1250 lines
55 KiB
C++
1250 lines
55 KiB
C++
/*******************************************************************************
|
|
* 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 Yurii Shyrma (iuriish@yahoo.com), created on 14.03.2019
|
|
//
|
|
|
|
#ifndef LIBND4J_LOOPS_H
|
|
#define LIBND4J_LOOPS_H
|
|
|
|
#include <functional>
|
|
#include <system/pointercast.h>
|
|
#include <helpers/shape.h>
|
|
#include <helpers/LoopKind.h>
|
|
#include <helpers/OmpLaunchHelper.h>
|
|
#include <array/DataTypeUtils.h>
|
|
#include <ops/ops.h>
|
|
#include <loops/indexreduce.h>
|
|
#include <helpers/ConstantTadHelper.h>
|
|
#include <system/openmp_pragmas.h>
|
|
#include <execution/Threads.h>
|
|
|
|
namespace sd {
|
|
|
|
template <typename X, typename Z, typename E>
|
|
class ND4J_EXPORT ReductionLoops {
|
|
protected:
|
|
public:
|
|
|
|
template <typename OpType>
|
|
static FORCEINLINE void loopReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, E* extraParams, int64_t start, int64_t stop);
|
|
};
|
|
|
|
template <typename X, typename Z>
|
|
class ReductionFloatLoops : public ReductionLoops<X, Z, Z> {
|
|
public:
|
|
static void wrapper(const int opNum, X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, Z* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static void innerloopReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, Z* extraParams, int64_t start, int64_t stop);
|
|
};
|
|
|
|
template <typename X, typename Z>
|
|
class ND4J_EXPORT ReductionBoolLoops : public ReductionLoops<X, Z, X> {
|
|
public:
|
|
static void wrapper(const int opNum, X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static void innerloopReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop);
|
|
};
|
|
|
|
template <typename X, typename Z>
|
|
class ND4J_EXPORT ReductionLongLoops : public ReductionLoops<X, Z, X> {
|
|
public:
|
|
static void wrapper(const int opNum, X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static void innerloopReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop);
|
|
};
|
|
|
|
template <typename X>
|
|
class ND4J_EXPORT ReductionSameLoops : public ReductionLoops<X, X, X> {
|
|
public:
|
|
static void wrapper(const int opNum, X* x, Nd4jLong* xShapeInfo, X* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static void innerloopReduce(X* x, Nd4jLong* xShapeInfo, X* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams, int64_t start, int64_t stop);
|
|
};
|
|
|
|
|
|
template <typename X, typename Z>
|
|
class ND4J_EXPORT IndexReductionLoops {
|
|
private:
|
|
public:
|
|
static void wrapIndexReduce(const int opNum, void* x, Nd4jLong* xShapeInfo, void* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* extraParams);
|
|
|
|
template <typename OpType>
|
|
static void loopIndexReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams);
|
|
};
|
|
|
|
|
|
template <typename X, typename Z, typename E>
|
|
class ND4J_EXPORT TransformLoops {
|
|
|
|
public:
|
|
|
|
template<typename OpType>
|
|
static FORCEINLINE void loopTransform(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, E* extraParams, uint64_t threadId, uint64_t numThreads);
|
|
};
|
|
|
|
template <typename X, typename Z>
|
|
class ND4J_EXPORT Reduction3Loops {
|
|
public:
|
|
|
|
template <typename OpType>
|
|
static FORCEINLINE void loopReduce3(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static FORCEINLINE void loopReduce3All(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams, int64_t start, int64_t stop);
|
|
|
|
static void wrapper(const int opNum, X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams, int64_t start, int64_t stop);
|
|
|
|
static void wrapperAll(const int opNum, X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static void innerloopReduce3(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, int* dims, int dimsLen, Z* extraParams, int64_t start, int64_t stop);
|
|
|
|
template <typename OpType>
|
|
static void innerloopReduce3All(X* x, Nd4jLong* xShapeInfo, X* y, Nd4jLong* yShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets, Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets, Z* extraParams, int64_t start, int64_t stop);
|
|
};
|
|
|
|
|
|
|
|
|
|
/*
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Y, typename Z>
|
|
void Loops::loopXYZ(const X* x, const Nd4jLong* xShapeInfo,
|
|
const Y* y, const Nd4jLong* yShapeInfo,
|
|
Z* z, const Nd4jLong* zShapeInfo,
|
|
Z* extraParams,
|
|
std::function<Z(X,Y,Z*)> op) {
|
|
|
|
const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopXYZ(xShapeInfo, yShapeInfo, zShapeInfo);
|
|
|
|
const Nd4jLong* xShape = shape::shapeOf(xShapeInfo);
|
|
const Nd4jLong* xStride = shape::stride(xShapeInfo);
|
|
const Nd4jLong* yStride = shape::stride(yShapeInfo);
|
|
const Nd4jLong* zStride = shape::stride(zShapeInfo);
|
|
|
|
const Nd4jLong len = shape::length(xShapeInfo);
|
|
|
|
OmpLaunchHelper threadsInfo(len);
|
|
|
|
switch (kindOfLoop) {
|
|
|
|
case LoopKind::EWS1: {
|
|
PRAGMA_OMP_PARALLEL_THREADS(threadsInfo._numThreads)
|
|
{
|
|
const auto threadNum = omp_get_thread_num();
|
|
const auto threadOffset = threadsInfo.getThreadOffset(threadNum);
|
|
const auto lenPerThread = static_cast<uint>(threadsInfo.getItersPerThread(threadNum));
|
|
|
|
const auto xi = x + threadOffset;
|
|
const auto yi = y + threadOffset;
|
|
auto zi = z + threadOffset;
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (uint i = 0; i < lenPerThread; i++)
|
|
zi[i] = op(xi[i], yi[i], extraParams);
|
|
}
|
|
}
|
|
break;
|
|
|
|
case LoopKind::EWSNONZERO: {
|
|
const uint xEws = shape::elementWiseStride(xShapeInfo);
|
|
const uint yEws = shape::elementWiseStride(yShapeInfo);
|
|
const uint zEws = shape::elementWiseStride(zShapeInfo);
|
|
|
|
PRAGMA_OMP_PARALLEL_THREADS(threadsInfo._numThreads)
|
|
{
|
|
const auto threadNum = omp_get_thread_num();
|
|
const auto threadOffset = threadsInfo.getThreadOffset(threadNum);
|
|
const auto lenPerThread = static_cast<uint>(threadsInfo.getItersPerThread(threadNum));
|
|
const auto xi = x + threadOffset * xEws;
|
|
const auto yi = y + threadOffset * yEws;
|
|
auto zi = z + threadOffset * zEws;
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (uint i = 0; i < lenPerThread; i++)
|
|
zi[i*zEws] = op(xi[i*xEws], yi[i*yEws], extraParams);
|
|
}
|
|
}
|
|
break;
|
|
|
|
case LoopKind::RANK1: {
|
|
PRAGMA_OMP_PARALLEL_FOR
|
|
for (uint i0 = 0; i0 < len; ++i0)
|
|
z[i0 * zStride[0]] = op(x[i0 * xStride[0]], y[i0 * yStride[0]], extraParams);
|
|
}
|
|
break;
|
|
|
|
case LoopKind::RANK2: {
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD
|
|
for (uint i0 = 0; i0 < xShape[0]; ++i0)
|
|
for (uint i1 = 0; i1 < xShape[1]; ++i1)
|
|
z[i0 * zStride[0] + i1 * zStride[1]] = op(x[i0 * xStride[0] + i1 * xStride[1]], y[i0 * yStride[0] + i1 * yStride[1]], extraParams);
|
|
}
|
|
break;
|
|
|
|
case LoopKind::RANK3: {
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD_COLLAPSE(2)
|
|
for (uint i0 = 0; i0 < xShape[0]; ++i0)
|
|
for (uint i1 = 0; i1 < xShape[1]; ++i1)
|
|
for (uint i2 = 0; i2 < xShape[2]; ++i2)
|
|
z[i0*zStride[0]+i1*zStride[1]+i2*zStride[2]] = op(x[i0*xStride[0]+i1*xStride[1]+i2*xStride[2]], y[i0*yStride[0]+i1*yStride[1]+i2*yStride[2]], extraParams);
|
|
}
|
|
break;
|
|
|
|
case LoopKind::RANK4: {
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD_COLLAPSE(3)
|
|
for (uint i0 = 0; i0 < xShape[0]; ++i0)
|
|
for (uint i1 = 0; i1 < xShape[1]; ++i1)
|
|
for (uint i2 = 0; i2 < xShape[2]; ++i2)
|
|
for (uint i3 = 0; i3 < xShape[3]; ++i3)
|
|
z[i0*zStride[0]+i1*zStride[1]+i2*zStride[2]+i3*zStride[3]] = op(x[i0*xStride[0]+i1*xStride[1]+i2*xStride[2]+i3*xStride[3]], y[i0*yStride[0]+i1*yStride[1]+i2*yStride[2]+i3*yStride[3]], extraParams);
|
|
}
|
|
break;
|
|
|
|
case LoopKind::RANK5: {
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD_COLLAPSE(4)
|
|
for (uint i0 = 0; i0 < xShape[0]; ++i0)
|
|
for (uint i1 = 0; i1 < xShape[1]; ++i1)
|
|
for (uint i2 = 0; i2 < xShape[2]; ++i2)
|
|
for (uint i3 = 0; i3 < xShape[3]; ++i3)
|
|
for (uint i4 = 0; i4 < xShape[4]; ++i4)
|
|
z[i0*zStride[0]+i1*zStride[1]+i2*zStride[2]+i3*zStride[3]+i4*zStride[4]] = op(x[i0*xStride[0]+i1*xStride[1]+i2*xStride[2]+i3*xStride[3]+i4*xStride[4]], y[i0*yStride[0]+i1*yStride[1]+i2*yStride[2]+i3*yStride[3]+i4*yStride[4]], extraParams);
|
|
}
|
|
break;
|
|
|
|
default: {
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
uint yShapeInfoCast[MAX_RANK];
|
|
uint zShapeInfoCast[MAX_RANK];
|
|
|
|
bool canCastX = DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastY = DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
|
|
bool canCastZ = DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
|
|
|
|
PRAGMA_OMP_PARALLEL_THREADS(threadsInfo._numThreads)
|
|
{
|
|
auto threadNum = omp_get_thread_num();
|
|
auto threadOffset = threadsInfo.getThreadOffset(threadNum);
|
|
auto lenPerThread = static_cast<uint>(threadsInfo.getItersPerThread(threadNum));
|
|
PRAGMA_OMP_SIMD
|
|
for (uint i = 0; i < lenPerThread; i++) {
|
|
auto xOffset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, canCastX);
|
|
auto yOffset = shape::indexOffset(i + threadOffset, yShapeInfo, yShapeInfoCast, canCastY);
|
|
auto zOffset = shape::indexOffset(i + threadOffset, zShapeInfo, zShapeInfoCast, canCastZ);
|
|
z[zOffset] = op(x[xOffset], y[yOffset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
*/
|
|
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Z, typename E>
|
|
template <typename OpType>
|
|
void sd::ReductionLoops<X, Z, E>::loopReduce(X* x, Nd4jLong* xShapeInfo,
|
|
Z* z, Nd4jLong* zShapeInfo,
|
|
Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets,
|
|
E* extraParams, int64_t start, int64_t stop) {
|
|
|
|
const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopTadXZ(xShapeInfo, zShapeInfo, tadShapeInfo);
|
|
|
|
const Nd4jLong zLen = shape::length(zShapeInfo);
|
|
const Nd4jLong tadLen = shape::length(tadShapeInfo);
|
|
|
|
const uint tadEws = shape::elementWiseStride(tadShapeInfo);
|
|
const uint zEws = shape::elementWiseStride(zShapeInfo);
|
|
|
|
const Nd4jLong* tadShape = shape::shapeOf(tadShapeInfo);
|
|
const Nd4jLong* tadStride = shape::stride(tadShapeInfo);
|
|
|
|
int numThreads = OmpLaunchHelper::tadThreads(tadLen, zLen);
|
|
|
|
switch (kindOfLoop) {
|
|
|
|
//*********************************************//
|
|
// case LoopKind::SMALLARR2DX: {
|
|
// shape::printShapeInfoLinear(xShapeInfo);
|
|
// shape::printShapeInfoLinear(zShapeInfo);
|
|
// const auto xLen = zLen * tadLen;
|
|
// for (uint i = 0; i < xLen; ++i) {
|
|
// const auto zOffset = shape::subArrayOffset(i, xShapeInfo, zShapeInfo, dimsToExclude, dimsLen);
|
|
// const uint tadInd = (i / tadEws) % tadLen;
|
|
// auto startVal = tadInd ? z[zOffset] : static_cast<Z>(OpType::startingValue(x));
|
|
// z[zOffset] = OpType::update(startVal, OpType::op(x[i], extraParams), extraParams);
|
|
// if(tadInd == tadLen - 1)
|
|
// z[zOffset] = OpType::postProcess(z[zOffset], tadLen, extraParams);
|
|
// printf("%u - %lld\n", i, zOffset);
|
|
// }
|
|
// }
|
|
case LoopKind::SMALLARR2DX: {
|
|
const auto uTadLen = static_cast<uint>(tadLen);
|
|
const auto uZLenMinusOne = static_cast<uint>(zLen - 1);
|
|
const auto xLen = static_cast<uint>(zLen * uTadLen);
|
|
const auto sv = static_cast<Z>(OpType::startingValue(x));
|
|
|
|
for (uint i = 0; i <= uZLenMinusOne; i++)
|
|
z[i] = OpType::startingValue(x);
|
|
|
|
uint zOffset = 0;
|
|
for (uint i = 0; i < xLen; ++i) {
|
|
z[zOffset] = OpType::update(z[zOffset], OpType::op(x[i], extraParams), extraParams);
|
|
zOffset = zOffset == uZLenMinusOne ? 0 : zOffset + 1;
|
|
}
|
|
|
|
for (uint i = 0; i <= uZLenMinusOne; i++)
|
|
z[i] = OpType::postProcess(z[i], tadLen, extraParams);
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWS1: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; j++)
|
|
s = OpType::update(s, OpType::op(tad[j], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWSNONZERO: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; j++)
|
|
s = OpType::update(s, OpType::op(tad[j * tadEws], extraParams), extraParams);
|
|
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK1: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadLen; ++i0)
|
|
s = OpType::update(s, OpType::op(tad[i0 * tadStride[0]], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK2: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0)
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1)
|
|
s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1]], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK3: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0)
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1)
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2)
|
|
s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2]], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK4: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0)
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1)
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2)
|
|
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3)
|
|
s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2] + i3 * tadStride[3]], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK5: {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0)
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1)
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2)
|
|
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3)
|
|
for (Nd4jLong i4 = 0; i4 < tadShape[4]; ++i4)
|
|
s = OpType::update(s, OpType::op(tad[i0 * tadStride[0] + i1 * tadStride[1] + i2 * tadStride[2] + i3 * tadStride[3] + i4 * tadStride[4]], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::X_EWSNONZERO: {
|
|
uint castZShapeInfo[MAX_RANK];
|
|
const bool canCastZ = sd::DataTypeUtils::castShapeInfo<uint>(zShapeInfo, castZShapeInfo);
|
|
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; j++)
|
|
s = OpType::update(s, OpType::op(tad[j * tadEws], extraParams), extraParams);
|
|
|
|
auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, canCastZ);
|
|
z[zOffset] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::Z_EWSNONZERO: {
|
|
uint castTadShapeInfo[MAX_RANK];
|
|
const bool canCastTad = sd::DataTypeUtils::castShapeInfo<uint>(tadShapeInfo, castTadShapeInfo);
|
|
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; j++) {
|
|
auto tadOffset = shape::indexOffset(j, tadShapeInfo, castTadShapeInfo, canCastTad);
|
|
s = OpType::update(s, OpType::op(tad[tadOffset], extraParams), extraParams);
|
|
}
|
|
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
default: {
|
|
auto innertadOffsets = new Nd4jLong[tadLen];
|
|
shape::calcOffsets(tadShapeInfo, innertadOffsets);
|
|
|
|
uint castZShapeInfo[MAX_RANK];
|
|
const bool canCastZ = sd::DataTypeUtils::castShapeInfo<uint>(zShapeInfo, castZShapeInfo);
|
|
|
|
for (auto i = start; i < stop; i++) {
|
|
auto tad = x + tadOffsets[i];
|
|
auto s = OpType::startingValue(tad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; j++)
|
|
s = OpType::update(s, OpType::op(tad[innertadOffsets[j]], extraParams), extraParams);
|
|
|
|
auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, canCastZ);
|
|
z[zOffset] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
|
|
delete[] innertadOffsets;
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Z, typename E>
|
|
template <typename OpType>
|
|
void sd::TransformLoops<X, Z, E>::loopTransform(X* x, Nd4jLong* xShapeInfo,
|
|
Z* z, Nd4jLong* zShapeInfo,
|
|
E* extraParams, uint64_t threadId, uint64_t numThreads) {
|
|
|
|
const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopXZ(xShapeInfo, zShapeInfo);
|
|
|
|
const Nd4jLong* xShape = shape::shapeOf(const_cast<Nd4jLong*>(xShapeInfo));
|
|
const Nd4jLong* xStride = shape::stride(const_cast<Nd4jLong*>(xShapeInfo));
|
|
const Nd4jLong* zStride = shape::stride(const_cast<Nd4jLong*>(zShapeInfo));
|
|
|
|
const Nd4jLong len = shape::length(xShapeInfo);
|
|
|
|
if (len == 0)
|
|
return;
|
|
|
|
switch (kindOfLoop) {
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWS1: {
|
|
auto span = samediff::Span::build(threadId, numThreads, 0, len, 1);
|
|
int64_t start = span.startX(), stop = span.stopX();
|
|
|
|
for (auto i = start; i < stop; i++)
|
|
z[i] = OpType::op(x[i], extraParams);
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWSNONZERO: {
|
|
const uint xEws = shape::elementWiseStride(xShapeInfo);
|
|
const uint zEws = shape::elementWiseStride(zShapeInfo);
|
|
|
|
auto span = samediff::Span::build(threadId, numThreads, 0, len, 1);
|
|
int64_t start = span.startX(), stop = span.stopX();
|
|
|
|
for (auto i = start; i < stop; i++)
|
|
z[i * zEws] = OpType::op(x[i * xEws], extraParams);
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::Z_EWSNONZERO: {
|
|
const uint zEws = shape::elementWiseStride(zShapeInfo);
|
|
uint castXShapeInfo[MAX_RANK];
|
|
const bool canCastX = sd::DataTypeUtils::castShapeInfo<uint>(xShapeInfo, castXShapeInfo);
|
|
|
|
auto span = samediff::Span::build(threadId, numThreads, 0, len, 1);
|
|
int64_t start = span.startX(), stop = span.stopX();
|
|
|
|
if (zEws > 1) {
|
|
for (auto i = start; i < stop; i++) {
|
|
const auto xOffset = shape::indexOffset(i, xShapeInfo, castXShapeInfo, canCastX);
|
|
z[i * zEws] = OpType::op(x[xOffset], extraParams);
|
|
}
|
|
}
|
|
else {
|
|
for (auto i = start; i < stop; i++) {
|
|
const auto xOffset = shape::indexOffset(i, xShapeInfo, castXShapeInfo, canCastX);
|
|
z[i] = OpType::op(x[xOffset], extraParams);
|
|
}
|
|
}
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK1: {
|
|
auto span = samediff::Span::build(threadId, numThreads, 0, len, 1);
|
|
|
|
for (auto i0 = span.startX(); i0 < span.stopX(); i0++)
|
|
z[i0 * zStride[0]] = OpType::op(x[i0 * xStride[0]], extraParams);
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK2: {
|
|
auto uXShape0 = static_cast<uint>(xShape[0]);
|
|
auto uXShape1 = static_cast<uint>(xShape[1]);
|
|
|
|
auto loop = samediff::ThreadsHelper::pickLoop2d(numThreads, uXShape0, uXShape1);
|
|
auto span = samediff::Span2::build(loop, threadId, numThreads, 0, uXShape0, 1, 0, uXShape1, 1);
|
|
|
|
for (auto i0 = span.startX(); i0 < span.stopX(); i0++) {
|
|
auto z0 = i0 * zStride[0];
|
|
auto x0 = i0 * xStride[0];
|
|
|
|
for (auto i1 = span.startY(); i1 < span.stopY(); ++i1)
|
|
z[z0 + i1 * zStride[1]] = OpType::op(x[x0 + i1 * xStride[1]], extraParams);
|
|
}
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK3: {
|
|
auto uXShape0 = xShape[0];
|
|
auto uXShape1 = xShape[1];
|
|
auto uXShape2 = xShape[2];
|
|
|
|
auto loop = samediff::ThreadsHelper::pickLoop2d(numThreads, uXShape0, uXShape1);
|
|
auto span = samediff::Span2::build(loop, threadId, numThreads, 0, uXShape0, 1, 0, uXShape1, 1);
|
|
|
|
|
|
for (auto i0 = span.startX(); i0 < span.stopX(); i0++)
|
|
for (auto i1 = span.startY(); i1 < span.stopY(); i1++) {
|
|
auto z0 = i0 * zStride[0] + i1 * zStride[1];
|
|
auto x0 = i0 * xStride[0] + i1 * xStride[1];
|
|
|
|
for (Nd4jLong i2 = 0; i2 < uXShape2; ++i2)
|
|
z[z0 + i2 * zStride[2]] = OpType::op(x[x0 + i2 * xStride[2]], extraParams);
|
|
}
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK4: {
|
|
auto uXShape0 = xShape[0];
|
|
auto uXShape1 = xShape[1];
|
|
auto uXShape2 = xShape[2];
|
|
auto uXShape3 = xShape[3];
|
|
|
|
auto loop = samediff::ThreadsHelper::pickLoop3d(numThreads, uXShape0, uXShape1, uXShape2);
|
|
auto span = samediff::Span3::build(loop, threadId, numThreads, 0, uXShape0, 1, 0, uXShape1, 1, 0, uXShape2, 1);
|
|
|
|
for (auto i0 = span.startX(); i0 < span.stopX(); i0++)
|
|
for (auto i1 = span.startY(); i1 < span.stopY(); i1++)
|
|
for (auto i2 = span.startZ(); i2 < span.stopZ(); i2++) {
|
|
auto x0 = i0 * xStride[0] + i1 * xStride[1] + i2 * xStride[2];
|
|
auto z0 = i0 * zStride[0] + i1 * zStride[1] + i2 * zStride[2];
|
|
|
|
for (Nd4jLong i3 = 0; i3 < uXShape3; ++i3)
|
|
z[z0 + i3 * zStride[3]] = OpType::op(x[x0 + i3 * xStride[3]], extraParams);
|
|
}
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK5: {
|
|
auto uXShape0 = xShape[0];
|
|
auto uXShape1 = xShape[1];
|
|
auto uXShape2 = xShape[2];
|
|
auto uXShape3 = xShape[3];
|
|
auto uXShape4 = xShape[4];
|
|
|
|
auto loop = samediff::ThreadsHelper::pickLoop3d(numThreads, uXShape0, uXShape1, uXShape2);
|
|
auto span = samediff::Span3::build(loop, threadId, numThreads, 0, uXShape0, 1, 0, uXShape1, 1, 0, uXShape2, 1);
|
|
|
|
|
|
for (auto i0 = span.startX(); i0 < span.stopX(); i0++)
|
|
for (auto i1 = span.startY(); i1 < span.stopY(); i1++)
|
|
for (auto i2 = span.startZ(); i2 < span.stopZ(); i2++) {
|
|
auto z0 = i0 * zStride[0] + i1 * zStride[1] + i2 * zStride[2];
|
|
auto x0 = i0 * xStride[0] + i1 * xStride[1] + i2 * xStride[2];
|
|
|
|
for (Nd4jLong i3 = 0; i3 < uXShape3; ++i3) {
|
|
|
|
auto z1 = z0 + i3 * zStride[3];
|
|
auto x1 = x0 + i3 * xStride[3];
|
|
|
|
for (Nd4jLong i4 = 0; i4 < uXShape4; ++i4)
|
|
z[z1 + i4 * zStride[4]] = OpType::op(x[x1 + i4 * xStride[4]], extraParams);
|
|
|
|
}
|
|
}
|
|
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
default: {
|
|
uint xShapeInfoCast[MAX_RANK];
|
|
uint zShapeInfoCast[MAX_RANK];
|
|
|
|
bool canCastX = DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
|
|
bool canCastZ = DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
|
|
|
|
auto span = samediff::Span::build(threadId, numThreads, 0, len, 1);
|
|
|
|
for (auto i = span.startX(); i < span.stopX(); i++) {
|
|
auto xOffset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
|
|
auto zOffset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
|
|
z[zOffset] = OpType::op(x[xOffset], extraParams);
|
|
}
|
|
}
|
|
|
|
}
|
|
}
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Z>
|
|
template <typename OpType>
|
|
void sd::Reduction3Loops<X, Z>::loopReduce3(X* x, Nd4jLong* xShapeInfo,
|
|
X* y, Nd4jLong* yShapeInfo,
|
|
Z* z, Nd4jLong* zShapeInfo,
|
|
int* dims, int dimsLen,
|
|
Z* extraParameters, int64_t start, int64_t stop) {
|
|
|
|
// both tads have same shape, however strides and ews may differ
|
|
|
|
Z param0(OpType::startingValue(x)), param1(OpType::startingValue(x)), param2(extraParameters ? extraParameters[0] : OpType::startingValue(x));
|
|
|
|
const Nd4jLong xLen = shape::length(xShapeInfo);
|
|
const Nd4jLong yLen = shape::length(yShapeInfo);
|
|
|
|
Nd4jLong* xTadShapeInfo = nullptr, * yTadShapeInfo = nullptr, * xTadOffsets = nullptr, * yTadOffsets = nullptr;
|
|
TadPack tadPackX, tadPackY;
|
|
std::vector<Nd4jLong> zeroOffsets;
|
|
|
|
if (xLen == yLen) {
|
|
tadPackX = sd::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dims, dimsLen);
|
|
tadPackY = sd::ConstantTadHelper::getInstance()->tadForDimensions(yShapeInfo, dims, dimsLen);
|
|
xTadShapeInfo = tadPackX.primaryShapeInfo();
|
|
yTadShapeInfo = tadPackY.primaryShapeInfo();
|
|
xTadOffsets = tadPackX.primaryOffsets();
|
|
yTadOffsets = tadPackY.primaryOffsets();
|
|
}
|
|
else if (yLen > xLen) {
|
|
tadPackY = sd::ConstantTadHelper::getInstance()->tadForDimensions(yShapeInfo, dims, dimsLen);
|
|
xTadShapeInfo = xShapeInfo;
|
|
yTadShapeInfo = tadPackY.primaryShapeInfo();
|
|
yTadOffsets = tadPackY.primaryOffsets();
|
|
}
|
|
else {
|
|
tadPackX = sd::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dims, dimsLen);
|
|
yTadShapeInfo = yShapeInfo;
|
|
xTadShapeInfo = tadPackX.primaryShapeInfo();
|
|
xTadOffsets = tadPackX.primaryOffsets();
|
|
}
|
|
|
|
|
|
const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopTadXYZ(xTadShapeInfo, yTadShapeInfo, zShapeInfo);
|
|
|
|
const auto xTadEws = shape::elementWiseStride(xTadShapeInfo);
|
|
const auto yTadEws = shape::elementWiseStride(yTadShapeInfo);
|
|
const auto zEws = shape::elementWiseStride(zShapeInfo);
|
|
|
|
const auto zLen = shape::length(zShapeInfo);
|
|
const auto tadLen = shape::length(xTadShapeInfo);
|
|
|
|
const auto tadShape = shape::shapeOf(xTadShapeInfo);
|
|
const auto xTadStride = shape::stride(xTadShapeInfo);
|
|
const auto yTadStride = shape::stride(xTadShapeInfo);
|
|
|
|
int numThreads = OmpLaunchHelper::tadThreads(tadLen, zLen);
|
|
|
|
switch (kindOfLoop) {
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWS1: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j)
|
|
s = OpType::update(s, OpType::op(xTad[j], yTad[j], extraParams), extraParams);
|
|
|
|
z[i] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWSNONZERO: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j)
|
|
s = OpType::update(s, OpType::op(xTad[j * xTadEws], yTad[j * yTadEws], extraParams), extraParams);
|
|
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK1: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadLen; ++i0) {
|
|
const auto xTadOffset = i0 * xTadStride[0];
|
|
const auto yTadOffset = i0 * yTadStride[0];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK2: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK3: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1] + i2 * xTadStride[2];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1] + i2 * yTadStride[2];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
}
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK4: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
|
|
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1] + i2 * xTadStride[2] + i3 * xTadStride[3];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1] + i2 * yTadStride[2] + i3 * yTadStride[3];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK5: {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
|
|
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) {
|
|
for (Nd4jLong i4 = 0; i4 < tadShape[4]; ++i4) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1] + i2 * xTadStride[2] + i3 * xTadStride[3] + i4 * xTadStride[4];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1] + i2 * yTadStride[2] + i3 * yTadStride[3] + i4 * yTadStride[4];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
default: {
|
|
uint castXTadShapeInfo[MAX_RANK];
|
|
const bool canCastXTad = sd::DataTypeUtils::castShapeInfo<uint>(xTadShapeInfo, castXTadShapeInfo);
|
|
|
|
if (shape::haveSameShapeAndStrides(xTadShapeInfo, yTadShapeInfo)) {
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j) {
|
|
const auto tadOffset = shape::indexOffset(j, xTadShapeInfo, castXTadShapeInfo, canCastXTad);
|
|
s = OpType::update(s, OpType::op(xTad[tadOffset], yTad[tadOffset], extraParams), extraParams);
|
|
}
|
|
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
else {
|
|
uint castYTadShapeInfo[MAX_RANK];
|
|
const bool canCastYTad = sd::DataTypeUtils::castShapeInfo<uint>(yTadShapeInfo, castYTadShapeInfo);
|
|
|
|
Z extraParams[3];
|
|
for (auto i = start; i < stop; i++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = xTadOffsets ? x + xTadOffsets[i] : x;
|
|
const auto yTad = yTadOffsets ? y + yTadOffsets[i] : y;
|
|
auto s = OpType::startingValue(xTad);
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j) {
|
|
const auto xTadOffset = shape::indexOffset(j, xTadShapeInfo, castXTadShapeInfo, canCastXTad);
|
|
const auto yTadOffset = shape::indexOffset(j, yTadShapeInfo, castYTadShapeInfo, canCastYTad);
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
z[i * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
};
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Z>
|
|
template <typename OpType>
|
|
void sd::Reduction3Loops<X, Z>::loopReduce3All(X* x, Nd4jLong* xShapeInfo,
|
|
X* y, Nd4jLong* yShapeInfo,
|
|
Z* z, Nd4jLong* zShapeInfo,
|
|
Nd4jLong* xTadShapeInfo, Nd4jLong* xTadOffsets,
|
|
Nd4jLong* yTadShapeInfo, Nd4jLong* yTadOffsets,
|
|
Z* extraParameters, int64_t start, int64_t stop) {
|
|
|
|
// both tads have same shape, however strides and ews may differ
|
|
|
|
Z param0(OpType::startingValue(x)), param1(OpType::startingValue(x)), param2(extraParameters ? extraParameters[0] : OpType::startingValue(x));
|
|
|
|
const LoopKind::Kind kindOfLoop = LoopKind::deduceKindOfLoopTadXYZ(xTadShapeInfo, yTadShapeInfo, zShapeInfo);
|
|
|
|
const auto xTadEws = shape::elementWiseStride(xTadShapeInfo);
|
|
const auto yTadEws = shape::elementWiseStride(yTadShapeInfo);
|
|
const auto zEws = shape::elementWiseStride(zShapeInfo);
|
|
|
|
const auto zLen = shape::length(zShapeInfo);
|
|
const auto tadLen = shape::length(xTadShapeInfo);
|
|
|
|
const auto numXTads = shape::length(xShapeInfo) / tadLen;
|
|
const auto numYTads = shape::length(yShapeInfo) / tadLen;
|
|
|
|
const auto tadShape = shape::shapeOf(xTadShapeInfo);
|
|
const auto xTadStride = shape::stride(xTadShapeInfo);
|
|
const auto yTadStride = shape::stride(yTadShapeInfo);
|
|
|
|
const auto startVal = OpType::startingValue(x);
|
|
|
|
int numThreads = OmpLaunchHelper::tadThreads(tadLen, numXTads * numYTads);
|
|
|
|
switch (kindOfLoop) {
|
|
//*********************************************//
|
|
case LoopKind::EWS1: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j)
|
|
s = OpType::update(s, OpType::op(xTad[j], yTad[j], extraParams), extraParams);
|
|
|
|
z[zInd] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::EWSNONZERO: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j)
|
|
s = OpType::update(s, OpType::op(xTad[j * xTadEws], yTad[j * yTadEws], extraParams), extraParams);
|
|
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK1: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadLen; ++i0) {
|
|
const auto xTadOffset = i0 * xTadStride[0];
|
|
const auto yTadOffset = i0 * yTadStride[0];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK2: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK3: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1] + i2 * xTadStride[2];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1] + i2 * yTadStride[2];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
}
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK4: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
|
|
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1] + i2 * xTadStride[2] + i3 * xTadStride[3];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1] + i2 * yTadStride[2] + i3 * yTadStride[3];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
case LoopKind::RANK5: {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong i0 = 0; i0 < tadShape[0]; ++i0) {
|
|
for (Nd4jLong i1 = 0; i1 < tadShape[1]; ++i1) {
|
|
for (Nd4jLong i2 = 0; i2 < tadShape[2]; ++i2) {
|
|
for (Nd4jLong i3 = 0; i3 < tadShape[3]; ++i3) {
|
|
for (Nd4jLong i4 = 0; i4 < tadShape[4]; ++i4) {
|
|
const auto xTadOffset = i0 * xTadStride[0] + i1 * xTadStride[1] + i2 * xTadStride[2] + i3 * xTadStride[3] + i4 * xTadStride[4];
|
|
const auto yTadOffset = i0 * yTadStride[0] + i1 * yTadStride[1] + i2 * yTadStride[2] + i3 * yTadStride[3] + i4 * yTadStride[4];
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
z[zInd * zEws] = OpType::postProcess(start, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
break;
|
|
|
|
//*********************************************//
|
|
default: {
|
|
uint castXTadShapeInfo[MAX_RANK];
|
|
const bool canCastXTad = sd::DataTypeUtils::castShapeInfo<uint>(xTadShapeInfo, castXTadShapeInfo);
|
|
|
|
if (shape::haveSameShapeAndStrides(xTadShapeInfo, yTadShapeInfo)) {
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j) {
|
|
const auto tadOffset = shape::indexOffset(j, xTadShapeInfo, castXTadShapeInfo, canCastXTad);
|
|
s = OpType::update(s, OpType::op(xTad[tadOffset], yTad[tadOffset], extraParams), extraParams);
|
|
}
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
else {
|
|
uint castYTadShapeInfo[MAX_RANK];
|
|
const bool canCastYTad = sd::DataTypeUtils::castShapeInfo<uint>(yTadShapeInfo, castYTadShapeInfo);
|
|
|
|
Z extraParams[3];
|
|
for (Nd4jLong ix = 0; ix < numXTads; ix++) {
|
|
for (Nd4jLong iy = 0; iy < numYTads; iy++) {
|
|
extraParams[0] = param0;
|
|
extraParams[1] = param1;
|
|
extraParams[2] = param2;
|
|
|
|
const auto xTad = x + xTadOffsets[ix];
|
|
const auto yTad = y + yTadOffsets[iy];
|
|
const auto zInd = ix * numYTads + iy;
|
|
auto s = startVal;
|
|
|
|
for (Nd4jLong j = 0; j < tadLen; ++j) {
|
|
const auto xTadOffset = shape::indexOffset(j, xTadShapeInfo, castXTadShapeInfo, canCastXTad);
|
|
const auto yTadOffset = shape::indexOffset(j, yTadShapeInfo, castYTadShapeInfo, canCastYTad);
|
|
s = OpType::update(s, OpType::op(xTad[xTadOffset], yTad[yTadOffset], extraParams), extraParams);
|
|
}
|
|
|
|
z[zInd * zEws] = OpType::postProcess(s, tadLen, extraParams);
|
|
}
|
|
};
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
#endif //LIBND4J_LOOPS_H
|