* 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 "log_aicpu.h"
#include <cfloat>
#include "cpu_kernel_utils.h"
#include "cpu_types.h"
#include "log.h"
#include "status.h"
#include "utils/kernel_util.h"
#include "utils/eigen_tensor.h"
namespace {
constexpr uint32_t kLogInputNum = 1U;
constexpr uint32_t kLogOutputNum = 1U;
const char* const kLog = "Log";
constexpr int64_t kDataDefaultParallelNum = 60 * 1024;
constexpr int64_t kFloat16AndIntParallelNum = 35 * 1024;
constexpr int64_t kCompAndBf16ParallelNum = 18 * 1024;
constexpr float kLog2Base = 2.0F;
constexpr float kLog10Base = 10.0F;
}
namespace aicpu {
template <typename T,
typename std::enable_if<!std::is_same<T, std::complex<float>>::value &&
!std::is_same<T, std::complex<double>>::value &&
!std::is_same<T, Eigen::bfloat16>::value &&
!std::is_same<T, Eigen::half>::value, T>::type* = nullptr>
inline void Log2Compute(T *input, T *output, int64_t start, int64_t end) {
for (int64_t index = start; index < end; index++) {
output[index] = std::log2(input[index]);
}
}
template <typename T,
typename std::enable_if<std::is_same<T, Eigen::bfloat16>::value ||
std::is_same<T, Eigen::half>::value, T>::type* = nullptr>
inline void Log2Compute(T *input, T *output, int64_t start, int64_t end) {
for (int64_t index = start; index < end; index++) {
output[index] = (T)(std::log2(input[index]));
}
}
template <typename T,
typename std::enable_if<std::is_same<T, std::complex<float>>::value ||
std::is_same<T, std::complex<double>>::value, T>::type* = nullptr>
inline void Log2Compute(T *input, T *output, int64_t start, int64_t end) {
T log_2 = T(std::log(kLog2Base));
for (int64_t index = start; index < end; index++) {
output[index] = std::log(input[index]) / log_2;
}
}
template <typename T,
typename std::enable_if<std::is_same<T, Eigen::bfloat16>::value ||
std::is_same<T, Eigen::half>::value, T>::type* = nullptr>
inline void Log10Compute(T *input, T *output, int64_t start, int64_t end) {
for (int64_t index = start; index < end; index++) {
output[index] = (T)(std::log10(input[index]));
}
}
template <typename T,
typename std::enable_if<!std::is_same<T, Eigen::bfloat16>::value &&
!std::is_same<T, Eigen::half>::value, T>::type* = nullptr>
inline void Log10Compute(T *input, T *output, int64_t start, int64_t end) {
for (int64_t index = start; index < end; index++) {
output[index] = std::log10(input[index]);
}
}
template <typename Tin, typename Tout>
inline void ComputeNaturalLog(Tin* input, Tout* output, int64_t start, int64_t end) {
for (int64_t index = start; index < end; index++) {
output[index] = static_cast<Tout>(std::log(input[index]));
}
}
template <typename Tin, typename Tout>
inline void ComputeShiftLog(Tin* input, Tout* output, float shift, int64_t start, int64_t end) {
Tout shift_value = static_cast<Tout>(shift);
for (int64_t index = start; index < end; index++) {
output[index] = Eigen::numext::log(shift_value + input[index]);
}
}
bool NeedParallel(int64_t data_num, DataType dtype){
return (data_num > kDataDefaultParallelNum) || (data_num > kFloat16AndIntParallelNum &&
dtype == DT_FLOAT16) || (data_num > kCompAndBf16ParallelNum &&
(dtype == DT_BFLOAT16 || dtype == DT_COMPLEX64 || dtype == DT_COMPLEX128));
}
bool NaturalLogCondition(float base, float shift, float scale) {
return (std::fabs(base - (-1.0F)) < FLT_EPSILON) && (std::fabs(shift) < FLT_EPSILON) &&
(std::fabs(scale - 1.0F) < FLT_EPSILON);
}
bool Log2Condition(float base, float shift, float scale) {
return (std::fabs(base - kLog2Base) < FLT_EPSILON) && (std::fabs(shift) < FLT_EPSILON) &&
(std::fabs(scale - 1.0F) < FLT_EPSILON);
}
bool Log10Condition(float base, float shift, float scale) {
return (std::fabs(base - kLog10Base) < FLT_EPSILON) && (std::fabs(shift) < FLT_EPSILON) &&
(std::fabs(scale - 1.0F) < FLT_EPSILON);
}
template <typename Tin, typename Tout>
uint32_t LogCompute(CpuKernelContext &ctx, float base, float shift, float scale) {
Tin *input = reinterpret_cast<Tin *>(ctx.Input(0)->GetData());
Tout *output = reinterpret_cast<Tout *>(ctx.Output(0)->GetData());
auto shard_log = [&](int64_t start, int64_t end) {
if (NaturalLogCondition(base, shift, scale)) {
ComputeNaturalLog(input, output, start, end);
} else if (Log2Condition(base, shift, scale)) {
Log2Compute(input, output, start, end);
} else if (Log10Condition(base, shift, scale)) {
Log10Compute(input, output, start, end);
} else if ((std::fabs(base - (-1.0F)) < FLT_EPSILON) && (std::fabs(shift) >= FLT_EPSILON) &&
(std::fabs(scale - 1.0F) < FLT_EPSILON)) {
ComputeShiftLog(input, output, shift, start, end);
} else {
Tout shift_value = static_cast<Tout>(shift);
Tout scale_value = static_cast<Tout>(scale);
Tout log_base = (std::fabs(base - (-1.0F)) < FLT_EPSILON) ? static_cast<Tout>(1.0) : static_cast<Tout>(Eigen::numext::log(base));
for (int64_t index = start; index < end; index++) {
output[index] = Eigen::numext::log(shift_value + scale_value * input[index]) / log_base;
}
}
};
int64_t data_num = ctx.Input(0)->NumElements();
DataType dtype = ctx.Input(0)->GetDataType();
if (NeedParallel(data_num, dtype)) {
uint32_t max_core_num = std::max(1U, aicpu::CpuKernelUtils::GetCPUNum(ctx) - kResvCpuNum);
KERNEL_HANDLE_ERROR(CpuKernelUtils::ParallelFor(ctx, data_num,
CeilMultiple(data_num, max_core_num), shard_log),
"Log Compute failed.")
} else {
shard_log(0, data_num);
}
return KERNEL_STATUS_OK;
}
template <typename Tin, typename Tout>
uint32_t LogComputeIntegr(CpuKernelContext &ctx, float base, float shift, float scale) {
Tin *input = reinterpret_cast<Tin *>(ctx.Input(0)->GetData());
Tout *output = reinterpret_cast<Tout *>(ctx.Output(0)->GetData());
auto shard_log = [&](int64_t start, int64_t end) {
if (NaturalLogCondition(base, shift, scale)) {
ComputeNaturalLog(input, output, start, end);
} else if (Log2Condition(base, shift, scale)) {
for (int64_t index = start; index < end; index++) {
output[index] = static_cast<Tout>(std::log2(input[index]));
}
} else if (Log10Condition(base, shift, scale)) {
for (int64_t index = start; index < end; index++) {
output[index] = static_cast<Tout>(std::log10(input[index]));
}
} else if ((std::fabs(base - (-1.0F)) < FLT_EPSILON) && (std::fabs(shift) >= FLT_EPSILON) &&
(std::fabs(scale - 1.0F) < FLT_EPSILON)) {
Tout shift_value = static_cast<Tout>(shift);
for (int64_t index = start; index < end; index++) {
output[index] = std::log(shift_value + input[index]);
}
} else {
Tout shift_value = static_cast<Tout>(shift);
Tout scale_value = static_cast<Tout>(scale);
Tout log_base = (std::fabs(base - (-1.0F)) < FLT_EPSILON) ? log_base = static_cast<Tout>(1.0) : static_cast<Tout>(Eigen::numext::log(base));
for (int64_t index = start; index < end; index++) {
output[index] = std::log(shift_value + scale_value * input[index]) / log_base;
}
}
};
int64_t data_num = ctx.Input(0)->NumElements();
if (data_num > kFloat16AndIntParallelNum) {
uint32_t max_core_num = std::max(1U, aicpu::CpuKernelUtils::GetCPUNum(ctx) - kResvCpuNum);
KERNEL_HANDLE_ERROR(CpuKernelUtils::ParallelFor(ctx, data_num,
CeilMultiple(data_num, max_core_num), shard_log),
"Log Compute failed.")
} else {
shard_log(0, data_num);
}
return KERNEL_STATUS_OK;
}
static const std::map<uint32_t, std::map<uint32_t, std::function<uint32_t(
CpuKernelContext &ctx, float base_, float shift_, float scale_)>>> log_calls = {
{DT_UINT8, {{DT_FLOAT, LogComputeIntegr<uint8_t, float>}}},
{DT_INT8, {{DT_FLOAT, LogComputeIntegr<int8_t, float>}}},
{DT_INT16, {{DT_FLOAT, LogComputeIntegr<int16_t, float>}}},
{DT_INT32, {{DT_FLOAT, LogComputeIntegr<int32_t, float>}}},
{DT_INT64, {{DT_FLOAT, LogComputeIntegr<int64_t, float>}}},
{DT_FLOAT, {{DT_FLOAT, LogCompute<float, float>}}},
{DT_DOUBLE, {{DT_DOUBLE, LogCompute<double, double>}}},
{DT_FLOAT16, {{DT_FLOAT16, LogCompute<Eigen::half, Eigen::half>}}},
{DT_BFLOAT16, {{DT_BFLOAT16, LogCompute<Eigen::bfloat16, Eigen::bfloat16>}}},
{DT_BOOL, {{DT_FLOAT, LogComputeIntegr<bool, float>}}},
{DT_COMPLEX64, {{DT_COMPLEX64, LogCompute<std::complex<float>, std::complex<float>>}}},
{DT_COMPLEX128, {{DT_COMPLEX128, LogCompute<std::complex<double>, std::complex<double>>}}}
};
uint32_t LogCpuKernel::LogOpCheck(CpuKernelContext &ctx) {
auto input = ctx.Input(0);
auto output = ctx.Output(0);
DataType input_type = input->GetDataType();
DataType output_type = output->GetDataType();
if (input_type == DT_UINT8 || input_type == DT_INT8 || input_type == DT_INT16 ||
input_type == DT_INT32 || input_type == DT_INT64 || input_type == DT_BOOL) {
KERNEL_CHECK_FALSE((output_type == DT_FLOAT), KERNEL_STATUS_PARAM_INVALID,
"When the log op input data is [%s], the output type"
" must be [DT_FLOAT32]. Current ouput tye is [%s].",
DTypeStr(input_type).c_str(), DTypeStr(output_type).c_str())
} else {
KERNEL_CHECK_FALSE((input_type == output_type), KERNEL_STATUS_PARAM_INVALID,
"The data type of output [%s] need be same with input [%s].",
DTypeStr(output_type).c_str(), DTypeStr(input_type).c_str())
}
AttrValue *base_attr = ctx.GetAttr("base");
base_ = (base_attr == nullptr) ? -1.0f : (base_attr->GetFloat());
KERNEL_CHECK_FALSE((IsValueEqual<float>(base_, -1.0f) || (base_ > 0 && !IsValueEqual<float>(base_, 1.0f))),
KERNEL_STATUS_PARAM_INVALID, "The Log op base attribute "
"[%f] is invaild, please user -1.0, (0, 1) and (1, inf).", base_)
AttrValue *shift_attr = ctx.GetAttr("shift");
shift_ = (shift_attr == nullptr) ? 0.0f : (shift_attr->GetFloat());
AttrValue *scale_attr = ctx.GetAttr("scale");
scale_ = (scale_attr == nullptr) ? 1.0f : (scale_attr->GetFloat());
KERNEL_LOG_DEBUG("Log op check params successed. base = %f, shift = %f, scale = %f.",
base_, shift_, scale_);
return KERNEL_STATUS_OK;
}
std::uint32_t LogCpuKernel::Compute(CpuKernelContext &ctx) {
KERNEL_HANDLE_ERROR(NormalCheck(ctx, kLogInputNum, kLogOutputNum),
"Check Log op input and output failed.");
KERNEL_HANDLE_ERROR(LogOpCheck(ctx), "Check Log op params failed.");
if (ctx.Output(0)->NumElements() == 0) {
KERNEL_LOG_DEBUG("log op output shape element number is zero.");
return KERNEL_STATUS_OK;
}
auto input_type = ctx.Input(0)->GetDataType();
auto output_type = ctx.Output(0)->GetDataType();
uint32_t ret;
try {
ret = log_calls.at(input_type).at(output_type)(ctx, base_, shift_, scale_);
} catch (const std::out_of_range &e) {
KERNEL_LOG_ERROR("Log op params input x and output y doesn't support tensor "
"types: [%s][%s], e.what:[%s]", DTypeStr(input_type).c_str(),
DTypeStr(output_type).c_str(), e.what());
ret = static_cast<uint32_t>(KERNEL_STATUS_PARAM_INVALID);
}
return ret;
}
REGISTER_CPU_KERNEL(kLog, LogCpuKernel);
}
