280 lines
9.4 KiB
C++
280 lines
9.4 KiB
C++
/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
==============================================================================*/
|
|
|
|
#include "tensorflow/lite/micro/micro_utils.h"
|
|
|
|
#include <limits.h>
|
|
#include <math.h>
|
|
#include <stdint.h>
|
|
|
|
#include "tensorflow/lite/c/common.h"
|
|
#include "tensorflow/lite/kernels/op_macros.h"
|
|
|
|
namespace tflite {
|
|
|
|
namespace {
|
|
|
|
static const uint8_t kAsymmetricUInt8Min = 0;
|
|
static const uint8_t kAsymmetricUInt8Max = UINT8_MAX;
|
|
static const uint8_t kSymmetricUInt8Min = 1;
|
|
static const uint8_t kSymmetricUInt8Max = UINT8_MAX;
|
|
static const int8_t kAsymmetricInt8Min = INT8_MIN;
|
|
static const int8_t kAsymmetricInt8Max = INT8_MAX;
|
|
static const int kSymmetricInt8Scale = kAsymmetricInt8Max;
|
|
|
|
static const int16_t kAsymmetricInt16Min = INT16_MIN;
|
|
static const int16_t kAsymmetricInt16Max = INT16_MAX;
|
|
static const int kSymmetricInt16Scale = kAsymmetricInt16Max;
|
|
|
|
static const int32_t kAsymmetricInt32Max = INT32_MAX;
|
|
static const int kSymmetricInt32Scale = kAsymmetricInt32Max;
|
|
|
|
} // namespace
|
|
|
|
int ElementCount(const TfLiteIntArray& dims) {
|
|
int result = 1;
|
|
for (int i = 0; i < dims.size; ++i) {
|
|
result *= dims.data[i];
|
|
}
|
|
return result;
|
|
}
|
|
|
|
// Converts a float value into an unsigned eight-bit quantized value.
|
|
uint8_t FloatToAsymmetricQuantizedUInt8(const float value, const float scale,
|
|
const int zero_point) {
|
|
int32_t result = round(value / scale) + zero_point;
|
|
if (result < kAsymmetricUInt8Min) {
|
|
result = kAsymmetricUInt8Min;
|
|
}
|
|
if (result > kAsymmetricUInt8Max) {
|
|
result = kAsymmetricUInt8Max;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
uint8_t FloatToSymmetricQuantizedUInt8(const float value, const float scale) {
|
|
int32_t result = round(value / scale);
|
|
if (result < kSymmetricUInt8Min) {
|
|
result = kSymmetricUInt8Min;
|
|
}
|
|
if (result > kSymmetricUInt8Max) {
|
|
result = kSymmetricUInt8Max;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
int8_t FloatToAsymmetricQuantizedInt8(const float value, const float scale,
|
|
const int zero_point) {
|
|
int32_t result = round(value / scale) + zero_point;
|
|
if (result < kAsymmetricInt8Min) {
|
|
result = kAsymmetricInt8Min;
|
|
}
|
|
if (result > kAsymmetricInt8Max) {
|
|
result = kAsymmetricInt8Max;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
int16_t FloatToAsymmetricQuantizedInt16(const float value, const float scale,
|
|
const int zero_point) {
|
|
int32_t result = round(value / scale) + zero_point;
|
|
if (result < kAsymmetricInt16Min) {
|
|
result = kAsymmetricInt16Min;
|
|
}
|
|
if (result > kAsymmetricInt16Max) {
|
|
result = kAsymmetricInt16Max;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
int8_t FloatToSymmetricQuantizedInt8(const float value, const float scale) {
|
|
return FloatToAsymmetricQuantizedInt8(value, scale, 0.0f);
|
|
}
|
|
|
|
int32_t FloatToSymmetricQuantizedInt32(const float value, const float scale) {
|
|
float quantized = round(value / scale);
|
|
if (static_cast<int>(quantized) > INT_MAX) {
|
|
quantized = static_cast<float>(INT_MAX);
|
|
} else if (quantized < INT_MIN) {
|
|
quantized = static_cast<float> INT_MIN;
|
|
}
|
|
|
|
return static_cast<int>(quantized);
|
|
}
|
|
|
|
void AsymmetricQuantize(const float* input, int8_t* output, int num_elements,
|
|
float scale, int zero_point) {
|
|
for (int i = 0; i < num_elements; i++) {
|
|
output[i] = FloatToAsymmetricQuantizedInt8(input[i], scale, zero_point);
|
|
}
|
|
}
|
|
|
|
void AsymmetricQuantize(const float* input, uint8_t* output, int num_elements,
|
|
float scale, int zero_point) {
|
|
for (int i = 0; i < num_elements; i++) {
|
|
output[i] = FloatToAsymmetricQuantizedUInt8(input[i], scale, zero_point);
|
|
}
|
|
}
|
|
|
|
void AsymmetricQuantize(const float* input, int16_t* output, int num_elements,
|
|
float scale, int zero_point) {
|
|
for (int i = 0; i < num_elements; i++) {
|
|
output[i] = FloatToAsymmetricQuantizedInt16(input[i], scale, zero_point);
|
|
}
|
|
}
|
|
|
|
void SymmetricQuantize(const float* input, int32_t* output, int num_elements,
|
|
float scale) {
|
|
for (int i = 0; i < num_elements; i++) {
|
|
output[i] = FloatToSymmetricQuantizedInt32(input[i], scale);
|
|
}
|
|
}
|
|
|
|
void SymmetricPerChannelQuantize(const float* input, int32_t* output,
|
|
int num_elements, int num_channels,
|
|
float* scales) {
|
|
int elements_per_channel = num_elements / num_channels;
|
|
for (int i = 0; i < num_channels; i++) {
|
|
for (int j = 0; j < elements_per_channel; j++) {
|
|
output[i * elements_per_channel + j] = FloatToSymmetricQuantizedInt32(
|
|
input[i * elements_per_channel + j], scales[i]);
|
|
}
|
|
}
|
|
}
|
|
|
|
void SignedSymmetricPerChannelQuantize(const float* values,
|
|
TfLiteIntArray* dims,
|
|
int quantized_dimension,
|
|
int8_t* quantized_values,
|
|
float* scaling_factors) {
|
|
int input_size = ElementCount(*dims);
|
|
int channel_count = dims->data[quantized_dimension];
|
|
int per_channel_size = input_size / channel_count;
|
|
|
|
int stride;
|
|
int channel_stride;
|
|
if (quantized_dimension == 0) {
|
|
stride = 1;
|
|
channel_stride = per_channel_size;
|
|
} else if (quantized_dimension == 3) {
|
|
stride = channel_count;
|
|
channel_stride = 1;
|
|
} else {
|
|
TF_LITE_FATAL("quantized dimension must be 0 or 3");
|
|
}
|
|
|
|
// Calculate scales for each channel.
|
|
for (int channel = 0; channel < channel_count; channel++) {
|
|
float min = 0;
|
|
float max = 0;
|
|
|
|
for (int i = 0; i < per_channel_size; i++) {
|
|
int idx = channel * channel_stride + i * stride;
|
|
min = fminf(min, values[idx]);
|
|
max = fmaxf(max, values[idx]);
|
|
}
|
|
scaling_factors[channel] =
|
|
fmaxf(fabs(min), fabs(max)) / kSymmetricInt8Scale;
|
|
for (int i = 0; i < per_channel_size; i++) {
|
|
int idx = channel * channel_stride + i * stride;
|
|
const int32_t quantized_value =
|
|
static_cast<int32_t>(roundf(values[idx] / scaling_factors[channel]));
|
|
// Clamp: just in case some odd numeric offset.
|
|
quantized_values[idx] = fminf(
|
|
kSymmetricInt8Scale, fmaxf(-kSymmetricInt8Scale, quantized_value));
|
|
}
|
|
}
|
|
}
|
|
|
|
void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims,
|
|
int8_t* quantized_values, float* scaling_factor) {
|
|
int input_size = ElementCount(*dims);
|
|
|
|
float min = 0;
|
|
float max = 0;
|
|
for (int i = 0; i < input_size; i++) {
|
|
min = fminf(min, values[i]);
|
|
max = fmaxf(max, values[i]);
|
|
}
|
|
*scaling_factor = fmaxf(fabs(min), fabs(max)) / kSymmetricInt8Scale;
|
|
for (int i = 0; i < input_size; i++) {
|
|
const int32_t quantized_value =
|
|
static_cast<int32_t>(roundf(values[i] / *scaling_factor));
|
|
// Clamp: just in case some odd numeric offset.
|
|
quantized_values[i] = fminf(kSymmetricInt8Scale,
|
|
fmaxf(-kSymmetricInt8Scale, quantized_value));
|
|
}
|
|
}
|
|
|
|
void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims,
|
|
int16_t* quantized_values, float* scaling_factor) {
|
|
int input_size = ElementCount(*dims);
|
|
|
|
float min = 0;
|
|
float max = 0;
|
|
for (int i = 0; i < input_size; i++) {
|
|
min = fminf(min, values[i]);
|
|
max = fmaxf(max, values[i]);
|
|
}
|
|
*scaling_factor = fmaxf(fabs(min), fabs(max)) / kSymmetricInt16Scale;
|
|
for (int i = 0; i < input_size; i++) {
|
|
const int32_t quantized_value =
|
|
static_cast<int32_t>(roundf(values[i] / *scaling_factor));
|
|
// Clamp: just in case some odd numeric offset.
|
|
quantized_values[i] = fminf(kSymmetricInt16Scale,
|
|
fmaxf(-kSymmetricInt16Scale, quantized_value));
|
|
}
|
|
}
|
|
|
|
void SignedSymmetricQuantize(const float* values, TfLiteIntArray* dims,
|
|
int32_t* quantized_values, float* scaling_factor) {
|
|
int input_size = ElementCount(*dims);
|
|
|
|
float min = 0;
|
|
float max = 0;
|
|
for (int i = 0; i < input_size; i++) {
|
|
min = fminf(min, values[i]);
|
|
max = fmaxf(max, values[i]);
|
|
}
|
|
|
|
*scaling_factor =
|
|
fmaxf(fabs(min), fabs(max)) / static_cast<float>(kSymmetricInt32Scale);
|
|
for (int i = 0; i < input_size; i++) {
|
|
const int32_t quantized_value =
|
|
static_cast<int32_t>(roundf(values[i] / *scaling_factor));
|
|
// Clamp: just in case some odd numeric offset.
|
|
quantized_values[i] = fminf(
|
|
static_cast<float>(kSymmetricInt32Scale),
|
|
fmaxf(static_cast<float>(-kSymmetricInt32Scale), quantized_value));
|
|
}
|
|
}
|
|
|
|
void SymmetricQuantize(const float* values, TfLiteIntArray* dims,
|
|
uint8_t* quantized_values, float* scaling_factor) {
|
|
SignedSymmetricQuantize(values, dims,
|
|
reinterpret_cast<int8_t*>(quantized_values),
|
|
scaling_factor);
|
|
}
|
|
|
|
void SymmetricDequantize(const int8_t* values, const int size,
|
|
const float dequantization_scale,
|
|
float* dequantized_values) {
|
|
for (int i = 0; i < size; ++i) {
|
|
dequantized_values[i] = values[i] * dequantization_scale;
|
|
}
|
|
}
|
|
|
|
} // namespace tflite
|