cavis/libnd4j/include/ops/declarable/helpers/cpu/convolutions.cpp

2496 lines
145 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 18.09.2018
//
#include <ops/declarable/helpers/convolutions.h>
#include<ops/declarable/helpers/addBias.h>
#include <ops/declarable/helpers/im2col.h>
#include <ops/declarable/helpers/col2im.h>
#include <NDArrayFactory.h>
#include <MmulHelper.h>
namespace nd4j {
namespace ops {
#ifdef HAVE_MKLDNN
using namespace mkldnn;
void ConvolutionUtils::getMKLDNNMemoryDescPool2d(
int kH, int kW, int sH, int sW, int pH, int pW, int dH, int dW, int poolingMode, int extraParam0, bool isNCHW,
int bS, int iC, int iH, int iW, int oC, int oH, int oW,
const NDArray* src, const NDArray* diff_src, const NDArray* dst, mkldnn::algorithm& algorithm,
mkldnn::memory::desc* pool_src_md, mkldnn::memory::desc* pool_diff_src_md, mkldnn::memory::desc* pool_dst_md,
mkldnn::memory::desc* user_src_md, mkldnn::memory::desc* user_diff_src_md, mkldnn::memory::desc* user_dst_md,
mkldnn::memory::dims& pool_strides, mkldnn::memory::dims& pool_kernel, mkldnn::memory::dims& pool_padding, mkldnn::memory::dims& pool_padding_r) {
mkldnn::memory::dims pool_src_tz = { bS, iC, iH, iW };
mkldnn::memory::dims pool_dst_tz = { bS, oC, oH, oW };
pool_strides = { sH, sW };
pool_kernel = { kH, kW };
pool_padding = { pH, pW };
pool_padding_r = { (oH - 1) * sH - iH + kH - pH,
(oW - 1) * sW - iW + kW - pW };
algorithm = poolingMode == 0 ? pooling_max
: extraParam0 == 0 ? pooling_avg_exclude_padding
: pooling_avg_include_padding;
auto type = mkldnn::memory::data_type::f32;
auto format = isNCHW ? mkldnn::memory::format::nchw : mkldnn::memory::format::nhwc;
auto supposed_to_be_any_format = mkldnn::memory::format::nChw8c; // doesn't work with "any"
if (src != nullptr && src->getBuffer() != nullptr && pool_src_md != nullptr) {
*pool_src_md = mkldnn::memory::desc({ pool_src_tz }, type, supposed_to_be_any_format);
*user_src_md = mkldnn::memory::desc({ pool_src_tz }, type, format);
user_src_md->data.format = mkldnn_blocked; // overrides "format = isNCHW ? nchw : nhwc"
user_src_md->data.layout_desc.blocking.strides[0][0] = src->stridesOf()[isNCHW ? 0 : 0];
user_src_md->data.layout_desc.blocking.strides[0][1] = src->stridesOf()[isNCHW ? 1 : 3];
user_src_md->data.layout_desc.blocking.strides[0][2] = src->stridesOf()[isNCHW ? 2 : 1];
user_src_md->data.layout_desc.blocking.strides[0][3] = src->stridesOf()[isNCHW ? 3 : 2];
}
if (diff_src != nullptr && diff_src->getBuffer() != nullptr && pool_diff_src_md != nullptr) {
*pool_diff_src_md = mkldnn::memory::desc({ pool_src_tz }, type, supposed_to_be_any_format);
*user_diff_src_md = mkldnn::memory::desc({ pool_src_tz }, type, format);
user_diff_src_md->data.format = mkldnn_blocked; // overrides "format = isNCHW ? nchw : nhwc"
user_diff_src_md->data.layout_desc.blocking.strides[0][0] = diff_src->stridesOf()[isNCHW ? 0 : 0];
user_diff_src_md->data.layout_desc.blocking.strides[0][1] = diff_src->stridesOf()[isNCHW ? 1 : 3];
user_diff_src_md->data.layout_desc.blocking.strides[0][2] = diff_src->stridesOf()[isNCHW ? 2 : 1];
user_diff_src_md->data.layout_desc.blocking.strides[0][3] = diff_src->stridesOf()[isNCHW ? 3 : 2];
}
if (dst != nullptr && dst->getBuffer() != nullptr && pool_dst_md != nullptr) {
*pool_dst_md = mkldnn::memory::desc({ pool_dst_tz }, type, supposed_to_be_any_format);
*user_dst_md = mkldnn::memory::desc({ pool_dst_tz }, type, format);
user_dst_md->data.format = mkldnn_blocked; // overrides "format = isNCHW ? nchw : nhwc"
user_dst_md->data.layout_desc.blocking.strides[0][0] = dst->stridesOf()[isNCHW ? 0 : 0];
user_dst_md->data.layout_desc.blocking.strides[0][1] = dst->stridesOf()[isNCHW ? 1 : 3];
user_dst_md->data.layout_desc.blocking.strides[0][2] = dst->stridesOf()[isNCHW ? 2 : 1];
user_dst_md->data.layout_desc.blocking.strides[0][3] = dst->stridesOf()[isNCHW ? 3 : 2];
}
}
void ConvolutionUtils::getMKLDNNMemoryDescPool3d(
int kD, int kH, int kW, int sD, int sH, int sW, int pD, int pH, int pW, int dD, int dH, int dW, int poolingMode, int extraParam0, bool isNCDHW,
int bS, int iC, int iD, int iH, int iW, int oC, int oD, int oH, int oW,
const NDArray* src, const NDArray* diff_src, const NDArray* dst, mkldnn::algorithm& algorithm,
mkldnn::memory::desc* pool_src_md, mkldnn::memory::desc* pool_diff_src_md, mkldnn::memory::desc* pool_dst_md,
mkldnn::memory::desc* user_src_md, mkldnn::memory::desc* user_diff_src_md, mkldnn::memory::desc* user_dst_md,
mkldnn::memory::dims& pool_strides, mkldnn::memory::dims& pool_kernel, mkldnn::memory::dims& pool_padding, mkldnn::memory::dims& pool_padding_r) {
mkldnn::memory::dims pool_src_tz = { bS, iC, iD, iH, iW };
mkldnn::memory::dims pool_dst_tz = { bS, oC, oD, oH, oW };
pool_strides = { sD, sH, sW };
pool_kernel = { kD, kH, kW };
pool_padding = { pD, pH, pW };
pool_padding_r = { (oD - 1) * sD - iD + kD - pD,
(oH - 1) * sH - iH + kH - pH,
(oW - 1) * sW - iW + kW - pW };
algorithm = poolingMode == 0 ? pooling_max
: extraParam0 == 0 ? pooling_avg_exclude_padding
: pooling_avg_include_padding;
auto type = mkldnn::memory::data_type::f32;
auto format = isNCDHW ? mkldnn::memory::format::ncdhw : mkldnn::memory::format::ndhwc;
auto supposed_to_be_any_format = mkldnn::memory::format::nCdhw8c; // doesn't work with "any"
if (src != nullptr && src->getBuffer() != nullptr && pool_src_md != nullptr) {
*pool_src_md = mkldnn::memory::desc({ pool_src_tz }, type, supposed_to_be_any_format);
*user_src_md = mkldnn::memory::desc({ pool_src_tz }, type, format);
user_src_md->data.format = mkldnn_blocked; // overrides "format = isNCDHW ? ncdhw : ndhwc"
user_src_md->data.layout_desc.blocking.strides[0][0] = src->stridesOf()[isNCDHW ? 0 : 0];
user_src_md->data.layout_desc.blocking.strides[0][1] = src->stridesOf()[isNCDHW ? 1 : 4];
user_src_md->data.layout_desc.blocking.strides[0][2] = src->stridesOf()[isNCDHW ? 2 : 1];
user_src_md->data.layout_desc.blocking.strides[0][3] = src->stridesOf()[isNCDHW ? 3 : 2];
user_src_md->data.layout_desc.blocking.strides[0][4] = src->stridesOf()[isNCDHW ? 4 : 3];
}
if (diff_src != nullptr && diff_src->getBuffer() != nullptr && pool_diff_src_md != nullptr) {
*pool_diff_src_md = mkldnn::memory::desc({ pool_src_tz }, type, supposed_to_be_any_format);
*user_diff_src_md = mkldnn::memory::desc({ pool_src_tz }, type, format);
user_diff_src_md->data.format = mkldnn_blocked; // overrides "format = isNCDHW ? ncdhw : ndhwc"
user_diff_src_md->data.layout_desc.blocking.strides[0][0] = diff_src->stridesOf()[isNCDHW ? 0 : 0];
user_diff_src_md->data.layout_desc.blocking.strides[0][1] = diff_src->stridesOf()[isNCDHW ? 1 : 4];
user_diff_src_md->data.layout_desc.blocking.strides[0][2] = diff_src->stridesOf()[isNCDHW ? 2 : 1];
user_diff_src_md->data.layout_desc.blocking.strides[0][3] = diff_src->stridesOf()[isNCDHW ? 3 : 2];
user_diff_src_md->data.layout_desc.blocking.strides[0][4] = diff_src->stridesOf()[isNCDHW ? 4 : 3];
}
if (dst != nullptr && dst->getBuffer() != nullptr && pool_dst_md != nullptr) {
*pool_dst_md = mkldnn::memory::desc({ pool_dst_tz }, type, supposed_to_be_any_format);
*user_dst_md = mkldnn::memory::desc({ pool_dst_tz }, type, format);
user_dst_md->data.format = mkldnn_blocked; // overrides "format = isNCDHW ? ncdhw : ndhwc"
user_dst_md->data.layout_desc.blocking.strides[0][0] = dst->stridesOf()[isNCDHW ? 0 : 0];
user_dst_md->data.layout_desc.blocking.strides[0][1] = dst->stridesOf()[isNCDHW ? 1 : 4];
user_dst_md->data.layout_desc.blocking.strides[0][2] = dst->stridesOf()[isNCDHW ? 2 : 1];
user_dst_md->data.layout_desc.blocking.strides[0][3] = dst->stridesOf()[isNCDHW ? 3 : 2];
user_dst_md->data.layout_desc.blocking.strides[0][4] = dst->stridesOf()[isNCDHW ? 4 : 3];
}
}
#endif
//////////////////////////////////////////////////////////////////////////
// [bS, iC, iD, iH, iW] is convoluted to [bS, iC, kD, kH, kW, oD, oH, oW]
template <typename T>
static void vol2col_(const NDArray& volume, NDArray& columns, 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 bS = volume.sizeAt(0);
const int iC = volume.sizeAt(1);
const int iD = volume.sizeAt(2);
const int iH = volume.sizeAt(3);
const int iW = volume.sizeAt(4);
const int kD = columns.sizeAt(2);
const int kH = columns.sizeAt(3);
const int kW = columns.sizeAt(4);
const int oD = columns.sizeAt(5);
const int oH = columns.sizeAt(6);
const int oW = columns.sizeAt(7);
const Nd4jLong colStride0 = columns.stridesOf()[0];
const Nd4jLong colStride1 = columns.stridesOf()[1];
const Nd4jLong colStride2 = columns.stridesOf()[2];
const Nd4jLong colStride3 = columns.stridesOf()[3];
const Nd4jLong colStride4 = columns.stridesOf()[4];
const Nd4jLong colStride5 = columns.stridesOf()[5];
const Nd4jLong colStride6 = columns.stridesOf()[6];
const Nd4jLong colStride7 = columns.stridesOf()[7];
const Nd4jLong volStride0 = volume.stridesOf()[0];
const Nd4jLong volStride1 = volume.stridesOf()[1];
const Nd4jLong volStride2 = volume.stridesOf()[2];
const Nd4jLong volStride3 = volume.stridesOf()[3];
const Nd4jLong volStride4 = volume.stridesOf()[4];
T* colBuff = columns.bufferAsT<T>();
T* volBuff = const_cast<NDArray&>(volume).bufferAsT<T>();
T *col, *vol;
int volDep, volRow, volCol;
if (volume.ordering() == 'c' && columns.ordering() == 'c' && shape::strideDescendingCAscendingF(volume.getShapeInfo()) && shape::strideDescendingCAscendingF(columns.getShapeInfo()))
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(col, vol, volDep, volRow, volCol) collapse(2))
for (int b = 0; b < bS; ++b) {
for (int c = 0; c < iC; ++c) {
for (int kDep = 0; kDep < kD; ++kDep) {
for (int kRow = 0; kRow < kH; ++kRow) {
for (int kCol = 0; kCol < kW; ++kCol) {
for (int colD = 0; colD < oD; ++colD) {
for (int colH = 0; colH < oH; ++colH) {
for (int colW = 0; colW < oW; ++colW) {
volDep = (-pD + kDep * dD) + colD*sD;
volRow = (-pH + kRow * dH) + colH*sH;
volCol = (-pW + kCol * dW) + colW*sW;
col = colBuff + b*colStride0 + c*colStride1 + kDep*colStride2 + kRow*colStride3 + kCol*colStride4 + colD*colStride5 + colH*colStride6 + colW*colStride7;
if (static_cast<unsigned>(volDep) >= static_cast<unsigned>(iD) || static_cast<unsigned>(volRow) >= static_cast<unsigned>(iH) || static_cast<unsigned>(volCol) >= static_cast<unsigned>(iW))
*col = static_cast<T>(0.);
else {
vol = volBuff + b*volStride0 + c*volStride1 + volDep*volStride2 + volRow*volStride3 + volCol*volStride4;
*col = *vol;
}
}
}
}
}
}
}
}
}
else
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(vol, col, volDep, volRow, volCol))
for (int b = 0; b < bS; b++) {
for (int colD = 0; colD < oD; ++colD) {
for (int colH = 0; colH < oH; ++colH) {
for (int colW = 0; colW < oW; ++colW) {
for (int c = 0; c < iC; ++c) {
for (int kDep = 0; kDep < kD; ++kDep) {
for (int kRow = 0; kRow < kH; ++kRow) {
for (int kCol = 0; kCol < kW; ++kCol) {
volDep = (-pD + kDep * dD) + colD*sD;
volRow = (-pH + kRow * dH) + colH*sH;
volCol = (-pW + kCol * dW) + colW*sW;
col = colBuff + b*colStride0 + c*colStride1 + kDep*colStride2 + kRow*colStride3 + kCol*colStride4 + colD*colStride5 + colH*colStride6 + colW*colStride7;
if (static_cast<unsigned>(volDep) >= static_cast<unsigned>(iD) || static_cast<unsigned>(volRow) >= static_cast<unsigned>(iH) || static_cast<unsigned>(volCol) >= static_cast<unsigned>(iW))
*col = static_cast<T>(0.);
else {
vol = volBuff + b*volStride0 + c*volStride1 + volDep*volStride2 + volRow*volStride3 + volCol*volStride4;
*col = *vol;
}
}
}
}
}
}
}
}
}
}
//////////////////////////////////////////////////////////////////////////
// [bS, iC, kD, kH, kW, oD, oH, oW] is de-convoluted to [bS, iC, iD, iH, iW]
template <typename T>
static void col2vol_(const NDArray& columns, NDArray& volume, 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) {
// initial zeroing of volume content
volume.nullify();
const int bS = volume.sizeAt(0);
const int iC = volume.sizeAt(1);
const int iD = volume.sizeAt(2);
const int iH = volume.sizeAt(3);
const int iW = volume.sizeAt(4);
const int kD = columns.sizeAt(2);
const int kH = columns.sizeAt(3);
const int kW = columns.sizeAt(4);
const int oD = columns.sizeAt(5);
const int oH = columns.sizeAt(6);
const int oW = columns.sizeAt(7);
const Nd4jLong colStride0 = columns.stridesOf()[0];
const Nd4jLong colStride1 = columns.stridesOf()[1];
const Nd4jLong colStride2 = columns.stridesOf()[2];
const Nd4jLong colStride3 = columns.stridesOf()[3];
const Nd4jLong colStride4 = columns.stridesOf()[4];
const Nd4jLong colStride5 = columns.stridesOf()[5];
const Nd4jLong colStride6 = columns.stridesOf()[6];
const Nd4jLong colStride7 = columns.stridesOf()[7];
const Nd4jLong volStride0 = volume.stridesOf()[0];
const Nd4jLong volStride1 = volume.stridesOf()[1];
const Nd4jLong volStride2 = volume.stridesOf()[2];
const Nd4jLong volStride3 = volume.stridesOf()[3];
const Nd4jLong volStride4 = volume.stridesOf()[4];
T* volBuff = volume.bufferAsT<T>();
T* colBuff = const_cast<NDArray&>(columns).bufferAsT<T>();
T* col, *vol;
int volDep, volRow, volCol;
if (volume.ordering() == 'c' && columns.ordering() == 'c' && shape::strideDescendingCAscendingF(volume.getShapeInfo()) && shape::strideDescendingCAscendingF(columns.getShapeInfo()))
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(col, vol, volDep, volRow, volCol) collapse(2))
for (int b = 0; b < bS; b++) {
for (int c = 0; c < iC; ++c) {
for (int kDep = 0; kDep < kD; ++kDep) {
for (int kRow = 0; kRow < kH; ++kRow) {
for (int kCol = 0; kCol < kW; ++kCol) {
for (int colD = 0; colD < oD; ++colD) {
for (int colH = 0; colH < oH; ++colH) {
for (int colW = 0; colW < oW; ++colW) {
volDep = -pD + kDep * dD + colD * sD;
volRow = -pH + kRow * dH + colH * sH;
volCol = -pW + kCol * dW + colW * sW;
if (static_cast<unsigned>(volDep) < static_cast<unsigned>(iD) && static_cast<unsigned>(volRow) < static_cast<unsigned>(iH) && static_cast<unsigned>(volCol) < static_cast<unsigned>(iW)) {
col = colBuff + b*colStride0 + c*colStride1 + kDep*colStride2 + kRow*colStride3 + kCol*colStride4 + colD*colStride5 + colH*colStride6 + colW*colStride7;
vol = volBuff + b*volStride0 + c*volStride1 + volDep*volStride2 + volRow*volStride3 + volCol*volStride4;
*vol += *col;
}
}
}
}
}
}
}
}
}
else
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(vol, col, volDep, volRow, volCol))
for (int b = 0; b < bS; b++) {
for (int colD = 0; colD < oD; ++colD) {
for (int colH = 0; colH < oH; ++colH) {
for (int colW = 0; colW < oW; ++colW) {
for (int c = 0; c < iC; ++c) {
for (int kDep = 0; kDep < kD; ++kDep) {
for (int kRow = 0; kRow < kH; ++kRow) {
for (int kCol = 0; kCol < kW; ++kCol) {
volDep = (-pD + kDep * dD) + colD*sD;
volRow = (-pH + kRow * dH) + colH*sH;
volCol = (-pW + kCol * dW) + colW*sW;
if (static_cast<unsigned>(volDep) < static_cast<unsigned>(iD) && static_cast<unsigned>(volRow) < static_cast<unsigned>(iH) && static_cast<unsigned>(volCol) < static_cast<unsigned>(iW)) {
col = colBuff + b*colStride0 + c*colStride1 + kDep*colStride2 + kRow*colStride3 + kCol*colStride4 + colD*colStride5 + colH*colStride6 + colW*colStride7;
vol = volBuff + b*volStride0 + c*volStride1 + volDep*volStride2 + volRow*volStride3 + volCol*volStride4;
*vol += *col;
}
}
}
}
}
}
}
}
}
}
#ifdef HAVE_MKLDNN
using namespace mkldnn;
void ConvolutionUtils::getMKLDNNMemoryDescConv2d(
int kH, int kW, int sH, int sW, int pH, int pW, int dH, int dW, bool isSameMode, bool isNCHW,
int bS, int iC, int iH, int iW, int oC, int oH, int oW, const NDArray* src, const NDArray* diff_src,
const NDArray* weights, const NDArray* diff_weights, const NDArray* bias, const NDArray* dst,
mkldnn::memory::desc* conv_src_md, mkldnn::memory::desc* conv_diff_src_md, mkldnn::memory::desc* conv_weights_md,
mkldnn::memory::desc* conv_diff_weights_md, mkldnn::memory::desc* conv_bias_md, mkldnn::memory::desc* conv_dst_md,
mkldnn::memory::desc* user_src_md, mkldnn::memory::desc* user_diff_src_md, mkldnn::memory::desc* user_weights_md,
mkldnn::memory::desc* user_diff_weights_md, mkldnn::memory::desc* user_bias_md, mkldnn::memory::desc* user_dst_md,
mkldnn::memory::dims& conv_strides, mkldnn::memory::dims& conv_padding, mkldnn::memory::dims& conv_padding_r) {
mkldnn::memory::dims conv_src_tz = { bS, iC, iH, iW };
mkldnn::memory::dims conv_weights_tz = { oC, iC, kH, kW };
mkldnn::memory::dims conv_bias_tz = { oC };
mkldnn::memory::dims conv_dst_tz = { bS, oC, oH, oW };
conv_strides = { sH, sW };
conv_padding = { pH, pW };
conv_padding_r = { (oH - 1) * sH - iH + kH - pH,
(oW - 1) * sW - iW + kW - pW };
auto type = mkldnn::memory::data_type::f32;
auto format = isNCHW ? mkldnn::memory::format::nchw : mkldnn::memory::format::nhwc;
auto formatw = mkldnn::memory::format::hwio;
if (src != nullptr && conv_src_md != nullptr) {
*conv_src_md = mkldnn::memory::desc({ conv_src_tz }, type, mkldnn::memory::format::any);
*user_src_md = mkldnn::memory::desc({ conv_src_tz }, type, format);
user_src_md->data.format = mkldnn_blocked; // overrides "format = isNCHW ? nchw : nhwc"
user_src_md->data.layout_desc.blocking.strides[0][0] = src->stridesOf()[isNCHW ? 0 : 0];
user_src_md->data.layout_desc.blocking.strides[0][1] = src->stridesOf()[isNCHW ? 1 : 3];
user_src_md->data.layout_desc.blocking.strides[0][2] = src->stridesOf()[isNCHW ? 2 : 1];
user_src_md->data.layout_desc.blocking.strides[0][3] = src->stridesOf()[isNCHW ? 3 : 2];
}
if (diff_src != nullptr && conv_diff_src_md != nullptr) {
*conv_diff_src_md = mkldnn::memory::desc({ conv_src_tz }, type, mkldnn::memory::format::any);
*user_diff_src_md = mkldnn::memory::desc({ conv_src_tz }, type, format);
user_diff_src_md->data.format = mkldnn_blocked; // overrides "format = isNCHW ? nchw : nhwc"
user_diff_src_md->data.layout_desc.blocking.strides[0][0] = diff_src->stridesOf()[isNCHW ? 0 : 0];
user_diff_src_md->data.layout_desc.blocking.strides[0][1] = diff_src->stridesOf()[isNCHW ? 1 : 3];
user_diff_src_md->data.layout_desc.blocking.strides[0][2] = diff_src->stridesOf()[isNCHW ? 2 : 1];
user_diff_src_md->data.layout_desc.blocking.strides[0][3] = diff_src->stridesOf()[isNCHW ? 3 : 2];
}
if (weights != nullptr && conv_weights_md != nullptr) {
*conv_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, mkldnn::memory::format::any);
*user_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, formatw);
user_weights_md->data.format = mkldnn_blocked; // overrides "formatw = hwio"
user_weights_md->data.layout_desc.blocking.strides[0][0] = weights->stridesOf()[3];
user_weights_md->data.layout_desc.blocking.strides[0][1] = weights->stridesOf()[2];
user_weights_md->data.layout_desc.blocking.strides[0][2] = weights->stridesOf()[0];
user_weights_md->data.layout_desc.blocking.strides[0][3] = weights->stridesOf()[1];
}
if (diff_weights != nullptr && conv_diff_weights_md != nullptr) {
*conv_diff_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, mkldnn::memory::format::any);
*user_diff_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, formatw);
user_diff_weights_md->data.format = mkldnn_blocked; // overrides "formatw = hwio"
user_diff_weights_md->data.layout_desc.blocking.strides[0][0] = diff_weights->stridesOf()[3];
user_diff_weights_md->data.layout_desc.blocking.strides[0][1] = diff_weights->stridesOf()[2];
user_diff_weights_md->data.layout_desc.blocking.strides[0][2] = diff_weights->stridesOf()[0];
user_diff_weights_md->data.layout_desc.blocking.strides[0][3] = diff_weights->stridesOf()[1];
}
if (bias != nullptr && conv_bias_md != nullptr) {
*conv_bias_md = mkldnn::memory::desc({ conv_bias_tz }, type, mkldnn::memory::format::any);
*user_bias_md = mkldnn::memory::desc({ conv_bias_tz }, type, mkldnn::memory::format::x);
}
if (dst != nullptr && conv_dst_md != nullptr) {
*conv_dst_md = mkldnn::memory::desc({ conv_dst_tz }, type, mkldnn::memory::format::any);
*user_dst_md = mkldnn::memory::desc({ conv_dst_tz }, type, format);
user_dst_md->data.format = mkldnn_blocked; // overrides "format = isNCHW ? nchw : nhwc"
user_dst_md->data.layout_desc.blocking.strides[0][0] = dst->stridesOf()[isNCHW ? 0 : 0];
user_dst_md->data.layout_desc.blocking.strides[0][1] = dst->stridesOf()[isNCHW ? 1 : 3];
user_dst_md->data.layout_desc.blocking.strides[0][2] = dst->stridesOf()[isNCHW ? 2 : 1];
user_dst_md->data.layout_desc.blocking.strides[0][3] = dst->stridesOf()[isNCHW ? 3 : 2];
}
}
void ConvolutionUtils::getMKLDNNMemoryDescConv3d(
int kD, int kH, int kW, int sD, int sH, int sW, int pD, int pH, int pW, int dD, int dH, int dW, bool isSameMode, bool isNCDHW,
int bS, int iC, int iD, int iH, int iW, int oC, int oD, int oH, int oW, const NDArray* src, const NDArray* diff_src,
const NDArray* weights, const NDArray* diff_weights, const NDArray* bias, const NDArray* dst,
mkldnn::memory::desc* conv_src_md, mkldnn::memory::desc* conv_diff_src_md, mkldnn::memory::desc* conv_weights_md,
mkldnn::memory::desc* conv_diff_weights_md, mkldnn::memory::desc* conv_bias_md, mkldnn::memory::desc* conv_dst_md,
mkldnn::memory::desc* user_src_md, mkldnn::memory::desc* user_diff_src_md, mkldnn::memory::desc* user_weights_md,
mkldnn::memory::desc* user_diff_weights_md, mkldnn::memory::desc* user_bias_md, mkldnn::memory::desc* user_dst_md,
mkldnn::memory::dims& conv_strides, mkldnn::memory::dims& conv_padding, mkldnn::memory::dims& conv_padding_r) {
mkldnn::memory::dims conv_src_tz = { bS, iC, iD, iH, iW };
mkldnn::memory::dims conv_weights_tz = { oC, iC, kD, kH, kW };
mkldnn::memory::dims conv_bias_tz = { oC };
mkldnn::memory::dims conv_dst_tz = { bS, oC, oD, oH, oW };
conv_strides = { sD, sH, sW };
conv_padding = { pD, pH, pW };
conv_padding_r = { (oD - 1) * sD - iD + kD - pD,
(oH - 1) * sH - iH + kH - pH,
(oW - 1) * sW - iW + kW - pW };
auto type = mkldnn::memory::data_type::f32;
auto format = isNCDHW ? mkldnn::memory::format::ncdhw : mkldnn::memory::format::ndhwc;
auto formatw = mkldnn::memory::format::dhwio;
if (src != nullptr && conv_src_md != nullptr) {
*conv_src_md = mkldnn::memory::desc({ conv_src_tz }, type, mkldnn::memory::format::any);
*user_src_md = mkldnn::memory::desc({ conv_src_tz }, type, format);
user_src_md->data.format = mkldnn_blocked; // overrides "format = isNCDHW ? ncdhw : ndhwc"
user_src_md->data.layout_desc.blocking.strides[0][0] = src->stridesOf()[isNCDHW ? 0 : 0];
user_src_md->data.layout_desc.blocking.strides[0][1] = src->stridesOf()[isNCDHW ? 1 : 4];
user_src_md->data.layout_desc.blocking.strides[0][2] = src->stridesOf()[isNCDHW ? 2 : 1];
user_src_md->data.layout_desc.blocking.strides[0][3] = src->stridesOf()[isNCDHW ? 3 : 2];
user_src_md->data.layout_desc.blocking.strides[0][4] = src->stridesOf()[isNCDHW ? 4 : 3];
}
if (diff_src != nullptr && conv_diff_src_md != nullptr) {
*conv_diff_src_md = mkldnn::memory::desc({ conv_src_tz }, type, mkldnn::memory::format::any);
*user_diff_src_md = mkldnn::memory::desc({ conv_src_tz }, type, format);
user_diff_src_md->data.format = mkldnn_blocked; // overrides "format = isNCDHW ? ncdhw : ndhwc"
user_diff_src_md->data.layout_desc.blocking.strides[0][0] = diff_src->stridesOf()[isNCDHW ? 0 : 0];
user_diff_src_md->data.layout_desc.blocking.strides[0][1] = diff_src->stridesOf()[isNCDHW ? 1 : 4];
user_diff_src_md->data.layout_desc.blocking.strides[0][2] = diff_src->stridesOf()[isNCDHW ? 2 : 1];
user_diff_src_md->data.layout_desc.blocking.strides[0][3] = diff_src->stridesOf()[isNCDHW ? 3 : 2];
user_diff_src_md->data.layout_desc.blocking.strides[0][4] = diff_src->stridesOf()[isNCDHW ? 4 : 3];
}
if (weights != nullptr && conv_weights_md != nullptr) {
*conv_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, mkldnn::memory::format::any);
*user_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, formatw);
user_weights_md->data.format = mkldnn_blocked; // overrides "formatw = dhwio"
user_weights_md->data.layout_desc.blocking.strides[0][0] = weights->stridesOf()[4];
user_weights_md->data.layout_desc.blocking.strides[0][1] = weights->stridesOf()[3];
user_weights_md->data.layout_desc.blocking.strides[0][2] = weights->stridesOf()[0];
user_weights_md->data.layout_desc.blocking.strides[0][3] = weights->stridesOf()[1];
user_weights_md->data.layout_desc.blocking.strides[0][4] = weights->stridesOf()[2];
}
if (diff_weights != nullptr && conv_diff_weights_md != nullptr) {
*conv_diff_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, mkldnn::memory::format::any);
*user_diff_weights_md = mkldnn::memory::desc({ conv_weights_tz }, type, formatw);
user_diff_weights_md->data.format = mkldnn_blocked; // overrides "formatw = dhwio"
user_diff_weights_md->data.layout_desc.blocking.strides[0][0] = diff_weights->stridesOf()[4];
user_diff_weights_md->data.layout_desc.blocking.strides[0][1] = diff_weights->stridesOf()[3];
user_diff_weights_md->data.layout_desc.blocking.strides[0][2] = diff_weights->stridesOf()[0];
user_diff_weights_md->data.layout_desc.blocking.strides[0][3] = diff_weights->stridesOf()[1];
user_diff_weights_md->data.layout_desc.blocking.strides[0][4] = diff_weights->stridesOf()[2];
}
if (bias != nullptr && conv_bias_md != nullptr) {
*conv_bias_md = mkldnn::memory::desc({ conv_bias_tz }, type, mkldnn::memory::format::any);
*user_bias_md = mkldnn::memory::desc({ conv_bias_tz }, type, mkldnn::memory::format::x);
}
if (dst != nullptr && conv_dst_md != nullptr) {
*conv_dst_md = mkldnn::memory::desc({ conv_dst_tz }, type, mkldnn::memory::format::any);
*user_dst_md = mkldnn::memory::desc({ conv_dst_tz }, type, format);
user_dst_md->data.format = mkldnn_blocked; // overrides "format = isNCDHW ? ncdhw : ndhwc"
user_dst_md->data.layout_desc.blocking.strides[0][0] = dst->stridesOf()[isNCDHW ? 0 : 0];
user_dst_md->data.layout_desc.blocking.strides[0][1] = dst->stridesOf()[isNCDHW ? 1 : 4];
user_dst_md->data.layout_desc.blocking.strides[0][2] = dst->stridesOf()[isNCDHW ? 2 : 1];
user_dst_md->data.layout_desc.blocking.strides[0][3] = dst->stridesOf()[isNCDHW ? 3 : 2];
user_dst_md->data.layout_desc.blocking.strides[0][4] = dst->stridesOf()[isNCDHW ? 4 : 3];
}
}
#endif
//////////////////////////////////////////////////////////////////////////
template <typename X, typename Y>
static void conv2d_(nd4j::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 isSameMode, const int isNCHW) {
// input [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
// weights [kH, kW, iC, oC] always
// bias [oC]
// output [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW)
// kH filter(kernel) height
// kW filter(kernel) width
// sH strides height
// sW strides width
// pH paddings height
// pW paddings width
// dH dilations height
// dW dilations width
// isSameMode 0-VALID, 1-SAME
// isNCHW 1-NCHW, 0-NHWC
int bS, iC, iH, iW, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width;
int indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, *input, *output, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWoC, indWkH, indOoH);
if(isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
#ifdef HAVE_MKLDNN
if (block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported<X, Y>()) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("conv2d"));
}
if (streams[0].checkAndReset({input, weights, bias}, {output}, {}, {kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW})) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc conv_src_md(empty), conv_weights_md(empty), conv_bias_md(empty), conv_dst_md(empty);
mkldnn::memory::desc user_src_md(empty), user_weights_md(empty), user_bias_md(empty), user_dst_md(empty);
mkldnn::memory::dims conv_strides, conv_padding, conv_padding_r;
ConvolutionUtils::getMKLDNNMemoryDescConv2d(kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW,
bS, iC, iH, iW, oC, oH, oW, input, nullptr, weights, nullptr, bias, output,
&conv_src_md, nullptr, &conv_weights_md, nullptr, &conv_bias_md, &conv_dst_md,
&user_src_md, nullptr, &user_weights_md, nullptr, &user_bias_md, &user_dst_md,
conv_strides, conv_padding, conv_padding_r);
auto conv_desc = bias != nullptr
? convolution_forward::desc(prop_kind::forward,
convolution_direct, conv_src_md, conv_weights_md, conv_bias_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero)
: convolution_forward::desc(prop_kind::forward,
convolution_direct, conv_src_md, conv_weights_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero);
auto engine = streams[0].getEngine();
auto conv_prim_desc = convolution_forward::primitive_desc(conv_desc, engine);
auto user_src_memory = mkldnn::memory({user_src_md, engine}, const_cast<NDArray*>(input)->buffer());
auto user_weights_memory = mkldnn::memory({user_weights_md, engine}, const_cast<NDArray*>(weights)->buffer());
auto user_dst_memory = mkldnn::memory({user_dst_md, engine}, output->buffer());
auto conv_src_memory = user_src_memory;
streams[0].addMemory(user_src_memory);
if (mkldnn::memory::primitive_desc(conv_prim_desc.src_primitive_desc())
!= user_src_memory.get_primitive_desc()) {
conv_src_memory = mkldnn::memory(conv_prim_desc.src_primitive_desc());
streams[0].addMemory(conv_src_memory);
streams[0].addOperation(reorder(user_src_memory, conv_src_memory));
}
auto conv_weights_memory = user_weights_memory;
streams[0].addMemory(user_weights_memory);
if (mkldnn::memory::primitive_desc(conv_prim_desc.weights_primitive_desc())
!= user_weights_memory.get_primitive_desc()) {
conv_weights_memory = mkldnn::memory(conv_prim_desc.weights_primitive_desc());
streams[0].addMemory(conv_weights_memory);
streams[0].addOperation(reorder(user_weights_memory, conv_weights_memory));
}
auto conv_dst_memory = user_dst_memory;
streams[0].addMemory(user_dst_memory);
if (mkldnn::memory::primitive_desc(conv_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
conv_dst_memory = mkldnn::memory(conv_prim_desc.dst_primitive_desc());
streams[0].addMemory(conv_dst_memory);
}
if (bias != nullptr) {
auto conv_bias_memory = mkldnn::memory(conv_prim_desc.bias_primitive_desc(), const_cast<NDArray*>(bias)->buffer());
streams[0].addMemory(conv_bias_memory);
streams[0].addOperation(convolution_forward(conv_prim_desc, conv_src_memory, conv_weights_memory, conv_bias_memory, conv_dst_memory));
} else {
streams[0].addOperation(convolution_forward(conv_prim_desc, conv_src_memory, conv_weights_memory, conv_dst_memory));
}
if (mkldnn::memory::primitive_desc(conv_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(conv_dst_memory, user_dst_memory));
}
}
streams[0].submitAndWait();
return;
}
#endif
nd4j_debug("MKL-DNN is not used for conv2d!\n", 0);
std::vector<int> permutForOutput;
if(isNCHW)
permutForOutput = {0, 3, 1, 2}; // [bS, oH, oW, oC] -> [bS, oC, oH, oW]
else
input = new NDArray(input->permute({0, 3, 1, 2})); // [bS, iH, iW, iC] -> [bS, iC, iH, iW] if NHWC
NDArray col('c', {bS, oH, oW, kH, kW, iC}, input->dataType(), input->getContext());
NDArray colP = col.permute({0, 5, 3, 4, 1, 2}); // {bS, iC, kH, kW, oH, oW}
NDArray mmulResult('f', {bS*oH*oW, oC}, output->dataType(), output->getContext());
//----- calculation of output -----//
auto ctx = block.launchContext();
helpers::im2col(*ctx, *input, colP, kH, kW, sH, sW, pH, pW, dH, dW, NDArrayFactory::create(0.f, input->getContext())); // [bS, iC, iH, iW] is convoluted to [bS, iC, kH, kW, oH, oW]
MmulHelper::tensorDot(&col, weights, &mmulResult, {3,4,5}, {0,1,2}, {}); // [bS, oH, oW, kH, kW, iC] x [kH, kW, iC, oC] = [bS, oH, oW, oC]
//----- assign outTemp to output -----//
if(isNCHW) {
mmulResult.reshapei({bS, oH, oW, oC});
mmulResult.permutei(permutForOutput);
}
output->assign(mmulResult);
//----- add biases if required -----//
if(bias)
// output->applyBroadcast(broadcast::Add, {indIOioC}, bias);
helpers::addBias(*output, *bias, isNCHW);
if(!isNCHW)
delete input;
}
//////////////////////////////////////////////////////////////////////////
template <typename X, typename Y>
static void conv2dBP_(nd4j::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 isSameMode, const int isNCHW) {
// input [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
// weights [kH, kW, iC, oC] always
// bias [oC]
// gradO [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
// gradI [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
// gradW [kH, kW, iC, oC] always
// gradB [oC]
// kH filter(kernel) height
// kW filter(kernel) width
// sH strides height
// sW strides width
// pH paddings height
// pW paddings width
// dH dilations height
// dW dilations width
// isSameMode 0-VALID, 1-SAME
// isNCHW 0-NHWC, 1-NCHW
int bS, iC, iH, iW, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width;
int indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWoC, indWkH, indOoH);
if(isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
#ifdef HAVE_MKLDNN
if (block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported<X, Y>()) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("conv2d_bp_weights"));
streams.push_back(MKLDNNStream("conv2d_bp_data"));
}
bool resetW = streams[0].checkAndReset({input, weights, bias, gradO}, {gradI, gradW, gradB}, {}, {kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW});
bool resetI = streams[1].checkAndReset({input, weights, bias, gradO}, {gradI, gradW, gradB}, {}, {kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW});
if (resetW || resetI) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc conv_src_md(empty), conv_diff_src_md(empty), conv_weights_md(empty),
conv_diff_weights_md(empty), conv_bias_md(empty), conv_dst_md(empty);
mkldnn::memory::desc user_src_md(empty), user_diff_src_md(empty), user_weights_md(empty),
user_diff_weights_md(empty), user_bias_md(empty), user_dst_md(empty);
mkldnn::memory::dims conv_strides, conv_padding, conv_padding_r;
ConvolutionUtils::getMKLDNNMemoryDescConv2d(kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW,
bS, iC, iH, iW, oC, oH, oW, input, gradI, weights, gradW, gradB, gradO,
&conv_src_md, &conv_diff_src_md, &conv_weights_md, &conv_diff_weights_md, &conv_bias_md, &conv_dst_md,
&user_src_md, &user_diff_src_md, &user_weights_md, &user_diff_weights_md, &user_bias_md, &user_dst_md,
conv_strides, conv_padding, conv_padding_r);
auto conv_desc = gradB != nullptr
? convolution_forward::desc(prop_kind::forward,
convolution_direct, conv_src_md, conv_weights_md, conv_bias_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero)
: convolution_forward::desc(prop_kind::forward,
convolution_direct, conv_src_md, conv_weights_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero);
auto conv_prim_desc = convolution_forward::primitive_desc(conv_desc, streams[0].getEngine());
if (gradW != nullptr) {
auto convW_desc = gradB != nullptr
? convolution_backward_weights::desc(
convolution_direct, conv_src_md, conv_diff_weights_md, conv_bias_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero)
: convolution_backward_weights::desc(
convolution_direct, conv_src_md, conv_diff_weights_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero);
auto engine = streams[0].getEngine();
auto convW_prim_desc = convolution_backward_weights::primitive_desc(convW_desc, engine, conv_prim_desc);
auto userW_src_memory = mkldnn::memory({user_src_md, engine}, const_cast<NDArray*>(input)->buffer());
auto userW_weights_memory = mkldnn::memory({user_diff_weights_md, engine}, gradW->buffer());
auto userW_dst_memory = mkldnn::memory({user_dst_md, engine}, const_cast<NDArray*>(gradO)->buffer());
auto convW_src_memory = userW_src_memory;
streams[0].addMemory(userW_src_memory);
if (mkldnn::memory::primitive_desc(convW_prim_desc.src_primitive_desc())
!= userW_src_memory.get_primitive_desc()) {
convW_src_memory = mkldnn::memory(convW_prim_desc.src_primitive_desc());
streams[0].addMemory(convW_src_memory);
streams[0].addOperation(reorder(userW_src_memory, convW_src_memory));
}
auto convW_weights_memory = userW_weights_memory;
streams[0].addMemory(userW_weights_memory);
if (mkldnn::memory::primitive_desc(convW_prim_desc.diff_weights_primitive_desc())
!= userW_weights_memory.get_primitive_desc()) {
convW_weights_memory = mkldnn::memory(convW_prim_desc.diff_weights_primitive_desc());
streams[0].addMemory(convW_weights_memory);
}
auto convW_dst_memory = userW_dst_memory;
streams[0].addMemory(userW_dst_memory);
if (mkldnn::memory::primitive_desc(convW_prim_desc.diff_dst_primitive_desc())
!= userW_dst_memory.get_primitive_desc()) {
convW_dst_memory = mkldnn::memory(convW_prim_desc.diff_dst_primitive_desc());
streams[0].addMemory(convW_dst_memory);
streams[0].addOperation(reorder(userW_dst_memory, convW_dst_memory));
}
if (gradB != nullptr) {
auto convW_bias_memory = mkldnn::memory(convW_prim_desc.diff_bias_primitive_desc(), gradB->buffer());
streams[0].addMemory(convW_bias_memory);
streams[0].addOperation(convolution_backward_weights(convW_prim_desc, convW_src_memory, convW_dst_memory, convW_weights_memory, convW_bias_memory));
} else {
streams[0].addOperation(convolution_backward_weights(convW_prim_desc, convW_src_memory, convW_dst_memory, convW_weights_memory));
}
if (mkldnn::memory::primitive_desc(convW_prim_desc.diff_weights_primitive_desc())
!= userW_weights_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(convW_weights_memory, userW_weights_memory));
}
}
if (gradI != nullptr) {
auto convI_desc =
convolution_backward_data::desc(
convolution_direct, conv_diff_src_md, conv_weights_md,
conv_dst_md, conv_strides, conv_padding, conv_padding_r, padding_kind::zero);
auto engine = streams[1].getEngine();
auto convI_prim_desc = convolution_backward_data::primitive_desc(convI_desc, engine, conv_prim_desc);
auto userI_src_memory = mkldnn::memory({user_diff_src_md, engine}, gradI->buffer());
auto userI_weights_memory = mkldnn::memory({user_weights_md, engine}, const_cast<NDArray*>(weights)->buffer());
auto userI_dst_memory = mkldnn::memory({user_dst_md, engine}, const_cast<NDArray*>(gradO)->buffer());
auto convI_src_memory = userI_src_memory;
streams[1].addMemory(userI_src_memory);
if (mkldnn::memory::primitive_desc(convI_prim_desc.diff_src_primitive_desc())
!= userI_src_memory.get_primitive_desc()) {
convI_src_memory = mkldnn::memory(convI_prim_desc.diff_src_primitive_desc());
streams[1].addMemory(convI_src_memory);
}
auto convI_weights_memory = userI_weights_memory;
streams[1].addMemory(userI_weights_memory);
if (mkldnn::memory::primitive_desc(convI_prim_desc.weights_primitive_desc())
!= userI_weights_memory.get_primitive_desc()) {
convI_weights_memory = mkldnn::memory(convI_prim_desc.weights_primitive_desc());
streams[1].addMemory(convI_weights_memory);
streams[1].addOperation(reorder(userI_weights_memory, convI_weights_memory));
}
auto convI_dst_memory = userI_dst_memory;
streams[1].addMemory(userI_dst_memory);
if (mkldnn::memory::primitive_desc(convI_prim_desc.diff_dst_primitive_desc())
!= userI_dst_memory.get_primitive_desc()) {
convI_dst_memory = mkldnn::memory(convI_prim_desc.diff_dst_primitive_desc());
streams[1].addMemory(convI_dst_memory);
streams[1].addOperation(reorder(userI_dst_memory, convI_dst_memory));
}
streams[1].addOperation(convolution_backward_data(convI_prim_desc, convI_dst_memory, convI_weights_memory, convI_src_memory));
if (mkldnn::memory::primitive_desc(convI_prim_desc.diff_src_primitive_desc())
!= userI_src_memory.get_primitive_desc()) {
streams[1].addOperation(reorder(convI_src_memory, userI_src_memory));
}
}
}
if (gradW != nullptr) {
streams[0].submitAndWait();
}
if (gradI != nullptr) {
streams[1].submitAndWait();
}
return;
}
#endif
nd4j_debug("MKL-DNN is not used for conv2d_bp!\n", 0);
std::vector<int> gradOaxesForDot;
if(!isNCHW) {
gradOaxesForDot = {0, 1, 2}; // bS, oH, oW
input = new NDArray(input->permute({0, 3, 1, 2})); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradI = new NDArray(gradI->permute({0, 3, 1, 2})); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
} else {
gradOaxesForDot = {0, 2, 3}; // bS, oH, oW
}
NDArray columns(input->ordering(), {bS, iC, kH, kW, oH, oW}, input->dataType(), input->getContext());
// ----- calculation of gradW ----- //
if(gradW) {
auto ctx = block.launchContext();
helpers::im2col(*ctx, *input, columns, kH, kW, sH, sW, pH, pW, dH, dW, NDArrayFactory::create(0.f, input->getContext())); // [bS, iC, iH, iW] is convoluted to [bS, iC, kH, kW, oH, oW]
nd4j::MmulHelper::tensorDot(&columns, gradO, gradW, {0,4,5}, gradOaxesForDot, {2, 0, 1, 3}); // [bS, iC, kH, kW, oH, oW] x [bS, oH, oW, oC]/[bS, oC, oH, oW] = [iC, kH, kW, oC]
}
// ----- calculation of gradB ----- //
if(gradB) {
NDArray* gradBR = gradB;
if(gradB->rankOf() == 2)
gradBR = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
gradO->reduceAlongDimension(reduce::Sum, gradBR, gradOaxesForDot); // sum over bS, oH, oW
if(gradBR != gradB)
delete gradBR;
}
//----- calculation of gradI -----//
nd4j::MmulHelper::tensorDot(weights, gradO, &columns, {indWoC}, {indIOioC}, {2, 3, 1, 0, 4, 5}); // [kH, kW, iC, oC]/[oC, iC, kH, kW]] x [bS, oH, oW, oC]/[bS, oC, oH, oW] = [kH, kW, iC, bS, oH, oW]
helpers::col2im(*block.launchContext(), columns, *gradI, sH, sW, pH, pW, iH, iW, dH, dW); // [bS, iC, kH, kW, oH, oW] is de-convoluted to [bS, iC, iH, iW]
if(!isNCHW) {
delete input;
delete gradI;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename X, typename Y>
static void depthwiseConv2d_(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 isSameMode, const int isNCHW) {
// input [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
// weights [kH, kW, iC, mC] always
// bias [oC] = iC*mC
// output [bS, oH, oW, iC*mC] (NHWC) or [bS, iC*mC, oH, oW] (NCHW)
// kH filter(kernel) height
// kW filter(kernel) width
// sH strides height
// sW strides width
// pH paddings height
// pW paddings width
// dH dilations height
// dW dilations width
// isSameMode 0-VALID, 1-SAME
// isNCHW 0-NCHW, 1-NHWC
int bS, iC, iH, iW, mC, oC, oH, oW; // batch size, input channels, input height/width, channels multiplier(oC = iC*mC), output channels, output height/width
int indIOioC, indIiH, indWmC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, *input, *output, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weights->sizeAt(indWmC); // channels multiplier
std::vector<std::vector<Nd4jLong>> modifColumns = {{1,0,4,5,2,3}, {iC,bS*oH*oW,kH*kW}}; // [bS,iC,kH,kW,oH,oW] -> [iC,bS,oH,oW,kH,kW] -> [iC,bS*oH*oW,kH*kW]
std::vector<std::vector<Nd4jLong>> modifOutput;
std::vector<Nd4jLong> outReShape;
if(!isNCHW) {
outReShape = {bS, oH, oW, iC, mC}; // [bS,oH,oW,iC*mC] -> [bS,oH,oW,iC,mC]
modifOutput = {{3,0,1,2,4},{iC, bS*oH*oW, mC}}; // [bS,oH,oW,iC,mC] -> [iC,bS,oH,oW,mC] -> [iC,bS*oH*oW,mC]
input = new NDArray(input->permute({0, 3, 1, 2})); // [bS,iH,iW,iC] -> [bS,iC,iH,iW]
}
else {
outReShape = {bS, iC, mC, oH, oW}; // [bS,iC*mC,oH,oW] -> [bS,iC,mC,oH,oW]
modifOutput = {{1,0,3,4,2},{iC, bS*oH*oW, mC}}; // [bS,iC,mC,oH,oW] -> [iC,bS,oH,oW,mC] -> [iC,bS*oH*oW,mC]
}
if(isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
NDArray columns(input->ordering(), {bS, iC, kH, kW, oH, oW}, input->dataType(), input->getContext());
NDArray outputReshaped = output->reshape(output->ordering(), outReShape);
helpers::im2col(*output->getContext(), *input, columns, kH, kW, sH, sW, pH, pW, dH, dW, NDArrayFactory::create(0.f, input->getContext())); // [bS, iC, iH, iW] is convoluted to [bS, iC, kH, kW, oH, oW]
MmulHelper::tensorDot(&columns, weights, &outputReshaped, modifColumns, {{2,0,1,3},{iC,kH*kW,mC}}, modifOutput); // [iC, bS*oH*oW, kW*kH] x [iC, kH*kW, mC] = [iC, bS*oH*oW, mC]
if(bias)
output->applyBroadcast(broadcast::Add, {indIOioC}, bias);
if(!isNCHW)
delete input;
}
//////////////////////////////////////////////////////////////////////////
template <typename X, typename Y>
static void depthwiseConv2dBP_(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 isSameMode, const int isNCHW) {
// input [bS, iH, iW, iC] (NDHWC) or [bS, iC, iH, iW] (NCDHW)
// weights [kH, kW, iC, mC] always
// bias [oC] = [iC*mC]
// gradO [bS, oH, oW, oC] (NDHWC) or [bS, oC, oH, oW] (NCDHW), epsilon_next
// gradI [bS, iH, iW, iC] (NDHWC) or [bS, iC, iH, iW] (NCDHW), epsilon
// gradW [kH, kW, iC, mC] always
// gradB [oC]
// kH filter(kernel) height
// kW filter(kernel) width
// sH strides height
// sW strides width
// pH paddings height
// pW paddings width
// dH dilations height
// dW dilations width
// isSameMode 0-VALID, 1-SAME
// isNCHW 0-NHWC, 1-NCHW
int bS, iC, iH, iW, mC, oC, oH, oW; // batch size, input channels, input height/width, channels multiplier(oC = iC*mC), output channels, output height/width
int indIOioC, indIiH, indWmC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weights->sizeAt(indWmC); // channels multiplier
std::vector<std::vector<Nd4jLong>> modifColumns = {{1,2,3,0,4,5}, {iC, kH*kW, bS*oH*oW}}; // [bS,iC,kH,kW,oH,oW] -> [iC, kH*kW, bS*oH*oW]
std::vector<std::vector<Nd4jLong>> modifGradO1, modifGradO2;
std::vector<Nd4jLong> gradOreShape;
if(!isNCHW) {
gradOreShape = {bS, oH, oW, iC, mC}; // [bS,oH,oW,iC*mC] -> [bS,oH,oW,iC,mC]
modifGradO1 = {{3,0,1,2,4},{iC, bS*oH*oW, mC}}; // [bS,oH,oW,iC,mC] -> [iC,bS,oH,oW,mC] -> [iC,bS*oH*oW,mC]
modifGradO2 = {{3,0,1,2},{iC, mC, bS*oH*oW}}; // [bS,oH,oW,iC*mC] -> [iC*mC,bS,oH,oW] -> [iC,mC,bS*oH*oW]
input = new NDArray(input->permute({0, 3, 1, 2})); // [bS,iH,iW,iC] -> [bS,iC,iH,iW]
gradI = new NDArray(gradI->permute({0, 3, 1, 2})); // [bS,iH,iW,iC] -> [bS,iC,iH,iW]
}
else {
gradOreShape = {bS, iC, mC, oH, oW}; // [bS,iC*mC,oH,oW] -> [bS,iC,mC,oH,oW]
modifGradO1 = {{1,0,3,4,2},{iC, bS*oH*oW, mC}}; // [bS,iC,mC,oH,oW] -> [iC,bS,oH,oW,mC] -> [iC,bS*oH*oW,mC]
modifGradO2 = {{1,0,2,3},{iC, mC, bS*oH*oW}}; // [bS,iC*mC,oH,oW] -> [iC*mC,bS,oH,oW] -> [iC,mC,bS*oH*oW]
}
if(isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
NDArray columns(input->ordering(), {bS, iC, kH, kW, oH, oW}, input->dataType(), input->getContext());
NDArray gradOreshaped = gradO->reshape(gradO->ordering(), gradOreShape);
// ----- calculation of gradW and gradB ----- //
helpers::im2col(*input->getContext(), *input, columns, kH, kW, sH, sW, pH, pW, dH, dW, NDArrayFactory::create(0.f, input->getContext())); // [bS, iC, iH, iW] is convoluted to [bS, iC, kH, kW, oH, oW]
nd4j::MmulHelper::tensorDot(&columns, &gradOreshaped, gradW, modifColumns, modifGradO1, {{2,0,1,3},{iC,kH*kW,mC}}); // [iC, kW*kH, bS*oH*oW] x [iC, bS*oH*oW, mC] = [iC, kH*kW, mC]
// ----- calculation of gradB ----- //
if(gradB) {
NDArray* gradBR = gradB;
if(gradB->rankOf() == 2)
gradBR = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
gradO->reduceAlongDimension(reduce::Sum, gradBR, {0,indOoH,indOoH+1}); // sum over bS, oH, oW
if(gradBR != gradB)
delete gradBR;
}
//----- calculation of gradI -----//
nd4j::MmulHelper::tensorDot(weights, gradO, &columns, {{2,0,1,3},{iC,kH*kW,mC}}, modifGradO2, modifColumns); // [iC, kH*kW, mC] x [iC, mC, bS*oH*oW] = [iC, kW*kH, bS*oH*oW]
helpers::col2im(*input->getContext(), columns, *gradI, sH, sW, pH, pW, iH, iW, dH, dW); // [bS, iC, kH, kW, oH, oW] is de-convoluted to [bS, iC, iH, iW]
if(!isNCHW) {
delete input;
delete gradI;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename X, typename Y>
static void sconv2d_(nd4j::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 isSameMode, const int isNCHW) {
// input [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
// weightsDepth [kH, kW, iC, mC] always
// weightsPoint [1, 1, iC*mC, oC] always
// bias [oC], oC = iC*mC if weightsPoint=nullptr
// output is [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW)
// kH filter(kernel) height
// kW filter(kernel) width
// sH strides height
// sW strides width
// pH paddings height
// pW paddings width
// dH dilations height
// dW dilations width
// isSameMode 0-VALID, 1-SAME
// isNCHW 1-NCHW, 0-NHWC
int bS, iC, iH, iW, mC, oC, oH, oW; // batch size, input channels, input height/width, channels multiplier, output channels, output height/width
int indIOioC, indIiH, indWmC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, *input, *output, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weightsDepth->sizeAt(indWmC); // channels multiplier
NDArray* outputDepth = output;
if(weightsPoint) // if pointwise convolution is expected
outputDepth = new NDArray(output->ordering(), !isNCHW ? std::vector<Nd4jLong>({bS, oH, oW, iC*mC}) : std::vector<Nd4jLong>({bS, iC*mC, oH, oW}), input->dataType(), input->getContext());
// ----- perform depthwise convolution (if weightsPoint is absent then oC = iC*mC) ----- //
ConvolutionUtils::depthwiseConv2d(block, input, weightsDepth, weightsPoint ? nullptr : bias, outputDepth, kH,kW, sH,sW, pH,pW, dH,dW, isSameMode, isNCHW);
// ----- perform pointwise convolution (oH = iH, oW = iW) ----- //
if (weightsPoint) {
ConvolutionUtils::conv2d(block, outputDepth, weightsPoint, bias, output, 1,1, 1,1, 0,0, 1,1, isSameMode, isNCHW); // in this case oH=iH, oW=iW
delete outputDepth;
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void upsampling2d_(const NDArray& input, NDArray& output, const int factorH, const int factorW, const bool isNCHW) {
// input has shape [bS, iC, iH, iW] (NCHW) or [bS, iH, iW, iC] (NHWC)
// output has shape [bS, iC, factorH*iH, factorW*iW ] (NCHW) or [bS, factorH*iH, factorW*iW, iC] (NHWC)
const T* x = input.bufferAsT<T>();
T* z = output.bufferAsT<T>();
const uint dimIH = isNCHW ? 2 : 1;
const uint dimIC = isNCHW ? 1 : 3;
const uint bS = input.sizeAt(0);
const uint iC = input.sizeAt(dimIC);
const uint oH = output.sizeAt(dimIH);
const uint oW = output.sizeAt(dimIH + 1);
const Nd4jLong xStride0 = input.stridesOf()[0];
const Nd4jLong xStride1 = input.stridesOf()[dimIC];
const Nd4jLong xStride2 = input.stridesOf()[dimIH];
const Nd4jLong xStride3 = input.stridesOf()[dimIH + 1];
const Nd4jLong zStride0 = output.stridesOf()[0];
const Nd4jLong zStride1 = output.stridesOf()[dimIC];
const Nd4jLong zStride2 = output.stridesOf()[dimIH];
const Nd4jLong zStride3 = output.stridesOf()[dimIH + 1];
uint xCoord2, xCoord3;
// loop through output array
PRAGMA_OMP_PARALLEL_FOR_ARGS(collapse(4) private(xCoord2, xCoord3))
for(uint b = 0; b < bS; ++b) {
for(uint c = 0; c < iC; ++c) {
for(uint h = 0; h < oH ; ++h) {
for(uint w = 0; w < oW ; ++w) {
xCoord2 = h / factorH;
xCoord3 = w / factorW;
z[b*zStride0 + c*zStride1 + h*zStride2 + w*zStride3] = x[b*xStride0 + c*xStride1 + xCoord2*xStride2 + xCoord3*xStride3];
}
}
}
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void upsampling3d_(const NDArray& input, NDArray& output, const int factorD, const int factorH, const int factorW, const bool isNCDHW) {
// input has shape [bS, iC, iD, iH, iW] (NCDHW) or [bS, iD, iH, iW, iC] (NDHWC)
// output has shape [bS, iC, factorD*iD, factorH*iH, factorW*iW ] (NCDHW) or [bS, factorD*iD, factorH*iH, factorW*iW, iC] (NDHWC)
const T* x = input.bufferAsT<T>();
T* z = output.bufferAsT<T>();
const uint dimID = isNCDHW ? 2 : 1;
const uint dimIC = isNCDHW ? 1 : 4;
const uint bS = input.sizeAt(0);
const uint iC = input.sizeAt(dimIC);
const uint oD = output.sizeAt(dimID);
const uint oH = output.sizeAt(dimID + 1);
const uint oW = output.sizeAt(dimID + 2);
const Nd4jLong xStride0 = input.stridesOf()[0];
const Nd4jLong xStride1 = input.stridesOf()[dimIC];
const Nd4jLong xStride2 = input.stridesOf()[dimID];
const Nd4jLong xStride3 = input.stridesOf()[dimID + 1];
const Nd4jLong xStride4 = input.stridesOf()[dimID + 2];
const Nd4jLong zStride0 = output.stridesOf()[0];
const Nd4jLong zStride1 = output.stridesOf()[dimIC];
const Nd4jLong zStride2 = output.stridesOf()[dimID];
const Nd4jLong zStride3 = output.stridesOf()[dimID + 1];
const Nd4jLong zStride4 = output.stridesOf()[dimID + 2];
uint xCoord2, xCoord3, xCoord4;
// loop through output array
PRAGMA_OMP_PARALLEL_FOR_ARGS(collapse(5) private(xCoord2, xCoord3, xCoord4))
for(uint b = 0; b < bS; ++b) {
for(uint c = 0; c < iC; ++c) {
for(uint d = 0; d < oD ; ++d) {
for(uint h = 0; h < oH ; ++h) {
for(uint w = 0; w < oW ; ++w) {
xCoord2 = d / factorD;
xCoord3 = h / factorH;
xCoord4 = w / factorW;
z[b*zStride0 + c*zStride1 + d*zStride2 + h*zStride3 + w*zStride4] = x[b*xStride0 + c*xStride1 + xCoord2*xStride2 + xCoord3*xStride3 + xCoord4*xStride4];
}
}
}
}
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void upsampling2dBP_(const NDArray& gradO, NDArray& gradI, const bool isNCHW) {
// gradO has shape [bS, iC, factorH*iH, factorW*iW ] (NCHW) or [bS, factorH*iH, factorW*iW, iC] (NHWC)
// gradI has shape [bS, iC, iH, iW] (NCHW) or [bS, iH, iW, iC] (NHWC)
const T* x = gradO.bufferAsT<T>();
T* z = gradI.bufferAsT<T>();
const uint dimIH = isNCHW ? 2 : 1;
const uint dimIC = isNCHW ? 1 : 3;
const uint bS = gradI.sizeAt(0);
const uint iC = gradI.sizeAt(dimIC);
const uint iH = gradI.sizeAt(dimIH);
const uint iW = gradI.sizeAt(dimIH + 1);
const uint factorH = gradO.sizeAt(dimIH) / iH;
const uint factorW = gradO.sizeAt(dimIH + 1) / iW;
const Nd4jLong xStride0 = gradO.stridesOf()[0];
const Nd4jLong xStride1 = gradO.stridesOf()[dimIC];
const Nd4jLong xStride2 = gradO.stridesOf()[dimIH];
const Nd4jLong xStride3 = gradO.stridesOf()[dimIH + 1];
const Nd4jLong zStride0 = gradI.stridesOf()[0];
const Nd4jLong zStride1 = gradI.stridesOf()[dimIC];
const Nd4jLong zStride2 = gradI.stridesOf()[dimIH];
const Nd4jLong zStride3 = gradI.stridesOf()[dimIH + 1];
// loop through output array
PRAGMA_OMP_PARALLEL_FOR_ARGS(collapse(4))
for(uint b = 0; b < bS; ++b) {
for(uint c = 0; c < iC; ++c) {
for(uint h = 0; h < iH; ++h) {
for(uint w = 0; w < iW; ++w) {
const auto zOffset = b*zStride0 + c*zStride1 + h*zStride2 + w*zStride3;
z[zOffset] = 0;
for(uint xh = h * factorH; xh < h * factorH + factorH; ++xh)
for(uint xw = w * factorW; xw < w * factorW + factorW; ++xw)
z[zOffset] += x[b*xStride0 + c*xStride1 + xh*xStride2 + xw*xStride3];
}
}
}
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void upsampling3dBP_(const NDArray& gradO, NDArray& gradI, const bool isNCDHW) {
// input has shape [bS, iC, iD, iH, iW] (NCDHW) or [bS, iD, iH, iW, iC] (NDHWC)
// output has shape [bS, iC, factorD*iD, factorH*iH, factorW*iW ] (NCDHW) or [bS, factorD*iD, factorH*iH, factorW*iW, iC] (NDHWC)
const T* x = gradO.bufferAsT<T>();
T* z = gradI.bufferAsT<T>();
const uint dimID = isNCDHW ? 2 : 1;
const uint dimIC = isNCDHW ? 1 : 4;
const uint bS = gradI.sizeAt(0);
const uint iC = gradI.sizeAt(dimIC);
const uint iD = gradI.sizeAt(dimID);
const uint iH = gradI.sizeAt(dimID + 1);
const uint iW = gradI.sizeAt(dimID + 2);
const uint factorD = gradO.sizeAt(dimID) / iD;
const uint factorH = gradO.sizeAt(dimID + 1) / iH;
const uint factorW = gradO.sizeAt(dimID + 2) / iW;
const Nd4jLong xStride0 = gradO.stridesOf()[0];
const Nd4jLong xStride1 = gradO.stridesOf()[dimIC];
const Nd4jLong xStride2 = gradO.stridesOf()[dimID];
const Nd4jLong xStride3 = gradO.stridesOf()[dimID + 1];
const Nd4jLong xStride4 = gradO.stridesOf()[dimID + 2];
const Nd4jLong zStride0 = gradI.stridesOf()[0];
const Nd4jLong zStride1 = gradI.stridesOf()[dimIC];
const Nd4jLong zStride2 = gradI.stridesOf()[dimID];
const Nd4jLong zStride3 = gradI.stridesOf()[dimID + 1];
const Nd4jLong zStride4 = gradI.stridesOf()[dimID + 2];
// loop through output array
PRAGMA_OMP_PARALLEL_FOR_ARGS(collapse(5))
for(uint b = 0; b < bS; ++b) {
for(uint c = 0; c < iC; ++c) {
for(uint d = 0; d < iD; ++d) {
for(uint h = 0; h < iH; ++h) {
for(uint w = 0; w < iW; ++w) {
const auto zOffset = b*zStride0 + c*zStride1 + d*zStride2 + h*zStride3 + w*zStride4;
z[zOffset] = 0;
for(uint xd = d * factorD; xd < d * factorD + factorD; ++xd)
for(uint xh = h * factorH; xh < h * factorH + factorH; ++xh)
for(uint xw = w * factorW; xw < w * factorW + factorW; ++xw)
z[zOffset] += x[b*xStride0 + c*xStride1 + xd*xStride2 + xh*xStride3 + xw*xStride4];
}
}
}
}
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void pooling2d_(nd4j::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 int poolingMode, const int extraParam0) {
// input is [bS, iC, iH, iW]
// output is [bS, iC, oH, oW]
T* out = output.bufferAsT<T>();
T* in = const_cast<NDArray&>(input).bufferAsT<T>();
const int kHEff = kH + (kH-1)*(dH-1);
const int kWEff = kW + (kW-1)*(dW-1);
const int bS = input.sizeAt(0);
const int iC = input.sizeAt(1);
const int iH = input.sizeAt(2);
const int iW = input.sizeAt(3);
const int oC = output.sizeAt(1);
const int oH = output.sizeAt(2);
const int oW = output.sizeAt(3);
#ifdef HAVE_MKLDNN
if (poolingMode < 2 && block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported<T, T>()) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("pooling2d"));
}
if (streams[0].checkAndReset({&input}, {&output}, {}, {kH, kW, sH, sW, pH, pW, dH, dW, poolingMode, extraParam0})) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc pool_src_md(empty), pool_dst_md(empty);
mkldnn::memory::desc user_src_md(empty), user_dst_md(empty);
mkldnn::memory::dims pool_strides, pool_kernel, pool_padding, pool_padding_r;
mkldnn::algorithm algorithm;
ConvolutionUtils::getMKLDNNMemoryDescPool2d(kH, kW, sH, sW, pH, pW, dH, dW, poolingMode, extraParam0, true,
bS, iC, iH, iW, oC, oH, oW, &input, nullptr, &output, algorithm,
&pool_src_md, nullptr, &pool_dst_md, &user_src_md, nullptr, &user_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r);
auto pool_desc = pooling_forward::desc(prop_kind::forward_inference, algorithm, pool_src_md, pool_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r, padding_kind::zero);
auto engine = streams[0].getEngine();
auto pool_prim_desc = pooling_forward::primitive_desc(pool_desc, engine);
auto user_src_memory = mkldnn::memory({user_src_md, engine}, const_cast<NDArray&>(input).buffer());
auto user_dst_memory = mkldnn::memory({user_dst_md, engine}, output.buffer());
auto pool_src_memory = user_src_memory;
streams[0].addMemory(user_src_memory);
if (mkldnn::memory::primitive_desc(pool_prim_desc.src_primitive_desc())
!= user_src_memory.get_primitive_desc()) {
pool_src_memory = mkldnn::memory(pool_prim_desc.src_primitive_desc());
streams[0].addMemory(pool_src_memory);
streams[0].addOperation(reorder(user_src_memory, pool_src_memory));
}
auto pool_dst_memory = user_dst_memory;
streams[0].addMemory(user_dst_memory);
if (mkldnn::memory::primitive_desc(pool_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
pool_dst_memory = mkldnn::memory(pool_prim_desc.dst_primitive_desc());
streams[0].addMemory(pool_dst_memory);
}
streams[0].addOperation(pooling_forward(pool_prim_desc, pool_src_memory, pool_dst_memory));
if (mkldnn::memory::primitive_desc(pool_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(pool_dst_memory, user_dst_memory));
}
}
streams[0].submitAndWait();
return;
}
#endif
nd4j_debug("MKL-DNN is not used for pooling2d!\n", 0);
const Nd4jLong iStride0 = input.stridesOf()[0];
const Nd4jLong iStride1 = input.stridesOf()[1];
const Nd4jLong iStride2 = input.stridesOf()[2];
const Nd4jLong iStride3 = input.stridesOf()[3];
const Nd4jLong oStride0 = output.stridesOf()[0];
const Nd4jLong oStride1 = output.stridesOf()[1];
const Nd4jLong oStride2 = output.stridesOf()[2];
const Nd4jLong oStride3 = output.stridesOf()[3];
const Nd4jLong iStep2 = dH*iStride2;
const Nd4jLong iStep3 = dW*iStride3;
const int kProd = kH*kW;
Nd4jLong hstart, wstart, hend, wend;
T *pIn;
if(poolingMode == 0) { // max
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, hstart, wstart, hend, wend) collapse(2))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH); // (Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-hstart) / static_cast<T>(dH));
if(wstart < 0)
wstart += dW * ((-wstart + dW -1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-wstart) / static_cast<T>(dW));
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(hend-iH) / static_cast<T>(dH));
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(wend-iW) / static_cast<T>(dW));
hstart *= iStride2;
hend *= iStride2;
wstart *= iStride3;
wend *= iStride3;
T max = -DataTypeUtils::max<T>();
for (Nd4jLong kh = hstart; kh < hend; kh += iStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep3) {
T val = pIn[kh + kw];
if (val > max)
max = val;
}
out[b * oStride0 + c * oStride1 + oh * oStride2 + ow * oStride3] = max;
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 1) { // avg
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, hstart, wstart, hend, wend) collapse(2))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH); // (Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-hstart) / static_cast<T>(dH));
if(wstart < 0)
wstart += dW * ((-wstart + dW -1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-wstart) / static_cast<T>(dW));
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(hend-iH) / static_cast<T>(dH));
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(wend-iW) / static_cast<T>(dW));
hstart *= iStride2;
hend *= iStride2;
wstart *= iStride3;
wend *= iStride3;
T sum = static_cast<T>(0.f);
for (Nd4jLong kh = hstart; kh < hend; kh += iStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep3)
sum += pIn[kh + kw];
if (extraParam0 == 0) { //Exclude padding
int a = (hend-hstart)/iStep2 + ((hend-hstart) % iStep2 == 0 ? 0 : 1);
int b = (wend-wstart)/iStep3 + ((wend-wstart) % iStep3 == 0 ? 0 : 1);
sum /= static_cast<T>(a * b); // Accounts for dilation
}
else if (extraParam0 == 1) //Include padding
sum /= kProd;
out[b * oStride0 + c * oStride1 + oh * oStride2 + ow * oStride3] = sum;
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 2) { // pnorm
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, hstart, wstart, hend, wend) collapse(2))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH); // (Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-hstart) / static_cast<T>(dH));
if(wstart < 0)
wstart += dW * ((-wstart + dW -1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-wstart) / static_cast<T>(dW));
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(hend-iH) / static_cast<T>(dH));
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(wend-iW) / static_cast<T>(dW));
hstart *= iStride2;
hend *= iStride2;
wstart *= iStride3;
wend *= iStride3;
T sum = static_cast<T>(0.f);
for (Nd4jLong kh = hstart; kh < hend; kh += iStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep3)
sum += nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kh + kw]), extraParam0);
sum = nd4j::math::nd4j_pow<T,T,T>(sum, static_cast<T>((T)1.f) / extraParam0);
out[b * oStride0 + c * oStride1 + oh * oStride2 + ow * oStride3] = sum;
}
}
}
}
}
else {
nd4j_printf("ConvolutionUtils::pooling2d: pooling mode argument can take three values only: 0, 1, 2, but got %i instead !\n", poolingMode);
throw "";
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void pooling3d_(nd4j::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) {
// input is [bS, iC, iD, iH, iW]
// output is [bS, iC, oD, oH, oW]
T* out = output.bufferAsT<T>();
T* in = const_cast<NDArray&>(input).bufferAsT<T>();
const int kDEff = kD + (kD-1)*(dD-1);
const int kHEff = kH + (kH-1)*(dH-1);
const int kWEff = kW + (kW-1)*(dW-1);
const int bS = input.sizeAt(0);
const int iC = input.sizeAt(1);
const int iD = input.sizeAt(2);
const int iH = input.sizeAt(3);
const int iW = input.sizeAt(4);
const int oC = output.sizeAt(1);
const int oD = output.sizeAt(2);
const int oH = output.sizeAt(3);
const int oW = output.sizeAt(4);
#ifdef HAVE_MKLDNN
if (poolingMode < 2 && block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported<T, T>()) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("pooling3d"));
}
if (streams[0].checkAndReset({&input}, {&output}, {}, {kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, poolingMode, extraParam0})) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc pool_src_md(empty), pool_dst_md(empty);
mkldnn::memory::desc user_src_md(empty), user_dst_md(empty);
mkldnn::memory::dims pool_strides, pool_kernel, pool_padding, pool_padding_r;
mkldnn::algorithm algorithm;
ConvolutionUtils::getMKLDNNMemoryDescPool3d(kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, poolingMode, extraParam0, true,
bS, iC, iD, iH, iW, oC, oD, oH, oW, &input, nullptr, &output, algorithm,
&pool_src_md, nullptr, &pool_dst_md, &user_src_md, nullptr, &user_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r);
auto pool_desc = pooling_forward::desc(prop_kind::forward_inference, algorithm, pool_src_md, pool_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r, padding_kind::zero);
auto engine = streams[0].getEngine();
auto pool_prim_desc = pooling_forward::primitive_desc(pool_desc, engine);
auto user_src_memory = mkldnn::memory({user_src_md, engine}, const_cast<NDArray&>(input).buffer());
auto user_dst_memory = mkldnn::memory({user_dst_md, engine}, output.buffer());
auto pool_src_memory = user_src_memory;
streams[0].addMemory(user_src_memory);
if (mkldnn::memory::primitive_desc(pool_prim_desc.src_primitive_desc())
!= user_src_memory.get_primitive_desc()) {
pool_src_memory = mkldnn::memory(pool_prim_desc.src_primitive_desc());
streams[0].addMemory(pool_src_memory);
streams[0].addOperation(reorder(user_src_memory, pool_src_memory));
}
auto pool_dst_memory = user_dst_memory;
streams[0].addMemory(user_dst_memory);
if (mkldnn::memory::primitive_desc(pool_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
pool_dst_memory = mkldnn::memory(pool_prim_desc.dst_primitive_desc());
streams[0].addMemory(pool_dst_memory);
}
streams[0].addOperation(pooling_forward(pool_prim_desc, pool_src_memory, pool_dst_memory));
if (mkldnn::memory::primitive_desc(pool_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(pool_dst_memory, user_dst_memory));
}
}
streams[0].submitAndWait();
return;
}
#endif
nd4j_debug("MKL-DNN is not used for pooling3d!\n", 0);
const Nd4jLong iStride0 = input.stridesOf()[0];
const Nd4jLong iStride1 = input.stridesOf()[1];
const Nd4jLong iStride2 = input.stridesOf()[2];
const Nd4jLong iStride3 = input.stridesOf()[3];
const Nd4jLong iStride4 = input.stridesOf()[4];
const Nd4jLong oStride0 = output.stridesOf()[0];
const Nd4jLong oStride1 = output.stridesOf()[1];
const Nd4jLong oStride2 = output.stridesOf()[2];
const Nd4jLong oStride3 = output.stridesOf()[3];
const Nd4jLong oStride4 = output.stridesOf()[4];
const Nd4jLong iStep2 = dD*iStride2;
const Nd4jLong iStep3 = dH*iStride3;
const Nd4jLong iStep4 = dW*iStride4;
const int kProd = kD*kH*kW;
Nd4jLong dstart, hstart, wstart, dend, hend, wend;
T sum, *pIn;
if(poolingMode == 0) { // max
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, sum, dstart, hstart, wstart, dend, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int od = 0; od < oD; ++od) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
dstart = od * sD - pD;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
dend = dstart + kDEff;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(dstart < 0)
dstart += dD * ((-dstart + dD - 1) / dD);
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH);
if(wstart < 0)
wstart += dW * ((-wstart + dW - 1) / dW);
if(dend > iD)
dend -= dD * ((dend-iD + dD - 1) / dD);
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH);
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW);
dstart *= iStride2;
dend *= iStride2;
hstart *= iStride3;
hend *= iStride3;
wstart *= iStride4;
wend *= iStride4;
sum = -DataTypeUtils::max<T>();
for (Nd4jLong kd = dstart; kd < dend; kd += iStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += iStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep4) {
T val = pIn[kd + kh + kw];
if (val > sum)
sum = val;
}
out[b * oStride0 + c * oStride1 + od * oStride2 + oh * oStride3 + ow * oStride4] = sum;
}
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 1) { // avg
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, sum, dstart, hstart, wstart, dend, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int od = 0; od < oD; ++od) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
dstart = od * sD - pD;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
dend = dstart + kDEff;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(dstart < 0)
dstart += dD * ((-dstart + dD - 1) / dD);
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH);
if(wstart < 0)
wstart += dW * ((-wstart + dW - 1) / dW);
if(dend > iD)
dend -= dD * ((dend-iD + dD - 1) / dD);
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH);
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW);
dstart *= iStride2;
dend *= iStride2;
hstart *= iStride3;
hend *= iStride3;
wstart *= iStride4;
wend *= iStride4;
sum = static_cast<T>(0.);
for (Nd4jLong kd = dstart; kd < dend; kd += iStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += iStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep4)
sum += pIn[kd + kh + kw];
if (extraParam0 == 0) //Exclude padding
sum /= nd4j::math::nd4j_ceil<double,T>(static_cast<double>(dend-dstart) / static_cast<double>(iStep2)) * nd4j::math::nd4j_ceil<double,T>(static_cast<double>(hend-hstart) / static_cast<double>(iStep3)) * nd4j::math::nd4j_ceil<double,T>(static_cast<double>(wend-wstart) / static_cast<double>(iStep4)); //Accounts for dilation
else if (extraParam0 == 1) //Include padding
sum /= kProd;
out[b * oStride0 + c * oStride1 + od * oStride2 + oh * oStride3 + ow * oStride4] = sum;
}
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 2) { // pnorm
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, sum, dstart, hstart, wstart, dend, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int od = 0; od < oD; ++od) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
dstart = od * sD - pD;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
dend = dstart + kDEff;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(dstart < 0)
dstart += dD * ((-dstart + dD - 1) / dD);
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH);
if(wstart < 0)
wstart += dW * ((-wstart + dW - 1) / dW);
if(dend > iD)
dend -= dD * ((dend-iD + dD - 1) / dD);
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH);
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW);
dstart *= iStride2;
dend *= iStride2;
hstart *= iStride3;
hend *= iStride3;
wstart *= iStride4;
wend *= iStride4;
sum = static_cast<T>(0.);
for (Nd4jLong kd = dstart; kd < dend; kd += iStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += iStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep4)
sum += nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kd + kh + kw]), extraParam0);
sum = nd4j::math::nd4j_pow<T,T,T>(sum, (T) 1.f / extraParam0);
out[b * oStride0 + c * oStride1 + od * oStride2 + oh * oStride3 + ow * oStride4] = sum;
}
}
}
}
}
}
else {
nd4j_printf("ConvolutionUtils::pooling3d: pooling mode argument can take three values only: 0, 1, 2, but got %i instead !\n", poolingMode);
throw "";
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void pooling2dBP_(nd4j::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) {
// input [bS, iC, iH, iW]
// gradI [bS, iC, iH, iW] -> gradI is output in this function
// gradO [bS, iC, oH, oW]
// initial zeroing of gradI
gradI.nullify();
T* in = const_cast<NDArray&>(input).bufferAsT<T>();
T* gO = const_cast<NDArray&>(gradO).bufferAsT<T>();
T* gI = gradI.bufferAsT<T>();
const int kHEff = kH + (kH-1)*(dH-1);
const int kWEff = kW + (kW-1)*(dW-1);
const int bS = gradI.sizeAt(0);
const int iC = gradI.sizeAt(1);
const int iH = gradI.sizeAt(2);
const int iW = gradI.sizeAt(3);
const int oC = gradO.sizeAt(1);
const int oH = gradO.sizeAt(2);
const int oW = gradO.sizeAt(3);
#ifdef HAVE_MKLDNN
if (poolingMode < 2 && block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported<T, T>()) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("pooling2d_bp"));
}
if (streams[0].checkAndReset({&input, &gradO}, {&gradI}, {}, {kH, kW, sH, sW, pH, pW, dH, dW, poolingMode, extraParam0})) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc pool_src_md(empty), pool_diff_src_md(empty), pool_dst_md(empty);
mkldnn::memory::desc user_src_md(empty), user_diff_src_md(empty), user_dst_md(empty);
mkldnn::memory::dims pool_strides, pool_kernel, pool_padding, pool_padding_r;
mkldnn::algorithm algorithm;
ConvolutionUtils::getMKLDNNMemoryDescPool2d(kH, kW, sH, sW, pH, pW, dH, dW, poolingMode, extraParam0, true,
bS, iC, iH, iW, oC, oH, oW, &input, &gradI, &gradO, algorithm,
&pool_src_md, &pool_diff_src_md, &pool_dst_md, &user_src_md, &user_diff_src_md, &user_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r);
// input is sometimes null, so we can't rely on pool_src_md being valid
auto pool_desc = pooling_forward::desc(prop_kind::forward, algorithm,
const_cast<NDArray&>(input).buffer() != nullptr ? pool_src_md : pool_diff_src_md,
pool_dst_md, pool_strides, pool_kernel, pool_padding, pool_padding_r, padding_kind::zero);
auto engine = streams[0].getEngine();
auto pool_prim_desc = pooling_forward::primitive_desc(pool_desc, engine);
auto poolB_desc = pooling_backward::desc(algorithm, pool_diff_src_md, pool_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r, padding_kind::zero);
auto poolB_prim_desc = pooling_backward::primitive_desc(poolB_desc, engine, pool_prim_desc);
auto userB_src_memory = mkldnn::memory({user_src_md, engine}, gradI.buffer());
auto userB_dst_memory = mkldnn::memory({user_dst_md, engine}, const_cast<NDArray&>(gradO).buffer());
auto poolB_src_memory = userB_src_memory;
streams[0].addMemory(userB_src_memory);
if (mkldnn::memory::primitive_desc(poolB_prim_desc.diff_src_primitive_desc())
!= userB_src_memory.get_primitive_desc()) {
poolB_src_memory = mkldnn::memory(poolB_prim_desc.diff_src_primitive_desc());
streams[0].addMemory(poolB_src_memory);
}
auto poolB_dst_memory = userB_dst_memory;
streams[0].addMemory(userB_dst_memory);
if (mkldnn::memory::primitive_desc(poolB_prim_desc.diff_dst_primitive_desc())
!= userB_dst_memory.get_primitive_desc()) {
poolB_dst_memory = mkldnn::memory(poolB_prim_desc.diff_dst_primitive_desc());
streams[0].addMemory(poolB_dst_memory);
streams[0].addOperation(reorder(userB_dst_memory, poolB_dst_memory));
}
if (algorithm == mkldnn::pooling_max) {
auto user_src_memory = mkldnn::memory({user_src_md, engine}, const_cast<NDArray&>(input).buffer());
auto pool_src_memory = user_src_memory;
streams[0].addMemory(user_src_memory);
if (mkldnn::memory::primitive_desc(pool_prim_desc.src_primitive_desc())
!= user_src_memory.get_primitive_desc()) {
pool_src_memory = mkldnn::memory(pool_prim_desc.src_primitive_desc());
streams[0].addMemory(pool_src_memory);
streams[0].addOperation(reorder(user_src_memory, pool_src_memory));
}
auto pool_dst_memory = mkldnn::memory(pool_prim_desc.dst_primitive_desc());
streams[0].addMemory(pool_dst_memory);
auto pool_workspace_memory = mkldnn::memory(pool_prim_desc.workspace_primitive_desc());
streams[0].addMemory(pool_workspace_memory);
streams[0].addOperation(pooling_forward(pool_prim_desc, pool_src_memory, pool_dst_memory, pool_workspace_memory));
streams[0].addOperation(pooling_backward(poolB_prim_desc, poolB_dst_memory, pool_workspace_memory, poolB_src_memory));
} else {
streams[0].addOperation(pooling_backward(poolB_prim_desc, poolB_dst_memory, poolB_src_memory));
}
if (mkldnn::memory::primitive_desc(poolB_prim_desc.diff_src_primitive_desc())
!= userB_src_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(poolB_src_memory, userB_src_memory));
}
}
streams[0].submitAndWait();
return;
}
#endif
nd4j_debug("MKL-DNN is not used for pooling2d_bp!\n", 0);
const Nd4jLong iStride0 = input.stridesOf()[0];
const Nd4jLong iStride1 = input.stridesOf()[1];
const Nd4jLong iStride2 = input.stridesOf()[2];
const Nd4jLong iStride3 = input.stridesOf()[3];
const Nd4jLong gIStride0 = gradI.stridesOf()[0];
const Nd4jLong gIStride1 = gradI.stridesOf()[1];
const Nd4jLong gIStride2 = gradI.stridesOf()[2];
const Nd4jLong gIStride3 = gradI.stridesOf()[3];
const Nd4jLong oStride0 = gradO.stridesOf()[0];
const Nd4jLong oStride1 = gradO.stridesOf()[1];
const Nd4jLong oStride2 = gradO.stridesOf()[2];
const Nd4jLong oStride3 = gradO.stridesOf()[3];
const Nd4jLong iStep2 = dH*iStride2;
const Nd4jLong iStep3 = dW*iStride3;
const Nd4jLong gIStep2 = dH*gIStride2;
const Nd4jLong gIStep3 = dW*gIStride3;
const int kProd = kH*kW;
const bool sameStrides = iStride0 == gIStride0 && iStride1 == gIStride1 && iStride2 == gIStride2 && iStride3 == gIStride3;
Nd4jLong hstart, wstart,hend, wend, maxKH, maxKW;
T sum, valO, *pIn, *pgI;
if(poolingMode == 0) { // max
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, valO, sum, hstart, wstart, hend, wend, maxKH, maxKW))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH); // (Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-hstart) / static_cast<T>(dH));
if(wstart < 0)
wstart += dW * ((-wstart + dW -1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-wstart) / static_cast<T>(dW));
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(hend-iH) / static_cast<T>(dH));
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(wend-iW) / static_cast<T>(dW));
sum = -DataTypeUtils::max<T>();
valO = gO[b*oStride0 + c*oStride1 + oh*oStride2 + ow*oStride3];
if(sameStrides) {
hstart *= iStride2;
hend *= iStride2;
wstart *= iStride3;
wend *= iStride3;
// we set these to default values
maxKH = hstart;
maxKW = wstart;
for (Nd4jLong kh = hstart; kh < hend; kh += iStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep3) {
T valIn = pIn[kh + kw];
if (valIn > sum) {
sum = valIn;
maxKH = kh;
maxKW = kw;
}
}
gI[pIn - in + maxKH + maxKW] += valO;
}
else {
// we set these to default values
maxKH = hstart;
maxKW = wstart;
for (Nd4jLong kh = hstart; kh < hend; kh += dH)
for (Nd4jLong kw = wstart; kw < wend; kw += dW) {
T valIn = pIn[kh * iStride2 + kw * iStride3];
if (valIn > sum) {
sum = valIn;
maxKH = kh;
maxKW = kw;
}
}
gI[b * gIStride0 + c * gIStride1 + maxKH * gIStride2 + maxKW * gIStride3] += valO;
}
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 1) { // avg
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pgI, valO, hstart, wstart, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pgI = gI + b * gIStride0 + c * gIStride1;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH); // (Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-hstart) / static_cast<T>(dH));
if(wstart < 0)
wstart += dW * ((-wstart + dW -1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-wstart) / static_cast<T>(dW));
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(hend-iH) / static_cast<T>(dH));
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(wend-iW) / static_cast<T>(dW));
hstart *= gIStride2;
hend *= gIStride2;
wstart *= gIStride3;
wend *= gIStride3;
valO = gO[b*oStride0 + c*oStride1 + oh*oStride2 + ow*oStride3];
if ((int) extraParam0 == 0) //Exclude padding
valO /= static_cast<T>(nd4j::math::nd4j_ceil<double,T>(static_cast<double>(hend-hstart) / static_cast<double>(gIStep2))) * static_cast<T>(nd4j::math::nd4j_ceil<double,T>(static_cast<double>(wend-wstart) / static_cast<double>(gIStep3))); //Accounts for dilation
else if ((int) extraParam0 == 1) //Include padding
valO /= kProd;
for (Nd4jLong kh = hstart; kh < hend; kh += gIStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += gIStep3)
pgI[kh + kw] += valO;
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 2) { // pnorm
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, valO, pgI, sum, hstart, wstart, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
pgI = sameStrides ? gI + (pIn - in) : gI + b * gIStride0 + c * gIStride1;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH); // (Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-hstart) / static_cast<T>(dH));
if(wstart < 0)
wstart += dW * ((-wstart + dW -1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(-wstart) / static_cast<T>(dW));
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(hend-iH) / static_cast<T>(dH));
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW); //(Nd4jLong)nd4j::math::nd4j_ceil<T,T>(static_cast<T>(wend-iW) / static_cast<T>(dW));
sum = static_cast<T>(0.f);
valO = gO[b*oStride0 + c*oStride1 + oh*oStride2 + ow*oStride3];
if(sameStrides) {
hstart *= iStride2;
hend *= iStride2;
wstart *= iStride3;
wend *= iStride3;
for (Nd4jLong kh = hstart; kh < hend; kh += iStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep3)
sum += nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kh + kw]), extraParam0);
valO *= nd4j::math::nd4j_pow<T,T,T>(sum, ((T)1. - extraParam0) / extraParam0);
for (Nd4jLong kh = hstart; kh < hend; kh += iStep2)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep3)
pgI[kh + kw] += valO * nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kh + kw]), extraParam0 - 1.f) * nd4j::math::nd4j_sgn<T,T>(pIn[kh + kw]);
}
else {
for (Nd4jLong kh = hstart; kh < hend; kh += dH)
for (Nd4jLong kw = wstart; kw < wend; kw += dW)
sum += nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kh * iStride2 + kw * iStride3]), extraParam0);
valO *= nd4j::math::nd4j_pow<T,T,T>(sum, ((T)1. - extraParam0) / extraParam0);
for (Nd4jLong kh = hstart; kh < hend; kh += dH) {
for (Nd4jLong kw = wstart; kw < wend; kw += dW) {
const auto inVal = pIn[kh * iStride2 + kw * iStride3];
pgI[kh * gIStride2 + kw * gIStride3] += valO * nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(inVal), extraParam0 - 1.f) * nd4j::math::nd4j_sgn<T,T>(inVal);
}
}
}
}
}
}
}
}
else {
nd4j_printf("ConvolutionUtils::pooling2dBP: pooling mode argument can take three values only: 0, 1, 2, but got %i instead !\n", poolingMode);
throw "";
}
}
//////////////////////////////////////////////////////////////////////////
template <typename T>
static void pooling3dBP_(nd4j::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) {
// input [bS, iC, iD, iH, iW]
// gradI [bS, iC, iD, iH, iW] -> gradI is output in this function
// gradO [bS, iC, oD, oH, oW]
// initial zeroing of gradI
gradI.nullify();
T* in = const_cast<NDArray&>(input).bufferAsT<T>();
T* gO = const_cast<NDArray&>(gradO).bufferAsT<T>();
T* gI = gradI.bufferAsT<T>();
const int kDEff = kD + (kD-1)*(dD-1);
const int kHEff = kH + (kH-1)*(dH-1);
const int kWEff = kW + (kW-1)*(dW-1);
const int bS = gradI.sizeAt(0);
const int iC = gradI.sizeAt(1);
const int iD = gradI.sizeAt(2);
const int iH = gradI.sizeAt(3);
const int iW = gradI.sizeAt(4);
const int oC = gradO.sizeAt(1);
const int oD = gradO.sizeAt(2);
const int oH = gradO.sizeAt(3);
const int oW = gradO.sizeAt(4);
#ifdef HAVE_MKLDNN
if (poolingMode < 2 && block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported<T, T>()) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("pooling3d_bp"));
}
if (streams[0].checkAndReset({&input, &gradO}, {&gradI}, {}, {kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, poolingMode, extraParam0})) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc pool_src_md(empty), pool_diff_src_md(empty), pool_dst_md(empty);
mkldnn::memory::desc user_src_md(empty), user_diff_src_md(empty), user_dst_md(empty);
mkldnn::memory::dims pool_strides, pool_kernel, pool_padding, pool_padding_r;
mkldnn::algorithm algorithm;
ConvolutionUtils::getMKLDNNMemoryDescPool3d(kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, poolingMode, extraParam0, true,
bS, iC, iD, iH, iW, oC, oD, oH, oW, &input, &gradI, &gradO, algorithm,
&pool_src_md, &pool_diff_src_md, &pool_dst_md, &user_src_md, &user_diff_src_md, &user_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r);
// input is sometimes null, so we can't rely on pool_src_md being valid
if (const_cast<NDArray&>(input).buffer() == nullptr) {
pool_src_md = pool_diff_src_md;
user_src_md = user_diff_src_md;
}
auto pool_desc = pooling_forward::desc(prop_kind::forward, algorithm, pool_src_md,
pool_dst_md, pool_strides, pool_kernel, pool_padding, pool_padding_r, padding_kind::zero);
auto engine = streams[0].getEngine();
auto pool_prim_desc = pooling_forward::primitive_desc(pool_desc, engine);
auto poolB_desc = pooling_backward::desc(algorithm, pool_diff_src_md, pool_dst_md,
pool_strides, pool_kernel, pool_padding, pool_padding_r, padding_kind::zero);
auto poolB_prim_desc = pooling_backward::primitive_desc(poolB_desc, engine, pool_prim_desc);
auto userB_src_memory = mkldnn::memory({user_diff_src_md, engine}, gradI.buffer());
auto userB_dst_memory = mkldnn::memory({user_dst_md, engine}, const_cast<NDArray&>(gradO).buffer());
auto poolB_src_memory = userB_src_memory;
streams[0].addMemory(userB_src_memory);
if (mkldnn::memory::primitive_desc(poolB_prim_desc.diff_src_primitive_desc())
!= userB_src_memory.get_primitive_desc()) {
poolB_src_memory = mkldnn::memory(poolB_prim_desc.diff_src_primitive_desc());
streams[0].addMemory(poolB_src_memory);
}
auto poolB_dst_memory = userB_dst_memory;
streams[0].addMemory(userB_dst_memory);
if (mkldnn::memory::primitive_desc(poolB_prim_desc.diff_dst_primitive_desc())
!= userB_dst_memory.get_primitive_desc()) {
poolB_dst_memory = mkldnn::memory(poolB_prim_desc.diff_dst_primitive_desc());
streams[0].addMemory(poolB_dst_memory);
streams[0].addOperation(reorder(userB_dst_memory, poolB_dst_memory));
}
if (algorithm == mkldnn::pooling_max) {
auto user_src_memory = mkldnn::memory({user_src_md, engine}, const_cast<NDArray&>(input).buffer());
auto pool_src_memory = user_src_memory;
streams[0].addMemory(user_src_memory);
if (mkldnn::memory::primitive_desc(pool_prim_desc.src_primitive_desc())
!= user_src_memory.get_primitive_desc()) {
pool_src_memory = mkldnn::memory(pool_prim_desc.src_primitive_desc());
streams[0].addMemory(pool_src_memory);
streams[0].addOperation(reorder(user_src_memory, pool_src_memory));
}
auto pool_dst_memory = mkldnn::memory(pool_prim_desc.dst_primitive_desc());
streams[0].addMemory(pool_dst_memory);
auto pool_workspace_memory = mkldnn::memory(pool_prim_desc.workspace_primitive_desc());
streams[0].addMemory(pool_workspace_memory);
streams[0].addOperation(pooling_forward(pool_prim_desc, pool_src_memory, pool_dst_memory, pool_workspace_memory));
streams[0].addOperation(pooling_backward(poolB_prim_desc, poolB_dst_memory, pool_workspace_memory, poolB_src_memory));
} else {
streams[0].addOperation(pooling_backward(poolB_prim_desc, poolB_dst_memory, poolB_src_memory));
}
if (mkldnn::memory::primitive_desc(poolB_prim_desc.diff_src_primitive_desc())
!= userB_src_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(poolB_src_memory, userB_src_memory));
}
}
streams[0].submitAndWait();
return;
}
#endif
nd4j_debug("MKL-DNN is not used for pooling3d_bp!\n", 0);
const Nd4jLong iStride0 = input.stridesOf()[0];
const Nd4jLong iStride1 = input.stridesOf()[1];
const Nd4jLong iStride2 = input.stridesOf()[2];
const Nd4jLong iStride3 = input.stridesOf()[3];
const Nd4jLong iStride4 = input.stridesOf()[4];
const Nd4jLong gIStride0 = gradI.stridesOf()[0];
const Nd4jLong gIStride1 = gradI.stridesOf()[1];
const Nd4jLong gIStride2 = gradI.stridesOf()[2];
const Nd4jLong gIStride3 = gradI.stridesOf()[3];
const Nd4jLong gIStride4 = gradI.stridesOf()[4];
const Nd4jLong oStride0 = gradO.stridesOf()[0];
const Nd4jLong oStride1 = gradO.stridesOf()[1];
const Nd4jLong oStride2 = gradO.stridesOf()[2];
const Nd4jLong oStride3 = gradO.stridesOf()[3];
const Nd4jLong oStride4 = gradO.stridesOf()[4];
const Nd4jLong iStep2 = dD*iStride2;
const Nd4jLong iStep3 = dH*iStride3;
const Nd4jLong iStep4 = dW*iStride4;
const Nd4jLong gIStep2 = dD*gIStride2;
const Nd4jLong gIStep3 = dH*gIStride3;
const Nd4jLong gIStep4 = dW*gIStride4;
const int kProd = kD*kH*kW;
const bool sameStrides = iStride0 == gIStride0 && iStride1 == gIStride1 && iStride2 == gIStride2 && iStride3 == gIStride3 && iStride4 == gIStride4;
Nd4jLong dstart, hstart, wstart, dend, hend, wend, maxKD, maxKH, maxKW;
T sum, valO, *pIn, *pgI;
if(poolingMode == 0) { // max
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, valO, sum, dstart, hstart, wstart, dend, hend, wend, maxKD, maxKH, maxKW))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int od = 0; od < oD; ++od) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
dstart = od * sD - pD;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
dend = dstart + kDEff;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(dstart < 0)
dstart += dD * ((-dstart + dD - 1) / dD);
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH);
if(wstart < 0)
wstart += dW * ((-wstart + dW - 1) / dW);
if(dend > iD)
dend -= dD * ((dend-iD + dD - 1) / dD);
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH);
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW);
sum = -DataTypeUtils::max<T>();
valO = gO[b*oStride0 + c*oStride1+ od*oStride2 + oh*oStride3 + ow*oStride4];
if(sameStrides) {
dstart *= iStride2;
dend *= iStride2;
hstart *= iStride3;
hend *= iStride3;
wstart *= iStride4;
wend *= iStride4;
maxKD = dstart;
maxKH = hstart;
maxKW = wstart;
for (Nd4jLong kd = dstart; kd < dend; kd += iStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += iStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep4) {
T valIn = pIn[kd + kh + kw];
if (valIn > sum) {
sum = valIn;
maxKD = kd;
maxKH = kh;
maxKW = kw;
}
}
gI[pIn - in + maxKD + maxKH + maxKW] += valO;
}
else {
// we set these to default values
maxKH = hstart;
maxKW = wstart;
maxKD = dstart;
for (Nd4jLong kd = dstart; kd < dend; kd += dD)
for (Nd4jLong kh = hstart; kh < hend; kh += dH)
for (Nd4jLong kw = wstart; kw < wend; kw += dW) {
T valIn = pIn[kd * iStride2 + kh * iStride3 + kw * iStride4];
if (valIn > sum) {
sum = valIn;
maxKD = kd;
maxKH = kh;
maxKW = kw;
}
}
gI[b * gIStride0 + c * gIStride1 + maxKD * gIStride2 + maxKH * gIStride3 + maxKW * gIStride4] += valO;
}
}
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 1) { // avg
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pgI, valO, dstart, hstart, wstart, dend, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int od = 0; od < oD; ++od) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pgI = gI + b * gIStride0 + c * gIStride1;
dstart = od * sD - pD;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
dend = dstart + kDEff;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(dstart < 0)
dstart += dD * ((-dstart + dD - 1) / dD);
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH);
if(wstart < 0)
wstart += dW * ((-wstart + dW - 1) / dW);
if(dend > iD)
dend -= dD * ((dend-iD + dD - 1) / dD);
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH);
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW);
dstart *= gIStride2;
dend *= gIStride2;
hstart *= gIStride3;
hend *= gIStride3;
wstart *= gIStride4;
wend *= gIStride4;
valO = gO[b*oStride0 + c*oStride1+ od*oStride2 + oh*oStride3 + ow*oStride4];
if (extraParam0 == 0) //Exclude padding
valO /= nd4j::math::nd4j_ceil<double,T>(static_cast<double>(dend-dstart) / static_cast<double>(gIStep2)) * nd4j::math::nd4j_ceil<double,T>(static_cast<double>(hend-hstart) / static_cast<double>(gIStep3)) * nd4j::math::nd4j_ceil<double,T>(static_cast<double>(wend-wstart) / static_cast<double>(gIStep4)); //Accounts for dilation
else if (extraParam0 == 1) //Include padding
valO /= kProd;
for (Nd4jLong kd = dstart; kd < dend; kd += gIStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += gIStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += gIStep4)
pgI[kd + kh + kw] += valO;
}
}
}
}
}
}
/*************************************************************************/
else if(poolingMode == 2) { // pnorm
PRAGMA_OMP_PARALLEL_FOR_ARGS(private(pIn, pgI, valO, sum, dstart, hstart, wstart, dend, hend, wend))
for(int b = 0; b < bS; ++b) {
for(int c = 0; c < iC; ++c) {
for(int od = 0; od < oD; ++od) {
for(int oh = 0; oh < oH; ++oh) {
for(int ow = 0; ow < oW; ++ow) {
pIn = in + b * iStride0 + c * iStride1;
pgI = gI + (pIn - in);
dstart = od * sD - pD;
hstart = oh * sH - pH;
wstart = ow * sW - pW;
dend = dstart + kDEff;
hend = hstart + kHEff;
wend = wstart + kWEff;
if(dstart < 0)
dstart += dD * ((-dstart + dD - 1) / dD);
if(hstart < 0)
hstart += dH * ((-hstart + dH - 1) / dH);
if(wstart < 0)
wstart += dW * ((-wstart + dW - 1) / dW);
if(dend > iD)
dend -= dD * ((dend-iD + dD - 1) / dD);
if(hend > iH)
hend -= dH * ((hend-iH + dH - 1) / dH);
if(wend > iW)
wend -= dW * ((wend-iW + dW - 1) / dW);
sum = static_cast<T>(0.);
valO = gO[b*oStride0 + c*oStride1+ od*oStride2 + oh*oStride3 + ow*oStride4];
if(sameStrides) {
dstart *= iStride2;
dend *= iStride2;
hstart *= iStride3;
hend *= iStride3;
wstart *= iStride4;
wend *= iStride4;
for (Nd4jLong kd = dstart; kd < dend; kd += iStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += iStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep4)
sum += nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kd + kh + kw]), extraParam0);
valO *= nd4j::math::nd4j_pow<T,T,T>(sum, ((T)1.f - extraParam0) / extraParam0);
for (Nd4jLong kd = dstart; kd < dend; kd += iStep2)
for (Nd4jLong kh = hstart; kh < hend; kh += iStep3)
for (Nd4jLong kw = wstart; kw < wend; kw += iStep4)
pgI[kd + kh + kw] += valO * nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kd + kh + kw]), extraParam0 - (T)1.f) * nd4j::math::nd4j_sgn<T,T>(pIn[kd + kh + kw]);
}
else {
for (Nd4jLong kd = dstart; kd < dend; kd += dD)
for (Nd4jLong kh = hstart; kh < hend; kh += dH)
for (Nd4jLong kw = wstart; kw < wend; kw += dW)
sum += nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(pIn[kd * iStride2 + kh * iStride3 + kw * iStride4]), extraParam0);
valO *= nd4j::math::nd4j_pow<T,T,T>(sum, ((T)1.f - extraParam0) / extraParam0);
for (Nd4jLong kd = dstart; kd < dend; kd += dD)
for (Nd4jLong kh = hstart; kh < hend; kh += dH)
for (Nd4jLong kw = wstart; kw < wend; kw += dW) {
const auto inVal = pIn[kD * iStride2 + kh * iStride3 + kw * iStride4];
pgI[kd * gIStride2 + kh * gIStride3 + kw * gIStride4] += valO * nd4j::math::nd4j_pow<T,T,T>(nd4j::math::nd4j_abs<T>(inVal), extraParam0 - 1.f) * nd4j::math::nd4j_sgn<T,T>(inVal);
}
}
}
}
}
}
}
}
else {
nd4j_printf("ConvolutionUtils::pooling3dBP: pooling mode argument can take three values only: 0, 1, 2, but got %i instead !\n", poolingMode);
throw "";
}
}
void ConvolutionUtils::conv2d(nd4j::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 isSameMode, const int isNCHW) {
BUILD_SINGLE_SELECTOR_TWICE(input->dataType(), conv2d_, (block, input, weights, bias, output, kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::conv2dBP(nd4j::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 isSameMode, const int isNCHW) {
BUILD_SINGLE_SELECTOR_TWICE(input->dataType(), conv2dBP_, (block, input, weights, bias, gradO, gradI, gradW, gradB, kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::depthwiseConv2d(nd4j::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 isSameMode, const int isNCHW) {
BUILD_SINGLE_SELECTOR_TWICE(input->dataType(), depthwiseConv2d_, (input, weights, bias, output, kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::depthwiseConv2dBP(nd4j::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 isSameMode, const int isNCHW) {
BUILD_SINGLE_SELECTOR_TWICE(input->dataType(), depthwiseConv2dBP_, (input, weights, bias, gradO, gradI, gradW, gradB, kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::sconv2d(nd4j::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 isSameMode, const int isNCHW) {
BUILD_SINGLE_SELECTOR_TWICE(input->dataType(), sconv2d_, (block, input, weightsDepth, weightsPoint, bias, output, kH, kW, sH, sW, pH, pW, dH, dW, isSameMode, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::vol2col(nd4j::graph::Context& block, const NDArray& volume, NDArray& columns, 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) {
BUILD_SINGLE_SELECTOR(volume.dataType(), vol2col_, (volume, columns, sD, sH, sW, pD, pH, pW, dD, dH, dW), FLOAT_TYPES);
}
void ConvolutionUtils::col2vol(nd4j::graph::Context& block, const NDArray& columns, NDArray& volume, 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) {
BUILD_SINGLE_SELECTOR(volume.dataType(), col2vol_, (columns, volume, sD, sH, sW, pD, pH, pW, dD, dH, dW), FLOAT_TYPES);
}
void ConvolutionUtils::upsampling2d(nd4j::graph::Context& block, const NDArray& input, NDArray& output, const int factorH, const int factorW, const bool isNCHW) {
BUILD_SINGLE_SELECTOR(input.dataType(), upsampling2d_, (input, output, factorH, factorW, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::upsampling3d(nd4j::graph::Context& block, const NDArray& input, NDArray& output, const int factorD, const int factorH, const int factorW, const bool isNCDHW) {
BUILD_SINGLE_SELECTOR(input.dataType(), upsampling3d_, (input, output, factorD, factorH, factorW, isNCDHW), FLOAT_TYPES);
}
void ConvolutionUtils::upsampling2dBP(nd4j::graph::Context& block, const NDArray& gradO, NDArray& gradI, const bool isNCHW) {
BUILD_SINGLE_SELECTOR(gradO.dataType(), upsampling2dBP_, (gradO, gradI, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::upsampling3dBP(nd4j::graph::Context& block, const NDArray& gradO, NDArray& gradI, const bool isNCHW) {
BUILD_SINGLE_SELECTOR(gradO.dataType(), upsampling3dBP_, (gradO, gradI, isNCHW), FLOAT_TYPES);
}
void ConvolutionUtils::pooling2d(nd4j::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) {
BUILD_SINGLE_SELECTOR(input.dataType(), pooling2d_, (block, input, output, kH, kW, sH, sW, pH, pW, dH, dW, poolingMode, extraParam0), FLOAT_TYPES);
}
void ConvolutionUtils::pooling3d(nd4j::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) {
BUILD_SINGLE_SELECTOR(input.dataType(), pooling3d_, (block, input, output, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, poolingMode, extraParam0), FLOAT_TYPES);
}
void ConvolutionUtils::pooling2dBP(nd4j::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) {
BUILD_SINGLE_SELECTOR(input.dataType(), pooling2dBP_, (block, input, gradO, gradI, kH, kW, sH, sW, pH, pW, dH, dW, poolingMode, extraParam0), FLOAT_TYPES);
}
void ConvolutionUtils::pooling3dBP(nd4j::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) {
BUILD_SINGLE_SELECTOR(input.dataType(), pooling3dBP_, (block, input, gradO, gradI, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, poolingMode, extraParam0), FLOAT_TYPES);
}
}
}