1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
|
#include "./fused_nbit_rowwise_conversion.h"
#include <c10/util/Half.h>
#include <algorithm>
#include <cmath>
#include "common.h"
#ifdef USE_FBGEMM
#include "fbgemm/QuantUtils.h"
#endif
namespace caffe2 {
void FloatToFused8BitRowwiseQuantized__base(
const float* input,
size_t input_rows,
int input_columns,
std::uint8_t* output) {
constexpr float kEpsilon = 1e-8f;
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int output_columns = input_columns + 2 * sizeof(float);
for (std::size_t row = 0; row < input_rows; ++row) {
const float* input_row = input + row * input_columns;
std::uint8_t* output_row = output + row * output_columns;
float* output_row_scale_bias =
reinterpret_cast<float*>(output_row + input_columns);
float minimum_element =
*std::min_element(input_row, input_row + input_columns);
float maximum_element =
*std::max_element(input_row, input_row + input_columns);
float range = maximum_element - minimum_element;
output_row_scale_bias[0] = range / 255.0f;
output_row_scale_bias[1] = minimum_element;
const auto inverse_scale = 255.0f / (range + kEpsilon);
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (std::size_t col = 0; col < input_columns; ++col) {
output_row[col] =
std::lrintf((input_row[col] - minimum_element) * inverse_scale);
}
}
}
void Fused8BitRowwiseQuantizedToFloat__base(
const std::uint8_t* input,
size_t input_rows,
int input_columns,
float* output) {
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
int output_columns = input_columns - 2 * sizeof(float);
for (std::size_t row = 0; row < input_rows; ++row) {
const std::uint8_t* input_row = input + row * input_columns;
const float* input_row_scale_bias =
reinterpret_cast<const float*>(input_row + output_columns);
float* output_row = output + row * output_columns;
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (std::size_t col = 0; col < output_columns; ++col) {
output_row[col] =
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
input_row[col] * input_row_scale_bias[0] + input_row_scale_bias[1];
}
}
}
void FloatToFused8BitRowwiseQuantized(
const float* input,
size_t input_rows,
int input_columns,
std::uint8_t* output) {
#ifdef USE_FBGEMM
fbgemm::FloatOrHalfToFused8BitRowwiseQuantizedSBFloat<float>(
input, input_rows, input_columns, output);
#else
FloatToFused8BitRowwiseQuantized__base(
input, input_rows, input_columns, output);
#endif
}
void Fused8BitRowwiseQuantizedToFloat(
const std::uint8_t* input,
size_t input_rows,
int input_columns,
float* output) {
#ifdef USE_FBGEMM
fbgemm::Fused8BitRowwiseQuantizedSBFloatToFloatOrHalf<float>(
input, input_rows, input_columns, output);
#else
Fused8BitRowwiseQuantizedToFloat__base(
input, input_rows, input_columns, output);
#endif
}
void FloatToFusedNBitRowwiseQuantizedSBHalf__base(
int bit_rate,
const float* input,
size_t input_rows,
int input_columns,
std::uint8_t* output) {
int num_elem_per_byte = 8 / bit_rate;
int output_columns =
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
(input_columns + num_elem_per_byte - 1) / num_elem_per_byte +
2 * sizeof(at::Half);
for (std::size_t row = 0; row < input_rows; ++row) {
const float* input_row = input + row * input_columns;
std::uint8_t* output_row = output + row * output_columns;
at::Half* output_row_scale_bias = reinterpret_cast<at::Half*>(
output_row +
(input_columns + num_elem_per_byte - 1) / num_elem_per_byte);
float minimum_element =
*std::min_element(input_row, input_row + input_columns);
float maximum_element =
*std::max_element(input_row, input_row + input_columns);
minimum_element = static_cast<at::Half>(minimum_element);
const float range = maximum_element - minimum_element;
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
at::Half scale = range == 0 ? 1.0f : range / ((1 << bit_rate) - 1);
if (scale == 0) {
// Corner case handling when maximum_element == minimum_element
// Any scale would work because X - minimum_element will be 0 for all X
scale = 1.0f;
}
float inverse_scale = 1.0f / scale;
if (std::isinf(inverse_scale)) {
scale = 1.0f;
inverse_scale = 1.0f;
}
output_row_scale_bias[0] = scale;
output_row_scale_bias[1] = minimum_element;
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (std::size_t col = 0; col < input_columns; ++col) {
float X = input_row[col];
std::uint8_t quantized = std::max(
0,
std::min<int>(
std::lrintf((X - minimum_element) * inverse_scale),
(1 << bit_rate) - 1));
if (col % num_elem_per_byte == 0) {
output_row[col / num_elem_per_byte] = quantized;
} else {
output_row[col / num_elem_per_byte] |=
(quantized << ((col % num_elem_per_byte) * bit_rate));
}
}
}
}
void FusedNBitRowwiseQuantizedSBHalfToFloat__base(
int bit_rate,
const std::uint8_t* input,
size_t input_rows,
int input_columns,
float* output) {
int num_elem_per_byte = 8 / bit_rate;
int output_columns =
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
(input_columns - 2 * sizeof(at::Half)) * num_elem_per_byte;
for (std::size_t row = 0; row < input_rows; ++row) {
const std::uint8_t* input_row = input + row * input_columns;
const at::Half* input_row_scale_bias = reinterpret_cast<const at::Half*>(
input_row +
(output_columns + num_elem_per_byte - 1) / num_elem_per_byte);
float scale = input_row_scale_bias[0];
float bias = input_row_scale_bias[1];
float* output_row = output + row * output_columns;
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (std::size_t col = 0; col < output_columns; ++col) {
std::uint8_t quantized = input_row[col / num_elem_per_byte];
quantized >>= (col % num_elem_per_byte) * bit_rate;
quantized &= (1 << bit_rate) - 1;
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
output_row[col] = scale * quantized + bias;
}
}
}
void FloatToFusedNBitRowwiseQuantizedSBHalf(
int bit_rate,
const float* input,
size_t input_rows,
int input_columns,
std::uint8_t* output) {
#ifdef USE_FBGEMM
fbgemm::FloatOrHalfToFusedNBitRowwiseQuantizedSBHalf<float>(
bit_rate, input, input_rows, input_columns, output);
#else
FloatToFusedNBitRowwiseQuantizedSBHalf__base(
bit_rate, input, input_rows, input_columns, output);
#endif
}
void FusedNBitRowwiseQuantizedSBHalfToFloat(
int bit_rate,
const std::uint8_t* input,
size_t input_rows,
int input_columns,
float* output) {
#ifdef USE_FBGEMM
fbgemm::FusedNBitRowwiseQuantizedSBHalfToFloatOrHalf<float>(
bit_rate, input, input_rows, input_columns, output);
#else
FusedNBitRowwiseQuantizedSBHalfToFloat__base(
bit_rate, input, input_rows, input_columns, output);
#endif
}
} // namespace caffe2
|