* 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>
113 lines
6.2 KiB
C++
113 lines
6.2 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, created on 20.03.2018
|
|
//
|
|
|
|
#include <ops/declarable/CustomOperations.h>
|
|
#include <ops/declarable/helpers/convolutions.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
|
|
|
|
|
|
CUSTOM_OP_IMPL(pointwise_conv2d, 2, 1, false, 0, 0) {
|
|
|
|
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
|
|
auto weights = INPUT_VARIABLE(1); // [1, 1, iC, oC] always
|
|
auto bias = block.width() > 2 ? INPUT_VARIABLE(2) : nullptr; // [oC]
|
|
|
|
auto output = OUTPUT_VARIABLE(0); // [bS, iH, iW, oC] (NHWC) or [bS, oC, iH, iW] (NCHW)
|
|
|
|
REQUIRE_TRUE(input->rankOf() == 4, 0, "CUSTOM POINTWISECONV2D OP: rank of input array must be equal to 4, but got %i instead !", input->rankOf());
|
|
REQUIRE_TRUE(weights->rankOf() == 4, 0, "CUSTOM POINTWISECONV2D OP: rank of weights array must be equal to 4, but got %i instead !", weights->rankOf());
|
|
if(bias)
|
|
REQUIRE_TRUE(bias->rankOf() <= 2, 0, "CUSTOM POINTWISECONV2D OP: rank of biases array must be equal <= 2, but got %i instead !", bias->rankOf());
|
|
|
|
int kH = 1; // filter(kernel) height
|
|
int kW = 1; // filter(kernel) width
|
|
int sH = 1; // strides height
|
|
int sW = 1; // strides width
|
|
int pH = 0; // paddings height
|
|
int pW = 0; // paddings width
|
|
int dH = 1; // dilations height
|
|
int dW = 1; // dilations width
|
|
int isNCHW = block.getIArguments()->size() > 0 ? !INT_ARG(0) : 1; // INT_ARG(0): 0-NCHW, 1-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);
|
|
|
|
std::string expectedWeightsShape = ShapeUtils::shapeAsString({1, 1, iC, oC});
|
|
REQUIRE_TRUE(expectedWeightsShape == ShapeUtils::shapeAsString(weights), 0, "CUSTOM POINTWISECONV2D OP: wrong shape of weights array, expected is %s, but got %s instead !", expectedWeightsShape.c_str(), ShapeUtils::shapeAsString(weights).c_str());
|
|
if (bias)
|
|
REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM POINTWISECONV2D OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf());
|
|
|
|
ConvolutionUtils::conv2d(block, input, weights, bias, output, kH,kW, sH,sW, pH,pW, dH,dW, 1/*isSameMode*/, isNCHW);
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
DECLARE_TYPES(pointwise_conv2d) {
|
|
getOpDescriptor()
|
|
->setAllowedInputTypes(nd4j::DataType::ANY)
|
|
->setAllowedOutputTypes({ALL_FLOATS});
|
|
}
|
|
|
|
|
|
DECLARE_SHAPE_FN(pointwise_conv2d) {
|
|
|
|
Nd4jLong* inputShapeInfo = inputShape->at(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
|
|
Nd4jLong* weightsShapeInfo = inputShape->at(1); // [1, 1, iC, oC] always
|
|
Nd4jLong* biasShapeInfo = block.width() > 2 ? inputShape->at(2) : nullptr; // [oC]
|
|
|
|
const int rank = 4;
|
|
REQUIRE_TRUE(inputShapeInfo[0] == rank, 0, "CUSTOM POINTWISECONV2D OP: rank of input array must be equal to %i, but got %i instead !", rank, inputShapeInfo[0]);
|
|
REQUIRE_TRUE(weightsShapeInfo[0] == rank, 0, "CUSTOM POINTWISECONV2D OP: rank of weights array must be equal to %i, but got %i instead !", rank, weightsShapeInfo[0]);
|
|
|
|
int isNCHW = block.getIArguments()->size() > 0 ? !INT_ARG(0) : 1; // INT_ARG(0): 0-NCHW, 1-NHWC
|
|
|
|
int indIOioC, indWoC(3);
|
|
if(!isNCHW)
|
|
indIOioC = 3;
|
|
else
|
|
indIOioC = 1;
|
|
|
|
const int bS = inputShapeInfo[1]; // batch size
|
|
const int iC = inputShapeInfo[indIOioC+1]; // input channels
|
|
const int oC = weightsShapeInfo[indWoC+1]; // output channels
|
|
|
|
std::string expectedWeightsShape = ShapeUtils::shapeAsString({1, 1, iC, oC});
|
|
REQUIRE_TRUE(expectedWeightsShape == ShapeUtils::shapeAsString(weightsShapeInfo), 0, "POINTWISECONV2D OP: wrong shape of weights array, expected is %s, but got %s instead !", expectedWeightsShape.c_str(), ShapeUtils::shapeAsString(weightsShapeInfo).c_str());
|
|
if (biasShapeInfo)
|
|
REQUIRE_TRUE(biasShapeInfo[0] <= 2 && oC == shape::length(biasShapeInfo), 0, "POINTWISECONV2D OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, biasShapeInfo[0], shape::length(biasShapeInfo));
|
|
|
|
auto outputShapeInfo = ShapeBuilders::copyShapeInfoAndType(inputShapeInfo, weightsShapeInfo, true, block.getWorkspace());
|
|
|
|
// do not forget to put oC instead of iC in outputShapeInfo
|
|
outputShapeInfo[indIOioC + 1] = oC;
|
|
|
|
shape::updateStrides(outputShapeInfo, shape::order(inputShapeInfo));
|
|
|
|
return SHAPELIST(CONSTANT(outputShapeInfo));
|
|
}
|
|
|
|
|
|
}
|
|
}
|