Refactored fake_quant_with_min_max_vars op cuda implementation.

master
shugeo 2019-10-10 14:00:49 +03:00
parent c13e945a96
commit 753565145c
1 changed files with 30 additions and 41 deletions

View File

@ -34,58 +34,47 @@ namespace helpers {
// output - output tensor // output - output tensor
// //
template <typename T> template <typename T>
void fakeQuantWithMinMaxVars_(NDArray* input, NDArray* min, NDArray* max, int numBits, bool narrowed, NDArray* output) { static void Nudge(T min, T max, int quant_min, int quant_max, T* scale, T* nudged_min, T* nudged_max) {
int lowIntBound = narrowed?1:0; T quant_max_float = static_cast<T>(quant_max);
int upperIntBound = 1 << numBits - 1; T quant_min_float = static_cast<T>(quant_min);
min->syncToHost(); *scale = (max - min) / (quant_max_float - quant_min_float);
max->syncToHost(); auto zero_point_from_min = quant_min_float - min / *scale;
const float quant_min_float = static_cast<float>(lowIntBound); uint16_t const nudged_zero_point = [zero_point_from_min, quant_min, quant_max, quant_max_float, quant_min_float] {
const float quant_max_float = static_cast<float>(upperIntBound);
T scale = (max->t<T>(0) - min->t<T>(0)) / (quant_max_float - quant_min_float);
const T zero_point_from_min = quant_min_float - min->t<T>(0) / scale;
const uint16_t nudged_zero_point = [zero_point_from_min, lowIntBound,
quant_min_float, upperIntBound,
quant_max_float] {
if (zero_point_from_min < quant_min_float) { if (zero_point_from_min < quant_min_float) {
return static_cast<uint16_t>(lowIntBound); return static_cast<uint16_t>(quant_min);
} }
if (zero_point_from_min > quant_max_float) { if (zero_point_from_min > quant_max_float) {
return static_cast<uint16_t>(upperIntBound); return static_cast<uint16_t>(quant_max);
} }
return static_cast<uint16_t>(roundf(zero_point_from_min)); return nd4j::math::nd4j_round<T,uint16_t>(zero_point_from_min);
}(); }();
*nudged_min = (quant_min_float - nudged_zero_point) * (*scale);
*nudged_max = (quant_max_float - nudged_zero_point) * (*scale);
}
auto nudged_min = (quant_min_float - nudged_zero_point) * (scale); template <typename T>
auto nudged_max = (quant_max_float - nudged_zero_point) * (scale); void fakeQuantWithMinMaxVars_(NDArray* input, NDArray* min, NDArray* max, int numBits, bool narrowed, NDArray* output) {
int lowIntBound = narrowed?1:0;
int upperIntBound = (1 << numBits) - 1;
min->syncToHost();
max->syncToHost();
T scale, nudged_min, nudged_max;
Nudge(min->t<T>(0), max->t<T>(0), lowIntBound, upperIntBound, &scale, &nudged_min, &nudged_max);
auto wiseMax = LAMBDA_T(x, nudged_min) { auto wiseMinMaxAndSoOn = LAMBDA_T(x, nudged_min, nudged_max, scale) {
T val = x;
if (x < nudged_min) { if (x < nudged_min) {
return nudged_min; val = nudged_min;
} }
return x; else if (x > nudged_max) {
val = nudged_max;
}
else
val = x;
return (math::nd4j_floor<T,T>((val - nudged_min) / scale + T(0.5)) * scale + nudged_min);
}; };
auto wiseMin = LAMBDA_T(x, nudged_max) { input->applyLambda(wiseMinMaxAndSoOn, output);
if (x > nudged_max) {
return nudged_max;
}
return x;
};
auto scaleTensor(*input);
auto clamped(*input);
scaleTensor.assign(scale);
input->applyLambda(wiseMin, &clamped);
clamped.applyLambda(wiseMax, output);
*output -= nudged_min;
(*output) /= scaleTensor;
(*output) += T(0.5f);
output->applyTransform(transform::Floor, nullptr, nullptr);
(*output) *= scaleTensor;
(*output) += nudged_min;
} }
void fakeQuantWithMinMaxVars(NDArray* input, NDArray* min, NDArray* max, int numBits, bool narrowed, NDArray* output) { void fakeQuantWithMinMaxVars(NDArray* input, NDArray* min, NDArray* max, int numBits, bool narrowed, NDArray* output) {