* Copyright (c) 2024 Huawei Technologies Co., Ltd.
* 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 "transdata_operation.h"
#include "transdata_ops_runner.h"
#include "atb/utils/tensor_check.h"
#include "atb/utils/param_to_json.h"
#include "atb/utils/singleton.h"
#include "atb/operation/atb_operation_ir_cfg.h"
#include "atb/operation/op_param_funcs.h"
static const uint32_t IN_TENSOR_NUM = 1;
static const uint32_t OUT_TENSOR_NUM = 1;
static const uint32_t ALIGN_FLOAT16 = 16;
static const uint32_t ALIGN_INT8 = 32;
static const uint32_t ALIGN_BF16 = 16;
static constexpr int64_t DIM_1 = 1;
static constexpr int64_t DIM_2 = 2;
static constexpr int64_t DIM_3 = 3;
static constexpr int64_t DIM_4 = 4;
static const int64_t DEFAULT_ALIGN = 16;
namespace atb {
template <> Status CreateOperation(const infer::TransdataParam &opParam, Operation **operation)
{
if (operation == nullptr) {
return ERROR_INVALID_PARAM;
}
OP_PARAM_RSV_CHECK(opParam);
*operation = new (std::nothrow) TransdataOperation(opParam);
if (*operation == nullptr) {
ATB_LOG(ERROR) << "failed to new operation";
return ERROR_OUT_OF_HOST_MEMORY;
}
return NO_ERROR;
}
TransdataOperation::TransdataOperation(const infer::TransdataParam ¶m)
: OperationBase("TransdataOperation"), param_(param)
{
std::string opIrKey;
if (GetSingleton<Config>().Is310B()) {
opIrKey = param_.transdataType == atb::infer::TransdataParam::TransdataType::ND_TO_FRACTAL_NZ ?
"TransdataOperationNdToNzAtlas200I500A2" : "TransdataOperationNzToNdAtlas200I500A2";
} else {
opIrKey = param_.transdataType == atb::infer::TransdataParam::TransdataType::ND_TO_FRACTAL_NZ ?
"TransdataOperationNdToNz" : "TransdataOperationNzToNd";
}
operationIr_ = GetSingleton<AtbOperationIrCfg>().GetOperationIr(opIrKey);
}
TransdataOperation::~TransdataOperation() {}
uint32_t TransdataOperation::GetInputNum() const
{
return IN_TENSOR_NUM;
}
uint32_t TransdataOperation::GetOutputNum() const
{
return OUT_TENSOR_NUM;
}
Status TransdataOperation::InferShapeImpl(const SVector<TensorDesc> &inTensorDescs,
SVector<TensorDesc> &outTensorDescs) const
{
static std::map<aclDataType, uint32_t> alignTable = {
{ACL_FLOAT16, ALIGN_FLOAT16},
{ACL_INT8, ALIGN_INT8},
{ACL_BF16, ALIGN_BF16},
};
aclDataType inTensorDType = inTensorDescs.at(0).dtype;
std::map<aclDataType, uint32_t>::const_iterator it = alignTable.find(inTensorDType);
if (it == alignTable.end()) {
return ERROR_INVALID_PARAM;
}
int64_t align = static_cast<int64_t>(it->second);
outTensorDescs.at(0).dtype = inTensorDescs.at(0).dtype;
auto inTensorDims = inTensorDescs.at(0).shape.dimNum;
if (param_.transdataType == atb::infer::TransdataParam::ND_TO_FRACTAL_NZ) {
outTensorDescs.at(0).format = ACL_FORMAT_FRACTAL_NZ;
outTensorDescs.at(0).shape.dimNum = DIM_4;
const bool dim3Flag = (inTensorDims == DIM_3);
auto outTensorDim0 = dim3Flag ? inTensorDescs.at(0).shape.dims[0] : 1;
int64_t outTensorDim1 =
OperationUtil::RoundUp(inTensorDescs.at(0).shape.dims[dim3Flag ? DIM_2 : DIM_1], align) / align;
auto outTensorDim2 =
OperationUtil::RoundUp(inTensorDescs.at(0).shape.dims[dim3Flag ? DIM_1 : 0], DEFAULT_ALIGN);
int64_t outTensorDim3 = align;
outTensorDescs.at(0).shape.dims[0] = outTensorDim0;
outTensorDescs.at(0).shape.dims[DIM_1] = outTensorDim1;
outTensorDescs.at(0).shape.dims[DIM_2] = outTensorDim2;
outTensorDescs.at(0).shape.dims[DIM_3] = outTensorDim3;
} else {
outTensorDescs.at(0).format = ACL_FORMAT_ND;
outTensorDescs.at(0).shape.dimNum = DIM_3;
outTensorDescs.at(0).shape.dims[0] = inTensorDescs.at(0).shape.dims[0];
if (param_.outCrops[0] > 0 && param_.outCrops[1] > 0) {
outTensorDescs.at(0).shape.dims[1] = param_.outCrops[0];
outTensorDescs.at(0).shape.dims[DIM_2] = param_.outCrops[1];
} else {
ATB_LOG(ERROR) << "outCrops can not <= 0!";
return ERROR_INVALID_PARAM;
}
}
return NO_ERROR;
}
Status TransdataOperation::OutCropsCheck(const TensorDesc &inTensorDesc) const
{
int64_t outCrop0Lower = OperationUtil::RoundUp(inTensorDesc.shape.dims[DIM_2], ALIGN_FLOAT16) - ALIGN_FLOAT16;
int64_t outCrop0Upper = OperationUtil::RoundUp(inTensorDesc.shape.dims[DIM_2], ALIGN_FLOAT16);
int64_t outCrop1Lower =
OperationUtil::RoundUp(inTensorDesc.shape.dims[1] * inTensorDesc.shape.dims[DIM_3], ALIGN_FLOAT16) -
ALIGN_FLOAT16;
int64_t outCrop1Upper =
OperationUtil::RoundUp(inTensorDesc.shape.dims[1] * inTensorDesc.shape.dims[DIM_3], ALIGN_FLOAT16);
if (param_.outCrops[0] > outCrop0Upper || param_.outCrops[0] <= outCrop0Lower ||
param_.outCrops[1] > outCrop1Upper || param_.outCrops[1] <= outCrop1Lower) {
ATB_LOG(ERROR) << "outCrops not in the valid range!";
return ERROR_INVALID_PARAM;
}
return NO_ERROR;
}
Status TransdataOperation::InferShapeCheckImpl(const SVector<TensorDesc> &inTensorDescs) const
{
if (param_.transdataType == atb::infer::TransdataParam::TransdataType::ND_TO_FRACTAL_NZ) {
if (inTensorDescs.at(0).shape.dimNum != DIM_2 && inTensorDescs.at(0).shape.dimNum != DIM_3) {
ATB_LOG(ERROR) << GetLogPrefix() << "ND_TO_FRACTAL_NZ inTensor DimNum should be 2 or 3";
return ERROR_INVALID_TENSOR_DIM;
}
} else if (param_.transdataType == atb::infer::TransdataParam::TransdataType::FRACTAL_NZ_TO_ND) {
if (inTensorDescs.at(0).shape.dimNum != DIM_4) {
ATB_LOG(ERROR) << GetLogPrefix() << "FRACTAL_NZ_TO_ND inTensor should be DimNum 4";
return ERROR_INVALID_TENSOR_DIM;
}
if (param_.outCrops.size() != DIM_2) {
ATB_LOG(ERROR) << "outCrops size should be 2!";
return ERROR_INVALID_PARAM;
}
return OutCropsCheck(inTensorDescs.at(0));
} else {
ATB_LOG(ERROR) << "transdataType is not support, only support ND_TO_FRACTAL_NZ and "
"FRACTAL_NZ_TO_ND";
return ERROR_INVALID_PARAM;
}
return NO_ERROR;
}
Status TransdataOperation::SetupCheckImpl(const SVector<Tensor> &inTensors, const SVector<Tensor> &outTensors) const
{
if (param_.transdataType == atb::infer::TransdataParam::TransdataType::ND_TO_FRACTAL_NZ) {
if ((inTensors.at(0).desc.shape.dimNum != DIM_2 && inTensors.at(0).desc.shape.dimNum != DIM_3) ||
outTensors.at(0).desc.shape.dimNum != DIM_4) {
ATB_LOG(ERROR) << GetLogPrefix() << "ND_TO_FRACTAL_NZ inTensor DimNum should be 2 or 3, "
<< "outTensor DimNum should be 4.";
return ERROR_INVALID_TENSOR_DIM;
}
} else {
if (inTensors.at(0).desc.shape.dimNum != DIM_4 || outTensors.at(0).desc.shape.dimNum != DIM_3) {
ATB_LOG(ERROR) << GetLogPrefix() << "FRACTAL_NZ_TO_ND inTensor should be DimNum 4, "
<< "outTensor should be DimNum 3.";
return ERROR_INVALID_TENSOR_DIM;
}
if (param_.outCrops.size() != DIM_2) {
ATB_LOG(ERROR) << "outCrops size should be 2!";
return ERROR_INVALID_PARAM;
}
return OutCropsCheck(inTensors.at(0).desc);
}
return NO_ERROR;
}
std::shared_ptr<Runner> TransdataOperation::CreateRunner(Context &context) const
{
(void)context;
return std::make_shared<TransdataOpsRunner>(param_);
}
nlohmann::json TransdataOperation::GetParamJson() const
{
return OpParamToJson(param_);
}
}
