cavis/libnd4j/include/loops/cpu/scalar_int.cpp
raver119 320924278d
Legacy API changes (#441)
* initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* another initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* another initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* one more initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Refactored buffer() and shapeInfo() methods usage with NDArray class.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt Graph class methods to use const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt choose op to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt where op shape method to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt lstsq op to use constant empty shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt matrix_diag_part op shape routine to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt determinant ops to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt mean_pairwssqerr_loss ops to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt shape methods for loss ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt log_loss op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt shape methods for ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt dilation2d ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted deconv2d ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted dynamicRNN op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods for ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods for lstm layer ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* few updates

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* first cuda tweak

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Adopt constant shapes for sconv2d ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt constant shapes for gru ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt constant shapes with shape methods for segment ops and so on.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted constant shapes with unsorted_segment_* ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted constant shapes with gamma op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods of reduce_stddev ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods for reduce_* ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt shape method for squeeze op.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt strided_slice shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored concat op shape method to adopt constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape method for mirror_pad op.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted split op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted tile ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Added const cast for mkldnn routines handles.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored logSoftMaxForVector_ routine to conform with proper data and shape pointer casts.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Cosmetic changes to proper usage of constant pointers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored a couple shape comparators for strides and addBias helpers to proper use data pointers with inplace option.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored depthToSpace helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored histogram helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored im2col helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored gather and gatherND helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage on percentile helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed gather shape with helpers and range buffer usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with space to depth helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage and constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with LUP decomposition>

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored onehot_ helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pad and prefix to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactoed softmax helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed space to batch helpers to use buffers properly.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed stack and split helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with sparse to dense helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with mindistance_ helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with tile helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed constant shape usage with legacy pairwise bool ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored a couple of methods to adopt constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed broadcasting with constant shape."

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const usage with inplace reverse and constant shapes with legacy reduction.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored legacy ops with const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored sort to adopt constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected sort for constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed constant shape usage with special methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored Context to conform with constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* CUDA broadcasting headers

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* pairwise/indexreduce/random headers

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Refactored native ops to adopt constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* legacy reduce3/scalar headers

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Corrected pullRow signature and tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected routines to proper use of constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored tests to use constant shapes properly.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored legacy ops tests to use constant shapes properly.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored buffer usage with NDArray tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed native ops tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed special concat routine.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with test.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with a test.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored TAD.h and tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored calcStrides* routines to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed miscelaneous errors with constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* NativeOps const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Corrected definitions for declared functions.

Signed-off-by: shugeo <sgazeos@gmail.com>

* NativeOps const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* few more const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed const shapes with shape routines.

Signed-off-by: shugeo <sgazeos@gmail.com>

* few more const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed shape method for broadcastable case.

Signed-off-by: shugeo <sgazeos@gmail.com>

* few more const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* xw_plus_b BP shape fn restored

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed signatures with broadcasting.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Repaired backprops shape methods for a set of operations.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored broadcast bool for cuda.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored methods for 3 args with const qualifier.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed a couple of kernel signatures for broadcasting.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernels signatures for const buffers and shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pairwise methods to persistent buffers and shapes usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt const to buffers and shapes with kernels.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt const to buffers and shapes with scalar kernels.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored indexreduce kernels signatures to use const buffers and shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pairwise kernels to adopt cons shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pairwise bool kernels to adopt cons shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored random special ops to conform with const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored native ops to conform with const shapes and buffers under cuda platform.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Cosmetical changes only.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shapes and buffers error.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected start pos routine.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored methods to conform with const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored helpers to use proper methods instead.

Signed-off-by: shugeo <sgazeos@gmail.com>

* bunch of changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next bunch of changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next bunch of changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed execScalar declaration.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed execScalar declaration.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected const shape cases with sort and so on.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shapes for sort.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored kernel declarations to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernels declarations to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected kernel declarations to adopt const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernels declarations to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed segment helpers kernels declarations and so on to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shape usage with segment and solve helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernel declaration with adjustWeight helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed cuda implementations for constant shape helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted const shape usage with kernels.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted top_k kernels to use const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected kernels declarations to adopt const shapes with helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored NDArray definitions to adopt const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shapes with image suppression helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Slight improvement with buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored buffer usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored buffer usage with tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shape usage with definitions.

Signed-off-by: shugeo <sgazeos@gmail.com>

* minor updates on cpu side

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Refactored const shape usage with ConstantDescritor and native ops with cuda platform.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored tear and tile kernels to adopt with const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* softmax_loop fix

Signed-off-by: raver119 <raver119@gmail.com>

* update missing signature

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* softmax again

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* few more missing consts

Signed-off-by: raver119 <raver119@gmail.com>

* new methods updated

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-09 08:06:14 +03:00

204 lines
9.4 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 raver119@gmail.com
//
#include "../scalar_int.h"
#include <system/op_boilerplate.h>
#include <types/types.h>
#include <helpers/LoopKind.h>
#include <execution/Threads.h>
#include "../legacy_ops.h"
using namespace simdOps;
namespace functions {
namespace scalar {
template<typename X>
template<typename OpType>
void ScalarIntTransform<X>::transform(const void *vx, const Nd4jLong *xShapeInfo,
void *vextraParams,
void *vz, const Nd4jLong *zShapeInfo,
const void *vscalars,
int *dimension, int dimensionLength,
const Nd4jLong *xTadShapeInfo, const Nd4jLong *xTadOffsets,
const Nd4jLong *zTadShapeInfo, const Nd4jLong *zTadOffsets,
const uint64_t start, const uint64_t stop) {
auto x = reinterpret_cast<const X *>(vx);
auto z = reinterpret_cast<X *>(vz);
auto scalars = reinterpret_cast<const X *>(vscalars);
auto extraParams = reinterpret_cast<X *>(vextraParams);
if (zTadShapeInfo == nullptr) {
zTadShapeInfo = xTadShapeInfo;
zTadOffsets = xTadOffsets;
}
// tad preparation
const int xTadEws = shape::elementWiseStride(xTadShapeInfo);
const int zTadEws = shape::elementWiseStride(zTadShapeInfo);
const int tadLength = shape::tadLength(xShapeInfo, dimension, dimensionLength);
const int numTads = shape::length(xShapeInfo) / tadLength;
sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopXZ(xTadShapeInfo, zTadShapeInfo);
if (kindOfLoop != sd::LoopKind::EWS1 && kindOfLoop != sd::LoopKind::EWSNONZERO) {
printf("ScalarIntTransform<X>::transform: super-bad loop visited. Shouldn't ever happen\n");
return;
}
int num_threads = sd::math::nd4j_min<int>(numTads, sd::Environment::getInstance()->maxThreads());
if (kindOfLoop == sd::LoopKind::EWS1) {
for (auto r = start; r < stop; r++) {
auto oZ = z + zTadOffsets[r];
auto oX = x + xTadOffsets[r];
PRAGMA_OMP_SIMD
for (int f = 0; f < tadLength; f++)
oZ[f] = OpType::op(oX[f], scalars[r], extraParams);
};
}
else {
for (auto r = start; r < stop; r++) {
auto oZ = z + zTadOffsets[r];
auto oX = x + xTadOffsets[r];
PRAGMA_OMP_SIMD
for (int f = 0; f < tadLength; f++)
oZ[f * zTadEws] = OpType::op(oX[f * xTadEws], scalars[r], extraParams);
};
}
}
template<typename X>
void ScalarIntTransform<X>::transform(int opNum,
const void *x, const Nd4jLong *xShapeInfo,
void *extraParams,
void *z, const Nd4jLong *zShapeInfo,
const void *scalars,
int *dimension, int dimensionLength,
const Nd4jLong *xTadShapeInfo, const Nd4jLong *xTadOffsets,
const Nd4jLong *zTadShapeInfo, const Nd4jLong *zTadOffsets,
const uint64_t start, const uint64_t stop) {
DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, scalars, dimension, dimensionLength, xTadShapeInfo, xTadOffsets, zTadShapeInfo, zTadOffsets, start, stop), SCALAR_INT_OPS);
}
template<typename X>
void ScalarIntTransform<X>::transform(const int opNum,
const void *x, Nd4jLong xEws,
void *z, Nd4jLong zEws,
const void *scalar,
void *extraParams,
const uint64_t n,
const uint64_t start, const uint64_t stop) {
DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xEws, z, zEws, scalar, extraParams, n, start, stop), SCALAR_INT_OPS);
}
template<typename X>
void ScalarIntTransform<X>::transform(const int opNum,
const void *x, const Nd4jLong *xShapeInfo,
void *z, const Nd4jLong *zShapeInfo,
const void *scalar,
void *extraParams,
const uint64_t start, const uint64_t stop) {
DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xShapeInfo, z, zShapeInfo, scalar, extraParams, start, stop), SCALAR_INT_OPS);
}
template<typename X>
template<typename OpType>
void ScalarIntTransform<X>::transform(const void *vx, const Nd4jLong *xShapeInfo,
void *vz, const Nd4jLong *zShapeInfo,
const void *vscalar, void *vextraParams,
const uint64_t start, const uint64_t stop) {
auto x = reinterpret_cast<const X *>(vx);
auto z = reinterpret_cast<X *>(vz);
auto scalar = reinterpret_cast<const X *>(vscalar)[0];
auto extraParams = reinterpret_cast<X *>(vextraParams);
auto xEws = shape::elementWiseStride(xShapeInfo);
auto zEws = shape::elementWiseStride(zShapeInfo);
auto len = shape::length(xShapeInfo);
sd::LoopKind::Kind kindOfLoop = sd::LoopKind::deduceKindOfLoopXZ(xShapeInfo, zShapeInfo);
if (kindOfLoop == sd::LoopKind::EWS1 || kindOfLoop == sd::LoopKind::EWSNONZERO) {
transform<OpType>(x, xEws, z, zEws, vscalar, extraParams, len, start, stop);
return;
}
uint xShapeInfoCast[MAX_RANK];
const bool canCastX = sd::DataTypeUtils::castShapeInfo<uint>(xShapeInfo, xShapeInfoCast);
if(shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
PRAGMA_OMP_SIMD
for (auto i = start; i < stop; i++) {
auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
z[offset] = OpType::op(x[offset], scalar, extraParams);
};
}
else {
uint zShapeInfoCast[MAX_RANK];
const bool canCastZ = sd::DataTypeUtils::castShapeInfo<uint>(zShapeInfo, zShapeInfoCast);
PRAGMA_OMP_SIMD
for (auto i = start; i < stop; i++) {
auto xOffset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
auto zOffset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
z[zOffset] = OpType::op(x[xOffset], scalar, extraParams);
};
}
}
template<typename X>
template<typename OpType>
void ScalarIntTransform<X>::transform(const void *vx, Nd4jLong xEws,
void *vz, Nd4jLong zEws,
const void *vscalar,
void *vextraParams,
const uint64_t len, const uint64_t start, const uint64_t stop) {
auto x = reinterpret_cast<const X *>(vx);
auto z = reinterpret_cast<X *>(vz);
auto scalar = reinterpret_cast<const X *>(vscalar)[0];
auto extraParams = reinterpret_cast<X *>(vextraParams);
if (scalar < (sizeof(X) * 8)) {
if (xEws == 1 && zEws == 1) {
for (auto i = start; i < stop; i++)
z[i] = OpType::op(x[i], scalar, extraParams);
} else {
for (auto i = start; i < stop; i++)
z[i * zEws] = OpType::op(x[i * xEws], scalar, extraParams);
}
}
}
BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT ScalarIntTransform, , INTEGER_TYPES);
}
}