Added doc for fake_quant_with_min_max* op helpers implementations.

master
shugeo 2019-10-10 18:31:17 +03:00
parent c3f755d975
commit c890de5a7b
1 changed files with 34 additions and 22 deletions

View File

@ -25,43 +25,54 @@ namespace nd4j {
namespace ops {
namespace helpers {
//
// nudge - nudged min max over scale
// scale = (Max - Min) / (quantMax - quantMin)
// quantMin = 0 or 1, quantMax = 2^b - 1 == (1 << b) - 1
//
template <typename T>
static void nudge(T min, T max, int quant_min, int quant_max, T* scale, T* nudged_min, T* nudged_max) {
T quant_max_float = static_cast<T>(quant_max);
T quant_min_float = static_cast<T>(quant_min);
*scale = (max - min) / (quant_max_float - quant_min_float);
auto zero_point_from_min = quant_min_float - min / *scale;
uint16_t const nudged_zero_point = [zero_point_from_min, quant_min, quant_max, quant_max_float, quant_min_float] {
if (zero_point_from_min < quant_min_float) {
return static_cast<uint16_t>(quant_min);
static void nudge(T min, T max, int quantMin, int quantMax, T* scale, T* nudgedMin, T* nudgedMax) {
// floating point instead integers
T quantMaxF = static_cast<T>(quantMax);
T quantMinF = static_cast<T>(quantMin);
// compute scale
*scale = (max - min) / (quantMaxF - quantMinF);
// compute left bound point
auto zeroPointFromMin = quantMinF - min / *scale;
// bound zero point to conform with range [0 or 1, 2^b - 1]
uint16_t const nudged_zero_point = [zeroPointFromMin, quantMin, quantMax, quantMaxF, quantMinF] {
if (zeroPointFromMin < quantMinF) {
return static_cast<uint16_t>(quantMin);
}
if (zero_point_from_min > quant_max_float) {
return static_cast<uint16_t>(quant_max);
if (zeroPointFromMin > quantMaxF) {
return static_cast<uint16_t>(quantMax);
}
return nd4j::math::nd4j_round<T,uint16_t>(zero_point_from_min);
return nd4j::math::nd4j_round<T,uint16_t>(zeroPointFromMin);
}();
*nudged_min = (quant_min_float - nudged_zero_point) * (*scale);
*nudged_max = (quant_max_float - nudged_zero_point) * (*scale);
// compute nudged min and max with computed nudged zero point
*nudgedMin = (quantMinF - nudged_zero_point) * (*scale);
*nudgedMax = (quantMaxF - nudged_zero_point) * (*scale);
}
template <typename T>
void fakeQuantWithMinMaxVarsPerChannel_(NDArray* input, NDArray* min, NDArray* max, int numBits, bool narrowed, NDArray* output) {
int lowIntBound = narrowed ? 1 : 0;
int upperIntBound = (1 << numBits) - 1;
auto channels = input->sizeAt(-1);
int lowIntBound = narrowed ? 1 : 0; // 0 or 1
int upperIntBound = (1 << numBits) - 1; // 2^b - 1
auto channels = input->sizeAt(-1); // last dimension
PRAGMA_OMP_PARALLEL_FOR
for (auto i = 0; i < channels; i++) {
T scale, nudged_min, nudged_max;
// nudge min and max first, with scale computing
nudge<T>(min->t<T>(i), max->t<T>(i), lowIntBound, upperIntBound, &scale, &nudged_min, &nudged_max);
// slide using last dimension and process all for given channel
for (auto e = 0; e < input->lengthOf(); e += channels) {
T val = input->t<T>(e + i);
if ( val <= nudged_min)
val = nudged_min;
else if (val >= nudged_max)
val = nudged_max;
// quantization itself
output->t<T>(e + i) = math::nd4j_floor<T,T>((val - nudged_min)/scale + T(0.5)) * scale + nudged_min;
}
}
@ -73,16 +84,17 @@ namespace helpers {
int upperIntBound = (1 << numBits) - 1;
T nudgedMin, nudgedMax, scale;
// nudge with given min and max and compute scale and nudged min and max
nudge<T>(min->t<T>(0), max->t<T>(0), lowIntBound, upperIntBound, &scale, &nudgedMin, &nudgedMax);
// quantization as one
auto fakeQuantizationWithMinMax = LAMBDA_T(x, nudgedMin, nudgedMax, scale) {
T val = x;
T val = x; // boundign value between nudged min and max
if (val < nudgedMin) {
val = nudgedMin;
}
else if (val > nudgedMax)
val = nudgedMax;
// converse value with scale and shifted with nudged min
return (nd4j::math::nd4j_floor<T,T>((val - nudgedMin)/scale + T(0.5)) * scale + nudgedMin);
};