cavis/libnd4j/include/ops/declarable/helpers/convolutions.h

336 lines
18 KiB
C
Raw Normal View History

2021-02-01 13:31:45 +01:00
/* ******************************************************************************
*
2019-06-06 14:21:15 +02:00
*
* 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.
*
2021-02-01 13:31:45 +01:00
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
2019-06-06 14:21:15 +02:00
* 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
******************************************************************************/
//
// Based on PyTorch - https://github.com/pytorch/pytorch
//
#ifndef LIBND4J_CONVOLUTIONS_H
#define LIBND4J_CONVOLUTIONS_H
#include <array/NDArray.h>
2019-06-06 14:21:15 +02:00
#include <graph/Context.h>
#include <system/dll.h>
2019-06-06 14:21:15 +02:00
#include <execution/LaunchContext.h>
namespace sd {
2019-06-06 14:21:15 +02:00
namespace ops {
enum PoolingType {
MAX_POOL = 0,
AVG_POOL = 1,
PNORM_POOL = 2,
};
class ND4J_EXPORT ConvolutionUtils {
2019-06-06 14:21:15 +02:00
public:
static inline void calcOutSizePool2D(int& oH, int& oW, const int kH, const int kW, const int sH, const int sW, const int pH, const int pW, const int dH, const int dW, const int iH, const int iW, const int paddingMode) {
if(paddingMode == 0) { // valid
cuDNN integration (#150) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * one file Signed-off-by: raver119 <raver119@gmail.com> * few more includes Signed-off-by: raver119 <raver119@gmail.com> * m? Signed-off-by: raver119 <raver119@gmail.com> * const Signed-off-by: raver119 <raver119@gmail.com> * cudnn linkage in tests Signed-off-by: raver119 <raver119@gmail.com> * culibos Signed-off-by: raver119 <raver119@gmail.com> * static reminder Signed-off-by: raver119 <raver119@gmail.com> * platform engine tag Signed-off-by: raver119 <raver119@gmail.com> * HAVE_CUDNN moved to config.h.in Signed-off-by: raver119 <raver119@gmail.com> * include Signed-off-by: raver119 <raver119@gmail.com> * include Signed-off-by: raver119 <raver119@gmail.com> * skip cudnn handle creation if there's not cudnn Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * target device in context Signed-off-by: raver119 <raver119@gmail.com> * platform engines Signed-off-by: raver119 <raver119@gmail.com> * platform engines Signed-off-by: raver119 <raver119@gmail.com> * allow multiple -h args Signed-off-by: raver119 <raver119@gmail.com> * allow multiple -h args Signed-off-by: raver119 <raver119@gmail.com> * move mkldnn out of CPU block Signed-off-by: raver119 <raver119@gmail.com> * link to mkldnn on cuda Signed-off-by: raver119 <raver119@gmail.com> * less prints Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * next step Signed-off-by: raver119 <raver119@gmail.com> * conv2d NCHW draft Signed-off-by: raver119 <raver119@gmail.com> * conv2d biasAdd Signed-off-by: raver119 <raver119@gmail.com> * test for MKL/CUDNN combined use Signed-off-by: raver119 <raver119@gmail.com> * - provide additional code for conv2d ff based on cudnn api, not tested yet Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on conv2d helper based on using cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - fixing several cuda bugs which appeared after cudnn lib had been started to use Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of conv2d backprop op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - implementaion of conv3d and conv3d_bp ops based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - bugs fixing in conv3d/conv3d_bp ops (cudnn in use) Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of depthwiseConv2d (ff/bp) op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of batchnorm ff op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - disable cudnn batchnorm temporary Signed-off-by: Yurii <iuriish@yahoo.com> * - add minor change in cmake Signed-off-by: Yurii <iuriish@yahoo.com> * engine for depthwise mkldnn Signed-off-by: raver119 <raver119@gmail.com> * couple of includes Signed-off-by: raver119 <raver119@gmail.com> * - provide permutation to cudnn batchnorm ff when format is NHWC Signed-off-by: Yurii <iuriish@yahoo.com> * lgamma fix Signed-off-by: raver119 <raver119@gmail.com> * - eliminate memory leak in two tests Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: Yurii Shyrma <iuriish@yahoo.com>
2020-01-20 19:32:46 +01:00
// oH = (iH - (kH + (kH-1)*(dH-1)) + 2*pH)/sH + 1;
// oW = (iW - (kW + (kW-1)*(dW-1)) + 2*pW)/sW + 1;
oH = (iH - ((kH - 1) * dH + 1) + 2 * pH) / sH + 1;
oW = (iW - ((kW - 1) * dW + 1) + 2 * pW) / sW + 1;
}
else if (paddingMode == 1) { // same
2019-06-06 14:21:15 +02:00
oH = (int) math::nd4j_ceil<double, double>(iH * 1. / sH);
oW = (int) math::nd4j_ceil<double, double>(iW * 1. / sW);
}
else { // causal
oH = (iH - 1) / sH + 1; // 2*pH = (kH-1)*dH
oW = (iW - 1) / sW + 1;
2019-06-06 14:21:15 +02:00
}
}
static inline void calcOutSizePool3D(int& oD, int& oH, int& oW, const int kD, const int kH, const int kW, const int sD, const int sH, const int sW, const int pD, const int pH, const int pW, const int dD, const int dH, const int dW, const int iD, const int iH, const int iW, const int paddingMode) {
2019-06-06 14:21:15 +02:00
if(paddingMode == 0) { // valid
cuDNN integration (#150) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * one file Signed-off-by: raver119 <raver119@gmail.com> * few more includes Signed-off-by: raver119 <raver119@gmail.com> * m? Signed-off-by: raver119 <raver119@gmail.com> * const Signed-off-by: raver119 <raver119@gmail.com> * cudnn linkage in tests Signed-off-by: raver119 <raver119@gmail.com> * culibos Signed-off-by: raver119 <raver119@gmail.com> * static reminder Signed-off-by: raver119 <raver119@gmail.com> * platform engine tag Signed-off-by: raver119 <raver119@gmail.com> * HAVE_CUDNN moved to config.h.in Signed-off-by: raver119 <raver119@gmail.com> * include Signed-off-by: raver119 <raver119@gmail.com> * include Signed-off-by: raver119 <raver119@gmail.com> * skip cudnn handle creation if there's not cudnn Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * target device in context Signed-off-by: raver119 <raver119@gmail.com> * platform engines Signed-off-by: raver119 <raver119@gmail.com> * platform engines Signed-off-by: raver119 <raver119@gmail.com> * allow multiple -h args Signed-off-by: raver119 <raver119@gmail.com> * allow multiple -h args Signed-off-by: raver119 <raver119@gmail.com> * move mkldnn out of CPU block Signed-off-by: raver119 <raver119@gmail.com> * link to mkldnn on cuda Signed-off-by: raver119 <raver119@gmail.com> * less prints Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * next step Signed-off-by: raver119 <raver119@gmail.com> * conv2d NCHW draft Signed-off-by: raver119 <raver119@gmail.com> * conv2d biasAdd Signed-off-by: raver119 <raver119@gmail.com> * test for MKL/CUDNN combined use Signed-off-by: raver119 <raver119@gmail.com> * - provide additional code for conv2d ff based on cudnn api, not tested yet Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on conv2d helper based on using cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - fixing several cuda bugs which appeared after cudnn lib had been started to use Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of conv2d backprop op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - implementaion of conv3d and conv3d_bp ops based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - bugs fixing in conv3d/conv3d_bp ops (cudnn in use) Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of depthwiseConv2d (ff/bp) op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of batchnorm ff op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - disable cudnn batchnorm temporary Signed-off-by: Yurii <iuriish@yahoo.com> * - add minor change in cmake Signed-off-by: Yurii <iuriish@yahoo.com> * engine for depthwise mkldnn Signed-off-by: raver119 <raver119@gmail.com> * couple of includes Signed-off-by: raver119 <raver119@gmail.com> * - provide permutation to cudnn batchnorm ff when format is NHWC Signed-off-by: Yurii <iuriish@yahoo.com> * lgamma fix Signed-off-by: raver119 <raver119@gmail.com> * - eliminate memory leak in two tests Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: Yurii Shyrma <iuriish@yahoo.com>
2020-01-20 19:32:46 +01:00
oD = (iD - ((kD - 1) * dD + 1) + 2 * pD) / sD + 1;
oH = (iH - ((kH - 1) * dH + 1) + 2 * pH) / sH + 1;
oW = (iW - ((kW - 1) * dW + 1) + 2 * pW) / sW + 1;
2019-06-06 14:21:15 +02:00
}
else if(paddingMode == 1) { // same
oD = (int) sd::math::nd4j_ceil<double, double>(iD * 1. / sD);
oH = (int) sd::math::nd4j_ceil<double, double>(iH * 1. / sH);
oW = (int) sd::math::nd4j_ceil<double, double>(iW * 1. / sW);
}
else { // causal
oD = (iD - 1) / sD + 1;
oH = (iH - 1) / sH + 1; // 2*pH = (kH-1)*dH
oW = (iW - 1) / sW + 1;
2019-06-06 14:21:15 +02:00
}
}
static inline void calcPadding2D(int& pH, int& pW, int oH, int oW, int iH, int iW, int kH, int kW, int sH, int sW, int dH, int dW, const int paddingMode = 1 /* default is same mode*/) {
2019-06-06 14:21:15 +02:00
if(paddingMode == 0) // valid
return;
if(paddingMode == 1) { // same
2019-06-06 14:21:15 +02:00
const int eKH = (kH - 1) * dH + 1;
const int eKW = (kW - 1) * dW + 1;
pH = ((oH - 1) * sH + eKH - iH) / 2; //Note that padBottom is 1 bigger than this if bracketed term is not divisible by 2
pW = ((oW - 1) * sW + eKW - iW) / 2;
}
else { // causal
pH = (kH - 1) * dH;
pW = (kW - 1) * dW;
2019-06-06 14:21:15 +02:00
}
}
static inline void calcPadding3D(int& pD, int& pH, int& pW, const int oD, const int oH, const int oW, const int iD, const int iH, const int iW, const int kD, const int kH, const int kW, const int sD, const int sH, const int sW, const int dD, const int dH, const int dW, const int paddingMode = 1 /* default is same mode*/) {
if(paddingMode == 0) // valid
return;
if(paddingMode == 1) { // same
2019-06-06 14:21:15 +02:00
const int eKD = (kD - 1) * dD + 1;
const int eKH = (kH - 1) * dH + 1;
const int eKW = (kW - 1) * dW + 1;
2019-06-06 14:21:15 +02:00
pD = ((oD - 1) * sD + eKD - iD) / 2;
pH = ((oH - 1) * sH + eKH - iH) / 2; //Note that padBottom is 1 bigger than this if bracketed term is not divisible by 2
pW = ((oW - 1) * sW + eKW - iW) / 2;
}
else { // causal
pD = (kD - 1) * dD;
pH = (kH - 1) * dH;
pW = (kW - 1) * dW;
}
2019-06-06 14:21:15 +02:00
}
// calculation of output height and width in 2D deconvolution procedure
static inline void calcOutSizeDeconv2D(int& oH, int& oW, const int kH, const int kW, const int sH, const int sW, const int pH, const int pW, const int dH, const int dW, const int iH, const int iW, const int paddingMode) {
if (paddingMode) {
2019-06-06 14:21:15 +02:00
oH = sH * iH;
oW = sW * iW;
}
else {
const int ekH = (kH - 1) * dH + 1;
const int ekW = (kW - 1) * dW + 1;
2019-06-06 14:21:15 +02:00
oH = sH * (iH - 1) + ekH - 2 * pH;
oW = sW * (iW - 1) + ekW - 2 * pW;
}
}
// calculation of output height and width in 3D deconvolution procedure
static inline void calcOutSizeDeconv3D(int& oD, int& oH, int& oW, const int kD, const int kH, const int kW, const int sD, const int sH, const int sW, const int pD, const int pH, const int pW, const int dD, const int dH, const int dW, const int iD, const int iH, const int iW, const int paddingMode) {
if (paddingMode) {
2019-06-06 14:21:15 +02:00
oD = sD * iD;
oH = sH * iH;
oW = sW * iW;
}
else {
const int ekD = (kD - 1) * dD + 1;
const int ekH = (kH - 1) * dH + 1;
const int ekW = (kW - 1) * dW + 1;
2019-06-06 14:21:15 +02:00
oD = sD * (iD - 1) + ekD - 2 * pD;
oH = sH * (iH - 1) + ekH - 2 * pH;
oW = sW * (iW - 1) + ekW - 2 * pW;
}
}
// evaluates sizes values and indexes using input and output arrays depending on data format
static inline void getSizesAndIndexesConv2d(const bool isNCHW, const int wFormat, const NDArray& input, const NDArray& output, int& bS, int& iC, int& iH, int& iW, int& oC, int& oH, int& oW, int& indIOioC, int& indIiH, int& indWiC, int& indWoC, int& indWkH, int& indOoH) {
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 07:06:14 +02:00
getSizesAndIndexesConv2d(isNCHW, wFormat, input.shapeInfo(), output.shapeInfo(), bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWoC, indWkH, indOoH);
2019-06-06 14:21:15 +02:00
}
static inline void getSizesAndIndexesConv2d(const bool isNCHW, const int wFormat, const Nd4jLong* inShapeInfo, const Nd4jLong* outShapeInfo, int& bS, int& iC, int& iH, int& iW, int& oC, int& oH, int& oW, int& indIOioC, int& indIiH, int& indWiC, int& indWoC, int& indWkH, int& indOoH) {
2019-06-06 14:21:15 +02:00
// input [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
// weights [kH, kW, iC, oC] (wFormat = 0), [oC, iC, kH, kW] (wFormat = 1), [oC, kH, kW, iC] (wFormat = 2)
2019-06-06 14:21:15 +02:00
// output [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW)
if(0 == wFormat) {
indWkH = 0; indWiC = 2; indWoC = 3;
}
else if(1 == wFormat) {
indWkH = 2; indWiC = 1; indWoC = 0;
}
else {
indWkH = 1; indWiC = 3; indWoC = 0;
}
2019-06-06 14:21:15 +02:00
if(!isNCHW) {
indIOioC = 3; indIiH = 1; indOoH = 1;
}
else {
indIOioC = 1; indIiH = 2; indOoH = 2;
}
bS = inShapeInfo[1]; // batch size
iC = inShapeInfo[indIOioC+1]; // input channels
iH = inShapeInfo[indIiH+1]; // input height
iW = inShapeInfo[indIiH+2]; // input width
oC = outShapeInfo[indIOioC+1]; // output channels
oH = outShapeInfo[indOoH+1]; // output height
oW = outShapeInfo[indOoH+2]; // output width
}
// evaluates sizes values and indexes using input and output arrays depending on data format
static inline void getSizesAndIndexesConv3d(const bool isNCDHW, const int wFormat, const NDArray& input, const NDArray& output, int& bS, int& iC, int& iD, int& iH, int& iW, int& oC, int& oD, int& oH, int& oW, int& indIOioC, int& indIOioD, int& indWiC, int& indWoC, int& indWkD) {
2019-06-06 14:21:15 +02:00
// input [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
// weights [kD, kH, kW, iC, oC] (wFormat = 0), [oC, iC, kD, kH, kW] (wFormat = 1), [oC, kD, kH, kW, iC] (wFormat = 2)
2019-06-06 14:21:15 +02:00
// output [bS, oD, oH, oW, oC] (NDHWC) or [bS, oC, oD, oH, oW] (NCDHW)
if(0 == wFormat) {
indWkD = 0; indWiC = 3; indWoC = 4;
}
else if(1 == wFormat) {
indWkD = 2; indWiC = 1; indWoC = 0;
}
else {
indWkD = 1; indWiC = 4; indWoC = 0;
}
2019-06-06 14:21:15 +02:00
if(!isNCDHW) {
indIOioC = 4; indIOioD = 1;
}
else {
indIOioC = 1; indIOioD = 2;
}
bS = input.sizeAt(0); // batch size
iC = input.sizeAt(indIOioC); // input channels
iD = input.sizeAt(indIOioD); // input depth
iH = input.sizeAt(indIOioD+1); // input height
iW = input.sizeAt(indIOioD+2); // input width
oC = output.sizeAt(indIOioC); // output channels
oD = output.sizeAt(indIOioD); // output depth
oH = output.sizeAt(indIOioD+1); // output height
oW = output.sizeAt(indIOioD+2); // output width
}
// static inline void calcPaddingAndDilationForConv2DMKL(const int iH, const int iW, const int oH, const int oW, const int kH, const int kW, const int sH, const int sW, const int paddingMode, int& pH, int& pW, int& dH, int& dW) {
// if(kH != 1) {
// if(paddingMode) {
// pH = (oH - 1) * sH - iH + kH - pH;
// dH = dH - 1;
// }
// else
// dH = (iH + 2*pH - (oH - 1) * sH - kH) / (kH - 1);
// }
// if(kW != 1) {
// if(paddingMode) {
// pW = (oW - 1) * sW - iW + kW - pW;
// dW = dW - 1;
// }
// else
// dW = (iW + 2*pW - (oW - 1) * sW - kW) / (kW - 1);
// }
// }
// static inline void calcPaddingAndDilationForConv3DMKL(const int iD, const int iH, const int iW, const int oD, const int oH, const int oW, const int kD, const int kH, const int kW, const int sD, const int sH, const int sW, const int paddingMode, int& pD, int& pH, int& pW, int& dD, int& dH, int& dW) {
// if(kD != 1) {
// if(paddingMode) {
// pD = (oD - 1) * sD - iD + kD - pD;
// dD = dD - 1;
// }
// else
// dD = (iD + 2*pD - (oD - 1) * sD - kD) / (kD - 1);
// }
// if(kH != 1) {
// if(paddingMode) {
// pH = (oH - 1) * sH - iH + kH - pH;
// dH = dH - 1;
// }
// else
// dH = (iH + 2*pH - (oH - 1) * sH - kH) / (kH - 1);
// }
// if(kW != 1) {
// if(paddingMode) {
// pW = (oW - 1) * sW - iW + kW - pW;
// dW = dW - 1;
// }
// else
// dW = (iW + 2*pW - (oW - 1) * sW - kW) / (kW - 1);
// }
// }
static std::vector<Nd4jLong> expectWeightsShape(const int wFormat, const int kH, const int kW, const int iC, const int oC) {
if(0 == wFormat)
return std::vector<Nd4jLong>({kH, kW, iC, oC});
if(1 == wFormat)
return std::vector<Nd4jLong>({oC, iC, kH, kW});
return std::vector<Nd4jLong>({oC, kH, kW, iC});
}
static std::vector<Nd4jLong> expectWeightsShape(const int wFormat, const int kD, const int kH, const int kW, const int iC, const int oC) {
if(0 == wFormat)
return std::vector<Nd4jLong>({kD, kH, kW, iC, oC});
if(1 == wFormat)
return std::vector<Nd4jLong>({oC, iC, kD, kH, kW});
return std::vector<Nd4jLong>({oC, kD, kH, kW, iC});
}
static void conv2d(sd::graph::Context &context, const NDArray* input, const NDArray* weights, const NDArray* bias, NDArray* output, const int kH, const int kW, const int sH, const int sW, int pH, int pW, const int dH, const int dW, const int paddingMode, const int isNCHW, const int wFormat);
2019-06-06 14:21:15 +02:00
// static void conv2d(sd::graph::Context & block, const std::vector<NDArray*>& inArrs, NDArray* output, const std::vector<int>& intArgs);
2019-06-06 14:21:15 +02:00
// static void conv2dBP(sd::graph::Context & block, const std::vector<NDArray*>& inArrs, const std::vector<NDArray*>& outArrs, const std::vector<int>& intArgs);
2019-06-06 14:21:15 +02:00
static void conv2dBP(sd::graph::Context & block, const NDArray* input, const NDArray* weights, const NDArray* bias, const NDArray* gradO, NDArray* gradI, NDArray* gradW, NDArray* gradB, const int kH, const int kW, const int sH, const int sW, int pH, int pW, const int dH, const int dW, const int paddingMode, const int isNCHW, const int wFormat);
2019-06-06 14:21:15 +02:00
static void depthwiseConv2d(sd::graph::Context & block, const NDArray* input, const NDArray* weights, const NDArray* bias, NDArray* output, const int kH, const int kW, const int sH, const int sW, int pH, int pW, const int dH, const int dW, const int paddingMode, const int isNCHW, const int wFormat);
2019-06-06 14:21:15 +02:00
static void depthwiseConv2dBP(sd::graph::Context & block, const NDArray* input, const NDArray* weights, const NDArray* bias, const NDArray* gradO, NDArray* gradI, NDArray* gradW, NDArray* gradB, const int kH, const int kW, const int sH, const int sW, int pH, int pW, const int dH, const int dW, const int paddingMode, const int isNCHW, const int wFormat);
2019-06-06 14:21:15 +02:00
static void sconv2d(sd::graph::Context & block, const NDArray* input, const NDArray* weightsDepth, const NDArray* weightsPoint, const NDArray* bias, NDArray* output, const int kH, const int kW, const int sH, const int sW, int pH, int pW, const int dH, const int dW, const int paddingMode, const int isNCHW, const int wFormat);
2019-06-06 14:21:15 +02:00
static void vol2col(sd::graph::Context & block, const NDArray& vol, NDArray& col, const int sD, const int sH, const int sW, const int pD, const int pH, const int pW, const int dD, const int dH, const int dW);
2019-06-06 14:21:15 +02:00
static void col2vol(sd::graph::Context & block, const NDArray& col, NDArray& vol, const int sD, const int sH, const int sW, const int pD, const int pH, const int pW, const int dD, const int dH, const int dW);
2019-06-06 14:21:15 +02:00
static void upsampling2d(sd::graph::Context & block, const NDArray& input, NDArray& output, const int factorH, const int factorW, const bool isNCHW);
2019-06-06 14:21:15 +02:00
static void upsampling3d(sd::graph::Context & block, const NDArray& input, NDArray& output, const int factorD, const int factorH, const int factorW, const bool isNCDHW);
2019-06-06 14:21:15 +02:00
static void upsampling2dBP(sd::graph::Context & block, const NDArray& gradO, NDArray& gradI, const bool isNCHW);
2019-06-06 14:21:15 +02:00
static void upsampling3dBP(sd::graph::Context & block, const NDArray& gradO, NDArray& gradI, const bool isNCDHW);
2019-06-06 14:21:15 +02:00
static void pooling2d(sd::graph::Context & block, const NDArray& input, NDArray& output, const int kH, const int kW, const int sH, const int sW, const int pH, const int pW, const int dH, const int dW, const PoolingType poolingMode, const int extraParam0);
2019-06-06 14:21:15 +02:00
static void pooling3d(sd::graph::Context & block, const NDArray& input, NDArray& output, const int kD, const int kH, const int kW, const int sD, const int sH, const int sW, const int pD, const int pH, const int pW, const int dD, const int dH, const int dW, const int poolingMode, const int extraParam0);
2019-06-06 14:21:15 +02:00
static void pooling2dBP(sd::graph::Context & block, const NDArray& input, const NDArray& gradO, NDArray& gradI, const int kH, const int kW, const int sH, const int sW, const int pH, const int pW, const int dH, const int dW, const int poolingMode, const int extraParam0);
2019-06-06 14:21:15 +02:00
static void pooling3dBP(sd::graph::Context & block, const NDArray& input, const NDArray& gradO, NDArray& gradI, const int kD, const int kH, const int kW, const int sD, const int sH, const int sW, const int pD, const int pH, const int pW, const int dD, const int dH, const int dW, const int poolingMode, const int extraParam0);
2019-06-06 14:21:15 +02:00
};
}
}
#endif //LIBND4J_CONVOLUTIONS_H