cavis/libnd4j/include/ops/declarable/platform/mkldnn/conv2d.cpp

332 lines
20 KiB
C++
Raw Normal View History

Platform helpers (#8216) * platform helpers draft Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * disable platform cmake Signed-off-by: raver119 <raver119@gmail.com> * another draft Signed-off-by: raver119 <raver119@gmail.com> * mkldnn convolution refactored Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more safety check Signed-off-by: raver119 <raver119@gmail.com> * prototype works Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * force static library mode for mkldnn Signed-off-by: raver119 <raver119@gmail.com> * - ismax fix - experimental arg fix - don't enforce openblas on Apple hardware Signed-off-by: raver119 <raver119@gmail.com> * bunch of small fixes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * declare concurrent Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - MKLDNN version upgrade to 1.0.2 - avgpool2d/maxpool2d APIs update Signed-off-by: raver119 <raver119@gmail.com> * - avgpool2d_bp/maxpool2d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * - conv2d/batchnorm APIs update Signed-off-by: raver119 <raver119@gmail.com> * - lrn/conv2d_bp/conv3d/conv3d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * all ops converted to MKLDNN 1.x Signed-off-by: raver119 <raver119@gmail.com> * bunch of tweaks Signed-off-by: raver119 <raver119@gmail.com> * namespace for platform helpers Signed-off-by: raver119 <raver119@gmail.com> * make sure platform helpers aren't opimized out Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * more of cpu_features Signed-off-by: raver119 <raver119@gmail.com> * - mkldnn removed from java - cpu_features checks in CpuNDArrayFactory Signed-off-by: raver119 <raver119@gmail.com> * F16C definition renamed Signed-off-by: raver119 <raver119@gmail.com> * some mkldnn rearrangements Signed-off-by: raver119 <raver119@gmail.com> * check supported instructions before doing anything Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * missied impl Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC option Signed-off-by: raver119 <raver119@gmail.com> * conv2d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool2d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * maxpool bp leaks fixed Signed-off-by: raver119 <raver119@gmail.com> * printf removed Signed-off-by: raver119 <raver119@gmail.com> * batchnorm fix Signed-off-by: raver119 <raver119@gmail.com> * AVX warning/error polishing Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * More polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * Polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * remove previous MKL-DNN support layer Signed-off-by: raver119 <raver119@gmail.com> * avx2 tweak Signed-off-by: raver119 <raver119@gmail.com> * allow static for apple Signed-off-by: raver119@gmail.com <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * restore OPENBLAS_PATH use Signed-off-by: raver119 <raver119@gmail.com> * add runtime check for avx/avx2 support Signed-off-by: raver119 <raver119@gmail.com> * convolution_auto Signed-off-by: raver119 <raver119@gmail.com> * Add logic for helper argument * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * skip OpTracker props for non-x86 builds Signed-off-by: raver119 <raver119@gmail.com> * linux arm isn't x86 :) Signed-off-by: raver119 <raver119@gmail.com> * avx-512 Signed-off-by: raver119 <raver119@gmail.com> * CUDA presets fix Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC Signed-off-by: raver119 <raver119@gmail.com> * prefetchw for avx2 Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC again Signed-off-by: raver119 <raver119@gmail.com>
2019-09-11 20:50:28 +02:00
/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
*
* This program and the accompanying materials are made available under the
* terms of the Apache License, Version 2.0 which is available at
* https://www.apache.org/licenses/LICENSE-2.0.
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
* License for the specific language governing permissions and limitations
* under the License.
*
* SPDX-License-Identifier: Apache-2.0
******************************************************************************/
//
// @author saudet
// @author raver119@gmail.com
//
#include <ops/declarable/PlatformHelper.h>
#include <ops/declarable/OpRegistrator.h>
#include <platform_boilerplate.h>
#include <helpers/MKLDNNStream.h>
#include "mkldnnUtils.h"
#include <ops/declarable/helpers/convolutions.h>
using namespace dnnl;
Platform helpers (#8216) * platform helpers draft Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * disable platform cmake Signed-off-by: raver119 <raver119@gmail.com> * another draft Signed-off-by: raver119 <raver119@gmail.com> * mkldnn convolution refactored Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more safety check Signed-off-by: raver119 <raver119@gmail.com> * prototype works Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * force static library mode for mkldnn Signed-off-by: raver119 <raver119@gmail.com> * - ismax fix - experimental arg fix - don't enforce openblas on Apple hardware Signed-off-by: raver119 <raver119@gmail.com> * bunch of small fixes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * declare concurrent Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - MKLDNN version upgrade to 1.0.2 - avgpool2d/maxpool2d APIs update Signed-off-by: raver119 <raver119@gmail.com> * - avgpool2d_bp/maxpool2d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * - conv2d/batchnorm APIs update Signed-off-by: raver119 <raver119@gmail.com> * - lrn/conv2d_bp/conv3d/conv3d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * all ops converted to MKLDNN 1.x Signed-off-by: raver119 <raver119@gmail.com> * bunch of tweaks Signed-off-by: raver119 <raver119@gmail.com> * namespace for platform helpers Signed-off-by: raver119 <raver119@gmail.com> * make sure platform helpers aren't opimized out Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * more of cpu_features Signed-off-by: raver119 <raver119@gmail.com> * - mkldnn removed from java - cpu_features checks in CpuNDArrayFactory Signed-off-by: raver119 <raver119@gmail.com> * F16C definition renamed Signed-off-by: raver119 <raver119@gmail.com> * some mkldnn rearrangements Signed-off-by: raver119 <raver119@gmail.com> * check supported instructions before doing anything Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * missied impl Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC option Signed-off-by: raver119 <raver119@gmail.com> * conv2d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool2d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * maxpool bp leaks fixed Signed-off-by: raver119 <raver119@gmail.com> * printf removed Signed-off-by: raver119 <raver119@gmail.com> * batchnorm fix Signed-off-by: raver119 <raver119@gmail.com> * AVX warning/error polishing Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * More polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * Polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * remove previous MKL-DNN support layer Signed-off-by: raver119 <raver119@gmail.com> * avx2 tweak Signed-off-by: raver119 <raver119@gmail.com> * allow static for apple Signed-off-by: raver119@gmail.com <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * restore OPENBLAS_PATH use Signed-off-by: raver119 <raver119@gmail.com> * add runtime check for avx/avx2 support Signed-off-by: raver119 <raver119@gmail.com> * convolution_auto Signed-off-by: raver119 <raver119@gmail.com> * Add logic for helper argument * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * skip OpTracker props for non-x86 builds Signed-off-by: raver119 <raver119@gmail.com> * linux arm isn't x86 :) Signed-off-by: raver119 <raver119@gmail.com> * avx-512 Signed-off-by: raver119 <raver119@gmail.com> * CUDA presets fix Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC Signed-off-by: raver119 <raver119@gmail.com> * prefetchw for avx2 Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC again Signed-off-by: raver119 <raver119@gmail.com>
2019-09-11 20:50:28 +02:00
namespace nd4j {
namespace ops {
namespace platforms {
//////////////////////////////////////////////////////////////////////
static void conv2dMKLDNN(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 paddingMode,
const int isNCHW) {
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);
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW, paddingMode);
dnnl_memory_desc_t empty;
dnnl::memory::desc conv_src_md(empty), conv_weights_md(empty), conv_bias_md(empty), conv_dst_md(empty);
dnnl::memory::desc user_src_md(empty), user_weights_md(empty), user_bias_md(empty), user_dst_md(empty);
dnnl::memory::dims conv_strides, conv_padding, conv_padding_r, conv_dilation;
mkldnnUtils::getMKLDNNMemoryDescConv2d(kH, kW, sH, sW, pH, pW, dH, dW, paddingMode, 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, conv_dilation);
auto conv_desc = bias != nullptr ? convolution_forward::desc(prop_kind::forward,
algorithm::convolution_auto, conv_src_md,
conv_weights_md, conv_bias_md,
conv_dst_md, conv_strides, conv_dilation, conv_padding,
conv_padding_r)
: convolution_forward::desc(prop_kind::forward,
algorithm::convolution_auto, conv_src_md,
conv_weights_md,
conv_dst_md, conv_strides, conv_dilation, conv_padding,
conv_padding_r);
auto engine = mkldnnUtils::getEngine(LaunchContext::defaultContext()->engine());
dnnl::stream stream(engine);
auto conv_prim_desc = convolution_forward::primitive_desc(conv_desc, engine);
auto user_src_memory = dnnl::memory(user_src_md, engine, const_cast<NDArray *>(input)->buffer());
auto user_weights_memory = dnnl::memory(user_weights_md, engine,
const_cast<NDArray *>(weights)->buffer());
auto user_dst_memory = dnnl::memory(user_dst_md, engine, output->buffer());
auto conv_src_memory = user_src_memory;
if (conv_prim_desc.src_desc() != user_src_memory.get_desc()) {
conv_src_memory = dnnl::memory(conv_prim_desc.src_desc(), engine);
reorder(user_src_memory, conv_src_memory).execute(stream, user_src_memory, conv_src_memory);
}
auto conv_weights_memory = user_weights_memory;
if (conv_prim_desc.weights_desc() != user_weights_memory.get_desc()) {
conv_weights_memory = dnnl::memory(conv_prim_desc.weights_desc(), engine);
reorder(user_weights_memory, conv_weights_memory).execute(stream, user_weights_memory,
conv_weights_memory);
}
auto conv_dst_memory = user_dst_memory;
if (conv_prim_desc.dst_desc() != user_dst_memory.get_desc()) {
conv_dst_memory = dnnl::memory(conv_prim_desc.dst_desc(), engine);
}
if (bias != nullptr) {
auto conv_bias_memory = dnnl::memory(conv_prim_desc.bias_desc(), engine,
const_cast<NDArray *>(bias)->buffer());
convolution_forward(conv_prim_desc).execute(stream, {{DNNL_ARG_SRC, conv_src_memory},
{DNNL_ARG_WEIGHTS, conv_weights_memory},
{DNNL_ARG_BIAS, conv_bias_memory},
{DNNL_ARG_DST, conv_dst_memory}});
} else {
convolution_forward(conv_prim_desc).execute(stream, {{DNNL_ARG_SRC, conv_src_memory},
{DNNL_ARG_WEIGHTS, conv_weights_memory},
{DNNL_ARG_DST, conv_dst_memory}});
}
if (conv_prim_desc.dst_desc() != user_dst_memory.get_desc()) {
reorder(conv_dst_memory, user_dst_memory).execute(stream, conv_dst_memory, user_dst_memory);
}
stream.wait();
}
//////////////////////////////////////////////////////////////////////
static void conv2dBpMKLDNN(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 paddingMode, const int isNCHW) {
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);
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW, paddingMode);
dnnl_memory_desc_t empty;
dnnl::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);
dnnl::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);
dnnl::memory::dims conv_strides, conv_padding, conv_padding_r, conv_dilation;
mkldnnUtils::getMKLDNNMemoryDescConv2d(kH, kW, sH, sW, pH, pW, dH, dW, paddingMode, 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, conv_dilation);
auto conv_desc = gradB != nullptr
? convolution_forward::desc(prop_kind::forward, algorithm::convolution_auto, conv_src_md, conv_weights_md, conv_bias_md, conv_dst_md, conv_strides, conv_dilation, conv_padding, conv_padding_r)
: convolution_forward::desc(prop_kind::forward, algorithm::convolution_auto, conv_src_md, conv_weights_md, conv_dst_md, conv_strides, conv_dilation, conv_padding, conv_padding_r);
auto conv_prim_desc = convolution_forward::primitive_desc(conv_desc, mkldnnUtils::getEngine( LaunchContext::defaultContext()->engine()));
auto engine = mkldnnUtils::getEngine(LaunchContext::defaultContext()->engine());
dnnl::stream stream(engine);
if (gradW != nullptr) {
auto convW_desc = gradB != nullptr ? convolution_backward_weights::desc(algorithm::convolution_auto, conv_src_md, conv_diff_weights_md, conv_bias_md, conv_dst_md, conv_strides, conv_dilation, conv_padding, conv_padding_r)
: convolution_backward_weights::desc(algorithm::convolution_auto, conv_src_md, conv_diff_weights_md, conv_dst_md, conv_strides, conv_dilation, conv_padding, conv_padding_r);
auto convW_prim_desc = convolution_backward_weights::primitive_desc(convW_desc, engine, conv_prim_desc);
auto userW_src_memory = dnnl::memory(user_src_md, engine, const_cast<NDArray *>(input)->buffer());
auto userW_weights_memory = dnnl::memory(user_diff_weights_md, engine, gradW->buffer());
auto userW_dst_memory = dnnl::memory(user_dst_md, engine,const_cast<NDArray *>(gradO)->buffer());
auto convW_src_memory = userW_src_memory;
if (convW_prim_desc.src_desc() != userW_src_memory.get_desc()) {
convW_src_memory = dnnl::memory(convW_prim_desc.src_desc(), engine);
reorder(userW_src_memory, convW_src_memory).execute(stream, userW_src_memory,convW_src_memory);
}
auto convW_weights_memory = userW_weights_memory;
if (convW_prim_desc.diff_weights_desc() != userW_weights_memory.get_desc()) {
convW_weights_memory = dnnl::memory(convW_prim_desc.diff_weights_desc(), engine);
Platform helpers (#8216) * platform helpers draft Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * disable platform cmake Signed-off-by: raver119 <raver119@gmail.com> * another draft Signed-off-by: raver119 <raver119@gmail.com> * mkldnn convolution refactored Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more safety check Signed-off-by: raver119 <raver119@gmail.com> * prototype works Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * force static library mode for mkldnn Signed-off-by: raver119 <raver119@gmail.com> * - ismax fix - experimental arg fix - don't enforce openblas on Apple hardware Signed-off-by: raver119 <raver119@gmail.com> * bunch of small fixes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * declare concurrent Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - MKLDNN version upgrade to 1.0.2 - avgpool2d/maxpool2d APIs update Signed-off-by: raver119 <raver119@gmail.com> * - avgpool2d_bp/maxpool2d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * - conv2d/batchnorm APIs update Signed-off-by: raver119 <raver119@gmail.com> * - lrn/conv2d_bp/conv3d/conv3d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * all ops converted to MKLDNN 1.x Signed-off-by: raver119 <raver119@gmail.com> * bunch of tweaks Signed-off-by: raver119 <raver119@gmail.com> * namespace for platform helpers Signed-off-by: raver119 <raver119@gmail.com> * make sure platform helpers aren't opimized out Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * more of cpu_features Signed-off-by: raver119 <raver119@gmail.com> * - mkldnn removed from java - cpu_features checks in CpuNDArrayFactory Signed-off-by: raver119 <raver119@gmail.com> * F16C definition renamed Signed-off-by: raver119 <raver119@gmail.com> * some mkldnn rearrangements Signed-off-by: raver119 <raver119@gmail.com> * check supported instructions before doing anything Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * missied impl Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC option Signed-off-by: raver119 <raver119@gmail.com> * conv2d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool2d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * maxpool bp leaks fixed Signed-off-by: raver119 <raver119@gmail.com> * printf removed Signed-off-by: raver119 <raver119@gmail.com> * batchnorm fix Signed-off-by: raver119 <raver119@gmail.com> * AVX warning/error polishing Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * More polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * Polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * remove previous MKL-DNN support layer Signed-off-by: raver119 <raver119@gmail.com> * avx2 tweak Signed-off-by: raver119 <raver119@gmail.com> * allow static for apple Signed-off-by: raver119@gmail.com <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * restore OPENBLAS_PATH use Signed-off-by: raver119 <raver119@gmail.com> * add runtime check for avx/avx2 support Signed-off-by: raver119 <raver119@gmail.com> * convolution_auto Signed-off-by: raver119 <raver119@gmail.com> * Add logic for helper argument * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * skip OpTracker props for non-x86 builds Signed-off-by: raver119 <raver119@gmail.com> * linux arm isn't x86 :) Signed-off-by: raver119 <raver119@gmail.com> * avx-512 Signed-off-by: raver119 <raver119@gmail.com> * CUDA presets fix Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC Signed-off-by: raver119 <raver119@gmail.com> * prefetchw for avx2 Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC again Signed-off-by: raver119 <raver119@gmail.com>
2019-09-11 20:50:28 +02:00
}
auto convW_dst_memory = userW_dst_memory;
if (convW_prim_desc.diff_dst_desc() != userW_dst_memory.get_desc()) {
convW_dst_memory = dnnl::memory(convW_prim_desc.diff_dst_desc(), engine);
reorder(userW_dst_memory, convW_dst_memory).execute(stream, userW_dst_memory, convW_dst_memory);
}
if (gradB != nullptr) {
auto convW_bias_memory = dnnl::memory(convW_prim_desc.diff_bias_desc(), engine, gradB->buffer());
convolution_backward_weights(convW_prim_desc).execute(stream,
{{DNNL_ARG_SRC, convW_src_memory},
{DNNL_ARG_DIFF_DST, convW_dst_memory},
{DNNL_ARG_DIFF_WEIGHTS, convW_weights_memory},
{DNNL_ARG_DIFF_BIAS, convW_bias_memory}});
}
else {
convolution_backward_weights(convW_prim_desc).execute(stream,
{{DNNL_ARG_SRC, convW_src_memory},
{DNNL_ARG_DIFF_DST, convW_dst_memory},
{DNNL_ARG_DIFF_WEIGHTS, convW_weights_memory}});
}
if (convW_prim_desc.diff_weights_desc() != userW_weights_memory.get_desc()) {
reorder(convW_weights_memory, userW_weights_memory).execute(stream, convW_weights_memory,
userW_weights_memory);
}
stream.wait();
Platform helpers (#8216) * platform helpers draft Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * disable platform cmake Signed-off-by: raver119 <raver119@gmail.com> * another draft Signed-off-by: raver119 <raver119@gmail.com> * mkldnn convolution refactored Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more safety check Signed-off-by: raver119 <raver119@gmail.com> * prototype works Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * force static library mode for mkldnn Signed-off-by: raver119 <raver119@gmail.com> * - ismax fix - experimental arg fix - don't enforce openblas on Apple hardware Signed-off-by: raver119 <raver119@gmail.com> * bunch of small fixes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * declare concurrent Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - MKLDNN version upgrade to 1.0.2 - avgpool2d/maxpool2d APIs update Signed-off-by: raver119 <raver119@gmail.com> * - avgpool2d_bp/maxpool2d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * - conv2d/batchnorm APIs update Signed-off-by: raver119 <raver119@gmail.com> * - lrn/conv2d_bp/conv3d/conv3d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * all ops converted to MKLDNN 1.x Signed-off-by: raver119 <raver119@gmail.com> * bunch of tweaks Signed-off-by: raver119 <raver119@gmail.com> * namespace for platform helpers Signed-off-by: raver119 <raver119@gmail.com> * make sure platform helpers aren't opimized out Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * more of cpu_features Signed-off-by: raver119 <raver119@gmail.com> * - mkldnn removed from java - cpu_features checks in CpuNDArrayFactory Signed-off-by: raver119 <raver119@gmail.com> * F16C definition renamed Signed-off-by: raver119 <raver119@gmail.com> * some mkldnn rearrangements Signed-off-by: raver119 <raver119@gmail.com> * check supported instructions before doing anything Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * missied impl Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC option Signed-off-by: raver119 <raver119@gmail.com> * conv2d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool2d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * maxpool bp leaks fixed Signed-off-by: raver119 <raver119@gmail.com> * printf removed Signed-off-by: raver119 <raver119@gmail.com> * batchnorm fix Signed-off-by: raver119 <raver119@gmail.com> * AVX warning/error polishing Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * More polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * Polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * remove previous MKL-DNN support layer Signed-off-by: raver119 <raver119@gmail.com> * avx2 tweak Signed-off-by: raver119 <raver119@gmail.com> * allow static for apple Signed-off-by: raver119@gmail.com <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * restore OPENBLAS_PATH use Signed-off-by: raver119 <raver119@gmail.com> * add runtime check for avx/avx2 support Signed-off-by: raver119 <raver119@gmail.com> * convolution_auto Signed-off-by: raver119 <raver119@gmail.com> * Add logic for helper argument * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * skip OpTracker props for non-x86 builds Signed-off-by: raver119 <raver119@gmail.com> * linux arm isn't x86 :) Signed-off-by: raver119 <raver119@gmail.com> * avx-512 Signed-off-by: raver119 <raver119@gmail.com> * CUDA presets fix Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC Signed-off-by: raver119 <raver119@gmail.com> * prefetchw for avx2 Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC again Signed-off-by: raver119 <raver119@gmail.com>
2019-09-11 20:50:28 +02:00
}
if (gradI != nullptr) {
auto convI_desc = convolution_backward_data::desc(algorithm::convolution_auto, conv_diff_src_md, conv_weights_md, conv_dst_md, conv_strides, conv_dilation, conv_padding, conv_padding_r);
auto convI_prim_desc = convolution_backward_data::primitive_desc(convI_desc, engine, conv_prim_desc);
auto userI_src_memory = dnnl::memory(user_diff_src_md, engine, gradI->buffer());
auto userI_weights_memory = dnnl::memory(user_weights_md, engine,const_cast<NDArray *>(weights)->buffer());
auto userI_dst_memory = dnnl::memory(user_dst_md, engine, const_cast<NDArray *>(gradO)->buffer());
auto convI_src_memory = userI_src_memory;
if (convI_prim_desc.diff_src_desc() != userI_src_memory.get_desc()) {
convI_src_memory = dnnl::memory(convI_prim_desc.diff_src_desc(), engine);
}
auto convI_weights_memory = userI_weights_memory;
if (convI_prim_desc.weights_desc() != userI_weights_memory.get_desc()) {
convI_weights_memory = dnnl::memory(convI_prim_desc.weights_desc(), engine);
reorder(userI_weights_memory, convI_weights_memory).execute(stream, userI_weights_memory, convI_weights_memory);
}
auto convI_dst_memory = userI_dst_memory;
if (convI_prim_desc.diff_dst_desc() != userI_dst_memory.get_desc()) {
convI_dst_memory = dnnl::memory(convI_prim_desc.diff_dst_desc(), engine);
reorder(userI_dst_memory, convI_dst_memory).execute(stream, userI_dst_memory, convI_dst_memory);
}
convolution_backward_data(convI_prim_desc).execute(stream,
{{DNNL_ARG_DIFF_DST, convI_dst_memory},
{DNNL_ARG_WEIGHTS, convI_weights_memory},
{DNNL_ARG_DIFF_SRC, convI_src_memory}});
if (convI_prim_desc.diff_src_desc() != userI_src_memory.get_desc()) {
reorder(convI_src_memory, userI_src_memory).execute(stream, convI_src_memory, userI_src_memory);
}
stream.wait();
}
}
//////////////////////////////////////////////////////////////////////
PLATFORM_IMPL(conv2d, ENGINE_CPU) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC] always
auto bias = block.width() > 2 ? INPUT_VARIABLE(2) : nullptr; // [oC]
auto output = OUTPUT_VARIABLE(0); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW)
int sH = INT_ARG(2); // strides height
int sW = INT_ARG(3); // strides width
int pH = INT_ARG(4); // paddings height
int pW = INT_ARG(5); // paddings width
int dH = INT_ARG(6); // dilations height
int dW = INT_ARG(7); // dilations width
int paddingMode = INT_ARG(8); // 0-VALID, 1-SAME
bool isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 0-NCHW, 1-NHWC
int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast<int>(weights->sizeAt(0)); // filter(kernel) height
int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast<int>(weights->sizeAt(1)); // filter(kernel) width
conv2dMKLDNN(block, input, weights, bias, output, kH, kW, sH, sW, pH, pW, dH, dW, paddingMode, isNCHW);
return Status::OK();
}
PLATFORM_CHECK(conv2d, ENGINE_CPU) {
// we don't want to use mkldnn if cpu doesn't support avx/avx2
if (::optimalLevel() < 2)
return false;
auto input = INPUT_VARIABLE(0);
auto weights = INPUT_VARIABLE(1);
// conv2d is only available for float32 dtype
return block.isUseMKLDNN() && input->dataType() == nd4j::DataType::FLOAT32 &&
weights->dataType() == nd4j::DataType::FLOAT32;
}
//////////////////////////////////////////////////////////////////////
PLATFORM_IMPL(conv2d_bp, ENGINE_CPU) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC] always
auto bias = block.width() > 3 ? INPUT_VARIABLE(2) : nullptr; // [oC]
auto gradO = block.width() > 3 ? INPUT_VARIABLE(3) : INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
auto gradW = OUTPUT_VARIABLE(1); // [kH, kW, iC, oC] always
auto gradB = block.width() > 3 ? OUTPUT_VARIABLE(2) : nullptr; // [oC]
int kH = INT_ARG(0); // filter(kernel) height
int kW = INT_ARG(1); // filter(kernel) width
int sH = INT_ARG(2); // strides height
int sW = INT_ARG(3); // strides width
int pH = INT_ARG(4); // paddings height
int pW = INT_ARG(5); // paddings width
int dH = INT_ARG(6); // dilations height
int dW = INT_ARG(7); // dilations width
int paddingMode = INT_ARG(8); // 0-VALID, 1-SAME
int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 0-NCHW, 1-NHWC
REQUIRE_TRUE(input->rankOf() == 4, 0,"CUSTOM CONV2D_BP OP: rank of input array must be equal to 4, but got %i instead !",input->rankOf());
REQUIRE_TRUE(weights->rankOf() == 4, 0,"CUSTOM CONV2D_BP OP: rank of weights array must be equal to 4, but got %i instead !",weights->rankOf());
REQUIRE_TRUE(gradO->rankOf() == 4, 0,"CUSTOM CONV2D_BP OP: rank of output's gradients (next epsilon) array must be equal to 4, but got %i instead !",gradO->rankOf());
conv2dBpMKLDNN(block, input, weights, bias, gradO, gradI, gradW, gradB, kH, kW, sH, sW, pH, pW, dH, dW, paddingMode, isNCHW);
return Status::OK();
}
cuDNN integration (#150) * initial commit Signed-off-by: raver119 <raver119@gmail.com> * one file Signed-off-by: raver119 <raver119@gmail.com> * few more includes Signed-off-by: raver119 <raver119@gmail.com> * m? Signed-off-by: raver119 <raver119@gmail.com> * const Signed-off-by: raver119 <raver119@gmail.com> * cudnn linkage in tests Signed-off-by: raver119 <raver119@gmail.com> * culibos Signed-off-by: raver119 <raver119@gmail.com> * static reminder Signed-off-by: raver119 <raver119@gmail.com> * platform engine tag Signed-off-by: raver119 <raver119@gmail.com> * HAVE_CUDNN moved to config.h.in Signed-off-by: raver119 <raver119@gmail.com> * include Signed-off-by: raver119 <raver119@gmail.com> * include Signed-off-by: raver119 <raver119@gmail.com> * skip cudnn handle creation if there's not cudnn Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * target device in context Signed-off-by: raver119 <raver119@gmail.com> * platform engines Signed-off-by: raver119 <raver119@gmail.com> * platform engines Signed-off-by: raver119 <raver119@gmail.com> * allow multiple -h args Signed-off-by: raver119 <raver119@gmail.com> * allow multiple -h args Signed-off-by: raver119 <raver119@gmail.com> * move mkldnn out of CPU block Signed-off-by: raver119 <raver119@gmail.com> * link to mkldnn on cuda Signed-off-by: raver119 <raver119@gmail.com> * less prints Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * next step Signed-off-by: raver119 <raver119@gmail.com> * conv2d NCHW draft Signed-off-by: raver119 <raver119@gmail.com> * conv2d biasAdd Signed-off-by: raver119 <raver119@gmail.com> * test for MKL/CUDNN combined use Signed-off-by: raver119 <raver119@gmail.com> * - provide additional code for conv2d ff based on cudnn api, not tested yet Signed-off-by: Yurii <iuriish@yahoo.com> * - further work on conv2d helper based on using cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - fixing several cuda bugs which appeared after cudnn lib had been started to use Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of conv2d backprop op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - implementaion of conv3d and conv3d_bp ops based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - bugs fixing in conv3d/conv3d_bp ops (cudnn in use) Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of depthwiseConv2d (ff/bp) op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - implementation of batchnorm ff op based on cudnn api Signed-off-by: Yurii <iuriish@yahoo.com> * - disable cudnn batchnorm temporary Signed-off-by: Yurii <iuriish@yahoo.com> * - add minor change in cmake Signed-off-by: Yurii <iuriish@yahoo.com> * engine for depthwise mkldnn Signed-off-by: raver119 <raver119@gmail.com> * couple of includes Signed-off-by: raver119 <raver119@gmail.com> * - provide permutation to cudnn batchnorm ff when format is NHWC Signed-off-by: Yurii <iuriish@yahoo.com> * lgamma fix Signed-off-by: raver119 <raver119@gmail.com> * - eliminate memory leak in two tests Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: Yurii Shyrma <iuriish@yahoo.com>
2020-01-20 19:32:46 +01:00
PLATFORM_CHECK(conv2d_bp, ENGINE_CPU) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC] always
auto bias = block.width() > 3 ? INPUT_VARIABLE(2) : nullptr; // [oC]
auto gradO = block.width() > 3 ? INPUT_VARIABLE(3) : INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
auto gradW = OUTPUT_VARIABLE(1); // [kH, kW, iC, oC] always
auto gradB = block.width() > 3 ? OUTPUT_VARIABLE(2) : nullptr; // [oC]
return block.isUseMKLDNN() &&
nd4j::MKLDNNStream::isSupported({input, weights, bias, gradO, gradI, gradW, gradB});
}
}
}
Platform helpers (#8216) * platform helpers draft Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * disable platform cmake Signed-off-by: raver119 <raver119@gmail.com> * another draft Signed-off-by: raver119 <raver119@gmail.com> * mkldnn convolution refactored Signed-off-by: raver119 <raver119@gmail.com> * minor tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more safety check Signed-off-by: raver119 <raver119@gmail.com> * prototype works Signed-off-by: raver119 <raver119@gmail.com> * meh Signed-off-by: raver119 <raver119@gmail.com> * force static library mode for mkldnn Signed-off-by: raver119 <raver119@gmail.com> * - ismax fix - experimental arg fix - don't enforce openblas on Apple hardware Signed-off-by: raver119 <raver119@gmail.com> * bunch of small fixes Signed-off-by: raver119@gmail.com <raver119@gmail.com> * declare concurrent Signed-off-by: raver119@gmail.com <raver119@gmail.com> * - MKLDNN version upgrade to 1.0.2 - avgpool2d/maxpool2d APIs update Signed-off-by: raver119 <raver119@gmail.com> * - avgpool2d_bp/maxpool2d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * - conv2d/batchnorm APIs update Signed-off-by: raver119 <raver119@gmail.com> * - lrn/conv2d_bp/conv3d/conv3d_bp APIs update Signed-off-by: raver119 <raver119@gmail.com> * all ops converted to MKLDNN 1.x Signed-off-by: raver119 <raver119@gmail.com> * bunch of tweaks Signed-off-by: raver119 <raver119@gmail.com> * namespace for platform helpers Signed-off-by: raver119 <raver119@gmail.com> * make sure platform helpers aren't opimized out Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * build cpu_features on x86 systems Signed-off-by: raver119 <raver119@gmail.com> * more of cpu_features Signed-off-by: raver119 <raver119@gmail.com> * - mkldnn removed from java - cpu_features checks in CpuNDArrayFactory Signed-off-by: raver119 <raver119@gmail.com> * F16C definition renamed Signed-off-by: raver119 <raver119@gmail.com> * some mkldnn rearrangements Signed-off-by: raver119 <raver119@gmail.com> * check supported instructions before doing anything Signed-off-by: raver119 <raver119@gmail.com> * typo Signed-off-by: raver119 <raver119@gmail.com> * missied impl Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC option Signed-off-by: raver119 <raver119@gmail.com> * conv2d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool2d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * avgpool3d_bp leak fix Signed-off-by: raver119 <raver119@gmail.com> * maxpool bp leaks fixed Signed-off-by: raver119 <raver119@gmail.com> * printf removed Signed-off-by: raver119 <raver119@gmail.com> * batchnorm fix Signed-off-by: raver119 <raver119@gmail.com> * AVX warning/error polishing Signed-off-by: AlexDBlack <blacka101@gmail.com> * Fix Signed-off-by: AlexDBlack <blacka101@gmail.com> * More polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * Polish Signed-off-by: AlexDBlack <blacka101@gmail.com> * remove previous MKL-DNN support layer Signed-off-by: raver119 <raver119@gmail.com> * avx2 tweak Signed-off-by: raver119 <raver119@gmail.com> * allow static for apple Signed-off-by: raver119@gmail.com <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * exclude mkldnn in one more place Signed-off-by: raver119 <raver119@gmail.com> * restore OPENBLAS_PATH use Signed-off-by: raver119 <raver119@gmail.com> * add runtime check for avx/avx2 support Signed-off-by: raver119 <raver119@gmail.com> * convolution_auto Signed-off-by: raver119 <raver119@gmail.com> * Add logic for helper argument * minor test fix Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * few tweaks Signed-off-by: raver119 <raver119@gmail.com> * skip OpTracker props for non-x86 builds Signed-off-by: raver119 <raver119@gmail.com> * linux arm isn't x86 :) Signed-off-by: raver119 <raver119@gmail.com> * avx-512 Signed-off-by: raver119 <raver119@gmail.com> * CUDA presets fix Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC Signed-off-by: raver119 <raver119@gmail.com> * prefetchw for avx2 Signed-off-by: raver119 <raver119@gmail.com> * BUILD_PIC again Signed-off-by: raver119 <raver119@gmail.com>
2019-09-11 20:50:28 +02:00
}