* 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.
*/
#ifndef AICPU_KERNELS_ADD_H
#define AICPU_KERNELS_ADD_H
#define EIGEN_USE_THREADS
#define EIGEN_USE_SIMPLE_THREAD_POOL
#include "cpu_kernel.h"
#include "utils/bcast.h"
namespace aicpu {
class AddCpuKernel : public CpuKernel {
public:
AddCpuKernel() = default;
~AddCpuKernel() = default;
uint32_t Compute(CpuKernelContext &ctx) override;
private:
* @brief top-level dispatch for one dtype; chooses a fast-path
* (same-shape / scalar-bcast) or falls back to generic Eigen bcast.
*/
template <typename T>
uint32_t AddCompute(const CpuKernelContext &ctx) const;
* @brief Run a per-shard body either serially (small workloads) or via
* ParallelFor (large workloads). Centralizes the threshold check,
* per-unit sizing, dispatch and error logging shared by every
* fast-path variant. `elem_bytes` is sizeof(T) for the worker.
*/
template <typename Body>
uint32_t RunMaybeParallel(const CpuKernelContext &ctx, const char *tag,
int64_t total, int64_t elem_bytes,
const Body &body) const;
* @brief Validate raw shapes / declared output shape and populate
* broadcast info. On success `calc_info` contains a fully derived
* broadcast plan. Centralizes the rank/shape checks so AddCompute
* stays under the per-method line limit.
*/
uint32_t ValidateAndBroadcast(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const;
* @brief fast path: element-wise add with identical shapes.
* Parallelized when output-bytes exceed threshold.
*/
template <typename T>
uint32_t AddSameShape(const CpuKernelContext &ctx, const T *x0, const T *x1,
T *y, int64_t total) const;
* @brief fast path: one operand is a scalar (single element) broadcast
* over the other. Parallelized when output-bytes exceed threshold.
* @param scalar_side 0 -> x0 is scalar, 1 -> x1 is scalar
*/
template <typename T>
uint32_t AddScalarBcast(const CpuKernelContext &ctx, const T *vec,
T scalar_val, T *y, int64_t total) const;
* @brief generic broadcast path (keeps original Eigen tensor semantics).
*/
template <typename T>
uint32_t AddGenericBcast(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const;
* @brief check if input&output addr is aligned
*/
bool AlignedCheck(const BCalcInfo &calc_info) const;
template <typename T>
uint32_t AddCalculateWithRankCheck(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const;
template <int32_t RANK, typename T>
uint32_t AddCalculateWithAlignedCheck(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const;
* @brief Eigen calculate for all types (fallback generic broadcast).
*/
template <int32_t RANK, typename T, int32_t OPTION>
uint32_t AddCalculate(BCalcInfo &calc_info) const;
};
}
#endif