* 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.
*/
#ifndef LIB_REDUCE_REDUCE_H
#define LIB_REDUCE_REDUCE_H
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 3101 || __NPU_ARCH__ == 5102)
#include "reduce_common.h"
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
#include "kernel_pop_stack_buffer.h"
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2201)
#include "../../../impl/adv_api/detail/reduce/reduce_prod/reduce_prod_v220_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_max/reduce_max_v220_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_min/reduce_min_v220_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_sum/reduce_sum_v220_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_mean/reduce_mean_v220_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_any/reduce_any_v220_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_all/reduce_all_v220_impl.h"
#elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3101 || __NPU_ARCH__ == 5102)
#include "../../../impl/adv_api/detail/reduce/reduce_max/reduce_max_c310_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_min/reduce_min_c310_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_sum/reduce_sum_c310_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_mean/reduce_mean_c310_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_any/reduce_any_c310_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_all/reduce_all_c310_impl.h"
#include "../../../impl/adv_api/detail/reduce/reduce_prod/reduce_prod_c310_impl.h"
#endif
#if ASCENDC_CPU_DEBUG
#include "kernel_log.h"
#include <type_traits>
#endif
namespace AscendC {
#pragma begin_pipe(V)
* \brief This function calculates the mean of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceMean(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceMeanImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the mean of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceMean(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceMean<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the max value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/half
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceMax(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceMaxImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the max value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/half
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceMax(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceMax<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the min value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/half
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceMin(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceMinImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the min value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/half
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceMin(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceMin<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the sum of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceSum(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceSumImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the sum of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceSum(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceSum<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the product of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceProd(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceProdImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates the product of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceProd(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceProd<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates any true value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/uint8_t
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceAny(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceAnyImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates any true value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceAny(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceAny<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates all true value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/uint8_t
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] sharedTmpBuffer, input local temporary Tensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceAll(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
Internal::ReduceAllImpl<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
* \brief This function calculates all true value of all elements in the input tensor based on the reduction pattern.
*
* \note support data type: float/uint8_t
*
* \tparam [in] T, data type
* \tparam [in] pattern, reduce pattern of input tensor, each character is a dimension
* \tparam [in] isReuseSource, whether allows API to modify source data, usually for performance reason
* \param [in] srcTensor, input LocalTensor
* \param [in] srcShape, actual shape used to reduce information of input tensor, its dim must be matched with pattern
* \param [in] srcInnerPad, whether the last axis of input tensor is padded to 32B aligned up.
* \param [out] dstTensor, output LocalTensor
*/
template <class T, class pattern, bool isReuseSource = false>
__aicore__ inline void ReduceAll(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t srcShape[], bool srcInnerPad)
{
if ASCEND_IS_AIC {
return;
}
LocalTensor<uint8_t> sharedTmpBuffer;
bool ans = PopStackBuffer<uint8_t, TPosition::LCM>(sharedTmpBuffer);
ASCENDC_ASSERT((ans), { KERNEL_LOG(KERNEL_ERROR, "PopStackBuffer Error!"); });
ReduceAll<T, pattern, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, srcShape, srcInnerPad);
}
#pragma end_pipe
}
#endif
#endif
