* Copyright (c) 2025 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.
*/
* \file quant.cpp
* \brief
*/
#include "tilefwk/tilefwk.h"
#include "interface/inner/tilefwk.h"
using namespace npu::tile_fwk;
namespace npu::tile_fwk {
constexpr float F_1 = 1.0;
constexpr float F_127 = 127.0;
constexpr float F_255 = 255.0;
constexpr float F_1E_12 = 1e-12f;
std::tuple<Tensor, Tensor> Quant(const Tensor& input, bool isSymmetry, bool hasSmoothFactor, const Tensor& smoothFactor)
{
auto inputFp32 = Cast(input, DataType::DT_FP32, CAST_NONE);
if (hasSmoothFactor) {
inputFp32 = Mul(inputFp32, smoothFactor);
}
if (isSymmetry) {
auto absRes = Abs(inputFp32);
auto maxValue = Amax(absRes, -1, true);
auto scaleQuant = ScalarDivS(maxValue, Element(DataType::DT_FP32, F_127), true);
auto outFp32 = Mul(inputFp32, scaleQuant);
auto outInt32 = Cast(outFp32, DataType::DT_INT32, CAST_RINT);
auto outHalf = Cast(outInt32, DataType::DT_FP16, CAST_ROUND);
auto outInt8 = Cast(outHalf, DataType::DT_INT8, CAST_TRUNC, SaturationMode::ON);
auto scaleDeQuant = ScalarDivS(scaleQuant, Element(DataType::DT_FP32, F_1), true);
return std::tie(outInt8, scaleDeQuant);
} else {
auto maxValue = Amax(inputFp32, -1, true);
auto minValue = Amin(inputFp32, -1, true);
auto scaleDeQuant = ScalarMaxS(
ScalarDivS(ScalarSub(maxValue, minValue), Element(DataType::DT_FP32, F_255)),
Element(DataType::DT_FP32, F_1E_12));
auto offset = ScalarSubS(ScalarDiv(maxValue, scaleDeQuant), Element(DataType::DT_FP32, F_127), true);
auto scaleQuant = ScalarDivS(scaleDeQuant, Element(DataType::DT_FP32, F_1), true);
auto outFp32 = Mul(inputFp32, scaleQuant);
auto outInt32 = Cast(outFp32, DataType::DT_INT32, CAST_RINT);
auto outHalf = Cast(outInt32, DataType::DT_FP16, CAST_ROUND);
auto outInt8 = Cast(outHalf, DataType::DT_INT8, CAST_TRUNC, SaturationMode::ON);
return std::tie(outInt8, scaleDeQuant);
}
}
}
namespace npu::tile_fwk::Matrix {
Tensor QuantMM(const Tensor& operand1, const Tensor& operand2, const Tensor& dequantScaleW)
{
auto quantA = Quant(operand1);
auto quantizedA = std::get<0>(quantA);
auto dequantScaleA = std::get<1>(quantA);
Tensor res;
if (operand1.GetShape().size() == NUM_VALUE_2) {
res = Matmul(DataType::DT_INT32, quantizedA, operand2, false, false);
} else if (operand1.GetShape().size() == NUM_VALUE_3) {
res = BatchMatmul(DataType::DT_INT32, quantizedA, operand2);
} else {
ASSERT(VectorErrorCode::ERR_PARAM_INVALID, operand1.GetShape().size() <= NUM_VALUE_3)
<< "QuantMM only supports 2D or 3D tensors.";
}
res = Cast(res, DataType::DT_FP32);
res = Mul(res, dequantScaleA);
res = Mul(res, dequantScaleW);
res = Cast(res, DataType::DT_BF16, CAST_RINT);
return res;
}
}