* Copyright (c) Huawei Technologies Co., Ltd. 2024-2025. All rights reserved.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#include "graph_operation_builder.h"
#include <stdexcept>
#include "atb/utils/log.h"
namespace TorchAtb {
using namespace atb;
using namespace atb::infer;
static const size_t MAX_TENSOR_NUM = 256;
GraphOperationBuilder::GraphOperationBuilder(const std::string &graphName)
{
graphParam_.name = graphName;
}
GraphOperationBuilder &GraphOperationBuilder::SetInputOutput(const std::vector<std::string> &inTensorNames,
const std::vector<std::string> &outTensorNames)
{
uint32_t id = 0;
size_t inTensorNum = inTensorNames.size();
if (inTensorNum > MAX_TENSOR_NUM) {
throw std::runtime_error("inTensor num should be <= 256, but get " + std::to_string(inTensorNum));
}
for (const std::string &inTensorName : inTensorNames) {
inTensorIds_[inTensorName] = id++;
}
size_t outTensorNum = outTensorNames.size();
if (outTensorNum > MAX_TENSOR_NUM) {
throw std::runtime_error("outTensor num should be <= 256, but get " + std::to_string(outTensorNum));
}
for (const std::string &outTensorName : outTensorNames) {
outTensorIds_[outTensorName] = id++;
}
graphParam_.inTensorNum = static_cast<uint32_t>(inTensorNum);
graphParam_.outTensorNum = static_cast<uint32_t>(outTensorNum);
return *this;
}
GraphOperationBuilder &GraphOperationBuilder::AddOperation(OperationWrapper &opWrapper,
const std::vector<std::string> &inTensorNames,
const std::vector<std::string> &outTensorNames)
{
atb::Operation *rawPtr = opWrapper.ReleaseOperation();
if (rawPtr == nullptr) {
throw std::runtime_error("add operation is null.");
}
Node node;
node.operation = rawPtr;
node.inTensorIds.resize(0);
node.outTensorIds.resize(0);
node.inTensorReshapeFuncs.resize(0);
for (const std::string &inTensorName : inTensorNames) {
node.inTensorIds.push_back(GetTensorId(inTensorName));
if (internalTensorNum_ > MAX_TENSOR_NUM) {
throw std::runtime_error("internalTensor num should be <= 256, but get " +
std::to_string(internalTensorNum_));
}
if (reshapedTensorIds_.find(inTensorName) != reshapedTensorIds_.end()) {
node.inTensorReshapeFuncs.push_back(reshapedTensorIds_[inTensorName].second);
} else {
node.inTensorReshapeFuncs.push_back(nullptr);
}
}
for (const std::string &outTensorName : outTensorNames) {
node.outTensorIds.push_back(GetTensorId(outTensorName));
if (internalTensorNum_ > MAX_TENSOR_NUM) {
throw std::runtime_error("internalTensor num should be <= 256, but get " +
std::to_string(internalTensorNum_));
}
}
graphParam_.nodes.push_back(node);
return *this;
}
GraphOperationBuilder &GraphOperationBuilder::Reshape(const std::string &srcTensorName,
const ReshapeHandler &reshapeHandler,
const std::string &reshapedTensorName)
{
atb::ReshapeFunc reshapeFunc = [reshapeHandler](const Dims &oldShape, Dims &newShape) {
std::vector<int64_t> oldShapeVec(oldShape.dimNum);
for (size_t i = 0; i < oldShape.dimNum; ++i) {
oldShapeVec[i] = oldShape.dims[i];
}
std::vector<int64_t> newShapeVec = reshapeHandler(oldShapeVec);
ATB_LOG(INFO) << "oldShapeVec: " << oldShapeVec << ", newShapeVec: " << newShapeVec;
newShape.dimNum = newShapeVec.size();
for (uint64_t i = 0; i < newShape.dimNum; ++i) {
newShape.dims[i] = newShapeVec[i];
}
};
reshapedTensorIds_[reshapedTensorName] = {GetTensorId(srcTensorName), reshapeFunc};
return *this;
}
OperationWrapper GraphOperationBuilder::Build()
{
graphParam_.internalTensorNum = internalTensorNum_;
return OperationWrapper(graphParam_);
}
uint32_t GraphOperationBuilder::GetTensorId(const std::string &tensorName)
{
if (inTensorIds_.find(tensorName) != inTensorIds_.end()) {
return inTensorIds_[tensorName];
} else if (outTensorIds_.find(tensorName) != outTensorIds_.end()) {
return outTensorIds_[tensorName];
} else if (reshapedTensorIds_.find(tensorName) != reshapedTensorIds_.end()) {
return reshapedTensorIds_[tensorName].first;
} else if (internalTensorIds_.find(tensorName) != internalTensorIds_.end()) {
return internalTensorIds_[tensorName];
} else {
uint32_t internalTensorId = inTensorIds_.size() + outTensorIds_.size() + internalTensorNum_++;
internalTensorIds_[tensorName] = internalTensorId;
return internalTensorId;
}
}
}