* 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 trunc.h
* \brief Defines a series of interface used to do elementwise math Truncation calculation.
* Get the interger part of float value, towards zero.
* e.g. Trunc(3.1) = 3, Trunc(-3.1) = -3, Trunc(3.9) = 3, Trunc(-3.9) = -3
*/
#if defined(__NPU_COMPILER_INTERNAL_PURE_SIMT__)
#error "trunc.h cannot be used with compile flag --enable-simt enabled."
#endif
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_TRUNC_H__
#endif
#ifndef LIB_MATH_TRUNC_H
#define LIB_MATH_TRUNC_H
#include "kernel_tensor.h"
#if ( \
defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 2002 || __NPU_ARCH__ == 3510 || __NPU_ARCH__ == 9201 || \
__NPU_ARCH__ == 5102 || __NPU_ARCH__ == 5162 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113))
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2002 || __NPU_ARCH__ == 2201)
#include "../../../impl/adv_api/detail/math/trunc/trunc_common_impl.h"
#else
#include "../../../impl/adv_api/detail/math/trunc/trunc_3510_impl.h"
#endif
#endif
namespace AscendC {
#pragma begin_pipe(V)
* \ingroup Trunc
* \brief compute Truncation elementwisely
* \tparam T: half/float
* \tparam isReuseSource: whether allows API to modify source data, usually for performance reason
* \param [out] dstTensor: output LocalTensor
* \param [in] srcTensor: input 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 trunc_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 to be processed.
* \note src/dst Tensor must be 32B aligned, and it doesn't allow src/dst/sharedTmpBuffer tensor address overlap.
*/
template <typename T, bool isReuseSource = false>
__aicore__ inline void Trunc(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const uint32_t calCount)
{
#if ( \
defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 2002 || __NPU_ARCH__ == 3510 || __NPU_ARCH__ == 9201 || \
__NPU_ARCH__ == 5102 || __NPU_ARCH__ == 5162 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113))
TruncImpl<T, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, calCount);
#endif
}
* \ingroup Trunc
* \brief compute Truncation elementwisely
* \tparam T: half/float
* \tparam isReuseSource: whether allows API to modify source data, usually for performance reason
* \param [out] dstTensor: output LocalTensor
* \param [in] srcTensor: input LocalTensor
* \param [in] calCount: the number of elements to be processed.
* \note src/dst Tensor must be 32B aligned, and it doesn't allow src/dst/sharedTmpBuffer tensor address overlap.
*/
template <typename T, bool isReuseSource = false>
__aicore__ inline void Trunc(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t calCount)
{
#if ( \
defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 2002 || __NPU_ARCH__ == 3510 || __NPU_ARCH__ == 9201 || \
__NPU_ARCH__ == 5102 || __NPU_ARCH__ == 5162 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113))
TruncImpl<T, isReuseSource>(dstTensor, srcTensor, calCount);
#endif
}
* \ingroup Trunc
* \brief compute Truncation elementwisely for whole source tensor
* \tparam T: half/float
* \tparam isReuseSource: whether allows API to modify source data, usually for performance reason
* \param [out] dstTensor: output LocalTensor
* \param [in] srcTensor: input 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 trunc_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.
* \note src/dst Tensor must be 32B aligned, and it doesn't allow src/dst/sharedTmpBuffer tensor address overlap.
*/
template <typename T, bool isReuseSource = false>
__aicore__ inline void Trunc(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer)
{
Trunc<T, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcTensor.GetSize());
}
* \ingroup Trunc
* \brief compute Truncation elementwisely for whole source tensor
* \tparam T: half/float
* \tparam isReuseSource: whether allows API to modify source data, usually for performance reason
* \param [out] dstTensor: output LocalTensor
* \param [in] srcTensor: input LocalTensor
* \note src/dst Tensor must be 32B aligned, and it doesn't allow src/dst/sharedTmpBuffer tensor address overlap.
*/
template <typename T, bool isReuseSource = false>
__aicore__ inline void Trunc(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor)
{
Trunc<T, isReuseSource>(dstTensor, srcTensor, srcTensor.GetSize());
}
#pragma end_pipe
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_TRUNC_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_TRUNC_H__
#endif
