cavis/libnd4j/include/ops/impl/specials.cpp

397 lines
12 KiB
C++
Raw Normal View History

2019-06-06 14:21:15 +02:00
/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
*
* This program and the accompanying materials are made available under the
* terms of the Apache License, Version 2.0 which is available at
* https://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.
*
* SPDX-License-Identifier: Apache-2.0
******************************************************************************/
//
// @author raver119@gmail.com, created on 07.10.2017.
// @author Yurii Shyrma (iuriish@yahoo.com)
//
#include <pointercast.h>
#include <helpers/shape.h>
#include <helpers/TAD.h>
#include <specials.h>
#include <dll.h>
#include <NDArray.h>
#include <ops/declarable/CustomOperations.h>
#include <types/types.h>
namespace nd4j {
/**
* Concatneate multi array of the same shape together
* along a particular dimension
*/
template <typename T>
void SpecialMethods<T>::concatCpuGeneric(int dimension, int numArrays, Nd4jPointer *data, Nd4jPointer *inputShapeInfo, void *vresult, Nd4jLong *resultShapeInfo) {
auto result = reinterpret_cast<T *>(vresult);
std::vector<Nd4jLong> iArgs = {dimension};
std::vector<double> tArgs;
std::vector<bool> bArgsEmpty;
std::vector<NDArray*> inputs(numArrays);
std::vector<NDArray*> outputs(1);
outputs[0] = new NDArray(static_cast<void*>(result), static_cast<Nd4jLong*>(resultShapeInfo));
for(int i = 0; i < numArrays; ++i)
inputs[i] = new NDArray(static_cast<void *>(data[i]), static_cast<Nd4jLong*>(inputShapeInfo[i]));
nd4j::ops::concat op;
auto status = op.execute(inputs, outputs, tArgs, iArgs, bArgsEmpty);
if(status != Status::OK())
throw std::runtime_error("concatCpuGeneric fails to be executed !");
delete outputs[0];
for(int i = 0; i < numArrays; ++i)
delete inputs[i];
}
/**
* This kernel accumulates X arrays, and stores result into Z
*
* @tparam T
* @param x
* @param z
* @param n
* @param length
*/
template<typename T>
void SpecialMethods<T>::accumulateGeneric(void **vx, void *vz, Nd4jLong *zShapeInfo, int n, const Nd4jLong length) {
auto z = reinterpret_cast<T *>(vz);
auto x = reinterpret_cast<T **>(vx);
// aggregation step
#ifdef _OPENMP
int _threads = omp_get_max_threads();
#else
// we can use whatever we want here, this value won't be used if there's no omp
int _threads = 4;
#endif
PRAGMA_OMP_PARALLEL_FOR_SIMD
for (Nd4jLong i = 0; i < length; i++) {
for (Nd4jLong ar = 0; ar < n; ar++) {
z[i] += x[ar][i];
}
}
}
/**
* This kernel averages X input arrays, and stores result to Z
*
* @tparam T
* @param x
* @param z
* @param n
* @param length
* @param propagate
*/
template<typename T>
void SpecialMethods<T>::averageGeneric(void **vx, void *vz, Nd4jLong *zShapeInfo, int n, const Nd4jLong length, bool propagate) {
auto z = reinterpret_cast<T *>(vz);
auto x = reinterpret_cast<T **>(vx);
if (z == nullptr) {
//code branch for absent Z
z = x[0];
PRAGMA_OMP_SIMD
for (Nd4jLong i = 0; i < length; i++) {
z[i] /= n;
}
#ifdef _OPENNMP
int _threads = omp_get_max_threads(); //nd4j::math::nd4j_min<int>(omp_get_max_threads() / 2, 4);
#else
// we can use whatever we want here, this value won't be used if there's no omp
int _threads = 4;
#endif
PRAGMA_OMP_PARALLEL_FOR_SIMD
for (Nd4jLong i = 0; i < length; i++) {
for (Nd4jLong ar = 1; ar < n; ar++) {
z[i] += x[ar][i] / n;
}
}
// instead of doing element-wise propagation, we just issue memcpy to propagate data
for (Nd4jLong ar = 1; ar < n; ar++) {
memcpy(x[ar], z, length * sizeof(T));
}
} else {
// code branch for existing Z
// memset before propagation
memset(z, 0, length * sizeof(T));
// aggregation step
#ifdef _OPENNMP
int _threads = omp_get_max_threads(); //nd4j::math::nd4j_min<int>(omp_get_max_threads() / 2, 4);
#else
// we can use whatever we want here, this value won't be used if there's no omp
int _threads = 4;
#endif
PRAGMA_OMP_PARALLEL_FOR_SIMD
for (Nd4jLong i = 0; i < length; i++) {
for (Nd4jLong ar = 0; ar < n; ar++) {
z[i] += x[ar][i] / n;
}
}
// instead of doing element-wise propagation, we just issue memcpy to propagate data
for (Nd4jLong ar = 0; ar < n; ar++) {
memcpy(x[ar], z, length * sizeof(T));
}
}
}
template <typename T>
Nd4jLong SpecialMethods<T>::getPosition(Nd4jLong *xShapeInfo, Nd4jLong index) {
auto xEWS = shape::elementWiseStride(xShapeInfo);
if (xEWS == 1)
return index;
else if (xEWS > 1)
return index * xEWS;
else
return shape::getIndexOffset(index, xShapeInfo, shape::length(xShapeInfo));
}
template<typename T>
void SpecialMethods<T>::quickSort_parallel_internal(T* array, Nd4jLong *xShapeInfo, int left, int right, int cutoff, bool descending) {
int i = left, j = right;
T tmp;
T pivot = array[getPosition(xShapeInfo, (left + right) / 2)];
{
/* PARTITION PART */
while (i <= j) {
if (descending) {
while (array[getPosition(xShapeInfo, i)] > pivot)
i++;
while (array[getPosition(xShapeInfo, j)] < pivot)
j--;
if (i <= j) {
tmp = array[getPosition(xShapeInfo, i)];
array[getPosition(xShapeInfo, i)] = array[getPosition(xShapeInfo, j)];
array[getPosition(xShapeInfo, j)] = tmp;
i++;
j--;
}
} else {
while (array[getPosition(xShapeInfo, i)] < pivot)
i++;
while (array[getPosition(xShapeInfo, j)] > pivot)
j--;
if (i <= j) {
tmp = array[getPosition(xShapeInfo, i)];
array[getPosition(xShapeInfo, i)] = array[getPosition(xShapeInfo, j)];
array[getPosition(xShapeInfo, j)] = tmp;
i++;
j--;
}
}
}
}
//
if ( ((right-left)<cutoff) ){
if (left < j){ quickSort_parallel_internal(array, xShapeInfo, left, j, cutoff, descending); }
if (i < right){ quickSort_parallel_internal(array, xShapeInfo, i, right, cutoff, descending); }
}else{
#pragma omp task
{ quickSort_parallel_internal(array, xShapeInfo, left, j, cutoff, descending); }
#pragma omp task
{ quickSort_parallel_internal(array, xShapeInfo, i, right, cutoff, descending); }
}
}
template<typename T>
void SpecialMethods<T>::quickSort_parallel(void *varray, Nd4jLong *xShapeInfo, Nd4jLong lenArray, int numThreads, bool descending){
auto array = reinterpret_cast<T *>(varray);
int cutoff = 1000;
PRAGMA_OMP_PARALLEL_THREADS(numThreads)
{
#pragma omp single nowait
{
quickSort_parallel_internal(array, xShapeInfo, 0, lenArray-1, cutoff, descending);
}
}
}
template <typename T>
int SpecialMethods<T>::nextPowerOf2(int number) {
int pos = 0;
while (number > 0) {
pos++;
number = number >> 1;
}
return (int) pow(2, pos);
}
template <typename T>
int SpecialMethods<T>::lastPowerOf2(int number) {
int p = 1;
while (p <= number)
p <<= 1;
p >>= 1;
return p;
}
template<typename T>
void SpecialMethods<T>::sortGeneric(void *vx, Nd4jLong *xShapeInfo, bool descending) {
auto x = reinterpret_cast<T *>(vx);
quickSort_parallel(x, xShapeInfo, shape::length(xShapeInfo), omp_get_max_threads(), descending);
}
template<typename T>
void SpecialMethods<T>::sortTadGeneric(void *vx, Nd4jLong *xShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, bool descending) {
auto x = reinterpret_cast<T *>(vx);
//quickSort_parallel(x, xShapeInfo, shape::length(xShapeInfo), omp_get_max_threads(), descending);
Nd4jLong xLength = shape::length(xShapeInfo);
Nd4jLong xTadLength = shape::tadLength(xShapeInfo, dimension, dimensionLength);
int numTads = xLength / xTadLength;
PRAGMA_OMP_PARALLEL_FOR
for (int r = 0; r < numTads; r++) {
T *dx = x + tadOffsets[r];
quickSort_parallel(dx, tadShapeInfo, xTadLength, 1, descending);
}
}
template<typename T>
void SpecialMethods<T>::decodeBitmapGeneric(void *dx, Nd4jLong N, void *vz, Nd4jLong *zShapeInfo) {
auto dz = reinterpret_cast<T *>(vz);
auto x = reinterpret_cast<int *>(dx);
Nd4jLong lim = N / 16 + 5;
FloatBits2 fb;
fb.i_ = x[2];
float threshold = fb.f_;
PRAGMA_OMP_PARALLEL_FOR
for (Nd4jLong e = 4; e < lim; e++) {
for (int bitId = 0; bitId < 16; bitId++) {
bool hasBit = (x[e] & 1 << (bitId) ) != 0;
bool hasSign = (x[e] & 1 << (bitId + 16) ) != 0;
if (hasBit) {
if (hasSign)
dz[(e - 4) * 16 + bitId] -= threshold;
else
dz[(e - 4) * 16 + bitId] += threshold;
} else if (hasSign) {
dz[(e - 4) * 16 + bitId] -= threshold / 2;
}
}
}
}
template<typename S, typename T>
void SpecialTypeConverter::convertGeneric(Nd4jPointer * extras, void *dx, Nd4jLong N, void *dz) {
auto x = reinterpret_cast<S *>(dx);
auto z = reinterpret_cast<T *>(dz);
if (N < nd4j::Environment::getInstance()->elementwiseThreshold()) {
for (int i = 0; i < N; i++) {
z[i] = static_cast<T>(x[i]);
}
} else {
PRAGMA_OMP_PARALLEL_FOR
for (int i = 0; i < N; i++) {
z[i] = static_cast<T>(x[i]);
}
}
};
BUILD_DOUBLE_TEMPLATE(template void SpecialTypeConverter::convertGeneric, (Nd4jPointer * extras, void *dx, Nd4jLong N, void *dz), LIBND4J_TYPES, LIBND4J_TYPES);
template<typename T>
Nd4jLong SpecialMethods<T>::encodeBitmapGeneric(void *vx, Nd4jLong *xShapeInfo, Nd4jLong N, int *dz, float threshold) {
auto dx = reinterpret_cast<T *>(vx);
Nd4jLong retVal = 0L;
#pragma omp parallel for schedule(guided) proc_bind(close) reduction(+:retVal)
for (Nd4jLong x = 0; x < N; x += 16) {
int byte = 0;
int byteId = x / 16 + 4;
for (int f = 0; f < 16; f++) {
Nd4jLong e = x + f;
if (e >= N)
continue;
T val = dx[e];
T abs = nd4j::math::nd4j_abs<T>(val);
int bitId = e % 16;
if (abs >= (T) threshold) {
byte |= 1 << (bitId);
retVal++;
if (val < (T) 0.0f) {
byte |= 1 << (bitId + 16);
dx[e] += threshold;
} else {
dx[e] -= threshold;
}
} else if (abs >= (T) threshold / (T) 2.0f && val < (T) 0.0f) {
byte |= 1 << (bitId + 16);
dx[e] += threshold / 2;
retVal++;
}
}
dz[byteId] = byte;
}
return retVal;
}
BUILD_SINGLE_TEMPLATE(template class SpecialMethods, , LIBND4J_TYPES);
}