* 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 "squared_difference_aicpu.h"
#include "cpu_kernel_utils.h"
#include "utils/eigen_tensor.h"
#include "utils/kernel_util.h"
namespace {
const uint32_t kOutputNum = 1;
const uint32_t kInputNum = 2;
const char* const kSquaredDifference = "SquaredDifference";
const int64_t kParallelDataNum = 2 * 1024;
const int64_t kParallelDataNumMid = 32 * 1024;
const int64_t kParallelDataNumSameShape = 7 * 1024;
const int64_t kParallelDataNumSameShapeMid = 35 * 1024;
#define SQUAREDDIFFERENCE_COMPUTE_CASE(DTYPE, TYPE, CTX) \
case (DTYPE): { \
uint32_t result = SquaredDifferenceCompute<TYPE>(CTX); \
if (result != KERNEL_STATUS_OK) { \
KERNEL_LOG_ERROR("SquaredDifference kernel compute failed."); \
return result; \
} \
break; \
}
}
namespace aicpu {
uint32_t SquaredDifferenceCpuKernel::Compute(CpuKernelContext& ctx)
{
KERNEL_HANDLE_ERROR(
NormalCheck(ctx, kInputNum, kOutputNum), "SquaredDifference check input and output number failed.");
KERNEL_HANDLE_ERROR(SquaredDifferenceCheck(ctx), "SquaredDifference check params failed.");
DataType data_type = ctx.Input(0)->GetDataType();
switch (data_type) {
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_FLOAT16, Eigen::half, ctx)
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_FLOAT, float, ctx)
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_DOUBLE, double, ctx)
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_INT32, int32_t, ctx)
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_INT64, int64_t, ctx)
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_COMPLEX64, std::complex<float>, ctx)
SQUAREDDIFFERENCE_COMPUTE_CASE(DT_COMPLEX128, std::complex<double>, ctx)
default:
KERNEL_LOG_ERROR("SquaredDifference kernel data type [%s] not support.", DTypeStr(data_type).c_str());
return KERNEL_STATUS_PARAM_INVALID;
}
return KERNEL_STATUS_OK;
}
template <typename T>
auto CalcDiffByType(T diff) -> T
{
return diff;
}
template <>
std::complex<float> CalcDiffByType<std::complex<float>>(std::complex<float> diff)
{
return std::conj(diff);
}
template <>
std::complex<double> CalcDiffByType<std::complex<double>>(std::complex<double> diff)
{
return std::conj(diff);
}
uint32_t SquaredDifferenceCpuKernel::SquaredDifferenceCheck(const CpuKernelContext& ctx) const
{
Tensor* input_0 = ctx.Input(0);
Tensor* input_1 = ctx.Input(1);
Tensor* output = ctx.Output(0);
KERNEL_CHECK_NULLPTR(input_0->GetData(), KERNEL_STATUS_PARAM_INVALID, "Get input 0 data failed.")
KERNEL_CHECK_NULLPTR(input_1->GetData(), KERNEL_STATUS_PARAM_INVALID, "Get input 1 data failed.")
KERNEL_CHECK_NULLPTR(output->GetData(), KERNEL_STATUS_PARAM_INVALID, "Get output data failed")
DataType input0_type = input_0->GetDataType();
DataType input1_type = input_1->GetDataType();
KERNEL_CHECK_FALSE(
(input0_type == input1_type), KERNEL_STATUS_PARAM_INVALID,
"The data type of input0 [%s] need be same with "
"input1 [%s].",
DTypeStr(input0_type).c_str(), DTypeStr(input1_type).c_str())
KERNEL_LOG_DEBUG(
"SquaredDifferenceCpuKernel[%s], input0: size[%lu];"
"input1: size[%lu], output: size[%lu].",
ctx.GetOpType().c_str(), input_0->GetDataSize(), input_1->GetDataSize(), output->GetDataSize());
return KERNEL_STATUS_OK;
}
template <typename T>
void SquaredDifferenceCpuKernel::SpecialCompute(
BcastShapeType type, int64_t start, int64_t end, const T* input1, const T* input2, T* output)
{
switch (type) {
case BcastShapeType::SAME_SHAPE:
for (int64_t i = start; i < end; ++i) {
auto diff = *(input1 + i) - *(input2 + i);
*(output + i) = diff * CalcDiffByType(diff);
}
break;
case BcastShapeType::X_ONE_ELEMENT:
for (int64_t i = start; i < end; ++i) {
T temp = *input1;
auto diff = (temp) - *(input2 + i);
*(output + i) = diff * CalcDiffByType(diff);
}
break;
case BcastShapeType::Y_ONE_ELEMENT:
for (int64_t i = start; i < end; ++i) {
T temp = *input2;
auto diff = *(input1 + i) - (temp);
*(output + i) = diff * CalcDiffByType(diff);
}
break;
default:
KERNEL_LOG_WARN("Invalid type [%d]", static_cast<int32_t>(type));
break;
}
}
template <typename T>
uint32_t SquaredDifferenceCpuKernel::NoBcastCompute(const CpuKernelContext& ctx)
{
auto in0 = static_cast<T*>(ctx.Input(0)->GetData());
auto in1 = static_cast<T*>(ctx.Input(1)->GetData());
auto out = static_cast<T*>(ctx.Output(0)->GetData());
int64_t in0_elements_nums = ctx.Input(0)->NumElements();
int64_t in1_elements_nums = ctx.Input(1)->NumElements();
int64_t data_num = ctx.Output(0)->NumElements();
BcastShapeType type = in0_elements_nums == in1_elements_nums ?
BcastShapeType::SAME_SHAPE :
(in0_elements_nums == 1 ? BcastShapeType::X_ONE_ELEMENT : BcastShapeType::Y_ONE_ELEMENT);
if (data_num >= kParallelDataNumSameShape) {
uint32_t min_core_num = 1;
int64_t max_core_num = std::max(min_core_num, aicpu::CpuKernelUtils::GetCPUNum(ctx) - kResvCpuNum);
if (data_num <= kParallelDataNumSameShapeMid) {
max_core_num = std::min(max_core_num, static_cast<int64_t>(4));
}
if (max_core_num > data_num) {
max_core_num = data_num;
}
auto sharder_squareddifference = [this, &type, &in0, &in1, &out](int64_t start, int64_t end) {
SpecialCompute<T>(type, start, end, in0, in1, out);
};
if (static_cast<int>(max_core_num) == 0) {
return KERNEL_STATUS_PARAM_INVALID;
}
KERNEL_HANDLE_ERROR(
CpuKernelUtils::ParallelFor(ctx, data_num, data_num / max_core_num, sharder_squareddifference),
"SquaredDifference Compute failed.")
} else {
SpecialCompute<T>(type, 0, data_num, in0, in1, out);
}
return KERNEL_STATUS_OK;
}
template <typename T>
uint32_t SquaredDifferenceCpuKernel::BcastCompute(const CpuKernelContext& ctx, const Bcast& bcast)
{
auto in0 = static_cast<T*>(ctx.Input(0)->GetData());
auto in1 = static_cast<T*>(ctx.Input(1)->GetData());
auto out = static_cast<T*>(ctx.Output(0)->GetData());
int64_t data_num = ctx.Output(0)->NumElements();
if (data_num >= kParallelDataNum) {
uint32_t min_core_num = 1;
int64_t max_core_num = std::max(min_core_num, aicpu::CpuKernelUtils::GetCPUNum(ctx) - kResvCpuNum);
if (data_num <= kParallelDataNumMid) {
max_core_num = std::min(max_core_num, static_cast<int64_t>(4));
}
if (max_core_num > data_num) {
max_core_num = data_num;
}
auto sharder_squareddifference = [this, &in0, &in1, &out, &bcast](int64_t start, int64_t end) {
for (int64_t i = start; i < end; ++i) {
auto diff = *(in0 + bcast.GetBroadcastXIndex(i)) - *(in1 + bcast.GetBroadcastYIndex(i));
*(out + i) = diff * CalcDiffByType(diff);
}
};
if (static_cast<int>(max_core_num) == 0) {
return KERNEL_STATUS_PARAM_INVALID;
}
KERNEL_HANDLE_ERROR(
CpuKernelUtils::ParallelFor(ctx, data_num, data_num / max_core_num, sharder_squareddifference),
"SquaredDifference Compute failed.")
} else {
for (int64_t i = 0; i < data_num; ++i) {
auto diff = *(in0 + bcast.GetBroadcastXIndex(i)) - *(in1 + bcast.GetBroadcastYIndex(i));
*(out + i) = diff * CalcDiffByType(diff);
}
}
return KERNEL_STATUS_OK;
}
template <typename T>
uint32_t SquaredDifferenceCpuKernel::SquaredDifferenceCompute(const CpuKernelContext& ctx)
{
Tensor* input0_tensor = ctx.Input(0);
auto input0_shape = input0_tensor->GetTensorShape()->GetDimSizes();
int64_t input0_elements_nums = input0_tensor->NumElements();
Tensor* input1_tensor = ctx.Input(1);
auto input1_shape = input1_tensor->GetTensorShape()->GetDimSizes();
int64_t input1_elements_nums = input1_tensor->NumElements();
bool is_no_need_bcast =
(input0_shape == input1_shape) || (input0_elements_nums == 1) || (input1_elements_nums == 1);
if (is_no_need_bcast) {
return NoBcastCompute<T>(ctx);
} else {
Bcast bcast(input0_shape, input1_shape);
if (!bcast.IsValid()) {
KERNEL_LOG_ERROR("[%s] broadcast failed.", ctx.GetOpType().c_str());
return KERNEL_STATUS_PARAM_INVALID;
}
return BcastCompute<T>(ctx, bcast);
}
}
REGISTER_CPU_KERNEL(kSquaredDifference, SquaredDifferenceCpuKernel);
}