Float64 fallback for mkldnn subgraph and rnn op (#15853)

ZhennanQin · pengzhao-intel · commit a37a76c42bf7 · 2019-09-19T16:22:51.000+08:00
* Fix mkldnn subgraph with float64

* Fix ci

Change-Id: I0bb4e8d7a0a534aa661601887cc633cb9c4fcadf

* Fix test

Change-Id: I96c529abe7adb6def90a22f03b3432263ef12fda

* Update dmlc-core

Change-Id: I472fb7bbffc16ed8c36494ab49838b08c59b2f12

* pin to offical dmlc

Change-Id: I5a27dc83b892bf8fcb34bb089449d1d3b6e9beed

* Fix GPU CI

Change-Id: I285947e01bdb0651c2c7830ed4eb76931a09b754

* Fix GPU CI

Change-Id: I6f23b51d6bda44f6ae18766ebe390118740bb9c7
diff --git a/src/operator/nn/mkldnn/mkldnn_base-inl.h b/src/operator/nn/mkldnn/mkldnn_base-inl.h
@@ -47,17 +47,18 @@
 #define MXNET_OPERATOR_NN_MKLDNN_MKLDNN_BASE_INL_H_
 
 #if MXNET_USE_MKLDNN == 1
+#include <algorithm>
 #include <iterator>
+#include <memory>
 #include <string>
 #include <unordered_map>
-#include <vector>
 #include <utility>
-#include <algorithm>
-#include <memory>
+#include <vector>
 #include "mkldnn.hpp"
+#include "mxnet/graph_attr_types.h"
 #include "mxnet/ndarray.h"
-#include "mxnet/resource.h"
 #include "mxnet/op_attr_types.h"
+#include "mxnet/resource.h"
 using namespace mkldnn;
 namespace mxnet {
 
@@ -132,6 +133,11 @@ static inline bool SupportMKLDNN(int dtype, const mxnet::TShape &shape) {
   return dtype == mshadow::kFloat32 && (ndim == 1 || ndim == 2 || ndim == 4);
 }
 
+static inline bool SupportMKLDNNRNN(const NDArray &input) {
+  int ndim = input.shape().ndim();
+  return (input.dtype() == mshadow::kFloat32) && (ndim == 3);
+}
+
 static inline bool SupportMKLDNNQuantize(int dtype) {
   return dtype == mshadow::kFloat32 || dtype == mshadow::kInt8 ||
          dtype == mshadow::kUint8;
@@ -569,7 +575,8 @@ class MKLDNNMemory {
   }
 };
 
-void FallBackCompute(FCompute fn, const nnvm::NodeAttrs &attrs,
+template <typename Compute, typename AttrState>
+void FallBackCompute(Compute fn, const AttrState &attrs,
                      const OpContext &ctx,
                      const std::vector<NDArray> &inputs,
                      const std::vector<OpReqType> &req,
diff --git a/src/operator/nn/mkldnn/mkldnn_base.cc b/src/operator/nn/mkldnn/mkldnn_base.cc
@@ -420,7 +420,8 @@ mkldnn::memory::primitive_desc GetPrimitiveDesc(mkldnn::memory::primitive_desc p
   return mkldnn::memory::primitive_desc(data_md, pd.get_engine());
 }
 
-void FallBackCompute(FCompute fn, const nnvm::NodeAttrs &attrs,
+template <typename Compute, typename AttrState>
+void FallBackCompute(Compute fn, const AttrState &attrs_states,
                      const OpContext &ctx,
                      const std::vector<NDArray> &inputs,
                      const std::vector<OpReqType> &req,
@@ -461,7 +462,7 @@ void FallBackCompute(FCompute fn, const nnvm::NodeAttrs &attrs,
     out_blobs[i] = output.data();
   }
 
-  fn(attrs, ctx, in_blobs, req, out_blobs);
+  fn(attrs_states, ctx, in_blobs, req, out_blobs);
   for (size_t i = 0; i < out_blobs.size(); i++) {
     if (req[i] == kAddTo && outputs[i].IsMKLDNNData())
       mxnet::common::CastNonDefaultStorage(temp_src, temp_dst, ctx, false);
@@ -518,6 +519,24 @@ static bool SimilarArray(const mxnet::NDArray &arr1, const mxnet::NDArray &arr2,
   return success.load();
 }
 
+template void FallBackCompute(void (*)(nnvm::NodeAttrs const &, OpContext const &,
+                                       std::vector<TBlob, std::allocator<TBlob> > const &,
+                                       std::vector<OpReqType, std::allocator<OpReqType> > const &,
+                                       std::vector<TBlob, std::allocator<TBlob> > const &),
+                              nnvm::NodeAttrs const &, OpContext const &,
+                              std::vector<NDArray, std::allocator<NDArray> > const &,
+                              std::vector<OpReqType, std::allocator<OpReqType> > const &,
+                              std::vector<NDArray, std::allocator<NDArray> > const &);
+
+template void FallBackCompute(void (*)(OpStatePtr const &, OpContext const &,
+                                       std::vector<TBlob, std::allocator<TBlob> > const &,
+                                       std::vector<OpReqType, std::allocator<OpReqType> > const &,
+                                       std::vector<TBlob, std::allocator<TBlob> > const &),
+                              OpStatePtr const &, OpContext const &,
+                              std::vector<NDArray, std::allocator<NDArray> > const &,
+                              std::vector<OpReqType, std::allocator<OpReqType> > const &,
+                              std::vector<NDArray, std::allocator<NDArray> > const &);
+
 void OpCheck::Init(const std::vector<mxnet::NDArray> &inputs_,
                    const std::vector<mxnet::NDArray> &outputs_) {
   auto ctx = inputs_[0].ctx();
diff --git a/src/operator/rnn.cc b/src/operator/rnn.cc
@@ -633,6 +633,20 @@ static void RNNStatefulComputeCPU(const OpStatePtr& state_ptr,
     });
   });
 }
+
+static void RNNStatefulComputeExCPU(const OpStatePtr& state_ptr, const OpContext& ctx,
+                                    const std::vector<NDArray>& inputs,
+                                    const std::vector<OpReqType>& req,
+                                    const std::vector<NDArray>& outputs) {
+  if (SupportMKLDNNRNN(inputs[0])) {
+    RNNStatefulComputeCPU(state_ptr, ctx, inputs, req, outputs);
+    return;
+  }
+  int use_mkldnn_rnn = dmlc::GetEnv("MXNET_USE_MKLDNN_RNN", 1);
+  dmlc::SetEnv("MXNET_USE_MKLDNN_RNN", 0);
+  FallBackCompute(RNNStatefulCompute<cpu>, state_ptr, ctx, inputs, req, outputs);
+  dmlc::SetEnv("MXNET_USE_MKLDNN_RNN", use_mkldnn_rnn);
+}
 #endif
 
 NNVM_REGISTER_OP(RNN)
@@ -719,7 +733,7 @@ The definition of GRU here is slightly different from paper but compatible with
 .set_attr<FStatefulCompute>("FStatefulCompute<cpu>", RNNStatefulCompute<cpu>)
 #if MXNET_USE_MKLDNN == 1
 .set_attr<bool>("TIsMKLDNN", true)
-.set_attr<FStatefulComputeEx>("FStatefulComputeEx<cpu>", RNNStatefulComputeCPU)
+.set_attr<FStatefulComputeEx>("FStatefulComputeEx<cpu>", RNNStatefulComputeExCPU)
 #endif
 .set_attr<nnvm::FGradient>("FGradient", RNNGrad{"_backward_RNN"})
 .set_attr<FResourceRequestEx>("FResourceRequestEx", RNNResourceEx)
diff --git a/src/operator/subgraph/build_subgraph.cc b/src/operator/subgraph/build_subgraph.cc
@@ -24,7 +24,6 @@
  */
 #include <nnvm/graph.h>
 #include <nnvm/pass.h>
-#include <mxnet/op_attr_types.h>
 #include <unordered_set>
 #include <stack>
 #include <queue>
@@ -105,6 +104,28 @@ void ResetNodeLabels(const nnvm::Graph& g,
   subgraph_nodes->clear();
 }
 
+/*
+ * \brief Prepare NodeAttr for node. NodeAttr will be used in SubgraphSelectorV2.
+ */
+static const std::shared_ptr<NodeAttr> PrepareNodeAttr(const nnvm::Graph& g,
+                                                       const BiDirectedNode& node) {
+  const auto& indexed_graph = g.indexed_graph();
+  if (g.HasAttr("dtype") && g.HasAttr("shape") && g.HasAttr("dispatch_mode")) {
+    const auto& vdtype = g.GetAttr<nnvm::DTypeVector>("dtype");
+    const auto& vshape = g.GetAttr<mxnet::ShapeVector>("shape");
+    const auto& dispatch_modes = g.GetAttr<mxnet::DispatchModeVector>("dispatch_mode");
+    auto ret = std::make_shared<NodeAttr>();
+    ret->dispatch_mode = dispatch_modes[indexed_graph.node_id(node.node)];
+    for (const auto& e : node.node->inputs) {
+      ret->ishape.emplace_back(vshape[indexed_graph.entry_id(e)]);
+      ret->itype.emplace_back(vdtype[indexed_graph.entry_id(e)]);
+    }
+    return ret;
+  } else {
+    return nullptr;
+  }
+}
+
 /*!
  * \brief This function traverses the nodes in a computation graph from a starting
  * node following the input edges and output edges, and marks all nodes that
@@ -153,7 +174,7 @@ bool LabelSubgraph(const nnvm::Graph& g, SubgraphSelectorV2Ptr subgraph_selector
       CHECK_LT(nid, simple_nodes.size());
       const bool select_input =
           (snode->label == -1) && (!excluded_nodes || !excluded_nodes->count(snode)) &&
-          subgraph_selector->SelectInput(*cur_node, *snode);
+          subgraph_selector->SelectInput(*cur_node, *snode, PrepareNodeAttr(g, *snode));
       if (select_input) {
         // e.node is a subgraph node
         snode->label = label;
@@ -170,7 +191,7 @@ bool LabelSubgraph(const nnvm::Graph& g, SubgraphSelectorV2Ptr subgraph_selector
       CHECK_LT(nid, simple_nodes.size());
       const bool select_output =
           (snode->label == -1) && (!excluded_nodes || !excluded_nodes->count(snode)) &&
-          subgraph_selector->SelectOutput(*cur_node, *snode);
+          subgraph_selector->SelectOutput(*cur_node, *snode, PrepareNodeAttr(g, *snode));
       if (select_output) {
         // it->first is a subgraph node
         snode->label = label;
@@ -325,14 +346,16 @@ void SelectSubgraphNodes(nnvm::Graph* g, SubgraphSelectorV2Ptr subgraph_selector
                          std::vector<SubgraphSelectorV2Ptr>* subgraph_selectors,
                          const BiDirectedNode* node, const size_t snid, size_t* subgraph_id) {
   const auto& indexed_graph = g->indexed_graph();
+
   auto node_cmp = [&] (const BiDirectedNode* node1, const BiDirectedNode* node2) {
     return indexed_graph.node_id(node1->node) < indexed_graph.node_id(node2->node);
   };
-  if (simple_nodes[snid]->label == -1 && subgraph_selector->Select(*node)) {
+  if ((simple_nodes[snid]->label == -1) &&
+      subgraph_selector->Select(*node, PrepareNodeAttr(*g, *node))) {
     // pre-select nodes that can be grouped in a subgraph
     std::vector<BiDirectedNode*> preselected_nodes;
     PreSelectSubgraphNodes(*g, subgraph_selector, *subgraph_id, snid, simple_nodes,
-                           &preselected_nodes);
+                            &preselected_nodes);
 
     // filter out unqualified pre-selected nodes
     std::vector<BiDirectedNode*> filtered_nodes = subgraph_selector->Filter(preselected_nodes);
diff --git a/src/operator/subgraph/mkldnn/mkldnn_conv_property.h b/src/operator/subgraph/mkldnn/mkldnn_conv_property.h
@@ -28,7 +28,7 @@
 #include "../../nn/mkldnn/mkldnn_ops-inl.h"
 #include "../../tensor/matrix_op-inl.h"
 #include "../common.h"
-#include "../subgraph_property.h"
+#include "mkldnn_subgraph_base-inl.h"
 
 namespace mxnet {
 namespace op {
@@ -61,15 +61,15 @@ class SgMKLDNNConvSelector : public SubgraphSelector {
         disable_conv_sum_(dis_conv_sum),
         quantize_(quantize) {}
 
-  bool Select(const nnvm::Node &n) override {
+  bool Select(const nnvm::Node& n, const std::shared_ptr<NodeAttr>& node_attr) override {
     if (n.op() && n.op()->name == "Convolution") {
       const auto &param = nnvm::get<ConvolutionParam>(n.attrs.parsed);
-      if (param.kernel.ndim() == 2) {
+      if (param.kernel.ndim() == 2 && SupportMKLDNNAttr(node_attr)) {
         status_ = disable_all_ ? kSuccess : kStart;
         matched_list_.clear();
         matched_list_.push_back(&n);
         return true;
-      }
+        }
     }
     return false;
   }
@@ -161,7 +161,7 @@ class SgMKLDNNConvSelector : public SubgraphSelector {
     CHECK_GE(matched_list_.size(), 1);
     auto new_selector = SgMKLDNNConvSelector(disable_all_, disable_conv_bn_, disable_conv_act_,
                                              disable_conv_sum_, quantize_);
-    new_selector.Select(*matched_list_[0]);
+    new_selector.Select(*matched_list_[0], nullptr);
     *this = new_selector;
   }
 };
diff --git a/src/operator/subgraph/mkldnn/mkldnn_fc_post_quantize_property.h b/src/operator/subgraph/mkldnn/mkldnn_fc_post_quantize_property.h
@@ -33,7 +33,7 @@
 #include "../../nn/fully_connected-inl.h"
 #include "../../quantization/requantize-inl.h"
 #include "../common.h"
-#include "../subgraph_property.h"
+#include "mkldnn_subgraph_base-inl.h"
 
 namespace mxnet {
 namespace op {
diff --git a/src/operator/subgraph/mkldnn/mkldnn_fc_property.h b/src/operator/subgraph/mkldnn/mkldnn_fc_property.h
@@ -32,7 +32,7 @@
 #include <vector>
 #include "../common.h"
 #include "../../tensor/matrix_op-inl.h"
-#include "../subgraph_property.h"
+#include "mkldnn_subgraph_base-inl.h"
 #include "mkldnn_fc-inl.h"
 
 namespace mxnet {
@@ -58,8 +58,8 @@ class SgMKLDNNFCSelector : public SubgraphSelector {
       disable_fc_eltwise_(dis_fc_eltwise),
       quantized_(quantized) {}
 
-  bool Select(const nnvm::Node &n) override {
-    if (n.op() == Op::Get("FullyConnected")) {
+  bool Select(const nnvm::Node &n, const std::shared_ptr<NodeAttr>& node_attr) override {
+    if (n.op() == Op::Get("FullyConnected") && SupportMKLDNNAttr(node_attr)) {
       status_ = disable_fc_eltwise_ ? kSuccess : kStart;
       matched_list_.clear();
       matched_list_.push_back(&n);
@@ -150,7 +150,7 @@ class SgMKLDNNFCSelector : public SubgraphSelector {
   void Reset() override {
     CHECK_GE(matched_list_.size(), 1);
     auto new_selector = SgMKLDNNFCSelector(disable_fc_eltwise_, quantized_);
-    new_selector.Select(*matched_list_[0]);
+    new_selector.Select(*matched_list_[0], nullptr);
     *this = new_selector;
   }
 };
diff --git a/src/operator/subgraph/mkldnn/mkldnn_post_quantize_align_scale_property.h b/src/operator/subgraph/mkldnn/mkldnn_post_quantize_align_scale_property.h
@@ -24,7 +24,7 @@
 #include <string>
 #include <vector>
 #include "../common.h"
-#include "../subgraph_property.h"
+#include "mkldnn_subgraph_base-inl.h"
 
 namespace mxnet {
 namespace op {
diff --git a/src/operator/subgraph/mkldnn/mkldnn_post_quantize_property.h b/src/operator/subgraph/mkldnn/mkldnn_post_quantize_property.h
@@ -20,14 +20,14 @@
 #define MXNET_OPERATOR_SUBGRAPH_MKLDNN_MKLDNN_POST_QUANTIZE_PROPERTY_H_
 #if MXNET_USE_MKLDNN == 1
 
+#include <set>
 #include <string>
 #include <vector>
-#include <set>
-#include "../common.h"
-#include "../subgraph_property.h"
 #include "../../nn/mkldnn/mkldnn_convolution-inl.h"
-#include "mkldnn_conv-inl.h"
 #include "../../quantization/requantize-inl.h"
+#include "../common.h"
+#include "mkldnn_conv-inl.h"
+#include "mkldnn_subgraph_base-inl.h"
 
 namespace mxnet {
 namespace op {
diff --git a/src/operator/subgraph/mkldnn/mkldnn_subgraph_base-inl.h b/src/operator/subgraph/mkldnn/mkldnn_subgraph_base-inl.h
@@ -0,0 +1,42 @@
+/*
+ * 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.
+ */
+#ifndef MXNET_OPERATOR_SUBGRAPH_MKLDNN_MKLDNN_SUBGRAPH_BASE_INL_H_
+#define MXNET_OPERATOR_SUBGRAPH_MKLDNN_MKLDNN_SUBGRAPH_BASE_INL_H_
+#if MXNET_USE_MKLDNN == 1
+
+#include "../subgraph_property.h"
+
+namespace mxnet {
+namespace op {
+
+static inline bool SupportMKLDNNAttr(const std::shared_ptr<NodeAttr>& node_attr) {
+  if (node_attr) {
+    int ndim = node_attr->ishape[0].ndim();
+    return (node_attr->dispatch_mode == DispatchMode::kFComputeEx) &&
+           (node_attr->itype[0] == mshadow::kFloat32) && (ndim == 1 || ndim == 2 || ndim == 4);
+  } else {
+    return true;
+  }
+}
+
+}  // namespace op
+}  // namespace mxnet
+
+#endif  // MXNET_USE_MKLDNN == 1
+#endif  // MXNET_OPERATOR_SUBGRAPH_MKLDNN_MKLDNN_SUBGRAPH_BASE_INL_H_
diff --git a/src/operator/subgraph/subgraph_property.h b/src/operator/subgraph/subgraph_property.h
diff --git a/src/operator/tensor/matrix_op-inl.h b/src/operator/tensor/matrix_op-inl.h
diff --git a/tests/python/mkl/test_subgraph.py b/tests/python/mkl/test_subgraph.py
diff --git a/tests/python/unittest/test_operator.py b/tests/python/unittest/test_operator.py
