* 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 deepnorm.h
* \brief Use deepnorm for the normalization process. DeepNorm(x) = LayerNorm(alpha * x + Sublayer(x))
* LayerNorm(x) = gamma * (x - mean) / (variance ^ 2 + epsilon) + beta
* For more info of DeepNorm, please check https://arxiv.org/abs/2203.00555
*/
#ifndef LIB_NORMALIZATION_DEEPNORM_H
#define LIB_NORMALIZATION_DEEPNORM_H
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2002 || __NPU_ARCH__ == 2201)
#include "kernel_tensor.h"
#include "../../../impl/adv_api/detail/normalization/deepnorm/deepnorm_common_impl.h"
#elif defined(__NPU_ARCH__) && __NPU_ARCH__ == 3101
#include "../../../impl/adv_api/detail/normalization/deepnorm/deepnorm_c310_impl.h"
#endif
#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2201 || __NPU_ARCH__ == 2002 || __NPU_ARCH__ == 3101)
namespace AscendC {
#pragma begin_pipe(V)
* \ingroup DeepNorm
* \note User should make sure that memory on UB is big enough for input tiling.
* \tparam T: Data type to be calculated, half or float
* \tparam isReuseSrc: Whether to reuse the buffer of srcTensor.
* If the value is true, srcTensor can used as tmpBuffer and the data in it will be overwritten.
* If the value is false, srcTensor will not be used as tmpBuffer for calculation.
* \tparam isBasicBlock: Whether the srcTensor shape [B, S, H] meets the following requirement
* 1. H % 64 = 0 2. H <= 2040 3. B * S % 8 = 0
* If shape meets the rules above, isBasicBlock can be set true and the performance will be better.
* Otherwise, the performance remains the same.
* \param [out] dstLocal: output localTensor. Shape is [B, S, H]
* \param [out] meanLocal: output localTensor for mean result. Shape is [B, S]
* \param [out] rstdLocal: output localTensor for variance result.Shape is [B, S]
* \param [in] srcLocal: Input localTensor x for calculation. Shape is [B, S, H]
* \param [in] gxLocal: Input localTensor Sublayer(x) for calculation. Shape is [B, S, H]
* \param [in] betaLocal: Input localTensor beta used in layernorm. Shape is [H].
* \param [in] gammaLocal: Input localTensor gamma used in layernorm. Shape is [H].
* \param [in] sharedTmpBuffer: extra temporary shared space used for intermediate values among calculation process,
* whose required space size should refer to corresponding tiling API, which is defined at
* deepnorm_tiling.h. Generally, the more space you allocate, the better performance you will achieve,
* and the performance reaches peak when buffer size is maximum(calculated by tiling function). Moreover, it
* is not guaranteed that the shared space will be cleared after usage, the data could be anything.
* \param [in] alpha: alpha param in DeepNorm formula
* \param [in] epsilon: epsilon param in LayerNorm formula
* \param [in] tiling: Tiling information required for DeepNorm calculation.
*/
template <typename T, bool isReuseSrc = false, bool isBasicBlock = false>
__aicore__ inline void DeepNorm(const LocalTensor<T>& dstLocal, const LocalTensor<T>& meanLocal,
const LocalTensor<T>& rstdLocal, const LocalTensor<T>& srcLocal, const LocalTensor<T>& gxLocal,
const LocalTensor<T>& betaLocal, const LocalTensor<T>& gammaLocal, const LocalTensor<uint8_t>& sharedTmpBuffer,
const T alpha, const T epsilon, DeepNormTiling& tiling)
{
if ASCEND_IS_AIC {
return;
}
DeepNormAPI::DeepNormImpl<T, isReuseSrc, isBasicBlock>(dstLocal, meanLocal, rstdLocal, srcLocal, gxLocal, betaLocal,
gammaLocal, sharedTmpBuffer, alpha, epsilon, tiling);
}
* \ingroup DeepNorm
* \note User should make sure that memory on UB is big enough for input tiling.
* \tparam T: Data type to be calculated, half or float
* \tparam isReuseSrc: Whether to reuse the buffer of srcTensor.
* If the value is true, srcTensor can used as tmpBuffer and the data in it will be overwritten.
* If the value is false, srcTensor will not be used as tmpBuffer for calculation.
* \tparam isBasicBlock: Whether the srcTensor shape [B, S, H] meets the following requirement
* 1. H % 64 = 0 2. H <= 2040 3. B * S % 8 = 0
* If shape meets the rules above, isBasicBlock can be set true and the performance will be better.
* Otherwise, the performance remains the same.
* \param [out] dstLocal: output localTensor. Shape is [B, S, H]
* \param [out] meanLocal: output localTensor for mean result. Shape is [B, S]
* \param [out] rstdLocal: output localTensor for variance result.Shape is [B, S]
* \param [in] srcLocal: Input localTensor x for calculation. Shape is [B, S, H]
* \param [in] gxLocal: Input localTensor Sublayer(x) for calculation. Shape is [B, S, H]
* \param [in] betaLocal: Input localTensor beta used in layernorm. Shape is [H].
* \param [in] gammaLocal: Input localTensor gamma used in layernorm. Shape is [H].
* \param [in] alpha: alpha param in DeepNorm formula
* \param [in] epsilon: epsilon param in LayerNorm formula
* \param [in] tiling: Tiling information required for DeepNorm calculation.
*/
template <typename T, bool isReuseSrc = false, bool isBasicBlock = false>
__aicore__ inline void DeepNorm(const LocalTensor<T>& dstLocal, const LocalTensor<T>& meanLocal,
const LocalTensor<T>& rstdLocal, const LocalTensor<T>& srcLocal, const LocalTensor<T>& gxLocal,
const LocalTensor<T>& betaLocal, const LocalTensor<T>& gammaLocal, const T alpha, const T epsilon,
DeepNormTiling& tiling)
{
if ASCEND_IS_AIC {
return;
}
DeepNormAPI::DeepNormImpl<T, isReuseSrc, isBasicBlock>(dstLocal, meanLocal, rstdLocal, srcLocal, gxLocal, betaLocal,
gammaLocal, alpha, epsilon, tiling);
}
#pragma end_pipe
}
#endif
#endif
