* 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_float_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/power/power_float_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_FLOAT_IMPL_H__
#endif
#ifndef IMPL_MATH_POWER_POWER_FLOAT_IMPL_H
#define IMPL_MATH_POWER_POWER_FLOAT_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 PowerF
* \brief Intrinsics of Power(Float Input).
*/
__aicore__ inline void InitTmpScalar(const LocalTensor<float>& tmpScalar)
{
NotNumUnion notNum;
notNum.i = F32_NAN;
SetVectorMask<float>(0, ONE_BLK_SIZE / sizeof(float));
Duplicate<float, false>(tmpScalar, 1.0f, MASK_PLACEHOLDER, 1, 1, DEFAULT_REPEAT_STRIDE);
Duplicate<float, false>(
tmpScalar[ONE_BLK_SIZE / sizeof(float)], notNum.f, MASK_PLACEHOLDER, 1, 1, DEFAULT_REPEAT_STRIDE);
}
__aicore__ inline void InitDst(
const LocalTensor<float>& dst, const LocalTensor<float>& src0, const LocalTensor<float>& src1,
const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam)
{
Abs<float, false>(dst, src0, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
Ln<float, false>(dst, dst, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
Mul<float, false>(dst, src1, dst, MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
Exp<float, false>(dst, dst, MASK_PLACEHOLDER, 1, unaryParam);
}
__aicore__ inline void DetermineSign(
const LocalTensor<float>& src1, const AscPowerFParams& param, const UnaryRepeatParams& unaryParam,
const BinaryRepeatParams& binaryParam)
{
Abs<float, false>(param.tmpTensor1, src1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CastFloat2Float(param.tmpTensor1, param.tmpTensor1, RoundMode::CAST_RINT, unaryParam);
PipeBarrier<PIPE_V>();
Muls<float, false>(param.tmpTensor2, param.tmpTensor1, 0.5f, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CastFloat2Float(param.tmpTensor2, param.tmpTensor2, RoundMode::CAST_FLOOR, unaryParam);
PipeBarrier<PIPE_V>();
Muls<float, false>(param.tmpTensor2, param.tmpTensor2, 2.0f, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
Sub<float, false>(param.tmpTensor1, param.tmpTensor1, param.tmpTensor2, MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
Muls<float, false>(param.tmpTensor1, param.tmpTensor1, -2.0f, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
Adds<float, false>(param.tmpTensor1, param.tmpTensor1, 1.0f, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CastFloat2Float(param.tmpTensor1, param.tmpTensor1, RoundMode::CAST_RINT, unaryParam);
}
__aicore__ inline void GenMaskForOne(
const LocalTensor<float>& src0, const LocalTensor<float>& src1, const AscPowerFParams& param,
const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam, const uint32_t calCount)
{
NotNumUnion notNum;
uint8_t repeat = DivCeil(calCount * sizeof(float), ONE_REPEAT_BYTE_SIZE);
Abs<float, false>(param.tmpTensor1, src1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CompareScalar<float, uint8_t, false>(
param.tmpMask1, param.tmpTensor1, static_cast<float>(0), CMPMODE::NE, MASK_PLACEHOLDER, repeat, unaryParam);
PipeBarrier<PIPE_V>();
CompareScalar<float, uint8_t, false>(
param.tmpMask2, src0, static_cast<float>(1), CMPMODE::NE, MASK_PLACEHOLDER, repeat, unaryParam);
PipeBarrier<PIPE_V>();
SetVectorMask<float>(0, ConstCeil(calCount, sizeof(uint16_t) * ONE_BYTE_BIT_SIZE));
And<uint16_t, false>(
param.tmpMask1.ReinterpretCast<uint16_t>(), param.tmpMask2.ReinterpretCast<uint16_t>(),
param.tmpMask1.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
SetVectorMask<float>(0, calCount);
PipeBarrier<PIPE_V>();
CompareScalar<float, uint8_t, false>(
param.tmpMask2, src0, static_cast<float>(-1), CMPMODE::NE, MASK_PLACEHOLDER, repeat, unaryParam);
PipeBarrier<PIPE_V>();
notNum.i = F32_INF;
CompareScalar<float, uint8_t, false>(
param.tmpMask3, param.tmpTensor1, notNum.f, CMPMODE::NE, MASK_PLACEHOLDER, repeat, unaryParam);
PipeBarrier<PIPE_V>();
SetVectorMask<float>(0, ConstCeil(calCount, sizeof(uint16_t) * ONE_BYTE_BIT_SIZE));
Or<uint16_t, false>(
param.tmpMask2.ReinterpretCast<uint16_t>(), param.tmpMask3.ReinterpretCast<uint16_t>(),
param.tmpMask2.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
And<uint16_t, false>(
param.tmpMask1.ReinterpretCast<uint16_t>(), param.tmpMask2.ReinterpretCast<uint16_t>(),
param.tmpMask1.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
SetVectorMask<float>(0, calCount);
}
__aicore__ inline void GenMaskForNan(
const LocalTensor<float>& src0, const LocalTensor<float>& src1, const AscPowerFParams& param,
const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam, const uint32_t calCount)
{
NotNumUnion notNum;
notNum.i = F32_INF;
uint8_t repeat = DivCeil(calCount * sizeof(float), ONE_REPEAT_BYTE_SIZE);
Abs<float, false>(param.tmpTensor1, src1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CompareScalar<float, uint8_t, false>(
param.tmpMask1, param.tmpTensor1, notNum.f, CMPMODE::EQ, MASK_PLACEHOLDER, repeat, unaryParam);
PipeBarrier<PIPE_V>();
SetVectorMask<float>(0, ConstCeil(calCount, sizeof(uint16_t) * ONE_BYTE_BIT_SIZE));
Or<uint16_t, false>(
param.tmpMask1.ReinterpretCast<uint16_t>(), param.finiteIntegerYMask.ReinterpretCast<uint16_t>(),
param.tmpMask1.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
SetVectorMask<float>(0, calCount);
CompareScalar<float, uint8_t, false>(
param.tmpMask2, src0, static_cast<float>(0), CMPMODE::GE, MASK_PLACEHOLDER, repeat, unaryParam);
notNum.i = F32_NEG_INF;
CompareScalar<float, uint8_t, false>(
param.tmpMask3, src0, notNum.f, CMPMODE::EQ, MASK_PLACEHOLDER, repeat, unaryParam);
PipeBarrier<PIPE_V>();
SetVectorMask<float>(0, ConstCeil(calCount, sizeof(uint16_t) * ONE_BYTE_BIT_SIZE));
Or<uint16_t, false>(
param.tmpMask2.ReinterpretCast<uint16_t>(), param.tmpMask3.ReinterpretCast<uint16_t>(),
param.tmpMask2.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
Or<uint16_t, false>(
param.tmpMask1.ReinterpretCast<uint16_t>(), param.tmpMask2.ReinterpretCast<uint16_t>(),
param.tmpMask1.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
SetVectorMask<float>(0, calCount);
}
__aicore__ inline void GenMaskForSign(
const LocalTensor<float>& src0, const LocalTensor<float>& src1, const AscPowerFParams& param,
const UnaryRepeatParams& unaryParam, const BinaryRepeatParams& binaryParam, const uint32_t calCount)
{
constexpr float intThreshold = 0.00000001f;
const uint8_t repeat = DivCeil(calCount * sizeof(float), ONE_REPEAT_BYTE_SIZE);
GrepSignBit(param.tmpMask1, src0, param.tmpTensor2, param.tmpTensor4, unaryParam, binaryParam, calCount);
PipeBarrier<PIPE_V>();
Abs<float, false>(param.tmpTensor2, src1, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CastFloat2Float(param.tmpTensor3, param.tmpTensor2, RoundMode::CAST_RINT, unaryParam);
PipeBarrier<PIPE_V>();
Sub<float, false>(param.tmpTensor3, param.tmpTensor2, param.tmpTensor3, MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
Abs<float, false>(param.tmpTensor3, param.tmpTensor3, MASK_PLACEHOLDER, 1, unaryParam);
PipeBarrier<PIPE_V>();
CompareScalar<float, uint8_t, false>(
param.finiteIntegerYMask, param.tmpTensor3, static_cast<float>(intThreshold), CMPMODE::LT, MASK_PLACEHOLDER,
repeat, unaryParam);
PipeBarrier<PIPE_V>();
SetVectorMask<float>(0, ConstCeil(calCount, sizeof(uint16_t) * ONE_BYTE_BIT_SIZE));
And<uint16_t, false>(
param.tmpMask1.ReinterpretCast<uint16_t>(), param.finiteIntegerYMask.ReinterpretCast<uint16_t>(),
param.tmpMask1.ReinterpretCast<uint16_t>(), MASK_PLACEHOLDER, 1, binaryParam);
SetVectorMask<float>(0, calCount);
}
__aicore__ inline void CommonPowerF(
const LocalTensor<float>& dstTensor, const LocalTensor<float>& srcTensor0, const LocalTensor<float>& srcTensor1,
const LocalTensor<float>& tmpScalar, const AscPowerFParams& powerParam, const uint32_t calCount)
{
const UnaryRepeatParams unaryParam;
const BinaryRepeatParams binaryParam;
PipeBarrier<PIPE_V>();
InitDst(dstTensor, srcTensor0, srcTensor1, unaryParam, binaryParam);
PipeBarrier<PIPE_V>();
DetermineSign(srcTensor1, powerParam, unaryParam, binaryParam);
PipeBarrier<PIPE_V>();
GenMaskForSign(srcTensor0, srcTensor1, powerParam, unaryParam, binaryParam, calCount);
PipeBarrier<PIPE_V>();
VselPowerTensorScalar(
powerParam.tmpTensor2, powerParam.tmpMask1, powerParam.tmpTensor1, tmpScalar, SELMODE::VSEL_TENSOR_SCALAR_MODE,
1, binaryParam, calCount);
PipeBarrier<PIPE_V>();
Mul<float, false>(dstTensor, powerParam.tmpTensor2, dstTensor, MASK_PLACEHOLDER, 1, binaryParam);
PipeBarrier<PIPE_V>();
GenMaskForOne(srcTensor0, srcTensor1, powerParam, unaryParam, binaryParam, calCount);
PipeBarrier<PIPE_V>();
VselPowerTensorScalar(
dstTensor, powerParam.tmpMask1, dstTensor, tmpScalar, SELMODE::VSEL_TENSOR_SCALAR_MODE, 1, binaryParam,
calCount);
PipeBarrier<PIPE_V>();
GenMaskForNan(srcTensor0, srcTensor1, powerParam, unaryParam, binaryParam, calCount);
PipeBarrier<PIPE_V>();
VselPowerTensorScalar(
dstTensor, powerParam.tmpMask1, dstTensor, tmpScalar[ONE_BLK_SIZE / sizeof(float)],
SELMODE::VSEL_TENSOR_SCALAR_MODE, 1, binaryParam, calCount);
PipeBarrier<PIPE_V>();
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_POWER_POWER_FLOAT_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_POWER_POWER_FLOAT_IMPL_H__
#endif