type.c_str()); return INTERNAL_ERROR; } // set stream_label OpDescPtr var_desc = var->GetOpDesc(); GE_CHECK_NOTNULL(var_desc); std::string stream_label; (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label); if (!stream_label.empty()) { GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed"); } GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)), return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed."); GELOGD("Recover trans node %s type %s success", trans_name.c_str(), iter->node_type.c_str()); } if (road.empty()) { return SUCCESS; } return UpdateVarFormats(var, road.rbegin()->output); } Status RecoverTransRoadForVarRef(const std::set &nodes, const VarTransRoad &road) { for (auto &var : nodes) { GE_CHECK_NOTNULL(var); int index = 0; NodePtr last_node = var; GELOGI("Recover trans nodes for variable ref %s", var->GetName().c_str()); for (auto iter = road.rbegin(); iter != road.rend(); ++iter) { auto trans_name = var->GetName() + "_trans_" + std::to_string(index++); auto ret = RecoverOneTransNodeForVarRef(trans_name, *iter, last_node, last_node); if (ret != SUCCESS) { ErrorManager::GetInstance().ATCReportErrMessage( "E15001", {"variable", "index", "type"}, {var->GetName(), std::to_string(index), iter->node_type}); GELOGE(INTERNAL_ERROR, "Failed to recover trans node for variable %s, index %d, type %s", var->GetName().c_str(), index, iter->node_type.c_str()); return INTERNAL_ERROR; } // set stream_label OpDescPtr var_desc = var->GetOpDesc(); GE_CHECK_NOTNULL(var_desc); std::string stream_label; (void)AttrUtils::GetStr(var_desc, ATTR_NAME_STREAM_LABEL, stream_label); if (!stream_label.empty()) { GE_CHK_STATUS_RET(SetStreamLabel(last_node, stream_label), "set stream label failed"); } GE_CHK_BOOL_EXEC((ge::AttrUtils::SetBool(last_node->GetOpDesc(), ge::ATTR_INSERTED_BY_GE, true)), return INTERNAL_ERROR, "Set attr ATTR_INSERTED_BY_GE failed."); } if (!(road.empty()) && (UpdateVarFormats(var, road.rbegin()->output) != SUCCESS)) { return INTERNAL_ERROR; } } return SUCCESS; } using VarNamesToRefs = std::map>; VarNamesToRefs CollectVarNamesToRefs(const ComputeGraphPtr &graph) { VarNamesToRefs names_to_refs; std::string var_name; if (graph == nullptr) { GELOGE(PARAM_INVALID, "graph is null."); return names_to_refs; } for (auto &node : graph->GetAllNodes()) { if (node->GetType() != VARIABLE) { continue; } if (AttrUtils::GetStr(node->GetOpDesc(), REF_VAR_SRC_VAR_NAME, var_name)) { (void)names_to_refs[var_name].insert(node); } } return names_to_refs; } Status TransferShape2NC1HWC0(Format src_format, const std::vector &src_shape, DataType dt, Format dst_format, std::vector &dst_shape) { if (src_format == FORMAT_NCHW) { formats::FormatTransferNchwNc1hwc0 transfer; if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) { GELOGE(INTERNAL_ERROR, "TransShape failed"); return FAILED; } } else if (src_format == FORMAT_NHWC) { formats::FormatTransferNhwcNc1hwc0 transfer; if (transfer.TransShape(src_format, src_shape, dt, dst_format, dst_shape) != SUCCESS) { GELOGE(INTERNAL_ERROR, "TransShape failed"); return FAILED; } } return SUCCESS; } Status ModifyInputFormatAndShape(NodePtr &node_ptr) { GE_CHECK_NOTNULL(node_ptr); auto op_desc = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); const GeTensorDescPtr &input = op_desc->MutableInputDesc(0); GE_CHECK_NOTNULL(input); ge::Format old_format = input->GetFormat(); std::vector old_shape = input->GetShape().GetDims(); ge::DataType dt = input->GetDataType(); std::vector dst_shape_dims; if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Trans shape failed"); return FAILED; } input->SetFormat(FORMAT_NC1HWC0); input->SetShape(ge::GeShape(dst_shape_dims)); auto output = op_desc->MutableOutputDesc(0); GE_CHECK_NOTNULL(output); output->SetFormat(FORMAT_NC1HWC0); output->SetShape(ge::GeShape(dst_shape_dims)); int64_t size = 0; graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size); if (graph_status != ge::GRAPH_SUCCESS) { GELOGE(graph_status, "GetTensorSizeInBytes failed!"); return FAILED; } ge::TensorUtils::SetSize(*output, size); ge::TensorUtils::SetSize(*input, size); return SUCCESS; } Status ModifyFormatAndShapeForSingleTensor(const GeTensorDescPtr &input_output) { GE_CHECK_NOTNULL(input_output); ge::Format old_format = input_output->GetFormat(); std::vector old_shape = input_output->GetShape().GetDims(); ge::DataType dt = input_output->GetDataType(); std::vector dst_shape_dims; if (TransferShape2NC1HWC0(old_format, old_shape, dt, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Trans shape failed"); return FAILED; } input_output->SetFormat(FORMAT_NC1HWC0); input_output->SetShape(ge::GeShape(dst_shape_dims)); return SUCCESS; } Status ModifyDataNetOutputFormatAndShape(OpDescPtr &op_desc, uint32_t index, Format storage_format, vector &dst_shape_dims) { GE_CHECK_NOTNULL(op_desc); const GeTensorDescPtr &input = op_desc->MutableInputDesc(index); GE_CHECK_NOTNULL(input); ge::Format old_format = input->GetFormat(); std::vector old_shape = input->GetShape().GetDims(); input->SetShape(ge::GeShape(dst_shape_dims)); input->SetFormat(storage_format); auto output = op_desc->MutableOutputDesc(index); GE_CHECK_NOTNULL(output); output->SetShape(ge::GeShape(dst_shape_dims)); output->SetFormat(storage_format); if (!output->MutableShape().IsUnknownShape()) { int64_t size = 0; graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(*output, size); if (graph_status != ge::GRAPH_SUCCESS) { GELOGE(graph_status, "GetTensorSizeInBytes failed!"); return FAILED; } ge::TensorUtils::SetSize(*input, size); ge::TensorUtils::SetSize(*output, size); GELOGI("Modify Data NetOutput format and shape success, node:%s, index:%d, old_shape:%s, old_Format:%s, " "new_shape:%s, new_format:%s, new_size:%lu", op_desc->GetName().c_str(), index, formats::JoinToString(old_shape).c_str(), ge::TypeUtils::FormatToSerialString(old_format).c_str(), formats::JoinToString(dst_shape_dims).c_str(), ge::TypeUtils::FormatToSerialString(storage_format).c_str(), size); } return SUCCESS; } Status CheckIfDynamicBatchScene(NodePtr &data_node, bool &is_dynamic_batch, NodePtr &switchn_node) { is_dynamic_batch = false; std::string related_node_name; if (AttrUtils::GetStr(data_node->GetOpDesc(), kMbatchSwitchnName, related_node_name)) { if (related_node_name.empty()) { ErrorManager::GetInstance().ATCReportErrMessage( "E15002", {"opname", "value", "reason"}, {data_node->GetName(), "flag", "but the value is empty"}); GELOGE(INTERNAL_ERROR, "The data node %s has switchn node flag, but the value is empty", data_node->GetName().c_str()); return INTERNAL_ERROR; } for (const NodePtr &next_node : data_node->GetOutNodes()) { if (next_node->GetName() == related_node_name) { switchn_node = next_node; break; } } if (switchn_node == nullptr) { ErrorManager::GetInstance().ATCReportErrMessage( "E15002", {"opname", "value", "reason"}, {data_node->GetName(), related_node_name, "but can not find it on the graph"}); GELOGE(INTERNAL_ERROR, "The data node %s has switchn node %s, but can not find it on the graph", data_node->GetName().c_str(), related_node_name.c_str()); return INTERNAL_ERROR; } is_dynamic_batch = true; } return SUCCESS; } bool CheckOpType(const NodePtr &node, const std::string type) { if (node->GetType() == type) { return true; } return false; } Status CheckIfNeedSetNdFormat(const NodePtr &node_ptr) { auto op = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op); auto inputDescsPtr = op->GetAllInputsDescPtr(); auto outputDescsPtr = op->GetAllOutputsDescPtr(); ge::Format format = ge::FORMAT_ND; // if user set shape larger than 4, inferformat may set NCHW or NHWC, GE should set ND before FE // process, otherwise fe will insert transdata. for (auto &inputDescPtr : inputDescsPtr) { GE_CHECK_NOTNULL(inputDescPtr); if ((inputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) && ((inputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (inputDescPtr->GetFormat() == ge::FORMAT_NHWC))) { GELOGI("The node inputdesc [%s] format need to be set ND", op->GetName().c_str()); inputDescPtr->SetFormat(format); inputDescPtr->SetOriginFormat(format); } } for (auto &outputDescPtr : outputDescsPtr) { GE_CHECK_NOTNULL(outputDescPtr); if ((outputDescPtr->GetShape().GetDims().size() > ge::DIM_DEFAULT_SIZE) && ((outputDescPtr->GetFormat() == ge::FORMAT_NCHW) || (outputDescPtr->GetFormat() == ge::FORMAT_NHWC))) { GELOGI("The node outputdesc [%s] format need to be set ND", op->GetName().c_str()); outputDescPtr->SetFormat(format); outputDescPtr->SetOriginFormat(format); } } return SUCCESS; } // A new function ending in 'DynShape' has been added for the dynamic shape processing. // In the dynamic shape process, transnode insertion by FE is advanced to the stage of whole // graph optimization, GE only sets the final data_type/format/shape information for variable, // data and netoutput, and no longer inserts the transnode. Status ProcessInputDtDynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node, DataType &dt_set) { GE_CHECK_NOTNULL(node_ptr); auto op_desc = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); const GeTensorDescPtr &input = op_desc->MutableInputDesc(0); GE_CHECK_NOTNULL(input); ge::DataType src_dtype = input->GetDataType(); if (src_dtype == dt_set) { GELOGI("The node name, %s dtype is fp16", node_ptr->GetName().c_str()); return SUCCESS; } input->SetDataType(dt_set); int64_t input_shape_size = 0; int64_t output_shape_size = 0; ge::graphStatus input_graph_status = ge::TensorUtils::GetTensorSizeInBytes(*input, input_shape_size); ge::graphStatus output_graph_status = ge::TensorUtils::GetTensorMemorySizeInBytes(*input, output_shape_size); if (input_graph_status != ge::GRAPH_SUCCESS && output_graph_status != ge::GRAPH_SUCCESS) { GELOGE(GRAPH_FAILED, "GetTensorSize failed!"); return FAILED; } ge::TensorUtils::SetSize(*input, input_shape_size); const GeTensorDescPtr &output = op_desc->MutableOutputDesc(0); GE_CHECK_NOTNULL(output); output->SetDataType(dt_set); ge::TensorUtils::SetSize(*output, output_shape_size); if (is_dynamic_batch) { GELOGI("The node [%s] dtype set fp16", switchn_node->GetName().c_str()); auto switchn_op_desc = switchn_node->GetOpDesc(); GE_CHECK_NOTNULL(switchn_op_desc); auto switchn_input = switchn_op_desc->MutableInputDesc(0); GE_CHECK_NOTNULL(switchn_input); switchn_input->SetDataType(dt_set); for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) { const GeTensorDescPtr &switchn_output = switchn_op_desc->MutableOutputDesc(i); GE_CHECK_NOTNULL(switchn_output); switchn_output->SetDataType(dt_set); } } return SUCCESS; } Status ProcessInputNC1HWC0DynShape(NodePtr &node_ptr, bool &is_dynamic_batch, NodePtr &switchn_node) { GE_CHECK_NOTNULL(node_ptr); auto op_desc = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); const GeTensorDescPtr &input = op_desc->MutableInputDesc(0); GE_CHECK_NOTNULL(input); ge::Format old_format = input->GetFormat(); ge::GeShape old_shape = input->GetShape(); bool support = ((old_format == FORMAT_NC1HWC0) || (old_format == FORMAT_NCHW) || (old_format == FORMAT_NHWC)); if (!support) { ErrorManager::GetInstance().ATCReportErrMessage( "E19014", {"opname", "value", "reason"}, {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]", "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"}); GELOGE(INTERNAL_ERROR, "The format [%s] is unsupported", TypeUtils::FormatToSerialString(old_format).c_str()); return FAILED; } if (ModifyInputFormatAndShape(node_ptr) != SUCCESS) { GELOGE(INTERNAL_ERROR, "modify format and shape failed"); return FAILED; } if (is_dynamic_batch) { auto switchn_op_desc = switchn_node->GetOpDesc(); GE_CHECK_NOTNULL(switchn_op_desc); const GeTensorDescPtr &switchn_input = switchn_op_desc->MutableInputDesc(0); if (ModifyFormatAndShapeForSingleTensor(switchn_input) != SUCCESS) { GELOGE(INTERNAL_ERROR, "modify format and shape failed"); return FAILED; } for (uint32_t i = 0; i < switchn_node->GetAllOutDataAnchorsSize(); ++i) { auto switchn_output = switchn_op_desc->MutableOutputDesc(i); GE_CHECK_NOTNULL(switchn_output); old_format = switchn_output->GetFormat(); old_shape = switchn_output->GetShape(); if (ModifyFormatAndShapeForSingleTensor(switchn_output) != SUCCESS) { GELOGE(INTERNAL_ERROR, "modify format and shape failed"); return FAILED; } } } return SUCCESS; } Status ProcessDataNodeDynShape(NodePtr &node_ptr) { auto op_desc = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); string set_dt_str; if (!ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_DATATYPE, set_dt_str)) { return SUCCESS; } DataType dt_set = TypeUtils::SerialStringToDataType(set_dt_str); GELOGI("input_fp16 is found, the node name is %s.", node_ptr->GetName().c_str()); bool is_dynamic_batch = false; NodePtr switchn_node = nullptr; if (CheckIfDynamicBatchScene(node_ptr, is_dynamic_batch, switchn_node)) { GELOGE(INTERNAL_ERROR, "CheckIfDynamicBatchScene failed"); return FAILED; } if (ProcessInputDtDynShape(node_ptr, is_dynamic_batch, switchn_node, dt_set) != SUCCESS) { GELOGE(INTERNAL_ERROR, "ProcessInputFP16 failed"); return FAILED; } // check if need to set format string set_format; bool ret = ge::AttrUtils::GetStr(node_ptr->GetOpDesc(), ATTR_ATC_USER_DEFINE_FORMAT, set_format); if (ret && (!set_format.empty()) && TypeUtils::SerialStringToFormat(set_format) == FORMAT_NC1HWC0) { GELOGI("The format of node [%s] should be set NC1HWC0.", node_ptr->GetName().c_str()); if (ProcessInputNC1HWC0DynShape(node_ptr, is_dynamic_batch, switchn_node) != SUCCESS) { GELOGE(INTERNAL_ERROR, "ProcessInputNC1HWC0 failed"); return FAILED; } } return SUCCESS; } Status GetStorageFormatAndShape(OpDescPtr &op_desc, const GeTensorDescPtr &tensor_desc_ptr, Format &storage_format, vector &dst_shape_dims) { GE_CHECK_NOTNULL(op_desc); GE_CHECK_NOTNULL(tensor_desc_ptr); storage_format = FORMAT_RESERVED; int64_t format = FORMAT_RESERVED; dst_shape_dims.clear(); if (ge::AttrUtils::GetInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_FORMAT, format)) { storage_format = static_cast(format); vector storage_shape; if (ge::AttrUtils::GetListInt(*tensor_desc_ptr, ATTR_NAME_STORAGE_SHAPE, storage_shape)) { for (auto dim : storage_shape) { dst_shape_dims.push_back(static_cast(dim)); } GELOGI("Update node by storage format, node: [%s], storage_format: [%s], storage_shape:[%s]", op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str(), formats::JoinToString(storage_shape).c_str()); } else { ErrorManager::GetInstance().ATCReportErrMessage( "15003", {"opname", "format"}, {op_desc->GetName(), TypeUtils::FormatToSerialString(storage_format)}); GELOGE(PARAM_INVALID, "Update node by storage format failed, storage_shape not set. " "node: [%s], storage_format [%s]", op_desc->GetName().c_str(), TypeUtils::FormatToSerialString(storage_format).c_str()); return FAILED; } ge::Format old_format = tensor_desc_ptr->GetFormat(); auto old_shape = tensor_desc_ptr->GetShape().GetDims(); if (old_format == storage_format && old_shape == dst_shape_dims) { GELOGI("Update node by storage format, not changed."); storage_format = FORMAT_RESERVED; return SUCCESS; } } return SUCCESS; } Status ProcessNetoutputNodeFp16Nc1hwc0DynShape(GeTensorDesc &src_desc, GeTensorDescPtr &net_output_input_desc, NodePtr &node) { bool is_dynamic = CheckOpType(node, MERGE); auto src_op_desc = node->GetOpDesc(); GE_CHECK_NOTNULL(src_op_desc); ge::GeShape src_shape = src_desc.GetShape(); ge::Format src_format = src_desc.GetFormat(); net_output_input_desc->SetDataType(DT_FLOAT16); if (is_dynamic) { auto merge_output = src_op_desc->MutableOutputDesc(0); GE_CHECK_NOTNULL(merge_output); merge_output->SetDataType(DT_FLOAT16); for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) { auto merge_input = src_op_desc->MutableInputDesc(i); GE_CHECK_NOTNULL(merge_input); merge_input->SetDataType(DT_FLOAT16); } } std::vector dst_shape_dims; std::vector src_shape_dims = src_shape.GetDims(); if (TransferShape2NC1HWC0(src_format, src_shape_dims, DT_FLOAT16, FORMAT_NC1HWC0, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Trans shape failed"); return FAILED; } ge::GeShape dst_shape(dst_shape_dims); net_output_input_desc->SetFormat(FORMAT_NC1HWC0); net_output_input_desc->SetShape(dst_shape); if (is_dynamic) { auto merge_out = src_op_desc->MutableOutputDesc(0); GE_CHECK_NOTNULL(merge_out); if (ModifyFormatAndShapeForSingleTensor(merge_out) != SUCCESS) { GELOGE(INTERNAL_ERROR, "modify format and shape failed"); return FAILED; } for (uint32_t i = 0; i < node->GetAllInDataAnchorsSize(); ++i) { auto merge_in = src_op_desc->MutableInputDesc(i); GE_CHECK_NOTNULL(merge_in); if (ModifyFormatAndShapeForSingleTensor(merge_in) != SUCCESS) { GELOGE(INTERNAL_ERROR, "modify format and shape failed"); return FAILED; } } } return SUCCESS; } bool NeedUpdateDtByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index, ge::DataType &dt) { GE_CHECK_NOTNULL(netout_desc); vector output_dt_str; if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_DATATYPE, output_dt_str)) { for (auto dt_str : output_dt_str) { vector dt_str_split = StringUtils::Split(dt_str, ':'); if (dt_str_split.size() == kUserDefinedElementCount) { if (dt_str_split[0] == to_string(index)) { dt = TypeUtils::SerialStringToDataType(dt_str_split[1]); GELOGI("Find netoutput node output %u datatype should be set %s .", index, TypeUtils::DataTypeToSerialString(dt).c_str()); return true; } } } } return false; } bool NeedUpdateFormatByOutputTypeParm(OpDescPtr &netout_desc, uint32_t &index) { GE_CHECK_NOTNULL(netout_desc); vector output_format_str; if (ge::AttrUtils::GetListStr(netout_desc, ATTR_ATC_USER_DEFINE_FORMAT, output_format_str)) { for (auto format_str : output_format_str) { vector format_str_split = StringUtils::Split(format_str, ':'); if (format_str_split.size() == kUserDefinedElementCount) { if (format_str_split[0] == to_string(index)) { GELOGI("Find netoutput node output %u format should be set NC1HWC0.", index); return true; } } } } return false; } Status ProcessNetoutputNodeDynShape(NodePtr &node) { auto op_desc = node->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); ge::DataType output_data_type = ge::DT_FLOAT; for (const auto &in_anchor : node->GetAllInDataAnchors()) { auto index = static_cast(in_anchor->GetIdx()); auto peer_out = in_anchor->GetPeerOutAnchor(); GE_CHECK_NOTNULL(peer_out); auto src_node = peer_out->GetOwnerNode(); GE_CHECK_NOTNULL(src_node); bool is_dynamic = CheckOpType(src_node, MERGE); OpDescPtr src_op_desc = src_node->GetOpDesc(); GE_CHECK_NOTNULL(src_op_desc); auto net_output_input_desc = op_desc->MutableInputDesc(index); GE_CHECK_NOTNULL(net_output_input_desc); ge::GeShape old_shape = net_output_input_desc->GetShape(); ge::Format old_format = net_output_input_desc->GetFormat(); ge::DataType old_dtype = net_output_input_desc->GetDataType(); // Update datatype if (NeedUpdateDtByOutputTypeParm(op_desc, index, output_data_type)) { GELOGI("Enter into process output_type schedule"); net_output_input_desc->SetDataType(output_data_type); if (is_dynamic) { auto merge_output = src_op_desc->MutableOutputDesc(0); GE_CHECK_NOTNULL(merge_output); merge_output->SetDataType(output_data_type); for (uint32_t i = 0; i < src_node->GetAllInDataAnchorsSize(); ++i) { auto merge_input = src_op_desc->MutableInputDesc(i); GE_CHECK_NOTNULL(merge_input); merge_input->SetDataType(output_data_type); } } } // check if is_output_adjust_hw_layout is set if (NeedUpdateFormatByOutputTypeParm(op_desc, index)) { if ((old_format != FORMAT_NCHW) && (old_format != FORMAT_NHWC) && (old_format != FORMAT_NC1HWC0)) { ErrorManager::GetInstance().ATCReportErrMessage( "E19014", {"opname", "value", "reason"}, {op_desc->GetName(), "format[" + TypeUtils::FormatToSerialString(old_format) + "]", "only support FORMAT_NC1HWC0,FORMAT_NCHW,FORMAT_NHWC"}); GELOGE(INTERNAL_ERROR, "Format is not one of NCHW, NHWC, NC1HWC0."); return FAILED; } GeTensorDesc old_desc(old_shape, old_format, old_dtype); if (ProcessNetoutputNodeFp16Nc1hwc0DynShape(old_desc, net_output_input_desc, src_node) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Process netoutput fp16 nc1hwc0."); return FAILED; } } } return SUCCESS; } long StringToLongNoThrow(const string &str) { try { return std::stol(str); } catch (const std::invalid_argument) { GELOGE(PARAM_INVALID, "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: " "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", str.c_str()); return PARAM_INVALID; } catch (const std::out_of_range) { GELOGE(PARAM_INVALID, "Parse shape range of input failed when transfer from string to int64. Given %s, while correct example: " "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", str.c_str()); return PARAM_INVALID; } } /** * Parser shape_range from string to vector * shape_range from option normally is "[1~20,3,3~6,-1],[1~20,3,3~6,-1]" * @param shape_range */ Status ParseDynamicInputShapeRange(const std::string &shape_range, std::vector>> &range) { if (shape_range.size() < 2) { GELOGE(PARAM_INVALID, "Shape range %s is invalid.", shape_range.c_str()); return PARAM_INVALID; } // different shape_range of single input are split by ']' vector shape_range_set = ge::StringUtils::Split(shape_range, ']'); if (shape_range_set.empty()) { GELOGE(PARAM_INVALID, "Shape range %s is not valid. Correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", shape_range.c_str()); return PARAM_INVALID; } for (auto &shape_range_str : shape_range_set) { if (shape_range_str.size() < 3) { // shape_range_str should be "[2~3,1" // or ",[2~3,1". because we should trim '[' or ',[' // so shape_range_str.size() < 3 is invalid continue; } // trim start bytes, after that, single input should be "1~20,3,3~6,-1" if (ge::StringUtils::StartWith(shape_range_str, "[")) { shape_range_str = shape_range_str.substr(1, shape_range_str.size()); } if (ge::StringUtils::StartWith(shape_range_str, ",")) { shape_range_str = shape_range_str.substr(2, shape_range_str.size()); } // parse shape_range of single input. eg. "1~20,3,3~6,-1" std::vector> range_of_single_input; vector dim_range_set = ge::StringUtils::Split(shape_range_str, ','); for (const auto &range_pair_str : dim_range_set) { vector range_pair_set = ge::StringUtils::Split(range_pair_str, '~'); pair range_pair; if (range_pair_set.size() == 1) { // fix dim auto range_value = StringToLongNoThrow(range_pair_set.at(0).c_str()); if (range_value < 0) { range_pair = std::make_pair(1, range_value); } else { range_pair = std::make_pair(range_value, range_value); } } else if (range_pair_set.size() == 2) { // unknown dim, should get range. auto range_left = StringToLongNoThrow(range_pair_set.at(0).c_str()); auto range_right = StringToLongNoThrow(range_pair_set.at(1).c_str()); if (range_left < 0 || range_right < 0) { GELOGE(PARAM_INVALID, "Shape range of input is invalid. Given range pair [%ld,%ld], while correct example: " "\"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", range_left, range_right); return PARAM_INVALID; } range_pair = std::make_pair(range_left, range_right); } else { GELOGE(PARAM_INVALID, "Shape range of input is invalid. Given %s, while correct example: \"[1~20,3,3~6,-1],[1~20,3,3~6,-1]\"", shape_range.c_str()); return PARAM_INVALID; } range_of_single_input.emplace_back(range_pair); } range.emplace_back(range_of_single_input); } return SUCCESS; } Status GetDynamicInputShapeRange(const std::vector &user_input, const std::map &graph_option, vector>> &range_vec) { // check both mode and shape_range option are all enabled auto mode_iter = graph_option.find(OPTION_EXEC_DYNAMIC_EXECUTE_MODE); bool enable_dynamic_execute_mode = (mode_iter != graph_option.end()) && (mode_iter->second == "dynamic_execute"); if (!enable_dynamic_execute_mode) { GELOGD("Graph Option: Can not find %s option in graph options or option value is empty", OPTION_EXEC_DYNAMIC_EXECUTE_MODE); } auto iter = graph_option.find(OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE); bool enable_input_shape_range = (iter != graph_option.end()) && (!iter->second.empty()); if (!enable_input_shape_range) { GELOGD("Graph Option: Can not find %s option in graph options or option value is empty", OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE); } if (enable_dynamic_execute_mode && enable_input_shape_range) { GELOGD("GraphOption: %s value is dynamic_execute, %s value is %s.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE, iter->second.c_str()); } else if (!enable_dynamic_execute_mode && !enable_input_shape_range) { return SUCCESS; } else { GELOGE(PARAM_INVALID, "Graph option: %s and %s should be enabled at the same time.", OPTION_EXEC_DYNAMIC_EXECUTE_MODE, OPTION_EXEC_DATA_INPUTS_SHAPE_RANGE); return PARAM_INVALID; } auto ret = ParseDynamicInputShapeRange(iter->second, range_vec); GE_CHK_STATUS_RET(ret, "Parse dynamic input shape range failed."); if (range_vec.size() != user_input.size()) { GELOGE(PARAM_INVALID, "Dynamic input shape range size is %zu, inputs size is %zu. Not match.", range_vec.size(), user_input.size()); return PARAM_INVALID; } return SUCCESS; } Status UpdateDynamicInputShapeRange(const ge::GeAttrValue::INT index, const vector>> &range_vec, OpDescPtr &op, GeTensorDesc &desc) { auto origin_shape = desc.GetShape(); auto current_shape_range_vec = range_vec.at(index); if (current_shape_range_vec.size() != origin_shape.GetDimNum()) { GELOGE(PARAM_INVALID, "Given shape_range dim num is %zu, current dim num is %zu, not match.Pleace Check.", current_shape_range_vec.size(), origin_shape.GetDimNum()); return PARAM_INVALID; } for (size_t i = 0; i < origin_shape.GetDimNum(); ++i) { auto curr_dim = origin_shape.GetDim(i); auto left_range = current_shape_range_vec.at(i).first; auto right_range = current_shape_range_vec.at(i).second; if (left_range == right_range) { // given shape_range is known dim, check is same as origin or not if (curr_dim != left_range) { GELOGE(PARAM_INVALID, "Given shape range is %ld, current dim shape is %ld, not match.Pleace Check.", left_range, curr_dim); return PARAM_INVALID; } origin_shape.SetDim(i, left_range); } else { // given shape_range is fix range, check input_shape is in this range or not if (right_range != UNKNOWN_DIM) { if ((curr_dim < left_range) || (curr_dim > right_range)) { GELOGE(PARAM_INVALID, "Given shape range is [%ld~%ld], current dim shape is %ld, out of range.Pleace Check.", left_range, right_range, curr_dim); return PARAM_INVALID; } } origin_shape.SetDim(i, UNKNOWN_DIM); } } desc.SetShape(origin_shape); desc.SetShapeRange(current_shape_range_vec); graphStatus graph_ret = op->UpdateInputDesc(0, desc); GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret); graph_ret = op->UpdateOutputDesc(0, desc); GE_CHK_STATUS_RET(graph_ret, "UpdateInputDesc fail, graph ret: %u", graph_ret); return SUCCESS; } } // namespace GraphPrepare::GraphPrepare() : compute_graph_(nullptr) {} GraphPrepare::~GraphPrepare() {} /** * @param graph * @return */ Status GraphPrepare::UpdateVariableFormats(ComputeGraphPtr &graph) { GE_CHECK_NOTNULL(graph); auto var_names_to_refs = CollectVarNamesToRefs(graph); for (auto &node : graph->GetAllNodes()) { if (node == nullptr) { continue; } if (node->GetType() != VARIABLE) { continue; } auto trans_road = VarManager::Instance(graph->GetSessionID())->GetTransRoad(node->GetName()); if (trans_road == nullptr) { GELOGD("The variable %s does not have any trans road", node->GetName().c_str()); continue; } GELOGI("Recover the trans road for var %s reversely", node->GetName().c_str()); auto ret = RecoverTransRoadForVar(node, *trans_road); if (ret != SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var %s", node->GetName().c_str()); return INTERNAL_ERROR; } auto iter = var_names_to_refs.find(node->GetName()); if (iter != var_names_to_refs.end()) { ret = RecoverTransRoadForVarRef(iter->second, *trans_road); if (ret != SUCCESS) { GELOGE(INTERNAL_ERROR, "Failed to recovery trans road for var ref %s", node->GetName().c_str()); return INTERNAL_ERROR; } } } return SUCCESS; } void GraphPrepare::SetOptions(const ge::GraphManagerOptions &options) { options_ = options; } Status GraphPrepare::Init(const ge::Graph &graph, uint64_t session_id) { compute_graph_ = GraphUtils::GetComputeGraph(graph); if (compute_graph_ != nullptr) { compute_graph_->SetSessionID(session_id); } session_id_ = session_id; Status ret = CheckGraph(); if (ret != SUCCESS) { GELOGE(ret, "RunGraph graph check fail, ret:%u", ret); return ret; } (void)compute_graph_->TopologicalSorting(); ret = CheckRefOp(); if (ret != SUCCESS) { GELOGE(ret, "RunGraph check ref op fail, ret:%u", ret); return ret; } return SUCCESS; } Status GraphPrepare::CheckGraph() { if (compute_graph_ == nullptr) { GELOGE(GE_GRAPH_INIT_FAILED, "Graph prepare init compute graph is NULLPTR"); return GE_GRAPH_INIT_FAILED; } auto nodes = compute_graph_->GetAllNodes(); if (nodes.empty()) { GELOGE(GE_GRAPH_INIT_FAILED, "Invalid graph, no nodes in this graph."); return GE_GRAPH_INIT_FAILED; } for (const NodePtr &node : compute_graph_->GetAllNodes()) { GE_CHECK_NOTNULL(node); if (node->GetOpDesc() == nullptr) { GELOGE(GE_GRAPH_INIT_FAILED, "Check Graph node opdesc is NULL"); return GE_GRAPH_INIT_FAILED; } } return SUCCESS; } Status GraphPrepare::CheckRefInputNode(const NodePtr &node, const std::string &input_name, const std::set &ref_nodes) { // Acceptable input types should be ref node, variable or Switch operator, which is issued by ME for dynamic // lossscale and would be optimized in SwitchToStreamSwitchPass. // Since ME dont differentiate between RefSwitch and Switch, and only issue Switch. static std::set acceptable_types = {ge::VARIABLE, ge::VARIABLEV2, ge::VARHANDLEOP, ge::REFSWITCH, ge::REFMERGE, ge::REFENTER, ge::REFNEXTITERATION, ge::REFEXIT, ge::SWITCH}; GE_CHECK_NOTNULL(node); const auto &op_desc = node->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); const auto input_index = op_desc->GetInputIndexByName(input_name); const auto &in_anchor = node->GetInDataAnchor(input_index); GE_CHECK_NOTNULL(in_anchor); const auto &peer_out_anchor = in_anchor->GetPeerOutAnchor(); GE_CHECK_NOTNULL(peer_out_anchor); const auto &input_node = peer_out_anchor->GetOwnerNode(); GE_CHECK_NOTNULL(input_node); const auto &input_op_desc = input_node->GetOpDesc(); GE_CHECK_NOTNULL(input_op_desc); bool is_ref = (ref_nodes.find(input_node) != ref_nodes.end()); if (is_ref) { return SUCCESS; } auto input_type = input_op_desc->GetType(); if (input_type == ge::FRAMEWORKOP) { if (!ge::AttrUtils::GetStr(input_op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, input_type)) { GELOGE(PARAM_INVALID, "Get original type failed."); return PARAM_INVALID; } } bool is_acceptable = (acceptable_types.find(input_type) != acceptable_types.end()); if (!is_acceptable) { ErrorManager::GetInstance().ATCReportErrMessage( "E15005", {"opname", "optype", "opname1", "optype1"}, {op_desc->GetName(), node->GetType(), input_op_desc->GetName(), input_op_desc->GetType()}); GELOGE(PARAM_INVALID, "The ref input of ref node %s[%s] must be ref node or variable, but %s[%s]isn't.", node->GetName().c_str(), node->GetType().c_str(), input_op_desc->GetName().c_str(), input_op_desc->GetType().c_str()); return PARAM_INVALID; } return SUCCESS; } Status GraphPrepare::CheckRefOp() { GE_CHECK_NOTNULL(compute_graph_); std::set ref_nodes; for (const NodePtr &node : compute_graph_->GetDirectNode()) { if (node == nullptr) { GELOGE(PARAM_INVALID, "param [node] must not be null."); return PARAM_INVALID; } auto op_desc = node->GetOpDesc(); if (op_desc == nullptr) { GELOGE(PARAM_INVALID, "OpDesc of param [node] must not be null."); return PARAM_INVALID; } auto input_name_index = op_desc->GetAllInputName(); auto outputs = op_desc->GetAllOutputName(); for (const auto &name_index : input_name_index) { if (op_desc->GetOutputIndexByName(name_index.first) != -1) { if (CheckRefInputNode(node, name_index.first, ref_nodes) != SUCCESS) { GELOGE(PARAM_INVALID, "CheckRefInputNode failed."); return PARAM_INVALID; } (void)ref_nodes.insert(node); // no need to check value } } } return SUCCESS; }; Status GraphPrepare::SetRtContext(rtContext_t rt_context, rtCtxMode_t mode) { GE_CHECK_NOTNULL(compute_graph_); GELOGI("set rt_context, session id: %lu, graph id: %u, mode %d, device id:%u.", session_id_, compute_graph_->GetGraphID(), static_cast(mode), ge::GetContext().DeviceId()); GE_CHK_RT_RET(rtCtxCreate(&rt_context, mode, ge::GetContext().DeviceId())); GE_CHK_RT_RET(rtCtxSetCurrent(rt_context)); RtContextUtil::GetInstance().AddRtContext(session_id_, compute_graph_->GetGraphID(), rt_context); return SUCCESS; } Status GraphPrepare::AdjustDataOpOutput(const NodePtr &node) { if (node == nullptr) { GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node is NULL"); return GE_GRAPH_GRAPH_NODE_NULL; } OpDescPtr op_desc_ptr = node->GetOpDesc(); if (op_desc_ptr == nullptr) { GELOGE(GE_GRAPH_GRAPH_NODE_NULL, "Input node opdesc is NULL"); return GE_GRAPH_GRAPH_NODE_NULL; } GeTensorDesc output = op_desc_ptr->GetOutputDesc(0); int64_t tensor_size = 0; graphStatus graph_status = TensorUtils::GetTensorMemorySizeInBytes(output, tensor_size); if (graph_status != GRAPH_SUCCESS) { ErrorManager::GetInstance().ATCReportErrMessage( "E19012", {"function", "reason"}, {"GetTensorMemorySizeInBytes", "opname is " + node->GetName()}); GELOGE(graph_status, "GetTensorMemorySizeInBytes failed!"); return FAILED; } TensorUtils::SetSize(output, tensor_size); graphStatus graph_ret = op_desc_ptr->UpdateOutputDesc(0, output); if (graph_ret != GRAPH_SUCCESS) { GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret); return graph_ret; } return SUCCESS; } Status GraphPrepare::UpdateInput(const std::vector &user_input, const std::map &graph_option) { // Get shape range of input in dynamic_execute mode vector>> dynamic_shape_range_vec; auto ret = GetDynamicInputShapeRange(user_input, graph_option, dynamic_shape_range_vec); GE_CHK_STATUS_RET(ret, "Graph option is not right on Dynamic execute mode."); compute_graph_->SaveDataFormat(ge::TypeUtils::DomiFormatToFormat(GetLocalOmgContext().format)); for (NodePtr &input_node : compute_graph_->GetDirectNode()) { GE_CHECK_NOTNULL(input_node); OpDescPtr op = input_node->GetOpDesc(); GE_CHECK_NOTNULL(op); if (op->GetType() == DATA) { GeAttrValue::INT index = 0; if ((!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) || (GetLocalOmgContext().is_dynamic_input)) { GELOGW("Get index from data attr failed"); continue; } if ((index < 0) || (static_cast(index) >= user_input.size())) { std::string situation = "data op index[" + std::to_string(index) + "]"; std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]"; ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason}); GELOGE(PARAM_INVALID, "user_input size = %zu, graph data op index = %ld.", user_input.size(), index); return FAILED; } if (IsDynamicDims(input_node)) { continue; } GeTensorDesc desc(user_input[index].GetTensorDesc()); auto format = desc.GetFormat(); auto origin_format = desc.GetOriginFormat(); // data maybe internal format [FRACTAL_NZ] at singleop process such as GEMM. auto tune_flag = (options_.build_mode == BUILD_MODE_TUNING) && (options_.build_step == BUILD_STEP_AFTER_BUILDER); bool need_check_internal_format = (!IsTansDataOpData(input_node)) && (!options_.is_single_op) && (!tune_flag); if (need_check_internal_format) { bool is_internal = TypeUtils::IsInternalFormat(format) || TypeUtils::IsInternalFormat(origin_format); if (is_internal) { ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {"Input format[" + TypeUtils::FormatToSerialString(format) + "] or origin_format[" + TypeUtils::FormatToSerialString(origin_format) + "]", "it is not support"}); GELOGE(PARAM_INVALID, "Input format %s or origin_format %s is not support.", TypeUtils::FormatToSerialString(format).c_str(), TypeUtils::FormatToSerialString(origin_format).c_str()); return FAILED; } } auto data_type = desc.GetDataType(); uint32_t length = 1; bool type_ret = TypeUtils::GetDataTypeLength(data_type, length); if (!type_ret) { ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"}); GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str()); return FAILED; } int64_t desc_shape = desc.GetShape().GetShapeSize(); FMK_INT64_UINT32_MULCHECK(desc_shape, length); int64_t shape_size = desc_shape * length; GE_IF_BOOL_EXEC(shape_size == 0 && desc.GetShape().GetDimNum() == 0, shape_size = static_cast(length)); int64_t size = 0; GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(desc, size) != GRAPH_SUCCESS, GELOGE(INTERNAL_ERROR, "TensorUtils GetSize failed"); return FAILED); bool size_check = (size != 0 && shape_size != size); if (size_check) { std::string situation = "input data size[" + std::to_string(size) + "] and shape_size[" + std::to_string(size) + "]"; std::string reason = "because size != 0 and shape_size != size"; ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason}); GELOGE(PARAM_INVALID, "input data size =%ld, shape_size =%ld.", size, shape_size); return FAILED; } ge::TensorUtils::SetSize(desc, shape_size); if (!tune_flag) { graphStatus graph_ret = op->UpdateInputDesc(0, desc); if (graph_ret != GRAPH_SUCCESS) { GELOGE(graph_ret, "UpdateInputDesc fail, graph_ret:%u", graph_ret); return graph_ret; } // Size will be recalculated in the build stage ge::TensorUtils::SetSize(desc, 0); graph_ret = op->UpdateOutputDesc(0, desc); if (graph_ret != GRAPH_SUCCESS) { GELOGE(graph_ret, "UpdateOutputDesc fail, graph_ret:%u", graph_ret); return graph_ret; } } else { GELOGI("data %s skip update info in tune mode", op->GetName().c_str()); } if (!dynamic_shape_range_vec.empty()) { ret = UpdateDynamicInputShapeRange(index, dynamic_shape_range_vec, op, desc); GE_CHK_STATUS_RET(ret, "Fail to update dynamic input shape range on %s.", op->GetName().c_str()); continue; } if (!options_.train_graph_flag) { Status ret = AdjustDataOpOutput(input_node); GE_IF_BOOL_EXEC(ret != SUCCESS, GELOGE(ret, "AdjustDataOpOutput fail, ret:%u", ret); return ret); } } } return SUCCESS; } Status GraphPrepare::TryDoAipp() { // infer and with aipp configure file, then call aipp insert if ((!options_.train_graph_flag) && (!options_.insert_op_file.empty())) { GE_DUMP(compute_graph_, "Before_insert_aipp"); Status ret = ge::InsertNewOpUtil::Instance().Init(); if (ret != SUCCESS) { GELOGE(INTERNAL_ERROR, "TryDoAipp: InsertNewOpUtil instance failed."); return INTERNAL_ERROR; } ret = ge::InsertNewOpUtil::Instance().Parse(options_.insert_op_file.c_str()); if (ret != SUCCESS) { GELOGE(GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED, "TryDoAipp: parse config file %s failed", options_.insert_op_file.c_str()); return GE_GRAPH_OPTIMIZE_INSERT_OP_PARSE_FAILED; } ret = ge::InsertNewOpUtil::Instance().InsertAippOps(compute_graph_, options_.insert_op_file); if (ret != SUCCESS) { GELOGE(GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED, "TryDoAipp: insert aipp op ret failed, ret:%u", ret); return GE_GRAPH_OPTIMIZE_INSERT_DYN_OP_FAILED; } } return SUCCESS; } Status GraphPrepare::FormatAndShapeProcess() { Status ret = ResourcePairProcess("add"); if (ret != SUCCESS) { GELOGE(ret, "ResourcePairProcess failed"); return ret; } GE_TIMESTAMP_START(InferOriginFormat1); ret = compute_graph_->InferOriginFormat(); GE_TIMESTAMP_END(InferOriginFormat1, "GraphPrepare::InferOriginFormat1"); GE_DUMP(compute_graph_, "after_first_inferformat"); if (ret != SUCCESS) { GELOGE(ret, "Prepare Graph first inferformat failed"); return ret; } GE_TIMESTAMP_START(InferShapeForPreprocess); ret = InferShapeForPreprocess(); GE_TIMESTAMP_END(InferShapeForPreprocess, "GraphPrepare::InferShapeForPreprocess"); GE_DUMP(compute_graph_, "after_infershape"); if (ret != SUCCESS) { GELOGE(GE_GRAPH_INFERSHAPE_FAILED, "Prepare Graph infershape failed"); return GE_GRAPH_INFERSHAPE_FAILED; } GE_TIMESTAMP_START(InferOriginFormat2); ret = compute_graph_->InferOriginFormat(); GE_TIMESTAMP_END(InferOriginFormat2, "GraphPrepare::InferOriginFormat2"); if (ret != SUCCESS) { GELOGE(ret, "Prepare Graph inferformat failed"); return ret; } ret = ResourcePairProcess("remove"); if (ret != SUCCESS) { return ret; } return ret; } Status GraphPrepare::ResourcePairProcess(const std::string &action) { PassManager control_pass; // Graph pass tmp logic for resource infershape if (options_.train_graph_flag) { try { if (action == "add") { (void)control_pass.AddPass("ResourcePairProcess::ResourcePairAddControlPass", new ResourcePairAddControlPass); } else { (void)control_pass.AddPass("ResourcePairProcess::ResourcePairRemoveControlPass", new ResourcePairRemoveControlPass); } } catch (std::bad_alloc &e) { GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occur, action:%s.", action.c_str()); return INTERNAL_ERROR; } } Status ret = control_pass.Run(compute_graph_); if (ret != SUCCESS && ret != NOT_CHANGED) { GELOGE(ret, "Run ResourcePairControlPass failed, action:%s, ret:%u.", action.c_str(), ret); return ret; } return SUCCESS; } Status GraphPrepare::UpdateDataNetOutputByStorageFormat() { for (auto &node_ptr : compute_graph_->GetAllNodes()) { GE_CHECK_NOTNULL(node_ptr); if (node_ptr->GetType() == DATA) { uint32_t index = 0; auto op_desc = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); const GeTensorDescPtr input = op_desc->MutableInputDesc(index); Format storage_format = FORMAT_RESERVED; vector dst_shape_dims; if (GetStorageFormatAndShape(op_desc, input, storage_format, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Get storage format for input failed"); return FAILED; } if (storage_format == FORMAT_RESERVED) { continue; } if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Modify format and shape for inputfailed"); return FAILED; } } if (node_ptr->GetType() == ge::NETOUTPUT) { auto op_desc = node_ptr->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); for (uint32_t index = 0; index < op_desc->GetOutputsSize(); index++) { const GeTensorDescPtr output = op_desc->MutableOutputDesc(index); Format storage_format = FORMAT_RESERVED; vector dst_shape_dims; if (GetStorageFormatAndShape(op_desc, output, storage_format, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Get storage format from output failed"); return FAILED; } if (storage_format == FORMAT_RESERVED) { continue; } if (ModifyDataNetOutputFormatAndShape(op_desc, index, storage_format, dst_shape_dims) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Modify format and shape for output failed"); return FAILED; } } } } return SUCCESS; } Status GraphPrepare::SaveOriginalGraphToOmModel() { if (options_.save_original_model == "true") { ModelHelper model_helper; Status ret = model_helper.SaveOriginalGraphToOmModel(ge::GraphUtils::CreateGraphFromComputeGraph(compute_graph_), options_.original_model_file); if (ret != SUCCESS) { // If save original model fail, process continue GELOGW("SaveOriginalGraphToOmModel fail"); } } return SUCCESS; } #define PP_RUN_AND_DUMP(name, func, ...) \ do { \ GE_RUN(Prepare, func, __VA_ARGS__); \ GE_DUMP(compute_graph, "PrepareAfter" name); \ GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \ } while (0) #define PP_RUN(name, func, ...) \ do { \ GE_RUN(Prepare, func, __VA_ARGS__); \ GELOGI("Prepare %s on graph %s success.", name, compute_graph->GetName().c_str()); \ } while (0) Status GraphPrepare::PrepareDynShape(const GraphNodePtr &graph_node, const std::vector &user_input, ge::ComputeGraphPtr &compute_graph, uint64_t session_id) { GE_CHECK_NOTNULL(graph_node->GetGraph()); GE_CHECK_NOTNULL(compute_graph); GetLocalOmgContext().type = static_cast(options_.framework_type); const Graph &const_graph = *graph_node->GetGraph(); PP_RUN("Init", Init, const_graph, session_id); PP_RUN("SetRtContext", SetRtContext, rtContext_t(), RT_CTX_GEN_MODE); PP_RUN_AND_DUMP("CheckAndUpdateInput", CheckAndUpdateInput, user_input, graph_node->GetOptions()); PP_RUN_AND_DUMP("GraphEquivalentTransformation", GraphEquivalentTransformation); PP_RUN_AND_DUMP("ProcessOutput", ProcessNetOutput); PP_RUN_AND_DUMP("ProcessMultiBatch", multibatch::ProcessMultiBatch, compute_graph_); PP_RUN_AND_DUMP("InsertAipp", TryDoAipp); PP_RUN_AND_DUMP("ProcessBeforeInfershape", ProcessBeforeInfershape); PP_RUN_AND_DUMP("InferFormatAndShape", FormatAndShapeProcess); PP_RUN_AND_DUMP("GetDynamicOutputShape", multibatch::GetDynamicOutputShape, compute_graph_); PP_RUN_AND_DUMP("ProcessAippStage2", InsertNewOpUtil::Instance().UpdateDataNodeByAipp, compute_graph_); PP_RUN("SaveOriginalGraphToOmModel", SaveOriginalGraphToOmModel); PP_RUN_AND_DUMP("PrepareOptimize", PrepareOptimize); return SUCCESS; } Status GraphPrepare::RecordAIPPInfo(ge::ComputeGraphPtr &compute_graph) { PP_RUN("RecordAIPPInfo", InsertNewOpUtil::Instance().RecordAIPPInfoToData, compute_graph_); return SUCCESS; } Status GraphPrepare::PrepareRunningFormatRefiner() { auto compute_graph = compute_graph_; PassManager pass_manager; GE_CHK_STATUS_RET(pass_manager.AddPass("PrepareRunningFormatRefiner::VariablePrepareOpPass", new (std::nothrow) VariablePrepareOpPass)) GE_TIMESTAMP_START(pass_manager); auto ret = pass_manager.Run(compute_graph); GE_TIMESTAMP_END(pass_manager, "GraphPrepare::PrepareRunningFormatRefiner"); if (ret != SUCCESS && ret != NOT_CHANGED) { GELOGE(ret, "Run passes for running format refiner failed, ret:%u.", ret); return ret; } PP_RUN_AND_DUMP("UpdateInputOutputByUserOptions", UpdateInputOutputByOptions); PP_RUN_AND_DUMP("UpdateVariableFormats", UpdateVariableFormats, compute_graph_); return SUCCESS; } Status GraphPrepare::SwitchOpOptimize(ComputeGraphPtr &compute_graph) { if (compute_graph == nullptr) { GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL"); return GE_GRAPH_NULL_INPUT; } GEPass ge_passes(compute_graph); NamesToPass hccl_group; HcclGroupPass hccl_group_pass; GELOGD("Add hccl group pass success"); hccl_group.emplace_back("HcclGroupPass", &hccl_group_pass); auto ret = ge_passes.Run(hccl_group); if (ret != SUCCESS) { GELOGE(ret, "Run HcclGroupPass pass for preprocess failed, ret:%u.", ret); return ret; } ret = compute_graph->TopologicalSorting(); if (ret != SUCCESS) { GELOGE(ret, "Graph topological sort failed, ret:%u.", ret); return ret; } return SUCCESS; } #undef PP_RUN_AND_DUMP #undef PP_RUN Status GraphPrepare::GenerateInfershapeGraph(ConstGraphPtr graph) { if (graph == nullptr) { GELOGE(GE_GRAPH_NULL_INPUT, "Input Graph is NULL"); return GE_GRAPH_NULL_INPUT; } const Graph &const_graph = *graph; Status ret = Init(const_graph, 0); if (ret != SUCCESS) { GELOGE(ret, "Init graph_prepare fail, ret:%u", ret); return ret; } GE_DUMP(compute_graph_, "after_parser"); GELOGI("Start infershape for dump json process."); ret = compute_graph_->InferOriginFormat(); GE_DUMP(compute_graph_, "after_inferformat"); if (ret != SUCCESS) { GELOGE(ret, "Prepare Graph inferformat failed"); return ret; } InferShapePass infer_shape_pass; NamesToPass names_to_passes; names_to_passes.emplace_back("InferShapePass", &infer_shape_pass); GEPass ge_passes(compute_graph_); ret = ge_passes.Run(names_to_passes); GE_DUMP(compute_graph_, "after_infershape"); if (ret != SUCCESS) { GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret); return ret; } ShapeRefiner::ClearContextMap(); return SUCCESS; } Status GraphPrepare::CheckConstOp() { for (auto &node_ptr : compute_graph_->GetAllNodes()) { GE_CHECK_NOTNULL(node_ptr); if (node_ptr->GetType() == CONSTANT) { Status ret = VerifyConstOp(node_ptr); GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed"); } else if (node_ptr->GetType() == FRAMEWORKOP) { auto op_desc = node_ptr->GetOpDesc(); if (op_desc == nullptr) { GELOGE(PARAM_INVALID, "Get op desc failed"); return PARAM_INVALID; } std::string original_type; GE_IF_BOOL_EXEC(ge::AttrUtils::GetStr(op_desc, ATTR_NAME_FRAMEWORK_ORIGINAL_TYPE, original_type), GELOGI("Get FrameWorkOp original type [%s]", original_type.c_str())); GELOGI("original type is %s", original_type.c_str()); if (original_type == CONSTANT) { Status ret = VerifyConstOp(node_ptr); GE_CHK_BOOL_RET_STATUS(ret == SUCCESS, ret, "Const Op Check failed"); } } } return SUCCESS; } Status GraphPrepare::VerifyConstOp(const NodePtr &node) { GE_CHECK_NOTNULL(node); auto op_desc = node->GetOpDesc(); GE_CHECK_NOTNULL(op_desc); ConstGeTensorPtr ge_tensor_ptr; if (!(AttrUtils::GetTensor(op_desc, ATTR_NAME_WEIGHTS, ge_tensor_ptr))) { GELOGE(PARAM_INVALID, "Get value from const attr failed"); return PARAM_INVALID; } GE_CHECK_NOTNULL(ge_tensor_ptr); auto data_size = ge_tensor_ptr->GetData().GetSize(); auto ge_tensor_desc = ge_tensor_ptr->GetTensorDesc(); int64_t shape_size = ge_tensor_desc.GetShape().GetShapeSize(); auto data_type = ge_tensor_desc.GetDataType(); if (data_type == DT_STRING) { return SUCCESS; } uint32_t length = 1; bool type_ret = TypeUtils::GetDataTypeLength(data_type, length); if (!type_ret) { ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {"Input datatype[" + TypeUtils::DataTypeToSerialString(data_type) + "]", "it is not support"}); GELOGE(PARAM_INVALID, "Input datatype %s is not support.", TypeUtils::DataTypeToSerialString(data_type).c_str()); return FAILED; } FMK_INT64_UINT32_MULCHECK(shape_size, length); GELOGI("Const real value Size:%zu, op_desc Shape Size:%ld, data_type:%s.", data_size, shape_size * length, TypeUtils::DataTypeToSerialString(data_type).c_str()); if (shape_size == 0) { if (ge_tensor_desc.GetShape().GetDims().size() == 0) { // shape = [], means it's a sclar tensor. GE_CHK_BOOL_EXEC(data_size / length == 1, ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid scalar tensor."}); return PARAM_INVALID, "Const is invalid scalar tensor."); } else { // shape = [x, y, 0,...], means it's a vector tensor that value is []. GE_CHK_BOOL_EXEC(data_size == 0, ErrorManager::GetInstance().ATCReportErrMessage("E10043", {"reason"}, {"Const is invalid vector scalar."}); return PARAM_INVALID, "Const is invalid vector scalar."); } } else { GE_CHK_BOOL_EXEC(data_size == static_cast(shape_size * length) && data_size != 0, ErrorManager::GetInstance().ATCReportErrMessage( "E10043", {"reason"}, {"Const input data size is not equal with tensor desc shape"}); return PARAM_INVALID, "Const input data size is not equal with tensor desc shape"); } return SUCCESS; } bool GraphPrepare::IsDynamicDims(const NodePtr &input_node) { auto data_shape = NodeUtils::GetOutputDesc(*input_node, kDataOutIndex).GetShape(); const auto &dims = data_shape.GetDims(); bool all_is_positive = false; if (std::all_of(dims.begin(), dims.end(), [](int64_t val) { return val >= 0; })) { all_is_positive = true; } if (!all_is_positive && !options_.input_shape.empty() && !options_.dynamic_dims.empty() && options_.dynamic_node_type != kInvalidDynaimcDimsType) { GELOGI("No need to check and update desc info, the dims of %s is %s.", input_node->GetName().c_str(), formats::JoinToString(dims).c_str()); return true; } return false; } Status GraphPrepare::CheckUserInput(const std::vector &user_input) { if (GetLocalOmgContext().is_dynamic_input) { return SUCCESS; } unsigned int node_num = 0; unsigned int data_num = 0; for (NodePtr &input_node : compute_graph_->GetDirectNode()) { GE_CHECK_NOTNULL(input_node); OpDescPtr op = input_node->GetOpDesc(); GE_CHECK_NOTNULL(op); node_num++; if (op->GetType() == DATA || op->GetType() == AIPPDATA) { data_num++; GeAttrValue::INT index = 0; if (!(AttrUtils::GetInt(op, ATTR_NAME_INDEX, index))) { GELOGE(GE_GRAPH_INIT_FAILED, "Get index from attr failed"); return GE_GRAPH_INIT_FAILED; } if ((index < 0) || (static_cast(index) >= user_input.size())) { std::string situation = "data op index[" + std::to_string(index) + "]"; std::string reason = "it must less than user_input size[" + std::to_string(user_input.size()) + "]"; ErrorManager::GetInstance().ATCReportErrMessage("E19025", {"situation", "reason"}, {situation, reason}); GELOGE(GE_GRAPH_INIT_FAILED, "user_input size:%zu, data op index:%ld.", user_input.size(), index); return GE_GRAPH_INIT_FAILED; } if (IsDynamicDims(input_node)) { continue; } GeTensorDesc desc(user_input[index].GetTensorDesc()); for (size_t i = 0; i < desc.GetShape().GetDimNum(); ++i) { int64_t dim = desc.GetShape().GetDim(i); if (dim < UNKNOWN_DIM_NUM) { std::string situation = "data dim[" + std::to_string(i) + "][" + std::to_string(dim) + "]" ; std::string reason = "it need >= -2"; REPORT_INPUT_ERROR("E19025", std::vector({"situation", "reason"}), std::vector({situation, reason})); GELOGE(GE_GRAPH_INIT_FAILED, "[Check][InputDim]data dim %zu is not supported, need >= -2, real:%ld.", i, dim); return GE_GRAPH_INIT_FAILED; } } } } if (node_num <= data_num) { GELOGW("Prepare check user input, data_num = %u, node_num = %u", data_num, node_num); } return SUCCESS; } Status GraphPrepare::InferShapeForPreprocess() { GELOGI("Start infershape for preprocess."); GEPass ge_passes(compute_graph_); NamesToPass names_to_passes; AssertPass assert_pass; if (!options_.train_graph_flag) { names_to_passes.emplace_back("AssertPass", &assert_pass); } SwitchDeadBranchElimination switch_dead_branch_elimination; names_to_passes.emplace_back("SwitchDeadBranchElimination", &switch_dead_branch_elimination); MergePass merge_pass; names_to_passes.emplace_back("MergePass", &merge_pass); InferShapePass infer_shape_pass; names_to_passes.emplace_back("InferShapePass", &infer_shape_pass); ReplaceWithEmptyConstPass replace_with_empty_const_pass; names_to_passes.emplace_back("ReplaceWithEmptyConstPass", &replace_with_empty_const_pass); DimensionComputePass dimension_compute_pass; names_to_passes.emplace_back("DimensionComputePass", &dimension_compute_pass); ConstantFoldingPass constant_folding_pass; names_to_passes.emplace_back("ConstantFoldingPass", &constant_folding_pass); int32_t dev_count = 0; AicpuConstantFoldingPass aicpu_constant_folding_pass; const char *aicpu_constant_folding_on = std::getenv("AICPU_CONSTANT_FOLDING_ON"); rtError_t rt_err = RT_ERROR_NONE; if (aicpu_constant_folding_on != nullptr) { rt_err = rtGetDeviceCount(&dev_count); if (rt_err == RT_ERROR_NONE) { Status result = SetRtContext(rtContext_t(), RT_CTX_NORMAL_MODE); if (result != SUCCESS) { GELOGE(result, "Set rt context failed."); return result; } names_to_passes.emplace_back("AicpuConstantFoldingPass", &aicpu_constant_folding_pass); } } Status ret = ge_passes.Run(names_to_passes); if (aicpu_constant_folding_on != nullptr) { if (rt_err == RT_ERROR_NONE) { Status result = SetRtContext(rtContext_t(), RT_CTX_GEN_MODE); if (result != SUCCESS) { GELOGE(result, "Set rt context failed."); return result; } } } ShapeRefiner::ClearContextMap(); if (ret != SUCCESS) { GELOGE(ret, "Run ge_passes infershape for preprocess failed, ret:%u.", ret); return ret; } return SUCCESS; } Status GraphPrepare::PrepareOptimize() { GELOGI("Start optimize for preprocess."); // check rw type GraphOptimize graph_optimize; bool has_conflict = false; graph_optimize.CheckRWConflict(compute_graph_, has_conflict); if (has_conflict) { GELOGE(GRAPH_PARAM_INVALID, "There has rw conflict.Stop optimize."); return FAILED; } PassManager original_graph_passes; // Graph pass try { (void)original_graph_passes.AddPass("PrepareOptimize::ShapeOperateOpRemovePass", new ShapeOperateOpRemovePass); (void)original_graph_passes.AddPass("PrepareOptimize::ReplaceTransShapePass", new ReplaceTransShapePass); (void)original_graph_passes.AddPass("PrepareOptimize::MarkAgnosticPass", new MarkAgnosticPass); } catch (std::bad_alloc &e) { GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs."); return INTERNAL_ERROR; } GE_TIMESTAMP_START(original_graph_passes); Status ret = original_graph_passes.Run(compute_graph_); GE_TIMESTAMP_END(original_graph_passes, "GraphPrepare::OriginalGraphPasses"); if (ret != SUCCESS && ret != NOT_CHANGED) { GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret); return ret; } // New pass GEPass ge_passes(compute_graph_); NamesToPass names_to_passes; EnterPass enter_pass; names_to_passes.emplace_back("EnterPass", &enter_pass); CondPass cond_pass; names_to_passes.emplace_back("CondPass", &cond_pass); PrintOpPass print_pass; if (options_.enable_print_op_pass) { names_to_passes.emplace_back("PrintOpPass", &print_pass); } NoUseReshapeRemovePass no_use_reshape_remove_pass; names_to_passes.emplace_back("NoUseReshapeRemovePass", &no_use_reshape_remove_pass); DropOutPass dropout_pass; AssertPass assert_pass; UnusedConstPass unused_const_pass; StopGradientPass stop_gradient_pass; PreventGradientPass prevent_gradient_pass; PlaceholderWithDefaultPass placeholder_with_default_pass; GuaranteeConstPass guarantee_const_pass; VarIsInitializedOpPass var_is_initialized_pass; ParallelConcatStartOpPass parallel_concat_start_op_pass; IdentityPass identity_pass(false); SnapshotPass snapshot_pass; if (!options_.train_graph_flag) { names_to_passes.emplace_back("DropOutPass", &dropout_pass); names_to_passes.emplace_back("AssertPass", &assert_pass); } names_to_passes.emplace_back("UnusedConstPass", &unused_const_pass); names_to_passes.emplace_back("StopGradientPass", &stop_gradient_pass); names_to_passes.emplace_back("PreventGradientPass", &prevent_gradient_pass); names_to_passes.emplace_back("PlaceholderWithDefaultPass", &placeholder_with_default_pass); names_to_passes.emplace_back("SnapshotPass", &snapshot_pass); names_to_passes.emplace_back("GuaranteeConstPass", &guarantee_const_pass); names_to_passes.emplace_back("VarIsInitializedOpPass", &var_is_initialized_pass); names_to_passes.emplace_back("ParallelConcatStartOpPass", ¶llel_concat_start_op_pass); names_to_passes.emplace_back("IdentityPass", &identity_pass); GE_TIMESTAMP_START(names_to_passes); ret = ge_passes.Run(names_to_passes); GE_TIMESTAMP_END(names_to_passes, "GraphPrepare::NamesToPasses"); if (ret != SUCCESS) { GELOGE(ret, "Run ge_passes optimize for preprocess failed, ret:%u.", ret); return ret; } PassManager graph_pass; try { (void)graph_pass.AddPass("PrepareOptimize::PrunePass", new PrunePass); // can't move to optimize1/2 directly, may cause more identity insert, cause CI fail (void)graph_pass.AddPass("PrepareOptimize::HcclMemcpyPass", new HcclMemcpyPass); } catch (std::bad_alloc &e) { GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs."); return INTERNAL_ERROR; } GE_TIMESTAMP_START(graph_passes); ret = graph_pass.Run(compute_graph_); GE_TIMESTAMP_END(graph_passes, "GraphPrepare::GraphPasses"); if (ret != SUCCESS && ret != NOT_CHANGED) { GELOGE(ret, "Run graph passes optimize for preprocess failed, ret:%u.", ret); return ret; } // The constant for train is CONSTANTOP, and is CONSTANT for inference. They will be unified in future. TypeConversionOfConstant(); ret = compute_graph_->TopologicalSorting(); if (ret != SUCCESS) { GELOGE(ret, "Graph topological sort failed, ret:%u.", ret); return ret; } GELOGI("End optimize for preprocess."); return SUCCESS; } void GraphPrepare::TypeConversionOfConstant() { bool is_acl_compile = false; for (ge::NodePtr &n : compute_graph_->GetAllNodes()) { // This can ensure that n is not a null pointer // No Conversion when called by aclOpCompile (void)AttrUtils::GetBool(n->GetOpDesc(), ATTR_SINGLE_OP_SCENE, is_acl_compile); if (is_acl_compile) { return; } } if (options_.train_graph_flag) { GELOGD("trans CONSTANT to CONSTANTOP in train."); for (ge::NodePtr &n : compute_graph_->GetAllNodes()) { // This can ensure that n is not a null pointer if (n->GetOpDesc()->GetType() == CONSTANT) { n->GetOpDesc()->SetType(CONSTANTOP); } } } else { GELOGD("trans CONSTANTOP to CONSTANT in inferrence."); for (ge::NodePtr &n : compute_graph_->GetAllNodes()) { // This can ensure that n is not a null pointer if (n->GetOpDesc()->GetType() == CONSTANTOP) { n->GetOpDesc()->SetType(CONSTANT); } } } } Status GraphPrepare::GraphEquivalentTransformation() { NamesToPass names_to_pass; ForPass for_pass; names_to_pass.emplace_back("ForToWhilePass", &for_pass); return GEPass(compute_graph_).Run(names_to_pass); } Status GraphPrepare::ProcessBeforeInfershape() { NamesToPass names_to_passes; CondRemovePass condition_remove_pass; names_to_passes.emplace_back("CondRemovePass", &condition_remove_pass); GE_TIMESTAMP_START(ProcessCondRemove); auto ret = GEPass(compute_graph_).Run(names_to_passes); GE_TIMESTAMP_END(ProcessCondRemove, "GraphManager::ProcessCondRemove"); if (ret != SUCCESS) { GELOGE(ret, "Run ge_passes optimize for OptimizeAfterMergeSubGraph failed, ret:%d.", ret); return ret; } return SUCCESS; } Status GraphPrepare::ProcessNetOutput() { PassManager graph_passes_before_infershape; try { if (options_.train_graph_flag) { graph_passes_before_infershape.AddPass("ProcessNetOutput::SavePass", new (std::nothrow) SavePass); } graph_passes_before_infershape.AddPass("ProcessNetOutput::NetOutputPass", new (std::nothrow) NetOutputPass); graph_passes_before_infershape.AddPass("ProcessNetOutput::DataPass", new (std::nothrow) DataPass); // Add NetOutput first. } catch (std::bad_alloc) { GELOGE(INTERNAL_ERROR, "Add pass failed, bad memory allocation occurs."); return INTERNAL_ERROR; } auto ret = graph_passes_before_infershape.Run(compute_graph_); if ((ret != SUCCESS) && (ret != NOT_CHANGED)) { GELOGE(ret, "Run graph_passes_before_infershape failed, ret:%d.", ret); return ret; } return SUCCESS; } Status GraphPrepare::CheckAndUpdateInput(const std::vector &user_input, const std::map &graph_option) { compute_graph_->SetInputSize(user_input.size()); if (user_input.empty()) { return SUCCESS; } auto ret = CheckUserInput(user_input); if (ret != SUCCESS) { GELOGE(ret, "Check user input failed."); return ret; } ret = UpdateInput(user_input, graph_option); if (ret != SUCCESS) { GELOGE(ret, "UpdateInput fail, ret:%u", ret); return ret; } if (user_input.size() != 0) { ret = CheckConstOp(); if (ret != SUCCESS) { GELOGE(ret, "CheckConstOp fail, ret:%u", ret); return ret; } } else { ret = compute_graph_->TopologicalSorting(); if (ret != SUCCESS) { GELOGE(ret, "graph prepare error: compute_graph_->Topological Sorting"); return FAILED; } } return SUCCESS; } Status GraphPrepare::UpdateInputOutputByOptions() { auto ret = UpdateDataNetOutputByStorageFormat(); if (ret != SUCCESS) { GELOGE(ret, "Update format acoording to storage format failed."); return ret; } if (options_.train_graph_flag) { GELOGI("This is train mode, no need to do this schedule."); return SUCCESS; } for (auto &node_ptr : compute_graph_->GetDirectNode()) { GE_CHECK_NOTNULL(node_ptr); if (CheckIfNeedSetNdFormat(node_ptr) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Set node [%s] format ND failed", node_ptr->GetName().c_str()); return FAILED; } if (node_ptr->GetType() == DATA) { if (ProcessDataNodeDynShape(node_ptr) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Process data node failed"); return FAILED; } } if (node_ptr->GetType() == ge::NETOUTPUT) { if (ProcessNetoutputNodeDynShape(node_ptr) != SUCCESS) { GELOGE(INTERNAL_ERROR, "Process netoutput node failed"); return FAILED; } } } return SUCCESS; } bool GraphPrepare::IsTansDataOpData(const ge::NodePtr &var_node) { for (auto &out_anchor : var_node->GetAllOutDataAnchors()) { GE_RT_FALSE_CHECK_NOTNULL(out_anchor); for (auto &in_anchor : out_anchor->GetPeerInDataAnchors()) { GE_RT_FALSE_CHECK_NOTNULL(in_anchor); ge::NodePtr dst_node = in_anchor->GetOwnerNode(); GE_RT_FALSE_CHECK_NOTNULL(dst_node); if (dst_node->GetType() == TRANSDATA) { return true; } } } return false; } } // namespace ge