* 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 tanh_common_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/tanh/tanh_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/tanh.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_TANH_TANH_COMMON_IMPL_H__
#endif
#ifndef IMPL_MATH_TANH_TANH_COMMON_IMPL_H
#define IMPL_MATH_TANH_TANH_COMMON_IMPL_H
#include "kernel_basic_intf.h"
#include "kernel_tensor.h"
#include "kernel_pop_stack_buffer.h"
#include "../../common/check.h"
#include "../math_common_impl.h"
#ifdef ASCENDC_CPU_DEBUG
#include "../../api_check/kernel_check/math/tanh/tanh_check.h"
#endif
#include "../../api_check/kernel_api_check.h"
namespace AscendC {
constexpr float FP32_MIN_V2 = -8.8;
constexpr float FP32_MAX_V2 = 8.8;
constexpr float DOUBLE_X = 2;
const uint8_t TANH_HALF_CALC_PROCEDURE = 2;
const uint8_t TANH_FLOAT_CALC_PROCEDURE = 1;
* Formula is y= (e^(2x)-1)/(e^(2x)+1)
*/
__aicore__ inline void TanhFormulaImpl(
const LocalTensor<float>& dstTensor, const LocalTensor<float>& srcTensor, const TanhParams<float>& params)
{
const LocalTensor<float>& tmpClip = params.tmpClip;
const UnaryRepeatParams unaryParams;
const BinaryRepeatParams binaryParams;
Mins<float, false>(tmpClip, srcTensor, FP32_MAX_V2, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Maxs<float, false>(tmpClip, tmpClip, FP32_MIN_V2, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(tmpClip, tmpClip, DOUBLE_X, MASK_PLACEHOLDER, params.repeatTimes, unaryParams);
PipeBarrier<PIPE_V>();
Exp<float, false>(tmpClip, tmpClip, MASK_PLACEHOLDER, params.repeatTimes, unaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(dstTensor, tmpClip, -1.0, MASK_PLACEHOLDER, params.repeatTimes, unaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(tmpClip, tmpClip, 1.0, MASK_PLACEHOLDER, params.repeatTimes, unaryParams);
PipeBarrier<PIPE_V>();
Div<float, false>(dstTensor, dstTensor, tmpClip, MASK_PLACEHOLDER, params.repeatTimes, binaryParams);
PipeBarrier<PIPE_V>();
}
template <typename T>
__aicore__ inline void TanhCompute(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const TanhParams<float>& params)
{
TanhFormulaImpl(dstTensor, srcTensor, params);
}
template <>
__aicore__ inline void TanhCompute(
const LocalTensor<half>& dstTensor, const LocalTensor<half>& srcTensor, const TanhParams<float>& params)
{
const LocalTensor<float>& tempTensorConv = params.tempTensorConv;
Cast<float, half, false>(
tempTensorConv, srcTensor, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
{1, 1, DEFAULT_REPEAT_STRIDE, HALF_DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
TanhFormulaImpl(tempTensorConv, tempTensorConv, params);
Cast<half, float, false>(
dstTensor, tempTensorConv, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
{1, 1, HALF_DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
}
template <typename T>
__aicore__ inline void TanhFormulasTmpCalc(TanhParams<float>& params, uint32_t tmpBufferSize)
{
uint32_t tmpUbIndex = 0;
if constexpr (sizeof(T) == sizeof(half)) {
params.stackSize = params.tmpBufferSize / TANH_HALF_CALC_PROCEDURE / ONE_BLK_SIZE * ONE_BLK_SIZE;
} else {
params.stackSize = params.tmpBufferSize / TANH_FLOAT_CALC_PROCEDURE / ONE_BLK_SIZE * ONE_BLK_SIZE;
}
CheckTmpBufferSize(params.stackSize, 0, tmpBufferSize);
if constexpr (sizeof(T) == sizeof(half)) {
params.tempTensorConv = params.sharedTmpBuffer[params.stackSize * (tmpUbIndex++)];
}
params.tmpClip = params.sharedTmpBuffer[params.stackSize * (tmpUbIndex++)];
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void TanhImpl(
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(Tanh, (T, isReuseSource), (dstTensor, srcTensor, sharedTmpBuffer, calCount));
TanhParams<float> params;
params.calCount = calCount;
uint32_t tmpBufferSize = sharedTmpBuffer.GetSize();
params.tmpBufferSize = tmpBufferSize / sizeof(float);
CheckTmpBufferSize(params.tmpBufferSize, 0, tmpBufferSize);
params.sharedTmpBuffer = sharedTmpBuffer.ReinterpretCast<float>();
TanhFormulasTmpCalc<T>(params, tmpBufferSize);
const uint32_t round = params.calCount / params.stackSize;
const uint32_t tail = params.calCount % params.stackSize;
SetMaskCount();
SetVectorMask<T, MaskMode::COUNTER>(0, params.stackSize);
uint32_t offset = 0;
for (uint32_t i = 0; i < round; i++) {
TanhCompute(dstTensor[offset], srcTensor[offset], params);
offset = offset + params.stackSize;
}
if (tail != 0) {
SetVectorMask<T, MaskMode::COUNTER>(0, tail);
TanhCompute(dstTensor[offset], srcTensor[offset], params);
}
SetMaskNorm();
ResetMask();
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void TanhImpl(
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!"); });
TanhImpl<T, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, calCount);
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_TANH_TANH_COMMON_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_TANH_TANH_COMMON_IMPL_H__
#endif