* 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.
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/reduce/reduce_common_util_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use \"#include \"adv_api/reduce/reduce.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_REDUCE_COMMON_UTIL_IMPL_H__
#endif
#ifndef IMPL_REDUCE_REDUCE_COMMON_UTIL_IMPL_H
#define IMPL_REDUCE_REDUCE_COMMON_UTIL_IMPL_H
#include "kernel_basic_intf.h"
#include "kernel_tensor.h"
#include "../common/check.h"
namespace AscendC {
namespace Pattern {
namespace Detail {
constexpr int32_t DIM_TWO = 2;
constexpr int32_t DIM_THREE = 3;
constexpr int32_t DIM_FOUR = 4;
constexpr int32_t DIM_FIVE = 5;
constexpr int32_t DIM_SIX = 6;
constexpr int32_t DIM_SEVEN = 7;
constexpr int32_t DIM_EIGHT = 8;
constexpr int32_t DIM_NINE = 9;
constexpr int32_t PATTERN_R = 0;
constexpr int32_t PATTERN_RA = 1;
constexpr int32_t PATTERN_AR = 2;
constexpr int32_t PATTERN_ARA = 3;
constexpr int32_t PATTERN_ARAR = 4;
constexpr int32_t PATTERN_ARARA = 5;
constexpr int32_t PATTERN_ARARAR = 6;
constexpr int32_t PATTERN_ARARARA = 7;
constexpr int32_t PATTERN_ARARARAR = 8;
constexpr int32_t PATTERN_ARARARARA = 9;
constexpr int32_t PATTERN_RAR = 10;
constexpr int32_t PATTERN_RARA = 11;
constexpr int32_t PATTERN_RARAR = 12;
constexpr int32_t PATTERN_RARARA = 13;
constexpr int32_t PATTERN_RARARAR = 14;
constexpr int32_t PATTERN_RARARARA = 15;
template <int32_t id, bool firstA, bool tailA, int32_t dim>
struct PatternConstInfo {
constexpr static int32_t ID = id;
constexpr static bool FirstA = firstA;
constexpr static bool TailA = tailA;
constexpr static int32_t Dim = dim;
};
}
}
namespace Internal {
enum class ApiMode : uint8_t { API_MODE_SUM = 0, API_MODE_MIN, API_MODE_MAX, API_MODE_ANY, API_MODE_ALL };
__aicore__ inline uint32_t FindClosestPowerOfTwo(uint32_t n)
{
ASCENDC_ASSERT(n != 0, { KERNEL_LOG(KERNEL_ERROR, "input n must be non-zero!"); });
constexpr uint32_t totalShiftBits = 63;
return totalShiftBits - ScalarCountLeadingZero(n);
}
template <class T>
__aicore__ inline void ComputeMaskBit(
uint32_t oneBlkMask, uint32_t oneBlkElems, uint32_t blkNum, uint64_t& maskLow, uint64_t& maskHigh)
{
if constexpr (sizeof(T) == sizeof(half)) {
uint32_t maskLBlkNum = blkNum > HALF_DEFAULT_REPEAT_STRIDE ? HALF_DEFAULT_REPEAT_STRIDE : blkNum;
uint32_t maskHBlkNum = blkNum - maskLBlkNum;
for (int32_t k = 0; k < maskLBlkNum; k++) {
maskLow += (((1ULL << oneBlkMask) - 1ULL) << (k * oneBlkElems));
}
for (int32_t k = 0; k < maskHBlkNum; k++) {
maskHigh += (((1ULL << oneBlkMask) - 1ULL) << (k * oneBlkElems));
}
} else if constexpr (sizeof(T) == sizeof(float)) {
for (int32_t k = 0; k < blkNum; k++) {
maskLow += (((1ULL << oneBlkMask) - 1ULL) << (k * oneBlkElems));
}
}
}
template <class T, ApiMode apiMode, MaskMode maskMode = MaskMode::NORMAL>
__aicore__ inline void BlockReduceCompute(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const int32_t repeat, const uint64_t mask[],
const int32_t blkStride, const int32_t repStride)
{
if constexpr (maskMode == MaskMode::NORMAL) {
if constexpr (apiMode == ApiMode::API_MODE_SUM) {
BlockReduceSum(dstTensor, srcTensor, repeat, mask, 1, 1, DEFAULT_REPEAT_STRIDE);
} else if constexpr (apiMode == ApiMode::API_MODE_MIN || apiMode == ApiMode::API_MODE_ALL) {
BlockReduceMin(dstTensor, srcTensor, repeat, mask, 1, 1, DEFAULT_REPEAT_STRIDE);
} else if constexpr (apiMode == ApiMode::API_MODE_MAX || apiMode == ApiMode::API_MODE_ANY) {
BlockReduceMax(dstTensor, srcTensor, repeat, mask, 1, 1, DEFAULT_REPEAT_STRIDE);
}
} else {
if constexpr (apiMode == ApiMode::API_MODE_MIN || apiMode == ApiMode::API_MODE_ALL) {
BlockReduceMin<T, false>(dstTensor, srcTensor, 1, MASK_PLACEHOLDER, 1, blkStride, repStride);
} else if constexpr (apiMode == ApiMode::API_MODE_MAX || apiMode == ApiMode::API_MODE_ANY) {
BlockReduceMax<T, false>(dstTensor, srcTensor, 1, MASK_PLACEHOLDER, 1, blkStride, repStride);
}
}
}
template <class T, ApiMode apiMode>
__aicore__ inline void WholeReduceCompute(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const int32_t repeat, const int32_t mask,
const int32_t repStride)
{
if constexpr (apiMode == ApiMode::API_MODE_MIN || apiMode == ApiMode::API_MODE_ALL) {
WholeReduceMin(dstTensor, srcTensor, mask, repeat, 1, 1, repStride, ReduceOrder::ORDER_ONLY_VALUE);
} else if constexpr (apiMode == ApiMode::API_MODE_MAX || apiMode == ApiMode::API_MODE_ANY) {
WholeReduceMax(dstTensor, srcTensor, mask, repeat, 1, 1, repStride, ReduceOrder::ORDER_ONLY_VALUE);
}
}
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_REDUCE_COMMON_UTIL_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_REDUCE_COMMON_UTIL_IMPL_H__
#endif