/******************************************************************************* * 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) // #include #include #include #include #include "mkldnnUtils.h" #include namespace sd { namespace ops { namespace platforms { ////////////////////////////////////////////////////////////////////////// static void deconv2TFdBackPropMKLDNN(const NDArray* weights, const NDArray* gradO, NDArray* gradI, const int bS, const int iC, const int iH, const int iW, const int oC, const int oH, const int oW, const int kH, const int kW, const int sH, const int sW, const int pH, const int pW, const int dH, const int dW, const bool isNCHW, const int wFormat) { // gradI [bS, iH, iW, iC], mkl doesn't support ndhwc format // weights [oC, iC, kH, kW] always, mkl doesn't support weights format [kH, kW, iC, oC] // gradO [bS, oH, oW, oC] dnnl::memory::dims strides = { sH, sW }; dnnl::memory::dims dilation = { dH - 1, dW - 1 }; dnnl::memory::dims padding = { pH, pW }; dnnl::memory::dims padding_r = { (oH - 1) * sH - iH + kH - pH, (oW - 1) * sW - iW + kW - pW }; // weights type dnnl::memory::data_type wType = weights->dataType() == DataType::FLOAT32 ? dnnl::memory::data_type::f32 : dnnl::memory::data_type::bf16; // gradO type dnnl::memory::data_type gradOType = gradO->dataType() == DataType::FLOAT32 ? dnnl::memory::data_type::f32 : dnnl::memory::data_type::bf16; // gradI type dnnl::memory::data_type gradIType = gradI->dataType() == DataType::FLOAT32 ? dnnl::memory::data_type::f32 : dnnl::memory::data_type::bf16; dnnl::memory::format_tag xFormatMkl = isNCHW ? dnnl::memory::format_tag::nchw : dnnl::memory::format_tag::nhwc; dnnl::memory::format_tag wFormatMkl = dnnl::memory::format_tag::oihw; dnnl::memory::dims xDims = {bS, iC, iH, iW}; dnnl::memory::dims wDims = {oC, iC, kH, kW}; dnnl::memory::dims zDims = {bS, oC, oH, oW}; // memory descriptors for arrays // input dnnl::memory::desc x_mkl_md = dnnl::memory::desc(xDims, gradOType, dnnl::memory::format_tag::any); // weights dnnl::memory::desc w_mkl_md = dnnl::memory::desc(wDims, wType, dnnl::memory::format_tag::any); dnnl::memory::desc w_user_md = dnnl::memory::desc(wDims, wType, wFormatMkl); w_user_md.data.format_kind = dnnl_blocked; // overrides format w_user_md.data.format_desc.blocking.strides[0] = weights->strideAt(3); // permute [kH, kW, iC, oC] -> [oC, iC, kH, kW] w_user_md.data.format_desc.blocking.strides[1] = weights->strideAt(2); w_user_md.data.format_desc.blocking.strides[2] = weights->strideAt(0); w_user_md.data.format_desc.blocking.strides[3] = weights->strideAt(1); // gradO dnnl::memory::desc gradO_mkl_md = dnnl::memory::desc(zDims, gradOType, dnnl::memory::format_tag::any); dnnl::memory::desc gradO_user_md = dnnl::memory::desc(zDims, gradOType, xFormatMkl); mkldnnUtils::setBlockStrides(gradO, gradO_user_md); // gradI dnnl::memory::desc gradI_mkl_md = dnnl::memory::desc(xDims, gradIType, dnnl::memory::format_tag::any); dnnl::memory::desc gradI_user_md = dnnl::memory::desc(xDims, gradIType, xFormatMkl); mkldnnUtils::setBlockStrides(gradI, gradI_user_md); auto engine = mkldnnUtils::getEngine(LaunchContext::defaultContext()->engine()); // forward primitive description dnnl::convolution_forward::desc op_ff_desc(dnnl::prop_kind::forward_inference, dnnl::algorithm::convolution_auto, x_mkl_md, w_mkl_md, gradO_mkl_md, strides, dilation, padding, padding_r); dnnl::convolution_forward::primitive_desc op_ff_prim_desc(op_ff_desc, engine); // backward data primitive description dnnl::convolution_backward_data::desc op_data_bp_desc(dnnl::algorithm::convolution_auto, gradI_mkl_md, w_mkl_md, gradO_mkl_md, strides, dilation, padding, padding_r); dnnl::convolution_backward_data::primitive_desc op_data_bp_prim_desc(op_data_bp_desc, engine, op_ff_prim_desc); // arguments (memory buffers) necessary for calculations std::unordered_map args; dnnl::stream stream(engine); // provide memory buffers and check whether reorder is required // weights mkldnnUtils::loadDataToMklStream(weights, engine, stream, w_user_md, op_data_bp_prim_desc.weights_desc(), args[DNNL_ARG_WEIGHTS]); // gradO mkldnnUtils::loadDataToMklStream(gradO, engine, stream, gradO_user_md, op_data_bp_prim_desc.diff_dst_desc(), args[DNNL_ARG_DIFF_DST]); // gradI auto gradI_user_mem = dnnl::memory(gradI_user_md, engine, gradI->getBuffer()); const bool gradIReorder = op_data_bp_prim_desc.diff_src_desc() != gradI_user_mem.get_desc(); auto gradI_mkl_mem = gradIReorder ? dnnl::memory(op_data_bp_prim_desc.diff_src_desc(), engine) : gradI_user_mem; args[DNNL_ARG_DIFF_SRC] = gradI_mkl_mem; // run backward data calculations dnnl::convolution_backward_data(op_data_bp_prim_desc).execute(stream, args); // reorder gradI if necessary if (gradIReorder) dnnl::reorder(gradI_mkl_mem, gradI_user_mem).execute(stream, gradI_mkl_mem, gradI_user_mem); stream.wait(); // shape::printArray(z_mkl_mem.map_data(),8); } ////////////////////////////////////////////////////////////////////////// PLATFORM_IMPL(deconv2d_tf, ENGINE_CPU) { auto gradO = INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC], [oC, iC, kH, kW], [oC, kH, kW, iC] auto gradIShape = INPUT_VARIABLE(0); // [4] - shape of input of conv2d (that is shape of gradI) auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast(weights->sizeAt(0));// filter(kernel) height int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast(weights->sizeAt(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 isSameMode = INT_ARG(8); // 0-VALID, 1-SAME int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 1-NHWC, 0-NCHW int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, iC, oC], 1 - [oC, iC, kH, kW], 2 - [oC, kH, kW, iC] const int rank = gradO->rankOf(); REQUIRE_TRUE(weights->rankOf() == rank, 0, "CUSTOM DECONV2D_TF MKLDNN OP: rank of weights array must be equal to 4, but got %i instead !", weights->rankOf()); REQUIRE_TRUE(gradIShape->rankOf() == 1, 0, "CUSTOM DECONV2D_TF MKLDNN OP: rank of array with output shape must be equal to 1, but got %i instead !", gradIShape->rankOf()); REQUIRE_TRUE(gradIShape->lengthOf() == rank, 0, "CUSTOM DECONV2D_TF MKLDNN OP: length of array with output shape must be equal to 4, but got %i instead !", gradIShape->lengthOf()); int indIOioC, indIiH, indWoC(3), indOoH; if(!isNCHW) { indIOioC = 3; indIiH = 1; indOoH = 1; } else { indIOioC = 1; indIiH = 2; indOoH = 2; } std::vector gradIShapeVector = gradIShape->template asVectorT(); const int bS = gradIShapeVector[0]; // batch size const int iH = gradIShapeVector[indIiH]; // input height const int iW = gradIShapeVector[indIiH+1]; // input width const int iC = gradIShapeVector[indIOioC]; // input channels const int oC = weights->sizeAt(indWoC); // output channels const int oH = gradO->sizeAt(indOoH); // input height const int oW = gradO->sizeAt(indOoH); // input width int trueoH, trueoW; // true output height, width ConvolutionUtils::calcOutSizePool2D(trueoH, trueoW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode); std::vector expectedGradOShape = ShapeUtils::composeShapeUsingDimsAndIdx({bS,oC,trueoH,trueoW, 0,indIOioC,indOoH,indOoH+1}); std::vector expectedWeightsShape = {kH, kW, iC, oC}; REQUIRE_TRUE(gradO->isSameShape(expectedGradOShape), 0, "CUSTOM DECONV2D_TF MKLDNN OP: wrong shape of input array, basing on array with output shape expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str()); REQUIRE_TRUE(weights->isSameShape(expectedWeightsShape), 0, "CUSTOM DECONV2D_TF MKLDNN OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weights).c_str()); if(isSameMode) // SAME ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW); // // mkl supports only [oC, iC, kH, kW] for weights // weights = new NDArray(weights->permute({3,2,0,1})); // [kH, kW, iC, oC] -> [oC, iC, kH, kW] // // mkl supports NCHW format only // if(!isNCHW) { // gradI = new NDArray(gradI->permute({0,3,1,2})); // [bS, iH, iW, iC] -> [bS, iC, iH, iW] // gradO = new NDArray(gradO->permute({0,3,1,2})); // [bS, oH, oW, oC] -> [bS, oC, oH, oW] // } deconv2TFdBackPropMKLDNN(weights, gradO, gradI, bS, iC, iH, iW, oC, oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, isNCHW, wFormat); // delete weights; // if(!isNCHW) { // delete gradI; // delete gradO; // } // ConvolutionUtils::conv2dBP(block, &input, weights, nullptr, gradO, gradI, nullptr, nullptr, kH,kW,sH,sW,pH,pW,dH,dW,isSameMode,isNCHW); return Status::OK(); } PLATFORM_CHECK(deconv2d_tf, ENGINE_CPU) { auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC] always auto gradO = INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCDHW), epsilon_next auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCDHW), gradI const DataType wType = weights->dataType(); const DataType gradOType = gradO->dataType(); const DataType gradIType = gradI->dataType(); return block.isUseMKLDNN() && ((wType==DataType::FLOAT32 || wType==DataType::BFLOAT16) && (gradOType==DataType::FLOAT32 || gradOType==DataType::BFLOAT16) && (gradIType==DataType::FLOAT32 || gradIType==DataType::BFLOAT16)); } } } }