* 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.
*/
#include "add_aicpu.h"
#include <algorithm>
#include <complex>
#include "utils/eigen_tensor.h"
#include "utils/kernel_util.h"
#include "cpu_kernel_utils.h"
namespace {
const char *const kAdd = "Add";
constexpr int64_t kParallelBytesThresh = 192LL * 1024LL;
constexpr int64_t kBytesPerShard = 256LL * 1024LL;
}
namespace aicpu {
uint32_t CheckPermissionType(const CpuKernelContext &ctx) {
const DataType input0_data_type = ctx.Input(kFirstInputIndex)->GetDataType();
const DataType input1_data_type = ctx.Input(kSecondInputIndex)->GetDataType();
const DataType output_data_type = ctx.Output(kFirstOutputIndex)->GetDataType();
if (input0_data_type != output_data_type) {
KERNEL_LOG_ERROR(
"[Add] Output must have the same dtype as input, but got "
"input0[%s] input1[%s] output[%s].",
DTypeStr(input0_data_type).c_str(),
DTypeStr(input1_data_type).c_str(),
DTypeStr(output_data_type).c_str());
return KERNEL_STATUS_PARAM_INVALID;
}
return KERNEL_STATUS_OK;
}
uint32_t AddCpuKernel::Compute(CpuKernelContext &ctx) {
KERNEL_HANDLE_ERROR(NormalMathCheck(ctx), "Normal match check params failed.");
KERNEL_HANDLE_ERROR(CheckPermissionType(ctx), "Type permission check failed.");
Tensor *input0 = ctx.Input(kFirstInputIndex);
Tensor *input1 = ctx.Input(kSecondInputIndex);
if ((input0->GetDataSize() == 0) || (input1->GetDataSize() == 0)) {
KERNEL_LOG_INFO("[%s] Input is empty tensor.", ctx.GetOpType().c_str());
return KERNEL_STATUS_OK;
}
const DataType input0_data_type = input0->GetDataType();
KERNEL_LOG_INFO("[%s] Compute begin, dtype=%d, in0_bytes=%lu, in1_bytes=%lu, out_bytes=%lu.",
ctx.GetOpType().c_str(), static_cast<int>(input0_data_type),
input0->GetDataSize(), input1->GetDataSize(),
ctx.Output(kFirstOutputIndex)->GetDataSize());
switch (input0_data_type) {
case DT_FLOAT16:
return AddCompute<Eigen::half>(ctx);
case DT_FLOAT:
return AddCompute<float>(ctx);
case DT_DOUBLE:
return AddCompute<double>(ctx);
case DT_INT8:
return AddCompute<int8_t>(ctx);
case DT_INT16:
return AddCompute<int16_t>(ctx);
case DT_INT32:
return AddCompute<int32_t>(ctx);
case DT_INT64:
return AddCompute<int64_t>(ctx);
case DT_UINT8:
return AddCompute<uint8_t>(ctx);
case DT_UINT16:
return AddCompute<uint16_t>(ctx);
case DT_UINT32:
return AddCompute<uint32_t>(ctx);
case DT_UINT64:
return AddCompute<uint64_t>(ctx);
case DT_COMPLEX64:
return AddCompute<std::complex<float>>(ctx);
case DT_COMPLEX128:
return AddCompute<std::complex<double>>(ctx);
default:
KERNEL_LOG_ERROR(
"[%s] Data type of input is not supported, got dtype=[%s]. "
"Expect one of {FP16,FP32,FP64,INT8..INT64,UINT8..UINT64,CPX64,CPX128}.",
ctx.GetOpType().c_str(), DTypeStr(input0_data_type).c_str());
return KERNEL_STATUS_PARAM_INVALID;
}
}
template <typename Body>
uint32_t AddCpuKernel::RunMaybeParallel(const CpuKernelContext &ctx,
const char *tag, int64_t total,
int64_t elem_bytes,
const Body &body) const {
const int64_t total_bytes = total * elem_bytes;
if (total_bytes < kParallelBytesThresh) {
KERNEL_LOG_INFO("[%s] %s serial path, total_bytes=%ld.",
ctx.GetOpType().c_str(), tag, total_bytes);
body(0, total);
return KERNEL_STATUS_OK;
}
if (__builtin_expect(elem_bytes <= 0, 0)) {
KERNEL_LOG_ERROR("[%s] %s invalid elem_bytes=%ld, fallback to serial.",
ctx.GetOpType().c_str(), tag, elem_bytes);
body(0, total);
return KERNEL_STATUS_OK;
}
const int64_t per_unit = std::max<int64_t>(1, kBytesPerShard / elem_bytes);
KERNEL_LOG_INFO("[%s] %s parallel path, total=%ld, per_unit=%ld.",
ctx.GetOpType().c_str(), tag, total, per_unit);
const uint32_t rc = CpuKernelUtils::ParallelFor(ctx, total, per_unit, body);
if (__builtin_expect(rc != KERNEL_STATUS_OK, 0)) {
KERNEL_LOG_ERROR("[%s] %s ParallelFor failed, rc=%u, total=%ld, per_unit=%ld.",
ctx.GetOpType().c_str(), tag, rc, total, per_unit);
return rc;
}
return KERNEL_STATUS_OK;
}
template <typename T>
uint32_t AddCpuKernel::AddSameShape(const CpuKernelContext &ctx, const T *x0,
const T *x1, T *y, int64_t total) const {
auto body = [x0, x1, y](int64_t beg, int64_t end) {
const T *__restrict__ a = x0 + beg;
const T *__restrict__ b = x1 + beg;
T *__restrict__ o = y + beg;
const int64_t n = end - beg;
for (int64_t i = 0; i < n; ++i) {
o[i] = a[i] + b[i];
}
};
return RunMaybeParallel(ctx, "same-shape", total,
static_cast<int64_t>(sizeof(T)), body);
}
template <typename T>
uint32_t AddCpuKernel::AddScalarBcast(const CpuKernelContext &ctx, const T *vec,
T scalar_val, T *y,
int64_t total) const {
auto body = [vec, scalar_val, y](int64_t beg, int64_t end) {
const T *__restrict__ a = vec + beg;
T *__restrict__ o = y + beg;
const T s = scalar_val;
const int64_t n = end - beg;
for (int64_t i = 0; i < n; ++i) {
o[i] = a[i] + s;
}
};
return RunMaybeParallel(ctx, "scalar-bcast", total,
static_cast<int64_t>(sizeof(T)), body);
}
template <typename T>
uint32_t AddCpuKernel::AddGenericBcast(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const {
(void)ctx;
return AddCalculateWithRankCheck<T>(ctx, calc_info);
}
uint32_t AddCpuKernel::ValidateAndBroadcast(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const {
const auto &raw0 = calc_info.input_0->GetTensorShape()->GetDimSizes();
const auto &raw1 = calc_info.input_1->GetTensorShape()->GetDimSizes();
const auto &raw_out = calc_info.output->GetTensorShape()->GetDimSizes();
if (raw0.size() > static_cast<size_t>(kRank8) ||
raw1.size() > static_cast<size_t>(kRank8) ||
raw_out.size() > static_cast<size_t>(kRank8)) {
KERNEL_LOG_ERROR(
"[%s] Rank of input/output must be <= 8, got in0_rank=%zu, "
"in1_rank=%zu, out_rank=%zu.",
ctx.GetOpType().c_str(), raw0.size(), raw1.size(), raw_out.size());
return KERNEL_STATUS_PARAM_INVALID;
}
Bcast bcast;
if (bcast.GenerateBcastInfo(calc_info) != KERNEL_STATUS_OK) {
KERNEL_LOG_ERROR("[%s] Generate broadcast info failed.",
ctx.GetOpType().c_str());
return KERNEL_STATUS_PARAM_INVALID;
}
bcast.GetBcastVec(calc_info);
if (raw_out.size() != calc_info.shape_out.size()) {
KERNEL_LOG_ERROR(
"[%s] Output rank [%zu] does not match broadcast rank [%zu].",
ctx.GetOpType().c_str(), raw_out.size(), calc_info.shape_out.size());
return KERNEL_STATUS_PARAM_INVALID;
}
for (size_t i = 0; i < raw_out.size(); ++i) {
if (raw_out[i] != calc_info.shape_out[i]) {
KERNEL_LOG_ERROR(
"[%s] Output dim[%zu]=%ld mismatches broadcast dim=%ld.",
ctx.GetOpType().c_str(), i, raw_out[i], calc_info.shape_out[i]);
return KERNEL_STATUS_PARAM_INVALID;
}
}
return KERNEL_STATUS_OK;
}
template <typename T>
uint32_t AddCpuKernel::AddCompute(const CpuKernelContext &ctx) const {
BCalcInfo calc_info;
calc_info.input_0 = ctx.Input(kFirstInputIndex);
calc_info.input_1 = ctx.Input(kSecondInputIndex);
calc_info.output = ctx.Output(kFirstOutputIndex);
KERNEL_CHECK_NULLPTR(calc_info.input_0->GetData(),
KERNEL_STATUS_PARAM_INVALID, "[%s] Get input[0] data failed",
ctx.GetOpType().c_str())
KERNEL_CHECK_NULLPTR(calc_info.input_1->GetData(),
KERNEL_STATUS_PARAM_INVALID, "[%s] Get input[1] data failed",
ctx.GetOpType().c_str())
KERNEL_CHECK_NULLPTR(calc_info.output->GetData(), KERNEL_STATUS_PARAM_INVALID,
"[%s] Get output data failed", ctx.GetOpType().c_str())
T *const x0_ptr = PtrToPtr<void, T>(calc_info.input_0->GetData());
T *const x1_ptr = PtrToPtr<void, T>(calc_info.input_1->GetData());
T *const y_ptr = PtrToPtr<void, T>(calc_info.output->GetData());
const int64_t n0 = calc_info.input_0->NumElements();
const int64_t n1 = calc_info.input_1->NumElements();
const int64_t ny = calc_info.output->NumElements();
KERNEL_LOG_INFO("[%s] Input[0] size=%lu, Input[1] size=%lu, Output size=%lu.",
ctx.GetOpType().c_str(), calc_info.input_0->GetDataSize(),
calc_info.input_1->GetDataSize(),
calc_info.output->GetDataSize());
const uint32_t vrc = ValidateAndBroadcast(ctx, calc_info);
if (vrc != KERNEL_STATUS_OK) {
return vrc;
}
const auto &raw0 = calc_info.input_0->GetTensorShape()->GetDimSizes();
const auto &raw1 = calc_info.input_1->GetTensorShape()->GetDimSizes();
const auto &raw_out = calc_info.output->GetTensorShape()->GetDimSizes();
if (raw0 == raw1 && raw0 == raw_out && n0 == n1 && n0 == ny) {
KERNEL_LOG_INFO("[%s] same-shape branch selected, elems=%ld.",
ctx.GetOpType().c_str(), ny);
return AddSameShape<T>(ctx, x0_ptr, x1_ptr, y_ptr, ny);
}
if (n0 == 1 && n1 == ny && raw1 == raw_out) {
KERNEL_LOG_INFO("[%s] x0-scalar bcast branch selected, elems=%ld.",
ctx.GetOpType().c_str(), ny);
return AddScalarBcast<T>(ctx, x1_ptr, *x0_ptr, y_ptr, ny);
}
if (n1 == 1 && n0 == ny && raw0 == raw_out) {
KERNEL_LOG_INFO("[%s] x1-scalar bcast branch selected, elems=%ld.",
ctx.GetOpType().c_str(), ny);
return AddScalarBcast<T>(ctx, x0_ptr, *x1_ptr, y_ptr, ny);
}
return AddGenericBcast<T>(ctx, calc_info);
}
template <typename T>
uint32_t AddCpuKernel::AddCalculateWithRankCheck(const CpuKernelContext &ctx,
BCalcInfo &calc_info) const {
const int32_t rank = static_cast<int32_t>(calc_info.shape_out.size());
switch (rank) {
case 0: {
const T v0 = *PtrToPtr<void, const T>(calc_info.input_0->GetData());
const T v1 = *PtrToPtr<void, const T>(calc_info.input_1->GetData());
T *value_out = PtrToPtr<void, T>(calc_info.output->GetData());
*value_out = v0 + v1;
return KERNEL_STATUS_OK;
}
case kRank1:
return AddCalculateWithAlignedCheck<kRank1, T>(ctx, calc_info);
case kRank2:
return AddCalculateWithAlignedCheck<kRank2, T>(ctx, calc_info);
case kRank3:
return AddCalculateWithAlignedCheck<kRank3, T>(ctx, calc_info);
case kRank4:
return AddCalculateWithAlignedCheck<kRank4, T>(ctx, calc_info);
case kRank5:
return AddCalculateWithAlignedCheck<kRank5, T>(ctx, calc_info);
case kRank6:
return AddCalculateWithAlignedCheck<kRank6, T>(ctx, calc_info);
case kRank7:
return AddCalculateWithAlignedCheck<kRank7, T>(ctx, calc_info);
case kRank8:
return AddCalculateWithAlignedCheck<kRank8, T>(ctx, calc_info);
default:
KERNEL_LOG_ERROR(
"[%s] Rank of output must be in [0,8], got rank=%zu.",
ctx.GetOpType().c_str(), calc_info.shape_out.size());
return KERNEL_STATUS_PARAM_INVALID;
}
}
template <int32_t RANK, typename T>
uint32_t AddCpuKernel::AddCalculateWithAlignedCheck(
const CpuKernelContext &ctx, BCalcInfo &calc_info) const {
(void)ctx;
if (AlignedCheck(calc_info)) {
return AddCalculate<RANK, T, Eigen::Aligned>(calc_info);
}
return AddCalculate<RANK, T, Eigen::Unaligned>(calc_info);
}
bool AddCpuKernel::AlignedCheck(const BCalcInfo &calc_info) const {
return AddrAlignedCheck(calc_info.input_0->GetData()) &&
AddrAlignedCheck(calc_info.input_1->GetData()) &&
AddrAlignedCheck(calc_info.output->GetData());
}
template <int32_t RANK, typename T, int32_t OPTION>
uint32_t AddCpuKernel::AddCalculate(BCalcInfo &calc_info) const {
Eigen::TensorMap<Eigen::Tensor<T, 1>, OPTION> input0(
PtrToPtr<void, T>(calc_info.input_0->GetData()),
calc_info.input_0->GetTensorShape()->NumElements());
Eigen::TensorMap<Eigen::Tensor<T, 1>, OPTION> input1(
PtrToPtr<void, T>(calc_info.input_1->GetData()),
calc_info.input_1->GetTensorShape()->NumElements());
Eigen::TensorMap<Eigen::Tensor<T, 1>, OPTION> output(
PtrToPtr<void, T>(calc_info.output->GetData()),
calc_info.output->GetTensorShape()->NumElements());
const auto &input_shape_0 = calc_info.input_0->GetTensorShape()->GetDimSizes();
const auto &input_shape_1 = calc_info.input_1->GetTensorShape()->GetDimSizes();
if (input_shape_0.empty()) {
const T v0 = *PtrToPtr<void, const T>(calc_info.input_0->GetData());
output = v0 + input1;
return KERNEL_STATUS_OK;
}
if (input_shape_1.empty()) {
const T v1 = *PtrToPtr<void, const T>(calc_info.input_1->GetData());
output = input0 + v1;
return KERNEL_STATUS_OK;
}
Eigen::DSizes<Eigen::DenseIndex, RANK> reshape0;
Eigen::DSizes<Eigen::DenseIndex, RANK> reshape1;
Eigen::DSizes<Eigen::DenseIndex, RANK> shape_out;
Eigen::array<Eigen::DenseIndex, RANK> bcast0;
Eigen::array<Eigen::DenseIndex, RANK> bcast1;
for (int32_t i = 0; i < RANK; i++) {
reshape0[(RANK - i) - 1] = calc_info.reshape_0[i];
reshape1[(RANK - i) - 1] = calc_info.reshape_1[i];
shape_out[(RANK - i) - 1] = calc_info.shape_out[i];
bcast0[(RANK - i) - 1] = calc_info.bcast_0[i];
bcast1[(RANK - i) - 1] = calc_info.bcast_1[i];
}
output.reshape(shape_out) = input0.reshape(reshape0).broadcast(bcast0) +
input1.reshape(reshape1).broadcast(bcast1);
return KERNEL_STATUS_OK;
}
REGISTER_CPU_KERNEL(kAdd, AddCpuKernel);
}