* 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/math/erfc/erfc_common_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/erfc.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_ERFC_ERFC_COMMON_IMPL_H__
#endif
#ifndef IMPL_MATH_ERFC_ERFC_COMMON_IMPL_H
#define IMPL_MATH_ERFC_ERFC_COMMON_IMPL_H
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
#include "../math_constant_util.h"
#include "../../common/check.h"
#ifdef ASCENDC_CPU_DEBUG
#include "../../api_check/kernel_check/math/erfc/erfc_check.h"
#endif
#include "../../api_check/kernel_api_check.h"
namespace AscendC {
constexpr float ERFC_BOUNDARY_MAX = 10;
constexpr uint8_t TMPBUF_IDX_3 = 2;
constexpr uint8_t TMPBUF_IDX_5 = 4;
constexpr uint8_t TMPBUF_IDX_6 = 5;
__aicore__ inline constexpr RoundMode GetErfcCastType()
{
#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 2201
return RoundMode::CAST_ROUND;
#else
return RoundMode::CAST_NONE;
#endif
}
__aicore__ inline void ErfcPreCompute(
const LocalTensor<float>& dstBuf1, const LocalTensor<float>& srcBuf1, const LocalTensor<float>& tmpCompBuf1,
uint32_t calSize)
{
BinaryRepeatParams binaryParams;
UnaryRepeatParams unaryParams;
Abs<float, false>(tmpCompBuf1, srcBuf1, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
constexpr float SCALAR_ERFC_FP32_MIN = 2.168404344971009e-19;
Adds<float, false>(
tmpCompBuf1, tmpCompBuf1, static_cast<float>(SCALAR_ERFC_FP32_MIN), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Div<float, false>(dstBuf1, srcBuf1, tmpCompBuf1, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void ErfcComputeR(const LocalTensor<float>& tmpCompBuf1, uint32_t calSize)
{
BinaryRepeatParams binaryParams;
UnaryRepeatParams unaryParams;
LocalTensor<float> tmpCompBuf3 = tmpCompBuf1[TMPBUF_IDX_3 * calSize];
LocalTensor<float> tmpCompBuf4 = tmpCompBuf3[calSize];
constexpr float R0 = 0.1735313680e-7;
constexpr float R1 = -0.9856738394e-6;
constexpr float R2 = 0.2517003236e-4;
constexpr float R3 = -0.3848015171e-3;
constexpr float R4 = 0.5681528564e0;
constexpr float R5 = 0.5245623129e1;
constexpr float R6 = 0.2107740710e2;
constexpr float R7 = 0.4212761755e2;
constexpr float R8 = 0.4380524149e2;
Muls<float, false>(tmpCompBuf3, tmpCompBuf1, static_cast<float>(R0), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R1), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R2), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R3), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R4), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R5), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R6), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R7), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf4, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf4, tmpCompBuf3, static_cast<float>(R8), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void ErfcComputeS(const LocalTensor<float>& tmpCompBuf1, uint32_t calSize)
{
BinaryRepeatParams binaryParams;
UnaryRepeatParams unaryParams;
LocalTensor<float> tmpCompBuf3 = tmpCompBuf1[TMPBUF_IDX_3 * calSize];
LocalTensor<float> tmpCompBuf5 = tmpCompBuf1[TMPBUF_IDX_5 * calSize];
constexpr float S1 = 0.9349684299e1;
constexpr float S2 = 0.3756930664e2;
constexpr float S3 = 0.8058268949e2;
constexpr float S4 = 0.9155653738e2;
constexpr float S5 = 0.4380524152e2;
Adds<float, false>(tmpCompBuf3, tmpCompBuf1, static_cast<float>(S1), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf5, tmpCompBuf1, tmpCompBuf3, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf3, tmpCompBuf5, static_cast<float>(S2), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf5, tmpCompBuf1, tmpCompBuf3, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf3, tmpCompBuf5, static_cast<float>(S3), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf5, tmpCompBuf1, tmpCompBuf3, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf3, tmpCompBuf5, static_cast<float>(S4), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf5, tmpCompBuf1, tmpCompBuf3, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf3, tmpCompBuf5, static_cast<float>(S5), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void ErfcPostCompute(
const LocalTensor<float>& dstBuf1, const LocalTensor<float>& srcBuf1, const LocalTensor<float>& tmpCompBuf1,
uint32_t calSize)
{
BinaryRepeatParams binaryParams;
UnaryRepeatParams unaryParams;
LocalTensor<float> tmpCompBuf2 = tmpCompBuf1[calSize];
LocalTensor<float> tmpCompBuf3 = tmpCompBuf2[calSize];
Muls<float, false>(tmpCompBuf2, srcBuf1, static_cast<float>(NEG_ONE), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpCompBuf2, tmpCompBuf2, static_cast<float>(NUM_ONE), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf3, srcBuf1, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Add<float, false>(dstBuf1, tmpCompBuf3, tmpCompBuf2, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void ErfcPublicSteps(const LocalTensor<float>& tmpCompBuf1, uint32_t calSize)
{
BinaryRepeatParams binaryParams;
UnaryRepeatParams unaryParams;
LocalTensor<float> tmpCompBuf2 = tmpCompBuf1[calSize];
LocalTensor<float> tmpCompBuf3 = tmpCompBuf2[calSize];
LocalTensor<float> tmpCompBuf4 = tmpCompBuf3[calSize];
LocalTensor<float> tmpCompBuf5 = tmpCompBuf4[calSize];
Mins<float, false>(
tmpCompBuf2, tmpCompBuf1, static_cast<float>(ERFC_BOUNDARY_MAX), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
ErfcComputeR(tmpCompBuf1, calSize);
ErfcComputeS(tmpCompBuf1, calSize);
Div<float, false>(tmpCompBuf3, tmpCompBuf4, tmpCompBuf3, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf1, tmpCompBuf1, tmpCompBuf1, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(tmpCompBuf1, tmpCompBuf1, static_cast<float>(NEG_ONE), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Exp<float, false>(tmpCompBuf1, tmpCompBuf1, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(tmpCompBuf3, tmpCompBuf1, tmpCompBuf3, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void ErfcClip(
const LocalTensor<float>& dst, const LocalTensor<float>& src, const LocalTensor<float>& tmpBuffer)
{
UnaryRepeatParams unaryParams;
Mins<float, false>(dst, src, static_cast<float>(ERFC_BOUNDARY_MAX), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Maxs<float, false>(tmpBuffer, dst, static_cast<float>(-ERFC_BOUNDARY_MAX), MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
}
template <typename T>
__aicore__ inline void ErfcCompute(
const LocalTensor<T>& dst, const LocalTensor<T>& src, const LocalTensor<T>& tmpBuffer, uint32_t calSize)
{
LocalTensor<T> tmpCompBuf1 = tmpBuffer;
ErfcClip(dst, src, dst);
ErfcPreCompute(dst, dst, tmpCompBuf1, calSize);
ErfcPublicSteps(tmpCompBuf1, calSize);
ErfcPostCompute(dst, dst, tmpCompBuf1, calSize);
}
template <>
__aicore__ inline void ErfcCompute<half>(
const LocalTensor<half>& dst, const LocalTensor<half>& src, const LocalTensor<half>& tmpBuffer, uint32_t calSize)
{
LocalTensor<float> tmpCompBuf1 = tmpBuffer.ReinterpretCast<float>();
LocalTensor<float> tmpCompBuf6 = tmpCompBuf1[TMPBUF_IDX_6 * calSize];
Cast<float, half, false>(
tmpCompBuf6, src, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
{1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE / 2});
PipeBarrier<PIPE_V>();
ErfcClip(tmpCompBuf1, tmpCompBuf6, tmpCompBuf6);
ErfcPreCompute(tmpCompBuf6, tmpCompBuf6, tmpCompBuf1, calSize);
ErfcPublicSteps(tmpCompBuf1, calSize);
ErfcPostCompute(tmpCompBuf1, tmpCompBuf6, tmpCompBuf1, calSize);
constexpr RoundMode castType = GetErfcCastType();
Cast<half, float, false>(
dst, tmpCompBuf1, castType, MASK_PLACEHOLDER, 1, {1, 1, DEFAULT_REPEAT_STRIDE / 2, DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void ErfcImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const uint32_t calCount)
{
if ASCEND_IS_AIC {
return;
}
CHECK_FUNC_HIGHLEVEL_API(Erfc, (T, isReuseSource), (dstTensor, srcTensor, sharedTmpBuffer, calCount));
LocalTensor<T> tmpBuffer = sharedTmpBuffer.ReinterpretCast<T>();
uint32_t bufferSize = sharedTmpBuffer.GetSize();
uint32_t tmpBufferSize = bufferSize / sizeof(T);
CheckTmpBufferSize(tmpBufferSize, 0, bufferSize);
constexpr uint8_t ERFC_HALF_CALC_PROCEDURE = 12;
constexpr uint8_t ERFC_FLOAT_CALC_PROCEDURE = 5;
uint32_t calSize = 0;
if constexpr (sizeof(T) == sizeof(half)) {
calSize = tmpBufferSize / ERFC_HALF_CALC_PROCEDURE / ONE_BLK_SIZE * ONE_BLK_SIZE;
} else {
calSize = tmpBufferSize / ERFC_FLOAT_CALC_PROCEDURE / ONE_BLK_SIZE * ONE_BLK_SIZE;
}
CheckTmpBufferSize(calSize, 0, bufferSize);
const uint32_t round = calCount / calSize;
const uint32_t tail = calCount % calSize;
SetMaskCount();
SetVectorMask<half, MaskMode::COUNTER>(0, calSize);
uint32_t offset = 0;
for (uint32_t i = 0; i < round; i++) {
ErfcCompute(dstTensor[offset], srcTensor[offset], tmpBuffer, calSize);
offset = offset + calSize;
}
if (tail != 0) {
SetVectorMask<half, MaskMode::COUNTER>(0, tail);
ErfcCompute(dstTensor[offset], srcTensor[offset], tmpBuffer, calSize);
}
SetMaskNorm();
ResetMask();
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void ErfcImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const uint32_t calCount)
{
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!"); });
ErfcImpl<T, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, calCount);
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_ERFC_ERFC_COMMON_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_ERFC_ERFC_COMMON_IMPL_H__
#endif