* 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 cos_common_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/cos/cos_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/cos.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_COS_COS_COMMON_IMPL_H__
#endif
#ifndef IMPL_MATH_COS_COS_COMMON_IMPL_H
#define IMPL_MATH_COS_COS_COMMON_IMPL_H
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
#include "kernel_pop_stack_buffer.h"
#include "../../common/check.h"
#ifdef ASCENDC_CPU_DEBUG
#include "../../api_check/kernel_check/math/cos/cos_check.h"
#endif
#include "../../api_check/kernel_api_check.h"
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102 || \
__NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
#include "cos_v220_impl.h"
#elif defined(__NPU_ARCH__) && __NPU_ARCH__ == 2002
#include "cos_v200_impl.h"
#endif
namespace AscendC {
const uint8_t COS_HALF_CALC_PROCEDURE = 4;
const uint8_t COS_FLOAT_NOREUSE_CALC_PROCEDURE = 3;
const uint8_t COS_FLOAT_REUSE_CALC_PROCEDURE = 2;
constexpr float COS_PI_FOR_X_TODIV = 0.3183098733425140380859375;
constexpr float PI_0 = 3.140625;
constexpr float COS_KPI_FIRS_PI_MULS = 0.0009670257568359375;
constexpr float COS_KPI_TWI_PI_MULS = 6.2771141529083251953125e-7;
constexpr float COS_KPI_THIR_PI_MULS = 1.21644916362129151821136474609375e-10;
constexpr float COS_KPI_FOR_PI_MULS = -1.0290623200529979163359041220560e-13;
constexpr float COS_PI_DOWN = 1.57079637050628662109375;
constexpr float COS_PI_RESDOWN_ADDS_NEG = -0.00000004371139000189375;
constexpr float COS_RES_MULTI_SCA = 2.604926501e-6;
constexpr float COS_RES_ADDICT_UP = -0.0001980894471;
constexpr float COS_2ADDS = 0.008333049340;
constexpr float COS_3ADDS = -0.1666665792;
constexpr float COS_POINT_FIVE = 0.5;
constexpr float COS_M4_SCA = 4.0;
constexpr float COS_K2_SCA = -2.0;
__aicore__ inline void KPI(
const LocalTensor<float>& inputX, const LocalTensor<float>& srcTensor, const LocalTensor<float>& roundTensor,
const LocalTensor<float>& kpi)
{
const UnaryRepeatParams unaryParams;
const BinaryRepeatParams binaryParams;
Muls<float, false>(kpi, roundTensor, PI_0, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Sub<float, false>(inputX, srcTensor, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(kpi, roundTensor, COS_KPI_FIRS_PI_MULS, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Sub<float, false>(inputX, inputX, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(inputX, inputX, COS_PI_DOWN, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(kpi, roundTensor, COS_KPI_TWI_PI_MULS, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Sub<float, false>(inputX, inputX, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(kpi, roundTensor, COS_KPI_THIR_PI_MULS, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Sub<float, false>(inputX, inputX, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(kpi, roundTensor, COS_KPI_FOR_PI_MULS, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Sub<float, false>(inputX, inputX, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(inputX, inputX, COS_PI_RESDOWN_ADDS_NEG, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void CosRound(
const LocalTensor<float>& inputX, const LocalTensor<float>& srcTensor, const LocalTensor<float>& roundTensor,
const LocalTensor<float>& kpi)
{
k=round(x/� + 1/2), x0=x-k�, x0 belongs to [-�, 0], (x0 + �/2) belongs to [-�/2, �/2]
�=�_0+�_1+�_2+�_3+�_4 achieve final precision compensation.
Final solution锛?
k = round(x * invpi + 1/2)
x -= k * pi_0
x -= k * pi_1
x = x + COS_PI_DOWN
x -= k * pi_2
x -= k * pi_3
x -= k * pi_4
x = x + COS_PI_RESDOWN_ADDS_NEG
*/
const UnaryRepeatParams unaryParams;
Muls<float, false>(roundTensor, srcTensor, COS_PI_FOR_X_TODIV, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(roundTensor, roundTensor, COS_POINT_FIVE, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
#if defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
CosCastFullMask(roundTensor, roundTensor, RoundMode::CAST_RINT);
#else
CosCast(roundTensor, roundTensor, RoundMode::CAST_RINT);
#endif
KPI(inputX, srcTensor, roundTensor, kpi);
}
__aicore__ inline void SignCompute(
const LocalTensor<float>& dstTensor, const LocalTensor<float>& inputX, const LocalTensor<float>& roundTensor,
const LocalTensor<float>& kpi)
{
const UnaryRepeatParams unaryParams;
const BinaryRepeatParams binaryParams;
Mul<float, false>(kpi, inputX, inputX, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(dstTensor, roundTensor, COS_POINT_FIVE, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
#if defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
CosCastFullMask(dstTensor, dstTensor, RoundMode::CAST_FLOOR);
#else
CosCast(dstTensor, dstTensor, RoundMode::CAST_FLOOR);
#endif
Muls<float, false>(dstTensor, dstTensor, COS_M4_SCA, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Muls<float, false>(roundTensor, roundTensor, COS_K2_SCA, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Add<float, false>(dstTensor, dstTensor, roundTensor, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(dstTensor, dstTensor, 1, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
}
__aicore__ inline void CosPolynomialApproximation(
const LocalTensor<float>& dstTensor, const LocalTensor<float>& inputX, const LocalTensor<float>& roundTensor,
const LocalTensor<float>& kpi)
{
cos(x) = (-1)^k*sin(x0 + �/2)
Finally, use sin(x) = xP(x) to calculate sin(x).
P(x) = (((x^2 * R0 + R1) * x^2 + R2) * x^2 + R3) * x^2 + 1.0
*/
const UnaryRepeatParams unaryParams;
const BinaryRepeatParams binaryParams;
SignCompute(dstTensor, inputX, roundTensor, kpi);
Muls<float, false>(roundTensor, kpi, COS_RES_MULTI_SCA, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(roundTensor, roundTensor, COS_RES_ADDICT_UP, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(roundTensor, roundTensor, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(roundTensor, roundTensor, COS_2ADDS, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(roundTensor, roundTensor, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(roundTensor, roundTensor, COS_3ADDS, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(roundTensor, roundTensor, kpi, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Adds<float, false>(roundTensor, roundTensor, 1.0, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(roundTensor, roundTensor, inputX, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Mul<float, false>(dstTensor, roundTensor, dstTensor, MASK_PLACEHOLDER, 1, binaryParams);
PipeBarrier<PIPE_V>();
Mins<float, false>(dstTensor, dstTensor, 1.0, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Maxs<float, false>(dstTensor, dstTensor, -1.0, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
}
template <typename T>
__aicore__ inline void CosCompute(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<float>& tmpBuffer,
const uint32_t calSize, bool isReuseSource)
{
const LocalTensor<T>& roundTensor = tmpBuffer;
const LocalTensor<T>& kpi = roundTensor[calSize];
LocalTensor<T> inputX = srcTensor;
if (!isReuseSource) {
inputX = roundTensor[calSize * 2];
}
CosRound(inputX, srcTensor, roundTensor, kpi);
CosPolynomialApproximation(dstTensor, inputX, roundTensor, kpi);
}
template <>
__aicore__ inline void CosCompute(
const LocalTensor<half>& dstTensor, const LocalTensor<half>& srcTensor, const LocalTensor<float>& tmpBuffer,
const uint32_t calSize, bool isReuseSource)
{
(void)isReuseSource;
const LocalTensor<float>& tempTensorConv = tmpBuffer;
const LocalTensor<float>& roundTensor = tempTensorConv[calSize];
const LocalTensor<float>& kpi = roundTensor[calSize];
const LocalTensor<float>& inputX = kpi[calSize];
Cast<float, half, false>(
tempTensorConv, srcTensor, RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
{1, 1, DEFAULT_REPEAT_STRIDE, HALF_DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
CosRound(inputX, tempTensorConv, roundTensor, kpi);
CosPolynomialApproximation(tempTensorConv, inputX, roundTensor, kpi);
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, bool isReuseSource = false>
__aicore__ inline void CosImpl(
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(Cos, (T, isReuseSource), (dstTensor, srcTensor, sharedTmpBuffer, calCount));
const uint32_t bufferSize = sharedTmpBuffer.GetSize();
const uint32_t tmpBufferSize = bufferSize / sizeof(float);
CheckTmpBufferSize(tmpBufferSize, 0, bufferSize);
LocalTensor<float> tmpBuffer = sharedTmpBuffer.ReinterpretCast<float>();
uint32_t calSize = 0;
if constexpr (sizeof(T) == sizeof(half)) {
calSize = tmpBufferSize / COS_HALF_CALC_PROCEDURE / ONE_BLK_SIZE * ONE_BLK_SIZE;
} else {
if constexpr (isReuseSource) {
calSize = tmpBufferSize / COS_FLOAT_REUSE_CALC_PROCEDURE / ONE_BLK_SIZE * ONE_BLK_SIZE;
} else {
calSize = tmpBufferSize / COS_FLOAT_NOREUSE_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<T, MaskMode::COUNTER>(0, calSize);
uint32_t offset = 0;
for (uint32_t i = 0; i < round; i++) {
CosCompute(dstTensor[offset], srcTensor[offset], tmpBuffer, calSize, isReuseSource);
offset = offset + calSize;
}
if (tail > 0) {
SetVectorMask<T, MaskMode::COUNTER>(0, tail);
CosCompute(dstTensor[offset], srcTensor[offset], tmpBuffer, calSize, isReuseSource);
}
SetMaskNorm();
ResetMask();
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void CosImpl(
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!"); });
CosImpl<T, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, calCount);
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_COS_COS_COMMON_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_COS_COS_COMMON_IMPL_H__
#endif