mirror of https://github.com/alibaba/MNN.git
82 lines
2.6 KiB
C++
82 lines
2.6 KiB
C++
//
|
|
// NPUInt8ToFloat.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/09/11.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "NPUInt8ToFloat.hpp"
|
|
#include "NPUBackend.hpp"
|
|
#include "../3rdParty/include/graph/op/all_ops.h"
|
|
#include <core/TensorUtils.hpp>
|
|
|
|
using namespace std;
|
|
|
|
namespace MNN {
|
|
|
|
using namespace ge;
|
|
NPUInt8ToFloat::NPUInt8ToFloat(Backend *b, const Op *op, const std::vector<Tensor *> &inputs,
|
|
const std::vector<Tensor *> &outputs)
|
|
: MNN::NPUCommonExecution(b,op) {}
|
|
|
|
|
|
ErrorCode NPUInt8ToFloat::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
|
|
mNpuBackend->setNetworkInput(inputs, mOp);
|
|
|
|
auto scale = mOp->main_as_QuantizedFloatParam()->tensorScale();
|
|
|
|
auto opName = mOp->name()->str();
|
|
shared_ptr<hiai::op::Scale> int8ToFloat(new hiai::op::Scale(opName));
|
|
|
|
auto xOp = mNpuBackend->getInputOps(mOp);
|
|
|
|
// om input filter const op
|
|
mConst_fliter = hiai::op::Const(opName + "_filter_const");
|
|
{
|
|
ge::TensorDesc fdesc(ge::Shape({1, scale->size(), 1, 1}), ge::FORMAT_NCHW, ge::DT_FLOAT);
|
|
ge::TensorPtr filter = std::make_shared<ge::Tensor>();
|
|
filter->SetTensorDesc(fdesc);
|
|
filter->SetData((uint8_t *)scale->data(), scale->size() * sizeof(float));
|
|
mConst_fliter.set_attr_value(filter);
|
|
}
|
|
|
|
mConstMin = hiai::op::Const(opName + "_clip_min");
|
|
{
|
|
float minData = -127;
|
|
ge::TensorDesc fdesc(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);
|
|
ge::TensorPtr constTensor = std::make_shared<ge::Tensor>();
|
|
constTensor->SetTensorDesc(fdesc);
|
|
constTensor->SetData((uint8_t *)(&minData), sizeof(float));
|
|
mConstMin.set_attr_value(constTensor);
|
|
}
|
|
|
|
mConstMax = hiai::op::Const(opName + "_clip_max");
|
|
{
|
|
float maxData = 127;
|
|
ge::TensorDesc fdesc(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);
|
|
ge::TensorPtr constTensor = std::make_shared<ge::Tensor>();
|
|
constTensor->SetTensorDesc(fdesc);
|
|
constTensor->SetData((uint8_t *)(&maxData), sizeof(float));
|
|
mConstMax.set_attr_value(constTensor);
|
|
}
|
|
|
|
shared_ptr<hiai::op::ClipByValue> clip(new hiai::op::ClipByValue(opName + "_clip"));
|
|
|
|
(*clip)
|
|
.set_input_x(*xOp)
|
|
.set_input_clip_value_min(mConstMin)
|
|
.set_input_clip_value_max(mConstMax);
|
|
|
|
(*int8ToFloat)
|
|
.set_input_x(*clip)
|
|
.set_input_scale(mConst_fliter);
|
|
|
|
mNpuBackend->setOutputOps(mOp, {clip, int8ToFloat}, outputs);
|
|
return NO_ERROR;
|
|
}
|
|
|
|
NPUCreatorRegister<TypedCreator<NPUInt8ToFloat>> __int8_to_float_op(OpType_Int8ToFloat);
|
|
|
|
} // namespace MNN
|