* 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 sin_3510_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/math/sin/sin_3510_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/sin.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_SIN_SIN_C310_IMPL_H__
#endif
#ifndef LIB_MATH_SIN_C310_IMPL_H
#define LIB_MATH_SIN_C310_IMPL_H
#include "kernel_basic_intf.h"
#include "kernel_tensor.h"
#if defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
#if __NPU_ARCH__ != 3003 && __NPU_ARCH__ != 3113
#include "../sincos/sincos_3510_impl.h"
#endif
#include "sin_common_utils.h"
#endif
namespace AscendC {
namespace Reg {
namespace Sin {
const uint8_t SIN_FLOAT_NOREUSE_CALC_PROCEDURE = 3;
const uint8_t SIN_FLOAT_REUSE_CALC_PROCEDURE = 2;
constexpr float SIN_PI_FOR_X_TODIV = 0.3183098733425140380859375;
constexpr float SIN_PI_V2 = 3.140625;
constexpr float SIN_KPI_FIRS_PI_MULS = 0.0009670257568359375;
constexpr float SIN_KPI_TWI_PI_MULS = 6.2771141529083251953125e-7;
constexpr float SIN_KPI_THIR_PI_MULS = 1.21644916362129151821136474609375e-10;
constexpr float SIN_RES_MULTI_SCA = 2.604926501e-6;
constexpr float SIN_RES_ADDICT_UP = -0.0001980894471;
constexpr float SIN_2ADDS = 0.008333049340;
constexpr float SIN_3ADDS = -0.1666665792;
constexpr float SIN_POINT_FIVE = 0.5;
constexpr float SIN_M4_SCA = 4.0;
constexpr float SIN_K2_SCA = -2.0;
constexpr float SIN_SCALAR_ONE = 1.0;
constexpr Reg::CastTrait sinCastTraitF16F32 = {
Reg::RegLayout::ZERO, Reg::SatMode::UNKNOWN, Reg::MaskMergeMode::ZEROING};
constexpr Reg::CastTrait sinCastTraitF32F16 = {
Reg::RegLayout::ZERO, Reg::SatMode::NO_SAT, Reg::MaskMergeMode::ZEROING, RoundMode::CAST_RINT};
__simd_callee__ inline void SinPolynomialApproximation(
Reg::RegTensor<float>& dstReg, Reg::RegTensor<float>& srcReg, Reg::RegTensor<float>& x,
Reg::RegTensor<float>& round, Reg::RegTensor<float>& kpi, Reg::MaskReg mask)
{
Reg::Muls(round, srcReg, SIN_PI_FOR_X_TODIV, mask);
Reg::Truncate<float, RoundMode::CAST_RINT, Reg::MaskMergeMode::ZEROING>(round, round, mask);
Reg::Muls(kpi, round, SIN_PI_V2, mask);
Reg::Sub(x, srcReg, kpi, mask);
Reg::Muls(kpi, round, SIN_KPI_FIRS_PI_MULS, mask);
Reg::Sub(x, x, kpi, mask);
Reg::Muls(kpi, round, SIN_KPI_TWI_PI_MULS, mask);
Reg::Sub(x, x, kpi, mask);
Reg::Muls(kpi, round, SIN_KPI_THIR_PI_MULS, mask);
Reg::Sub(x, x, kpi, mask);
sin(x) = (-1)^k*sin(x0)
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
*/
Reg::Mul(kpi, x, x, mask);
Reg::Muls(dstReg, round, SIN_POINT_FIVE, mask);
Reg::Truncate<float, RoundMode::CAST_FLOOR, Reg::MaskMergeMode::ZEROING>(dstReg, dstReg, mask);
Reg::Muls(dstReg, dstReg, SIN_M4_SCA, mask);
Reg::Muls(round, round, SIN_K2_SCA, mask);
Reg::Add(dstReg, dstReg, round, mask);
Reg::Adds(dstReg, dstReg, SIN_SCALAR_ONE, mask);
Reg::Muls(round, kpi, SIN_RES_MULTI_SCA, mask);
Reg::Adds(round, round, SIN_RES_ADDICT_UP, mask);
Reg::Mul(round, round, kpi, mask);
Reg::Adds(round, round, SIN_2ADDS, mask);
Reg::Mul(round, round, kpi, mask);
Reg::Adds(round, round, SIN_3ADDS, mask);
Reg::Mul(round, round, kpi, mask);
Reg::Adds(round, round, SIN_SCALAR_ONE, mask);
Reg::Mul(round, round, x, mask);
Reg::Mul(dstReg, round, dstReg, mask);
}
template <typename T>
__simd_vf__ inline void SinPolynomial(__ubuf__ T* dst, __ubuf__ T* src, uint32_t calCount, uint16_t repeat)
{
Reg::RegTensor<T> x;
Reg::RegTensor<float> xTmp;
Reg::RegTensor<float> round;
Reg::RegTensor<float> kpi;
Reg::RegTensor<T> srcReg;
Reg::RegTensor<float> srcTmp;
Reg::RegTensor<T> dstReg;
Reg::RegTensor<float> dstTmp;
Reg::MaskReg mask;
constexpr uint32_t oneRepSize = GetVecLen() / sizeof(float);
Reg::MaskReg maskAll = Reg::CreateMask<uint8_t>();
for (uint16_t i = 0; i < (uint16_t)repeat; i++) {
mask = Reg::UpdateMask<float>(calCount);
if constexpr (std::is_same<T, half>::value) {
Reg::LoadAlign<T, Reg::LoadDist::DIST_UNPACK_B16>(srcReg, src + i * oneRepSize);
Reg::Cast<float, half, sinCastTraitF16F32>(srcTmp, srcReg, mask);
SinPolynomialApproximation(dstTmp, srcTmp, xTmp, round, kpi, mask);
Reg::Cast<half, float, sinCastTraitF32F16>(dstReg, dstTmp, mask);
Reg::StoreAlign<T, Reg::StoreDist::DIST_PACK_B32>(dst + i * oneRepSize, dstReg, mask);
} else {
Reg::LoadAlign(srcReg, src + i * oneRepSize);
SinPolynomialApproximation(dstReg, srcReg, xTmp, round, kpi, mask);
Reg::StoreAlign(dst + i * oneRepSize, dstReg, mask);
}
}
}
template <typename T>
__aicore__ inline void SinPolynomialImpl(__ubuf__ T* dst, __ubuf__ T* src, uint32_t calCount)
{
constexpr uint32_t oneRepSize = GetVecLen() / sizeof(float);
uint16_t repeat = CeilDivision(calCount, oneRepSize);
SinPolynomial<T>(dst, src, calCount, repeat);
}
}
}
__aicore__ inline constexpr uint32_t GetSinTmpBufferLiveNode()
{
constexpr uint32_t tmpBufferLiveNode = sizeof(float) * 2;
return tmpBufferLiveNode;
}
template <typename T, bool isReuseSource = false, const SinConfig& config = defaultSinConfig>
__aicore__ inline void SinImpl(
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!"); });
SinImpl<T, isReuseSource, config>(dstTensor, srcTensor, sharedTmpBuffer, calCount);
}
template <typename T>
__aicore__ inline uint32_t GetSinTmpBufferSize(const LocalTensor<uint8_t>& sharedTmpBuffer)
{
uint32_t sharedTmpBufferSize = sharedTmpBuffer.GetSize() / GetSinTmpBufferLiveNode();
return AlignUp(sharedTmpBufferSize, GetDataBlockSizeInBytes()) / sizeof(T);
}
template <typename T, bool isReuseSource = false, const SinConfig& config = defaultSinConfig>
__aicore__ inline void SinImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const uint32_t calCount)
{
if ASCEND_IS_AIC {
return;
}
static_assert(SupportType<T, half, float>(), "current data type is not supported on current device!");
CheckTensorPos<T>(dstTensor, Hardware::UB, "dstTensor", "VECIN / VECCALC / VECOUT", "Sin");
CheckTensorPos<T>(srcTensor, Hardware::UB, "srcTensor", "VECIN / VECCALC / VECOUT", "Sin");
CheckTensorPos<uint8_t>(sharedTmpBuffer, Hardware::UB, "sharedTmpBuffer", "VECIN / VECCALC / VECOUT", "Sin");
ASCENDC_ASSERT((calCount <= srcTensor.GetSize()), {
KERNEL_LOG(
KERNEL_ERROR, "calCount is %u, which should not be larger than srcTensor length %u", calCount,
srcTensor.GetSize());
});
ASCENDC_ASSERT((calCount <= dstTensor.GetSize()), {
KERNEL_LOG(
KERNEL_ERROR, "calCount is %u, which should not be larger than dstTensor length %u", calCount,
dstTensor.GetSize());
});
if constexpr (config.algo == SinAlgo::POLYNOMIAL_APPROXIMATION) {
Reg::Sin::SinPolynomialImpl((__ubuf__ T*)dstTensor.GetPhyAddr(), (__ubuf__ T*)srcTensor.GetPhyAddr(), calCount);
} else if constexpr (config.algo == SinAlgo::RADIAN_REDUCTION) {
uint32_t sharedTmpBufferSize = GetSinTmpBufferSize<T>(sharedTmpBuffer);
uint32_t count = calCount;
uint16_t repeatTimes = static_cast<uint16_t>(CeilDivision(calCount, sharedTmpBufferSize));
for (uint16_t i = 0; i < repeatTimes; i++) {
uint32_t remainCount = count - sharedTmpBufferSize * i;
uint32_t oneRepSize = remainCount < sharedTmpBufferSize ? remainCount : sharedTmpBufferSize;
SinRadianReductionImpl(
(__ubuf__ T*)dstTensor.GetPhyAddr() + i * sharedTmpBufferSize,
(__ubuf__ T*)srcTensor.GetPhyAddr() + i * sharedTmpBufferSize,
(__ubuf__ uint32_t*)sharedTmpBuffer.GetPhyAddr(), oneRepSize);
}
}
}
__aicore__ inline void SinCastFullMask(
const LocalTensor<float>& dstTensor, const LocalTensor<float>& srcTensor, RoundMode castType)
{
uint64_t newMask = 64;
Cast<float, float, false>(
dstTensor, srcTensor, castType, newMask, 1, {1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_SIN_SIN_C310_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_MATH_SIN_SIN_C310_IMPL_H__
#endif