* 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 power_int_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/power/power_int_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/math/power.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_POWER_POWER_INT_IMPL_H__
#endif
#ifndef IMPL_MATH_POWER_POWER_INT_IMPL_H
#define IMPL_MATH_POWER_POWER_INT_IMPL_H
#include "kernel_basic_intf.h"
#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 2201
#include "power_v220_impl.h"
#elif defined(__NPU_ARCH__) && __NPU_ARCH__ == 2002
#include "power_v200_impl.h"
#endif
namespace AscendC {
* \ingroup PowerI
* \brief Intrinsics of Power(Int Input).
*/
__aicore__ inline void InitFinePowerI(
AscPowerIParams& param, const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam,
const uint8_t& repeat, const uint32_t calCount)
{
Sub<int32_t, false>(param.recordExpNode, param.oriAbsExp, param.recordExpNode, MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
CompareZeroPositive(param.mask, param.recordExpNode, unaryParam, repeat);
}
__aicore__ inline void FineProcessPowerI(
const LocalTensor<int32_t>& dst, const LocalTensor<int32_t>& src0, AscPowerIParams& param,
const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam, const uint8_t& repeat,
const uint32_t calCount)
{
Cast<float, int32_t, false>(
param.tmpTensor1.ReinterpretCast<float>(), param.recordExpNode, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
unaryParam);
PipeBarrier<PIPE_V>();
ReduceSumCount(
param.tmpScalar.ReinterpretCast<float>(), param.tmpTensor1.ReinterpretCast<float>(),
param.tmpTensor3.ReinterpretCast<float>(), calCount);
param.expIterateSum = param.tmpScalar.ReinterpretCast<float>().GetValue(0);
PipeBarrier<PIPE_V>();
if (param.expIterateSum != 0) {
Mul<int32_t, false>(param.tmpTensor1, dst, src0, MASK_PLACEHOLDER, 1, binaryParam);
Adds<int32_t, false>(param.tmpTensor2, param.recordExpNode, -1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
VselPowerTensorTensor(
dst.ReinterpretCast<float>(), param.mask, dst.ReinterpretCast<float>(),
param.tmpTensor1.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(),
SELMODE::VSEL_TENSOR_TENSOR_MODE, 1, binaryParam, calCount);
VselPowerTensorTensor(
param.recordExpNode.ReinterpretCast<float>(), param.mask, param.recordExpNode.ReinterpretCast<float>(),
param.tmpTensor2.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(),
SELMODE::VSEL_TENSOR_TENSOR_MODE, 1, binaryParam, calCount);
PipeBarrier<PIPE_V>();
CompareZeroPositive(param.mask, param.recordExpNode, unaryParam, repeat);
}
}
__aicore__ inline void BulkProcessPowerI(
const LocalTensor<int32_t>& dst, AscPowerIParams& param, const UnaryRepeatParams& unaryParam,
const BinaryRepeatParams& binaryParam, const uint8_t& repeat, const uint32_t calCount)
{
Cast<float, int32_t, false>(
param.tmpTensor1.ReinterpretCast<float>(), param.expUBIterate, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
unaryParam);
PipeBarrier<PIPE_V>();
ReduceSumCount(
param.tmpScalar.ReinterpretCast<float>(), param.tmpTensor1.ReinterpretCast<float>(),
param.tmpTensor3.ReinterpretCast<float>(), calCount);
param.expIterateSum = param.tmpScalar.ReinterpretCast<float>().GetValue(0);
PipeBarrier<PIPE_V>();
if (param.expIterateSum != 0) {
Mul<int32_t, false>(param.tmpTensor1, dst, dst, MASK_PLACEHOLDER, 1, binaryParam);
int32_t scalarValue = 2;
Muls<int32_t, false>(param.tmpTensor2, param.recordExpNode, scalarValue, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
VselPowerTensorTensor(
dst.ReinterpretCast<float>(), param.mask, dst.ReinterpretCast<float>(),
param.tmpTensor1.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(),
SELMODE::VSEL_TENSOR_TENSOR_MODE, 1, binaryParam, calCount);
VselPowerTensorTensor(
param.recordExpNode.ReinterpretCast<float>(), param.mask, param.recordExpNode.ReinterpretCast<float>(),
param.tmpTensor2.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(),
SELMODE::VSEL_TENSOR_TENSOR_MODE, 1, binaryParam, calCount);
ShiftRightOneBit(param.expUBIterate, param.tmpTensor3, unaryParam, binaryParam, calCount);
PipeBarrier<PIPE_V>();
CompareIntZero(param.mask, param.expUBIterate, param.tmpTensor3, unaryParam, repeat);
}
}
__aicore__ inline void InitBulkPowerI(
AscPowerIParams& param, const LocalTensor<int32_t>& src0, const LocalTensor<int32_t>& src1,
const LocalTensor<int32_t>& dst, const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam,
const uint8_t& repeat, const uint32_t calCount)
{
Cast<float, int32_t, false>(
param.tmpTensor1.ReinterpretCast<float>(), src1, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CompareScalar<float, uint8_t, false>(
param.negMask, param.tmpTensor1.ReinterpretCast<float>(), static_cast<float>(0), CMPMODE::LT, MASK_PLACEHOLDER,
repeat, unaryParam);
PipeBarrier<PIPE_V>();
Muls<int32_t, false>(param.tmpTensor1, src1, -1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
VselPowerTensorTensor(
param.expUBIterate.ReinterpretCast<float>(), param.negMask, param.tmpTensor1.ReinterpretCast<float>(),
src1.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(), SELMODE::VSEL_TENSOR_TENSOR_MODE, 1,
binaryParam, calCount);
PipeBarrier<PIPE_V>();
Muls<int32_t, false>(param.oriAbsExp, param.expUBIterate, 1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CompareZeroPositive(param.mask, param.oriAbsExp, unaryParam, repeat);
Duplicate<int32_t, false>(param.recordExpNode, 0, MASK_PLACEHOLDER, 1, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
Duplicate<int32_t, false>(param.tmpTensor2, 1, MASK_PLACEHOLDER, 1, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
VselPowerTensorTensor(
param.recordExpNode.ReinterpretCast<float>(), param.mask, param.recordExpNode.ReinterpretCast<float>(),
param.tmpTensor2.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(),
SELMODE::VSEL_TENSOR_TENSOR_MODE, 1, binaryParam, calCount);
VselPowerTensorTensor(
dst.ReinterpretCast<float>(), param.mask, param.tmpTensor2.ReinterpretCast<float>(),
src0.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(), SELMODE::VSEL_TENSOR_TENSOR_MODE, 1,
binaryParam, calCount);
ShiftRightOneBit(param.expUBIterate, param.tmpTensor3, unaryParam, binaryParam, calCount);
PipeBarrier<PIPE_V>();
CompareZeroPositive(param.mask, param.expUBIterate, unaryParam, repeat);
}
__aicore__ inline void HandleNegativeExpPowerI(
const LocalTensor<int32_t>& dst, const LocalTensor<int32_t>& src0, AscPowerIParams& param,
const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam, const uint8_t& repeat,
const uint32_t calCount)
{
CompareIntZero(param.mask, dst, param.tmpTensor3, unaryParam, repeat);
LocalTensor<float> resF32 = param.oriAbsExp.ReinterpretCast<float>();
LocalTensor<float> oneTensor = param.expUBIterate.ReinterpretCast<float>();
Cast<float, int32_t, false>(resF32, dst, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1, unaryParam);
Duplicate<float, false>(oneTensor, 1.0f, MASK_PLACEHOLDER, 1, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
Div<float, false>(resF32, oneTensor, resF32, MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
Cast<int32_t, float, false>(param.tmpTensor1, resF32, RoundMode::CAST_TRUNC, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
VselPowerTensorTensor(
dst.ReinterpretCast<float>(), param.negMask, param.tmpTensor1.ReinterpretCast<float>(),
dst.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(), SELMODE::VSEL_TENSOR_TENSOR_MODE, 1,
binaryParam, calCount);
Duplicate<int32_t, false>(param.tmpTensor2, 0, MASK_PLACEHOLDER, 1, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
VselPowerTensorTensor(
dst.ReinterpretCast<float>(), param.mask, param.tmpTensor2.ReinterpretCast<float>(),
dst.ReinterpretCast<float>(), param.tmpScalar.ReinterpretCast<float>(), SELMODE::VSEL_TENSOR_TENSOR_MODE, 1,
binaryParam, calCount);
}
__aicore__ inline void CommonPowerI(
const LocalTensor<int32_t>& dstTensor, const LocalTensor<int32_t>& srcTensor0,
const LocalTensor<int32_t>& srcTensor1, AscPowerIParams& param, const uint32_t calCount)
{
const UnaryRepeatParams unaryParam;
const BinaryRepeatParams binaryParam;
const uint8_t repeat = DivCeil(calCount * sizeof(int32_t), ONE_REPEAT_BYTE_SIZE);
PipeBarrier<PIPE_V>();
InitBulkPowerI(param, srcTensor0, srcTensor1, dstTensor, unaryParam, binaryParam, repeat, calCount);
PipeBarrier<PIPE_V>();
param.expIterateSum = 1;
do {
BulkProcessPowerI(dstTensor, param, unaryParam, binaryParam, repeat, calCount);
PipeBarrier<PIPE_V>();
} while (param.expIterateSum != 0);
InitFinePowerI(param, unaryParam, binaryParam, repeat, calCount);
PipeBarrier<PIPE_V>();
do {
FineProcessPowerI(dstTensor, srcTensor0, param, unaryParam, binaryParam, repeat, calCount);
PipeBarrier<PIPE_V>();
} while (param.expIterateSum != 0);
HandleNegativeExpPowerI(dstTensor, srcTensor0, param, unaryParam, binaryParam, repeat, calCount);
PipeBarrier<PIPE_V>();
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_POWER_POWER_INT_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_POWER_POWER_INT_IMPL_H__
#endif
