cavis/libnd4j/include/ops/declarable/platform/mkldnn/depthwiseConv2d.cpp

485 lines
28 KiB
C++

/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
* Copyright (c) 2019 Konduit K.K.
*
* 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 <ops/declarable/PlatformHelper.h>
#include <ops/declarable/OpRegistrator.h>
#include <system/platform_boilerplate.h>
#include <helpers/MKLDNNStream.h>
#include <ops/declarable/helpers/convolutions.h>
#include "mkldnnUtils.h"
using namespace dnnl;
namespace sd {
namespace ops {
namespace platforms {
//////////////////////////////////////////////////////////////////////////
static void depthwiseConv2dMKLDNN(const NDArray* input, const NDArray* weights, const NDArray* bias, 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 paddingMode, const bool isNCHW, const int wFormat) {
// mkl supports only following case: mC = 1, oC = iC
// input [bS, iC, iH, iW] nchw or [bS, iH, iW, iC] nhwc, since mkl doesn't support nhwc format we'll permute when nhwc is given
// weights {iC, mC, 1, kH, kW}
// bias [oC], may be nullptr
// output [bS, oC, oH, oW] nchw or [bS, oH, oW, oC] nhwc
// oC = iC*mC
int bS, iC, iH, iW, mC, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width;
int indIOioC, indIiH, indWmC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, wFormat, *input, *output, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weights->sizeAt(indWmC); // channels multiplier
const int pWSame = (paddingMode == 2 && dW > 1) ? ((oW - 1) * sW + (kW - 1) * dW + 1 - iW) / 2 : pW; // dH == 1 for causal mode in conv1d
dnnl::memory::dims strides = { sH, sW };
dnnl::memory::dims padding = { pH, pW };
dnnl::memory::dims padding_r = { (oH - 1) * sH - iH + kH - pH, (oW - 1) * sW - iW + kW - pWSame };
dnnl::memory::dims dilation = { dH-1, dW-1};
uint i0, i1, i2, i3;
if(0 == wFormat) {
i0 = 2; i1 = 3; i2 = 0; i3 = 1; // [kH, kW, iC, mC] -> [iC, mC, 1, kH, kW]
}
else if(1 == wFormat) {
i0 = 1; i1 = 0; i2 = 2; i3 = 3; // [mC, iC, kH, kW] -> [iC, mC, 1, kH, kW]
}
else {
i0 = 3; i1 = 0; i2 = 1; i3 = 2; // [mC, kH, kW, iC] -> [iC, mC, 1, kH, kW]
}
// input type
dnnl::memory::data_type xType;
if(input->dataType() == DataType::FLOAT32)
xType = dnnl::memory::data_type::f32;
else if(input->dataType() == DataType::HALF)
xType = dnnl::memory::data_type::f16;
else if(input->dataType() == DataType::UINT8)
xType = dnnl::memory::data_type::u8;
else
xType = dnnl::memory::data_type::s8;
// weights type
dnnl::memory::data_type wType = xType;
if(xType == dnnl::memory::data_type::u8)
wType = dnnl::memory::data_type::s8;
// output and bias type (have the same types)
dnnl::memory::data_type zType;
if(output->dataType() == DataType::FLOAT32)
zType = dnnl::memory::data_type::f32;
else if(output->dataType() == DataType::HALF)
zType = dnnl::memory::data_type::f16;
else if(output->dataType() == DataType::UINT8)
zType = dnnl::memory::data_type::u8;
else if(output->dataType() == DataType::INT8)
zType = dnnl::memory::data_type::s8;
else
zType = dnnl::memory::data_type::s32;
dnnl::memory::format_tag xzFormatMkl = isNCHW ? dnnl::memory::format_tag::nchw : dnnl::memory::format_tag::nhwc;
dnnl::memory::format_tag wFormatMkl = dnnl::memory::format_tag::goihw;
dnnl::memory::dims xDims = {bS, iC, iH, iW};
dnnl::memory::dims wDims = {iC, mC, 1, 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, xType, dnnl::memory::format_tag::any);
dnnl::memory::desc x_user_md = dnnl::memory::desc(xDims, xType, xzFormatMkl);
mkldnnUtils::setBlockStrides(*input, x_user_md);
// 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(i0); // permute
w_user_md.data.format_desc.blocking.strides[1] = weights->strideAt(i1);
w_user_md.data.format_desc.blocking.strides[2] = 0;
w_user_md.data.format_desc.blocking.strides[3] = weights->strideAt(i2);
w_user_md.data.format_desc.blocking.strides[4] = weights->strideAt(i3);
// bias
dnnl::memory::desc b_mkl_md;
if(bias != nullptr)
b_mkl_md = dnnl::memory::desc({oC}, zType, dnnl::memory::format_tag::x);
// output
dnnl::memory::desc z_mkl_md = dnnl::memory::desc(zDims, zType, dnnl::memory::format_tag::any);
dnnl::memory::desc z_user_md = dnnl::memory::desc(zDims, zType, xzFormatMkl);
mkldnnUtils::setBlockStrides(*output, z_user_md);
auto engine = mkldnnUtils::getEngine(LaunchContext::defaultContext()->engine());
// operation primitive description
dnnl::convolution_forward::desc op_desc(dnnl::prop_kind::forward_inference, dnnl::algorithm::convolution_auto,
x_mkl_md, w_mkl_md, b_mkl_md, z_mkl_md, strides, dilation, padding, padding_r);
dnnl::convolution_forward::primitive_desc op_prim_desc(op_desc, engine);
// arguments (memory buffers) necessary for calculations
std::unordered_map<int, dnnl::memory> args;
dnnl::stream stream(engine);
// provide memory buffers and check whether reorder is required
// input
mkldnnUtils::loadDataToMklStream(*input, engine, stream, x_user_md, op_prim_desc.src_desc(), args[DNNL_ARG_SRC]);
// weights
mkldnnUtils::loadDataToMklStream(*weights, engine, stream, w_user_md, op_prim_desc.weights_desc(), args[DNNL_ARG_WEIGHTS]);
// bias
if(bias != nullptr) {
auto b_mkl_mem = dnnl::memory(b_mkl_md, engine, const_cast<void*>(bias->buffer()));
args[DNNL_ARG_BIAS] = b_mkl_mem;
}
// output
auto z_user_mem = mkldnnUtils::loadDataToMklStream(*output, engine, stream, z_user_md, op_prim_desc.dst_desc(), args[DNNL_ARG_DST]);
// run calculations
dnnl::convolution_forward(op_prim_desc).execute(stream, args);
// reorder outputs if necessary
if (op_prim_desc.dst_desc() != z_user_mem.get_desc())
dnnl::reorder(args[DNNL_ARG_DST], z_user_mem).execute(stream, args[DNNL_ARG_DST], z_user_mem);
stream.wait();
// shape::printArray(z_mkl_mem.map_data<float>(),8);
}
//////////////////////////////////////////////////////////////////////////
static void depthwiseConv2dBpMKLDNN(const NDArray* input, const NDArray* weights, const NDArray* gradO, NDArray* gradI, NDArray* gradW, NDArray* gradB,
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 paddingMode, const bool isNCHW, const int wFormat) {
// mkl supports only following case: mC = 1, oC = iC
// input, gradI [bS, iC, iH, iW] nchw or [bS, iH, iW, iC] nhwc, since mkl doesn't support nhwc format we'll permute when nhwc is given
// weights/gradW {iC, mC, 1, kH, kW}
// gradB [oC], may be nullptr
// gradO [bS, oC, oH, oW] nchw or [bS, oH, oW, oC] nhwc
// oC = iC*mC
int bS, iC, iH, iW, mC, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width;
int indIOioC, indIiH, indWmC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, wFormat, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weights->sizeAt(indWmC);
const int pWSame = (paddingMode == 2 && dW > 1) ? ((oW - 1) * sW + (kW - 1) * dW + 1 - iW) / 2 : pW; // dH == 1 for causal mode in conv1d
dnnl::memory::dims strides = { sH, sW };
dnnl::memory::dims padding = { pH, pW };
dnnl::memory::dims padding_r = { (oH - 1) * sH - iH + kH - pH, (oW - 1) * sW - iW + kW - pWSame };
dnnl::memory::dims dilation = { dH-1, dW-1};
uint i0, i1, i2, i3;
if(0 == wFormat) {
i0 = 2; i1 = 3; i2 = 0; i3 = 1; // [kH, kW, iC, mC] -> [iC, mC, 1, kH, kW]
}
else if(1 == wFormat) {
i0 = 1; i1 = 0; i2 = 2; i3 = 3; // [mC, iC, kH, kW] -> [iC, mC, 1, kH, kW]
}
else {
i0 = 3; i1 = 0; i2 = 1; i3 = 2; // [mC, kH, kW, iC] -> [iC, mC, 1, kH, kW]
}
// input type
dnnl::memory::data_type xType = input->dataType() == DataType::FLOAT32 ? dnnl::memory::data_type::f32 : dnnl::memory::data_type::bf16;
// 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;
// gradW type
dnnl::memory::data_type gradWType = gradW->dataType() == DataType::FLOAT32 ? dnnl::memory::data_type::f32 : dnnl::memory::data_type::bf16;
// gradB type
dnnl::memory::data_type gradBType = gradB != nullptr ? (gradB->dataType() == DataType::FLOAT32 ? dnnl::memory::data_type::f32 : dnnl::memory::data_type::bf16) : dnnl::memory::data_type::f32;
dnnl::memory::format_tag xzFormatMkl = isNCHW ? dnnl::memory::format_tag::nchw : dnnl::memory::format_tag::nhwc;
dnnl::memory::format_tag wFormatMkl = dnnl::memory::format_tag::goihw;
dnnl::memory::dims xDims = {bS, iC, iH, iW};
dnnl::memory::dims wDims = {iC, mC, 1, 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, xType, dnnl::memory::format_tag::any);
dnnl::memory::desc x_user_md = dnnl::memory::desc(xDims, xType, xzFormatMkl);
mkldnnUtils::setBlockStrides(*input, x_user_md);
// 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(i0); // permute
w_user_md.data.format_desc.blocking.strides[1] = weights->strideAt(i1);
w_user_md.data.format_desc.blocking.strides[2] = 0;
w_user_md.data.format_desc.blocking.strides[3] = weights->strideAt(i2);
w_user_md.data.format_desc.blocking.strides[4] = weights->strideAt(i3);
// 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, xzFormatMkl);
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, xzFormatMkl);
mkldnnUtils::setBlockStrides(*gradI, gradI_user_md);
// gradW
dnnl::memory::desc gradW_mkl_md = dnnl::memory::desc(wDims, gradWType, dnnl::memory::format_tag::any);
dnnl::memory::desc gradW_user_md = dnnl::memory::desc(wDims, gradWType, wFormatMkl);
gradW_user_md.data.format_kind = dnnl_blocked; // overrides format
gradW_user_md.data.format_desc.blocking.strides[0] = gradW->strideAt(i0); // permute
gradW_user_md.data.format_desc.blocking.strides[1] = gradW->strideAt(i1);
gradW_user_md.data.format_desc.blocking.strides[2] = 0;
gradW_user_md.data.format_desc.blocking.strides[3] = gradW->strideAt(i2);
gradW_user_md.data.format_desc.blocking.strides[4] = gradW->strideAt(i3);
// gradB
dnnl::memory::desc gradB_mkl_md;
if(gradB != nullptr)
gradB_mkl_md = dnnl::memory::desc({oC}, gradBType, dnnl::memory::format_tag::x);
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, gradB_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);
// backward weights primitive description
dnnl::convolution_backward_weights::desc op_weights_bp_desc(dnnl::algorithm::convolution_auto, x_mkl_md, gradW_mkl_md, gradB_mkl_md, gradO_mkl_md, strides, dilation, padding, padding_r);
dnnl::convolution_backward_weights::primitive_desc op_weights_bp_prim_desc(op_weights_bp_desc, engine, op_ff_prim_desc);
// arguments (memory buffers) necessary for calculations
std::unordered_map<int, dnnl::memory> args;
dnnl::stream stream(engine);
// provide memory buffers and check whether reorder is required
// input
mkldnnUtils::loadDataToMklStream(*input, engine, stream, x_user_md, op_weights_bp_prim_desc.src_desc(), args[DNNL_ARG_SRC]);
// weights
mkldnnUtils::loadDataToMklStream(*weights, engine, stream, w_user_md, op_data_bp_prim_desc.weights_desc(), args[DNNL_ARG_WEIGHTS]);
// gradO
auto gradO_user_mem = dnnl::memory(gradO_user_md, engine, const_cast<void*>(gradO->buffer()));
const bool gradOReorderW = op_weights_bp_prim_desc.diff_dst_desc() != gradO_user_mem.get_desc();
const bool gradOReorderD = op_data_bp_prim_desc.diff_dst_desc() != gradO_user_mem.get_desc();
auto gradO_mkl_memW = gradOReorderW ? dnnl::memory(op_weights_bp_prim_desc.diff_dst_desc(), engine) : gradO_user_mem;
auto gradO_mkl_memD = gradOReorderD ? dnnl::memory(op_data_bp_prim_desc.diff_dst_desc(), engine) : gradO_user_mem;
if (gradOReorderW)
dnnl::reorder(gradO_user_mem, gradO_mkl_memW).execute(stream, gradO_user_mem, gradO_mkl_memW);
if (gradOReorderD)
dnnl::reorder(gradO_user_mem, gradO_mkl_memD).execute(stream, gradO_user_mem, gradO_mkl_memD);
args[DNNL_ARG_DIFF_DST] = gradO_mkl_memD;
// gradI
auto gradI_user_mem = mkldnnUtils::loadDataToMklStream(*gradI, engine, stream, gradI_user_md, op_data_bp_prim_desc.diff_src_desc(), args[DNNL_ARG_DIFF_SRC]);
// gradW
auto gradW_user_mem = mkldnnUtils::loadDataToMklStream(*gradW, engine, stream, gradW_user_md, op_weights_bp_prim_desc.diff_weights_desc(), args[DNNL_ARG_DIFF_WEIGHTS]);
// gradB
if(gradB != nullptr) {
auto gradB_mkl_mem = dnnl::memory(gradB_mkl_md, engine, gradB->buffer());
args[DNNL_ARG_DIFF_BIAS] = gradB_mkl_mem;
}
// run backward data calculations
dnnl::convolution_backward_data(op_data_bp_prim_desc).execute(stream, args);
if(gradOReorderW || gradOReorderD)
args[DNNL_ARG_DIFF_DST] = gradO_mkl_memW;
// run backward weights calculations
dnnl::convolution_backward_weights(op_weights_bp_prim_desc).execute(stream, args);
// reorder gradI if necessary
if (op_data_bp_prim_desc.diff_src_desc() != gradI_user_mem.get_desc())
dnnl::reorder(args[DNNL_ARG_DIFF_SRC], gradI_user_mem).execute(stream, args[DNNL_ARG_DIFF_SRC], gradI_user_mem);
if (op_weights_bp_prim_desc.diff_weights_desc() != gradW_user_mem.get_desc())
dnnl::reorder(args[DNNL_ARG_DIFF_WEIGHTS], gradW_user_mem).execute(stream, args[DNNL_ARG_DIFF_WEIGHTS], gradW_user_mem);
stream.wait();
// shape::printArray(z_mkl_mem.map_data<float>(),8);
}
//////////////////////////////////////////////////////////////////////
PLATFORM_IMPL(depthwise_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, mC], [mC, iC, kH, kW], [mC, kH, kW, iC]
auto bias = block.width() > 2 ? INPUT_VARIABLE(2) : nullptr; // [oC] = iC*mC
auto output = OUTPUT_VARIABLE(0); // [bS, oH, oW, iC*mC] (NHWC) or [bS, iC*mC, oH, oW] (NCHW)
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
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
int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, iC, mC], 1 - [mC, iC, kH, kW], 2 - [mC, kH, kW, iC]
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, wFormat, *input, *output, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weights->sizeAt(indWmC); // channels multiplier
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW, paddingMode);
std::vector<Nd4jLong> expectedWeightsShape = ConvolutionUtils::expectWeightsShape(wFormat, kH, kW, iC, mC);
REQUIRE_TRUE(weights->isSameShape(expectedWeightsShape), 0, "CUSTOM DEPTHWISECONV2D MKL OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weights).c_str());
REQUIRE_TRUE(output->sizeAt(indIOioC) == iC*mC, 0, "CUSTOM DEPTHWISECONV2D MKL OP: the output_channels must be equal to input_channels * channels_multiplier = %i !", iC*mC);
if (bias)
REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM DEPTHWISECONV2D MKL OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf());
depthwiseConv2dMKLDNN(input, weights, bias, output, kH, kW, sH, sW, pH, pW, dH, dW, paddingMode, isNCHW, wFormat);
return Status::OK();
}
//////////////////////////////////////////////////////////////////////
PLATFORM_CHECK(depthwise_conv2d, ENGINE_CPU) {
auto input = INPUT_VARIABLE(0);
auto weights = INPUT_VARIABLE(1);
auto bias = block.width() > 2 ? INPUT_VARIABLE(2) : nullptr;
auto output = INPUT_VARIABLE(0);
const DataType xType = input->dataType();
const DataType wType = weights->dataType();
const DataType zType = output->dataType();
const DataType bType = bias != nullptr ? bias->dataType() : zType;
const int mC = weights->sizeAt(3);
return block.isUseMKLDNN() && mC == 1 &&
(
(xType==DataType::FLOAT32 && wType==DataType::FLOAT32 && bType==DataType::FLOAT32 && zType==DataType::FLOAT32) ||
(xType==DataType::BFLOAT16 && wType==DataType::BFLOAT16 && bType==DataType::BFLOAT16 && zType==DataType::BFLOAT16) ||
((xType==DataType::UINT8 || xType==DataType::INT8) && wType==DataType::INT8 && (zType==DataType::UINT8 || zType==DataType::INT8 || zType==DataType::INT32 || zType==DataType::FLOAT32) && bType == zType)
);
}
//////////////////////////////////////////////////////////////////////////
PLATFORM_IMPL(depthwise_conv2d_bp, ENGINE_CPU) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NDHWC) or [bS, iC, iH, iW] (NCDHW)
auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, mC], [mC, iC, kH, kW], [mC, kH, kW, iC]
auto bias = block.width() > 3 ? INPUT_VARIABLE(2) : nullptr; // [oC] = [iC*mC]
auto gradO = block.width() > 3 ? INPUT_VARIABLE(3) : INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NDHWC) or [bS, oC, oH, oW] (NCDHW), epsilon_next
auto gradI = OUTPUT_NULLIFIED(0); // [bS, iH, iW, iC] (NDHWC) or [bS, iC, iH, iW] (NCDHW), epsilon
auto gradW = OUTPUT_NULLIFIED(1); // [kH, kW, iC, mC], [mC, iC, kH, kW], [mC, kH, kW, iC]
auto gradB = block.width() > 3 ? OUTPUT_NULLIFIED(2) : nullptr; // [oC]
REQUIRE_TRUE(input->rankOf() == 4, 0, "CUSTOM DEPTHWISECONV2D_BP MKL OP: rank of input array must be equal to 4, but got %i instead !", input->rankOf());
REQUIRE_TRUE(weights->rankOf() == 4, 0, "CUSTOM DEPTHWISECONV2D_BP MKL OP: rank of weights array must be equal to 4, but got %i instead !", weights->rankOf());
REQUIRE_TRUE(gradO->rankOf() == 4, 0, "CUSTOM DEPTHWISECONV2D_BP MKL OP: rank of output gradients (next epsilon) array must be equal to 4, but got %i instead !", gradO->rankOf());
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
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): 1-NHWC, 0-NCHW
int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, iC, mC], 1 - [mC, iC, kH, kW], 2 - [mC, kH, kW, iC]
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, wFormat, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWiC, indWmC, indWkH, indOoH);
mC = weights->sizeAt(indWmC); // channels multiplier
int trueoH, trueoW; // correct output height, width
ConvolutionUtils::calcOutSizePool2D(trueoH, trueoW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, paddingMode);
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW, paddingMode);
std::vector<Nd4jLong> expectedGradOShape = ShapeUtils::composeShapeUsingDimsAndIdx({bS,oC,trueoH,trueoW, 0,indIOioC,indOoH,indOoH+1});
std::vector<Nd4jLong> expectedWeightsShape = ConvolutionUtils::expectWeightsShape(wFormat, kH, kW, iC, mC);
REQUIRE_TRUE(gradO->isSameShape(expectedGradOShape), 0, "CUSTOM DEPTHWISECONV2D_BP MKL OP: wrong shape of output gradients (next epsilon) array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(weights->isSameShape(expectedWeightsShape), 0, "CUSTOM DEPTHWISECONV2D_BP MKL OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weights).c_str());
if(bias)
REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM DEPTHWISECONV2D_BP MKL OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf());
depthwiseConv2dBpMKLDNN(input, weights, gradO, gradI, gradW, gradB, kH, kW, sH, sW, pH, pW, dH, dW, paddingMode, isNCHW, wFormat);
return Status::OK();
}
//////////////////////////////////////////////////////////////////////
PLATFORM_CHECK(depthwise_conv2d_bp, ENGINE_CPU) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NDHWC) or [bS, iC, iH, iW] (NCDHW)
auto weights = INPUT_VARIABLE(1); // [kH, kW, iC, oC], [oC, iC, kH, kW], [oC, kH, kW, iC]
auto bias = block.width() > 3 ? INPUT_VARIABLE(2) : nullptr; // [oC] = [iC*mC]
auto gradO = block.width() > 3 ? INPUT_VARIABLE(3) : INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NDHWC) or [bS, oC, oH, oW] (NCDHW), epsilon_next
auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NDHWC) or [bS, iC, iH, iW] (NCDHW), epsilon
auto gradW = OUTPUT_VARIABLE(1); // [kH, kW, iC, oC], [oC, iC, kH, kW], [oC, kH, kW, iC]
auto gradB = block.width() > 3 ? OUTPUT_VARIABLE(2) : nullptr; // [oC]
const DataType xType = input->dataType();
const DataType wType = weights->dataType();
const DataType gradOType = gradO->dataType();
const DataType gradIType = gradI->dataType();
const DataType gradWType = gradW->dataType();
const DataType gradBType = gradB != nullptr ? gradB->dataType() : DataType::FLOAT32;
const int mC = weights->sizeAt(3);
return block.isUseMKLDNN() && mC == 1 && ((xType==DataType::FLOAT32 || xType==DataType::BFLOAT16) && (wType==DataType::FLOAT32 || wType==DataType::BFLOAT16) && (gradOType==DataType::FLOAT32 || gradOType==DataType::BFLOAT16) && (gradIType==DataType::FLOAT32 || gradIType==DataType::BFLOAT16) && (gradWType==DataType::FLOAT32 || gradWType==DataType::BFLOAT16) && (gradBType==DataType::FLOAT32 || gradBType==DataType::BFLOAT16) );
}
}
}
}