bartekkuncer commented on a change in pull request #20862: URL: https://github.com/apache/incubator-mxnet/pull/20862#discussion_r796590652
########## File path: src/operator/nn/dnnl/dnnl_where-inl.h ########## @@ -0,0 +1,73 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://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. + */ + +/*! + * \file dnnl_where-inl.h + */ + +#ifndef MXNET_OPERATOR_NN_DNNL_DNNL_WHERE_INL_H_ +#define MXNET_OPERATOR_NN_DNNL_DNNL_WHERE_INL_H_ + +#if MXNET_USE_ONEDNN == 1 +#include <memory> +#include <unordered_map> +#include <vector> +#include "./dnnl_base-inl.h" Review comment: Please avoid dots in includes. ########## File path: src/operator/nn/dnnl/dnnl_where.cc ########## @@ -0,0 +1,213 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://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. + */ + +/*! + * \file dnnl_where.cc + */ + +#if MXNET_USE_ONEDNN == 1 + +#include <algorithm> +#include <set> +#include <unordered_set> +#include "dnnl_where-inl.h" +#include "operator/operator_common.h" + +namespace mxnet { +namespace op { + +bool SupportDNNLWhere(const std::vector<NDArray>& inputs) { + static const std::set<int> supported_dtypes = { + mshadow::kFloat32, mshadow::kBfloat16, mshadow::kInt8, mshadow::kUint8}; + for (int i = 0; i < inputs.size(); ++i) { + if (!supported_dtypes.count(inputs[i].dtype()) || inputs[i].shape().Size() <= 0 || + inputs[i].shape().ndim() <= 0) { + return false; + } + } + return true; +} + +void DNNLWhereForward(const nnvm::NodeAttrs& attrs, + const OpContext& ctx, + const std::vector<NDArray>& inputs, + const std::vector<OpReqType>& req, + const std::vector<NDArray>& outputs) { + TmpMemMgr::Get()->Init(ctx.requested[0]); + const auto tensors = DNNLWhereFwd::Tensors(inputs, outputs[0]); + const auto fwd = DNNLWhereFwd::GetCached(tensors); + fwd.Execute(tensors, req, ctx); +} + +DNNLWhereFwd::Tensors::Tensors(const std::vector<NDArray>& inputs, const NDArray& output) + : condition(inputs[0]), left(inputs[1]), right(inputs[2]), output(output) {} + +DNNLWhereFwd DNNLWhereFwd::GetCached(const Tensors& tensors) { + using where_op_fwd_map = std::unordered_map<OpSignature, DNNLWhereFwd, OpHash>; +#if DMLC_CXX11_THREAD_LOCAL + static thread_local where_op_fwd_map fwds; +#else + static MX_THREAD_LOCAL where_op_fwd_map fwds; +#endif + + OpSignature key; + key.AddSign(tensors.condition); + key.AddSign(tensors.left); + key.AddSign(tensors.right); + key.AddSign(tensors.output); + + auto it = fwds.find(key); + if (it == fwds.end()) { + DNNLWhereFwd fwd(tensors); + it = AddToCache(&fwds, key, fwd); + } + return it->second; +} + +static mxnet::TShape GetBroadcastableShape(const mxnet::TShape& in_shape, + const mxnet::TShape& out_shape) { + if (in_shape == out_shape) { + return in_shape; + } + + mxnet::TShape broadcastable_in_shape(out_shape.ndim(), -1); Review comment: Maybe initialize TShape here to 1 instead of -1 and reduce the amount of iteration in the for loop below? ########## File path: src/operator/numpy/np_where_forward_op.cc ########## @@ -103,11 +137,20 @@ NNVM_REGISTER_OP(_npi_where) return std::vector<std::pair<int, int> >{{1, 0}, {2, 0}}; }) .set_attr<FCompute>("FCompute<cpu>", NumpyWhereOpForward<cpu>) +#if MXNET_USE_ONEDNN == 1 + .set_attr<FResourceRequest>("FResourceRequest", + [](const NodeAttrs& n) { + return std::vector<ResourceRequest>{ResourceRequest::kTempSpace}; + }) + .set_attr<FComputeEx>("FComputeEx<cpu>", WhereForwardEx) + .set_attr<bool>("TIsDNNL", true) + .set_attr<FInferStorageType>("FInferStorageType", WhereInferStorageType) +#endif .set_attr<nnvm::FGradient>( "FGradient", // Use the following lambda function instead of ElemwiseGradUseIn - // for best efficiency. grad[condition] = 0; to calculate grad[x] and grad[y] - // we need only condition from input. + // for best efficiency. grad[condition] = 0; to calculate grad[x] and Review comment: ```suggestion // grad[condition] = 0; to calculate grad[x] and grad[y] we need only condition from input. ``` ########## File path: src/operator/numpy/np_where_forward_op.cc ########## @@ -103,11 +137,20 @@ NNVM_REGISTER_OP(_npi_where) return std::vector<std::pair<int, int> >{{1, 0}, {2, 0}}; }) .set_attr<FCompute>("FCompute<cpu>", NumpyWhereOpForward<cpu>) +#if MXNET_USE_ONEDNN == 1 + .set_attr<FResourceRequest>("FResourceRequest", + [](const NodeAttrs& n) { + return std::vector<ResourceRequest>{ResourceRequest::kTempSpace}; + }) + .set_attr<FComputeEx>("FComputeEx<cpu>", WhereForwardEx) + .set_attr<bool>("TIsDNNL", true) + .set_attr<FInferStorageType>("FInferStorageType", WhereInferStorageType) +#endif .set_attr<nnvm::FGradient>( "FGradient", // Use the following lambda function instead of ElemwiseGradUseIn - // for best efficiency. grad[condition] = 0; to calculate grad[x] and grad[y] - // we need only condition from input. + // for best efficiency. grad[condition] = 0; to calculate grad[x] and + // grad[y] we need only condition from input. Review comment: ```suggestion ``` ########## File path: src/operator/numpy/np_where_forward_op.cc ########## @@ -103,11 +137,20 @@ NNVM_REGISTER_OP(_npi_where) return std::vector<std::pair<int, int> >{{1, 0}, {2, 0}}; }) .set_attr<FCompute>("FCompute<cpu>", NumpyWhereOpForward<cpu>) +#if MXNET_USE_ONEDNN == 1 + .set_attr<FResourceRequest>("FResourceRequest", + [](const NodeAttrs& n) { + return std::vector<ResourceRequest>{ResourceRequest::kTempSpace}; + }) + .set_attr<FComputeEx>("FComputeEx<cpu>", WhereForwardEx) + .set_attr<bool>("TIsDNNL", true) + .set_attr<FInferStorageType>("FInferStorageType", WhereInferStorageType) +#endif .set_attr<nnvm::FGradient>( "FGradient", // Use the following lambda function instead of ElemwiseGradUseIn Review comment: ```suggestion // Use the following lambda function instead of ElemwiseGradUseIn for best efficiency. ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
