[GitHub] [incubator-mxnet] mozga-intel commented on a change in pull request #20137: [v1.x][FEATURE][WIP] Add MKLDNN Deconvolution 1D and 3D support

[GitBox <gi...@apache.org>]

mozga-intel commented on a change in pull request #20137:
URL: https://github.com/apache/incubator-mxnet/pull/20137#discussion_r619104700



##########
File path: src/operator/nn/mkldnn/mkldnn_deconvolution-inl.h
##########
@@ -0,0 +1,377 @@
+/*
+ * 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 mkldnn_deconvolution-inl.h
+ * Naming convention:
+ *                 ________
+ *  (src) data --->|Deconv|
+ *     weights --->|  FWD |---> out (dst)
+ *        bias --->|______|
+ *                                 ________
+ *        (diff_src) data_grad <---|Deconv|<--- out_grad (diff_dst)
+ *  (diff_weight) weights_grad <---|  BWD |<--- data (src)
+ *       (diff_bias) bias_grad <---|      |<--- weight
+ *                                 |______|<--- bias
+ *
+ * "out" in this (and .cc) file will always refer to the output of Deconv FWD and
+ * "out_grad" to its gradient. The corresponding MKLDNN names are in parentheses.
+ */
+#ifndef MXNET_OPERATOR_NN_MKLDNN_MKLDNN_DECONVOLUTION_INL_H_
+#define MXNET_OPERATOR_NN_MKLDNN_MKLDNN_DECONVOLUTION_INL_H_
+
+#if MXNET_USE_MKLDNN == 1
+#include <utility>
+#include <vector>
+#include <numeric>
+
+#include "../deconvolution-inl.h"
+#include "./mkldnn_base-inl.h"
+#include "./mkldnn_ops-inl.h"
+
+namespace mxnet {
+namespace op {
+
+using deconv_fwd_t = mkldnn::deconvolution_forward;
+using deconv_fwd_pd_t = mkldnn::deconvolution_forward::primitive_desc;
+
+using deconv_bwd_data_t = mkldnn::deconvolution_backward_data;
+using deconv_bwd_data_pd_t = mkldnn::deconvolution_backward_data::primitive_desc;
+
+using deconv_bwd_weights_t = mkldnn::deconvolution_backward_weights;
+using deconv_bwd_weights_pd_t = mkldnn::deconvolution_backward_weights::primitive_desc;
+
+
+
+// Swaps the logical order of dimensions that in plain format would correspond to input and output
+// channels (for example: oihw => iohw, iohw => oihw, goihw => giohw).
+inline mkldnn::memory::desc IOLogicalSwapDesc(const mkldnn::memory::desc &desc,
+                                              const uint32_t num_group) {
+  std::vector<int> order(desc.data.ndims);
+  std::iota(std::begin(order), std::end(order), 0);
+  const int offset = static_cast<int>(num_group > 1);

Review comment:
       This offset returns true or false; Why do we need to do static_cast to int? Basically; O0-optimization will treat it like a:
    `mov  DWORD PTR [rbp-4], 0`, when a boolean type will have `mov  BYTE PTR [rbp-1], 0`.




-- 
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.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org