Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactored helpers both for cuda and cpu platforms.

master
shugeo 2019-10-10 18:02:49 +03:00
parent a09cb5e2be
commit c3f755d975
2 changed files with 48 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
libnd4j/include/ops/declarable/helpers/cpu/fake_quantization.cpp
View File

@ -26,7 +26,7 @@ namespace ops {
namespace helpers { namespace helpers {
template <typename T> template <typename T>
static void Nudge(T min, T max, int quant_min, int quant_max, T* scale, T* nudged_min, T* nudged_max) { 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_max_float = static_cast<T>(quant_max);
T quant_min_float = static_cast<T>(quant_min); T quant_min_float = static_cast<T>(quant_min);
*scale = (max - min) / (quant_max_float - quant_min_float); *scale = (max - min) / (quant_max_float - quant_min_float);
@ -53,7 +53,7 @@ namespace helpers {
PRAGMA_OMP_PARALLEL_FOR PRAGMA_OMP_PARALLEL_FOR
for (auto i = 0; i < channels; i++) { for (auto i = 0; i < channels; i++) {
T scale, nudged_min, nudged_max; T scale, nudged_min, nudged_max;
Nudge<T>(min->t<T>(i), max->t<T>(i), lowIntBound, upperIntBound, &scale, &nudged_min, &nudged_max); nudge<T>(min->t<T>(i), max->t<T>(i), lowIntBound, upperIntBound, &scale, &nudged_min, &nudged_max);
for (auto e = 0; e < input->lengthOf(); e += channels) { for (auto e = 0; e < input->lengthOf(); e += channels) {
T val = input->t<T>(e + i); T val = input->t<T>(e + i);
@ -67,37 +67,26 @@ namespace helpers {
} }
} }
template <typename T>
static void WiseMinMax(NDArray* input, T min, T max, NDArray* output) {
auto wiseMinMax = LAMBDA_T(x, min, max) {
if (x < min) {
return min;
}
else if (x > max)
return max;
return x;
};
input->applyLambda<T>(wiseMinMax, output);
}
template <typename T> template <typename T>
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) {
int lowIntBound = narrowed ? 1 : 0; int lowIntBound = narrowed ? 1 : 0;
int upperIntBound = (1 << numBits) - 1; int upperIntBound = (1 << numBits) - 1;
const float quant_min_float = static_cast<float>(lowIntBound); T nudgedMin, nudgedMax, scale;
const float quant_max_float = static_cast<float>(upperIntBound);
T nudged_min, nudged_max, scale;
Nudge<T>(min->t<T>(0), max->t<T>(0), quant_min_float, quant_max_float, &scale, &nudged_min, &nudged_max); nudge<T>(min->t<T>(0), max->t<T>(0), lowIntBound, upperIntBound, &scale, &nudgedMin, &nudgedMax);
WiseMinMax<T>(input, nudged_min, nudged_max, output);
*output -= nudged_min; auto fakeQuantizationWithMinMax = LAMBDA_T(x, nudgedMin, nudgedMax, scale) {
(*output) /= scale; T val = x;
(*output) += T(0.5f); if (val < nudgedMin) {
output->applyTransform(transform::Floor, nullptr, nullptr); val = nudgedMin;
(*output) *= scale; }
(*output) += nudged_min; else if (val > nudgedMax)
val = nudgedMax;
return (nd4j::math::nd4j_floor<T,T>((val - nudgedMin)/scale + T(0.5)) * scale + nudgedMin);
};
input->applyLambda<T>(fakeQuantizationWithMinMax, output);
} }
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) {

63
libnd4j/include/ops/declarable/helpers/cuda/fake_quantization.cu
View File

@ -34,44 +34,45 @@ namespace helpers {
// output - output tensor // output - output tensor
// //
template <typename T> template <typename T>
static __host__ __device__ void Nudge(T min, T max, int quant_min, int quant_max, T* scale, T* nudged_min, T* nudged_max) { static __host__ __device__ void
T quant_max_float = static_cast<T>(quant_max); nudge(T min, T max, int quantMin, int quantMax, T* scale, T* nudgedMin, T* nudgedMax) {
T quant_min_float = static_cast<T>(quant_min); T quantMaxF = static_cast<T>(quantMax);
*scale = (max - min) / (quant_max_float - quant_min_float); T quantMinF = static_cast<T>(quantMin);
auto zero_point_from_min = quant_min_float - min / *scale; *scale = (max - min) / (quantMaxF - quantMinF);
uint16_t const nudged_zero_point = [zero_point_from_min, quant_min, quant_max, quant_max_float, quant_min_float] { auto zeroPointFromMin = quantMinF - min / *scale;
if (zero_point_from_min < quant_min_float) { uint16_t const nudgedZeroPoint = [zeroPointFromMin, quantMin, quantMax, quantMaxF, quantMinF] {
return static_cast<uint16_t>(quant_min); if (zeroPointFromMin < quantMinF) {
return static_cast<uint16_t>(quantMin);
} }
if (zero_point_from_min > quant_max_float) { if (zeroPointFromMin > quantMaxF) {
return static_cast<uint16_t>(quant_max); 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); *nudgedMin = (quantMinF - nudgedZeroPoint) * (*scale);
*nudged_max = (quant_max_float - nudged_zero_point) * (*scale); *nudgedMax = (quantMaxF - nudgedZeroPoint) * (*scale);
} }
template <typename T> template <typename T>
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) {
int lowIntBound = narrowed?1:0; int lowIntBound = narrowed?1:0;
int upperIntBound = (1 << numBits) - 1; int upperIntBound = (1 << numBits) - 1;
min->syncToHost(); min->syncToHost(); // these are scalars, so nothing much happened
max->syncToHost(); max->syncToHost();
T scale, nudged_min, nudged_max; T scale, nudgedMin, nudgedMax;
Nudge(min->t<T>(0), max->t<T>(0), lowIntBound, upperIntBound, &scale, &nudged_min, &nudged_max); nudge(min->t<T>(0), max->t<T>(0), lowIntBound, upperIntBound, &scale, &nudgedMin, &nudgedMax);
auto wiseMinMaxAndSoOn = LAMBDA_T(x, nudged_min, nudged_max, scale) { auto wiseMinMaxAndSoOn = LAMBDA_T(x, nudgedMin, nudgedMax, scale) {
T val = x; T val = x;
if (x < nudged_min) { if (x < nudgedMin) {
val = nudged_min; val = nudgedMin;
} }
else if (x > nudged_max) { else if (x > nudgedMax) {
val = nudged_max; val = nudgedMax;
} }
else else
val = x; val = x;
return (math::nd4j_floor<T,T>((val - nudged_min) / scale + T(0.5)) * scale + nudged_min); return (math::nd4j_floor<T,T>((val - nudgedMin) / scale + T(0.5)) * scale + nudgedMin);
}; };
input->applyLambda(wiseMinMaxAndSoOn, output); input->applyLambda(wiseMinMaxAndSoOn, output);
@ -88,20 +89,20 @@ namespace helpers {
__syncthreads(); __syncthreads();
for (auto i = blockIdx.x; i < (int)channels; i += gridDim.x) { for (auto i = blockIdx.x; i < (int)channels; i += gridDim.x) {
T scale, nudged_min, nudged_max; T scale, nudgedMin, nudgedMax;
Nudge(min[i], max[i], lowIntBound, upperIntBound, &scale, &nudged_min, &nudged_max); nudge(min[i], max[i], lowIntBound, upperIntBound, &scale, &nudgedMin, &nudgedMax);
//auto wiseMinMaxAndSoOn = LAMBDA_T(x, nudged_min, nudged_max, scale) {
for (auto e = threadIdx.x; e < block; e += blockDim.x) { for (auto e = threadIdx.x; e < block; e += blockDim.x) {
T val = input[shape::getIndexOffset(e * channels + i, inputShape)]; T val = input[shape::getIndexOffset(e * channels + i, inputShape)];
if (val < nudged_min) { if (val < nudgedMin) {
val = nudged_min; val = nudgedMin;
} else if (val > nudged_max) { } else if (val > nudgedMax) {
val = nudged_max; val = nudgedMax;
} }
output[shape::getIndexOffset(e* channels + i, outputShape)] = (math::nd4j_floor<T, T>((val - nudged_min) / scale + T(0.5)) * scale + nudged_min); output[shape::getIndexOffset(e* channels + i, outputShape)] =
(math::nd4j_floor<T, T>((val - nudgedMin) / scale + T(0.5)) * scale + nudgedMin);
}; };
} }
} }
template <typename T> template <typename T>