* 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 ascend_dequant.h
* \brief
*/
#ifndef LIB_QUANTIZATION_ASCEND_DEQUANT_H
#define LIB_QUANTIZATION_ASCEND_DEQUANT_H
#include "include/adv_api/quantization/ascend_dequant_utils.h"
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 2002 || __NPU_ARCH__ == 3101 || \
__NPU_ARCH__ == 5102 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
#include "kernel_tensor.h"
#include "../../../impl/adv_api/detail/quantization/dequant/ascend_dequant_common_impl.h"
namespace AscendC {
#pragma begin_pipe(V)
* \ingroup AscendDequant
* \brief For DequantParams {m, n, calcount}, m means src tensor has m rows, each row has n num and the first calcount
* num will mul with corresponding num in deqScale.
* Ex: src(x, y) means yth num in the xth row in srcTensor
* Then dst(x, y) = src(x, y) * deqScale(y) if y is in range [0, calCount)
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor
* \param [in] deqScale: Input deqScale localTensor
* \param [in] sharedTmpBuffer： extra temporary shared space used for intermediate values among calculation process,
* whose required space size should refer to corresponding tiling API, which is defined at
* ascend_dequant_tiling.h. Generally, the more space you allocate, the better performance you will achieve,
* and the performance reaches peak when buffer size is maximum(calculated by tiling function). Moreover, it
* is not guaranteed that the shared space will be cleared after usage, the data could be anything.
* \param [in] params: DequantParams with m, n, calcount to describe the calculation process like above.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const LocalTensor<scaleT>& deqScale, const LocalTensor<uint8_t>& sharedTmpBuffer, DequantParams params)
{
AscendDequantImpl<dstT, scaleT, true, mode>(dstTensor, srcTensor, deqScale, sharedTmpBuffer, params,
params.m * params.n);
}
* \ingroup AscendDequant
* \brief For DequantParams {m, n, calcount}, m means src tensor has m rows, each row has n num and the first calcount
* num will mul with corresponding num in deqScale.
* Ex: src(x, y) means yth num in the xth row in srcTensor
* Then dst(x, y) = src(x, y) * deqScale(y) if y is in range [0, calCount)
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor
* \param [in] deqScale: Input deqScale localTensor
* \param [in] params: DequantParams with m, n, calcount to describe the calculation process like above.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const LocalTensor<scaleT>& deqScale, DequantParams params)
{
if ASCEND_IS_AIC {
return;
}
AscendDequantImpl<dstT, scaleT, mode>(dstTensor, srcTensor, deqScale, params);
}
* \ingroup AscendDequant
* \brief For AscendDequant function with calCount, assume that deqScale has n num, then for each n num in srcTensor,
* corresponding dst result = src(i) * deqScale(i) for index i (0 <= i < deqScale.GetSize())
* Note: must srcTensor.GetSize() % deqScale.GetSize() = 0
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor
* \param [in] deqScale: Input deqScale localTensor
* \param [in] sharedTmpBuffer：extra temporary shared space used for intermediate values among calculation process,
* whose required space size should refer to corresponding tiling API, which is defined at
* ascend_dequant_tiling.h. Generally, the more space you allocate, the better performance you will achieve,
* and the performance reaches peak when buffer size is maximum(calculated by tiling function). Moreover, it
* is not guaranteed that the shared space will be cleared after usage, the data could be anything.
* \param [in] calCount: The number of elements in srcTensor to be processed.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const LocalTensor<scaleT>& deqScale, const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t calCount)
{
AscendDequantCalcountImpl<dstT, scaleT, mode>(dstTensor, srcTensor, deqScale, sharedTmpBuffer, calCount);
}
* \ingroup AscendDequant
* \brief For AscendDequant function with calCount, assume that deqScale has n num, then for each n num in srcTensor,
* corresponding dst result = src(i) * deqScale(i) for index i (0 <= i < deqScale.GetSize())
* Note: must srcTensor.GetSize() % deqScale.GetSize() = 0
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor
* \param [in] deqScale: Input deqScale localTensor
* \param [in] calCount: The number of elements in srcTensor to be processed.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const LocalTensor<scaleT>& deqScale, const uint32_t calCount)
{
if ASCEND_IS_AIC {
return;
}
AscendDequantCalcountImpl<dstT, scaleT, mode>(dstTensor, srcTensor, deqScale, calCount);
}
* \ingroup AscendDequant
* \brief For AscendDequant function without calCount and dequantParams, assume that deqScale has n num, then for each
* n num in srcTensor, corresponding dst result = src(i) * deqScale(i) for index i (0 <= i < deqScale.GetSize())
* It is equivalent to calcount set as srcTensor.GetSize()
* Note: must srcTensor.GetSize() % deqScale.GetSize() = 0
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor
* \param [in] deqScale: Input deqScale localTensor
* \param [in] sharedTmpBuffer： extra temporary shared space used for intermediate values among calculation process,
* whose required space size should refer to corresponding tiling API, which is defined at
* ascend_dequant_tiling.h. Generally, the more space you allocate, the better performance you will achieve,
* and the performance reaches peak when buffer size is maximum(calculated by tiling function). Moreover, it
* is not guaranteed that the shared space will be cleared after usage, the data could be anything.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const LocalTensor<scaleT>& deqScale, const LocalTensor<uint8_t>& sharedTmpBuffer)
{
AscendDequantNoCalcountImpl<dstT, scaleT, mode>(dstTensor, srcTensor, deqScale, sharedTmpBuffer);
}
* \ingroup AscendDequant
* \brief For AscendDequant function without calCount and dequantParams, assume that deqScale has n num, then for each
* n num in srcTensor, corresponding dst result = src(i) * deqScale(i) for index i (0 <= i < deqScale.GetSize())
* It is equivalent to calcount set as srcTensor.GetSize()
* Note: must srcTensor.GetSize() % deqScale.GetSize() = 0
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor
* \param [in] deqScale: Input deqScale localTensor
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const LocalTensor<scaleT>& deqScale)
{
if ASCEND_IS_AIC {
return;
}
AscendDequantNoCalcountImpl<dstT, scaleT, mode>(dstTensor, srcTensor, deqScale);
}
* \ingroup AscendDequant
* \brief For DequantParams {m, n, calcount}, m means src tensor has m rows, each row has n num and the first calcount
* num will mul with scalar deqScale.
* Ex: src(x, y) means yth num in the xth row in srcTensor
* Then dst(x, y) = src(x, y) * deqScale if y is in range [0, calCount)
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor.
* \param [in] deqScale: Input deqScale scalar.
* \param [in] sharedTmpBuffer：extra temporary shared space used for intermediate values among calculation process,
* whose required space size should refer to corresponding tiling API, which is defined at
* ascend_dequant_tiling.h. Generally, the more space you allocate, the better performance you will achieve,
* and the performance reaches peak when buffer size is maximum(calculated by tiling function). Moreover, it
* is not guaranteed that the shared space will be cleared after usage, the data could be anything.
* \param [in] params: DequantParams with m, n, calcount to describe the calculation process like above.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const scaleT deqScale, const LocalTensor<uint8_t>& sharedTmpBuffer, DequantParams params)
{
if ASCEND_IS_AIC {
return;
}
AscendDequantScalarImpl<dstT, scaleT, true, mode>(dstTensor, srcTensor, deqScale, sharedTmpBuffer, params);
}
* \ingroup AscendDequant
* \brief For DequantParams {m, n, calcount}, m means src tensor has m rows, each row has n num and the first calcount
* num will mul with scalar deqScale.
* Ex: src(x, y) means yth num in the xth row in srcTensor
* Then dst(x, y) = src(x, y) * deqScale if y is in range [0, calCount)
* \tparam dstT: dstTensor data type
* \tparam scaleT: deqScale tensor data type.
* \tparam mode: deqScale calculate mode when dequantParams is in format {1, m*n, n}
* If mode = DEQUANT_WITH_SINGLE_ROW, then {1, m*n, n} will be converted to {m, n, n} and then process.
* If mode = DEQUANT_WITH_MULTI_ROW, then {1, m*n, n} will be transferred for following process.
* \param [out] dstTensor: Output localTensor.
* \param [in] srcTensor: Input src localTensor.
* \param [in] deqScale: Input deqScale scalar.
* \param [in] params: DequantParams with m, n, calcount to describe the calculation process like above.
*/
template <typename dstT, typename scaleT, DeQuantMode mode = DeQuantMode::DEQUANT_WITH_SINGLE_ROW>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<int32_t>& srcTensor,
const scaleT deqScale, DequantParams params)
{
if ASCEND_IS_AIC {
return;
}
AscendDequantScalarImpl<dstT, scaleT, mode>(dstTensor, srcTensor, deqScale, params);
}
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3101 || __NPU_ARCH__ == 5102)
template <typename dstT, typename srcT, typename scaleT, const AscendDeQuantConfig& config, const AscendDeQuantPolicy& policy>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<srcT>& srcTensor,
const LocalTensor<scaleT> &scaleTensor, const LocalTensor<scaleT> &offsetTensor,
const AscendDeQuantParam& para)
{
AscendDequantImpl<dstT, srcT, scaleT, config, policy>(dstTensor, srcTensor, scaleTensor, offsetTensor, para);
}
template <typename dstT, typename srcT, typename scaleT, const AscendDeQuantConfig& config, const AscendDeQuantPolicy& policy>
__aicore__ inline void AscendDequant(const LocalTensor<dstT>& dstTensor, const LocalTensor<srcT>& srcTensor,
const LocalTensor<scaleT> &scaleTensor, const LocalTensor<scaleT> &offsetTensor,
const LocalTensor<uint8_t> &sharedTmpBuffer, const AscendDeQuantParam& para)
{
AscendDequantImpl<dstT, srcT, scaleT, config, policy>(dstTensor, srcTensor, sharedTmpBuffer, scaleTensor, offsetTensor, para);
}
#endif
#pragma end_pipe
}
#endif
#endif
