Cherry-pick from PaddlePaddle:feature/anakin-engine: Anakin subgraph support.

CMakeLists.txt

@@ -64,6 +64,7 @@ option(WITH_DISTRIBUTE  "Compile with distributed support"              OFF)
 option(WITH_PSLIB       "Compile with pslib support"                    OFF)
 option(WITH_CONTRIB     "Compile the third-party contributation"        OFF)
 option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
+# TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
 option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
@@ -190,6 +191,7 @@ include(configure)          # add paddle env configuration
 if(WITH_GPU)
     include(cuda)
     include(tensorrt)
+    include(anakin_subgraph)
 endif()
 if(WITH_MKL OR WITH_MKLML)
     include(external/anakin)

cmake/anakin_subgraph.cmake

@@ -0,0 +1,32 @@
+if(NOT WITH_GPU)
+    return()
+endif()
+
+set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
+find_path(ANAKIN_INCLUDE_DIR anakin_config.h
+    PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
+    NO_DEFAULT_PATH
+)
+
+find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
+    PATHS ${ANAKIN_ROOT}
+    $ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
+    NO_DEFAULT_PATH
+    DOC "Path to ANAKIN library.")
+
+if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
+  if(WITH_DSO)
+    set(ANAKIN_FOUND ON)
+  endif(WITH_DSO)
+else()
+    set(ANAKIN_FOUND OFF)
+endif()
+
+if(ANAKIN_FOUND)
+    message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
+    include_directories(${ANAKIN_ROOT}/include)
+    include_directories(${ANAKIN_ROOT}/include/saber)
+    link_directories(${ANAKIN_ROOT})
+    add_definitions(-DPADDLE_WITH_ANAKIN)
+endif()

cmake/tensorrt.cmake

@@ -33,5 +33,6 @@ if(TENSORRT_FOUND)
     message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
         "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
     include_directories(${TENSORRT_INCLUDE_DIR})
+    link_directories(${TENSORRT_LIBRARY})
     add_definitions(-DPADDLE_WITH_TENSORRT)
 endif()

paddle/fluid/framework/ir/CMakeLists.txt

@@ -68,16 +68,22 @@ pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(identity_scale_op_clean_pass base)
 pass_library(sync_batch_norm_pass base)
 pass_library(runtime_context_cache_pass base)
+pass_library(simplify_anakin_detection_pattern_pass inference)
+pass_library(anakin_fillconstant_elementwisemul_fuse inference)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will
 # be detected by our pass. The index here represents the number of structures in the
 # pattern. We use index 3 ~ 6, because these quantities of structures are
 # common in the models.
-foreach (index RANGE 3 6)
+foreach (index RANGE 2 6)
    file(APPEND ${pass_file} "USE_PASS(transpose_flatten${index}_concat_fuse_pass);\n")
 endforeach()
 
+foreach (index RANGE 2 6)
+   file(APPEND ${pass_file} "USE_PASS(simplify_anakin_detection_pattern_pass${index});\n")
+endforeach()
+
 if(WITH_MKLDNN)
     pass_library(mkldnn_placement_pass base mkldnn)
     pass_library(depthwise_conv_mkldnn_pass base mkldnn)

paddle/fluid/framework/ir/anakin_fillconstant_elementwisemul_fuse.cc

@@ -0,0 +1,85 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed 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.
+
+#include <memory>
+#include <string>
+
+#include "paddle/fluid/framework/ir/anakin_fillconstant_elementwisemul_fuse.h"
+#include "paddle/fluid/framework/ir/graph_viz_pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+#define GET_IR_NODE(node__) GET_IR_NODE_FROM_SUBGRAPH(node__, node__, pattern);
+#define GET_NODES                 \
+  GET_IR_NODE(fill_constant);     \
+  GET_IR_NODE(fill_constant_out); \
+  GET_IR_NODE(elementwise_mul);   \
+  GET_IR_NODE(elementwise_mul_out);
+
+std::unique_ptr<ir::Graph> AnakinFillconstantElementwisemulFuse::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  const std::string pattern_name = "anakin_fillconstant_elementwisemul_fuse";
+  FusePassBase::Init(pattern_name, graph.get());
+
+  GraphPatternDetector gpd;
+  auto* x = gpd.mutable_pattern()
+                ->NewNode("x")
+                ->assert_is_op_input("elementwise_mul", "X")
+                ->AsInput();
+
+  patterns::AnakinFillConstantElementWiseMulFuse pattern(gpd.mutable_pattern(),
+                                                         pattern_name);
+  pattern(x);
+
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    GET_NODES;
+
+    PADDLE_ENFORCE(subgraph.count(x));
+    auto* elementwise_in = subgraph.at(x);
+    float constant_value =
+        boost::get<float>(fill_constant->Op()->GetAttr("value"));
+
+    framework::OpDesc new_op_desc;
+    new_op_desc.SetType("scale");
+    new_op_desc.SetInput("X", {elementwise_in->Name()});
+    new_op_desc.SetAttr("scale", constant_value);
+    new_op_desc.SetAttr("bias", static_cast<float>(0.0));
+    new_op_desc.SetAttr("bias_after_scale", true);
+    new_op_desc.SetOutput("Out", {elementwise_mul_out->Name()});
+    new_op_desc.Flush();
+
+    // Create a new node for the fused op.
+    auto* scale_op = graph->CreateOpNode(&new_op_desc);
+
+    IR_NODE_LINK_TO(elementwise_in, scale_op);       // Input
+    IR_NODE_LINK_TO(scale_op, elementwise_mul_out);  // Output
+
+    // Delete the unneeded nodes.
+    GraphSafeRemoveNodes(graph.get(),
+                         {fill_constant, fill_constant_out, elementwise_mul});
+  };
+
+  gpd(graph.get(), handler);
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(anakin_fillconstant_elementwisemul_fuse,
+              paddle::framework::ir::AnakinFillconstantElementwisemulFuse);

paddle/fluid/framework/ir/anakin_fillconstant_elementwisemul_fuse.h

@@ -0,0 +1,35 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed 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.
+
+#pragma once
+#include <memory>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+class AnakinFillconstantElementwisemulFuse : public FusePassBase {
+ public:
+  virtual ~AnakinFillconstantElementwisemulFuse() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -1470,6 +1470,171 @@ PDNode *patterns::TransposeFlattenConcat::operator()(
   return concat_out;
 }
 
+PDNode *patterns::AnakinDetectionPattern::operator()(
+    std::vector<PDNode *> conv_in, int times) {
+  // The times represents the repeat times of the
+  // {prior_box, prior_box_loc_out, flatten, prior_box_var_out, reshape}
+  const int kNumFields = 7;
+  const int kPriorBoxLocOffset = 1;
+  const int kReshape1Offset = 2;
+  const int kReshape1OutOffset = 3;
+  const int kPriorBoxVarOffset = 4;
+  const int kReshape2Offset = 5;
+  const int kReshape2OutOffset = 6;
+
+  const int kBoxCoderThirdInputOffset = times;
+  const int kMultiClassSecondInputNmsOffset = times + 1;
+
+  std::vector<PDNode *> nodes;
+
+  for (int i = 0; i < times; i++) {
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("prior_box" + std::to_string(i)))
+            ->assert_is_op("density_prior_box"));
+    nodes.push_back(pattern->NewNode(GetNodeName("box_out" + std::to_string(i)))
+                        ->assert_is_op_output("density_prior_box", "Boxes")
+                        ->assert_is_op_input("reshape2", "X")
+                        ->AsIntermediate());
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("reshape1" + std::to_string(i)))
+            ->assert_is_op("reshape2"));
+
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("reshape1_out" + std::to_string(i)))
+            ->assert_is_op_output("reshape2")
+            ->assert_is_op_nth_input("concat", "X", i)
+            ->AsIntermediate());
+
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("box_var_out" + std::to_string(i)))
+            ->assert_is_op_output("density_prior_box", "Variances")
+            ->assert_is_op_input("reshape2", "X")
+            ->AsIntermediate());
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("reshape2" + std::to_string(i)))
+            ->assert_is_op("reshape2"));
+
+    nodes.push_back(
+        pattern->NewNode(GetNodeName("reshape2_out" + std::to_string(i)))
+            ->assert_is_op_output("reshape2")
+            ->assert_is_op_nth_input("concat", "X", i)
+            ->AsIntermediate());
+  }
+
+  auto concat_op1 = pattern->NewNode(GetNodeName("concat1"))
+                        ->assert_is_op("concat")
+                        ->assert_op_has_n_inputs("concat", times);
+  auto concat_out1 = pattern->NewNode(GetNodeName("concat1_out"))
+                         ->assert_is_op_output("concat")
+                         ->AsIntermediate();
+
+  auto concat_op2 = pattern->NewNode(GetNodeName("concat2"))
+                        ->assert_is_op("concat")
+                        ->assert_op_has_n_inputs("concat", times);
+  auto concat_out2 = pattern->NewNode(GetNodeName("concat2_out"))
+                         ->assert_is_op_output("concat")
+                         ->AsIntermediate();
+
+  auto box_coder_op = pattern->NewNode(GetNodeName("box_coder"))
+                          ->assert_is_op("box_coder")
+                          ->assert_op_has_n_inputs("box_coder", 3);
+
+  auto box_coder_out = pattern->NewNode(GetNodeName("box_coder_out"))
+                           ->assert_is_op_output("box_coder")
+                           ->AsIntermediate();
+
+  auto transpose_before_nms =
+      pattern->NewNode(GetNodeName("transpose_before_nms"))
+          ->assert_is_op("transpose2");
+
+  auto transpose_before_nms_out =
+      pattern->NewNode(GetNodeName("transpose_before_nms_out"))
+          ->assert_is_op_output("transpose2")
+          ->assert_is_op_input("multiclass_nms", "Scores")
+          ->AsIntermediate();
+
+  auto multiclass_nms_op = pattern->NewNode(GetNodeName("multiclass_nms"))
+                               ->assert_is_op("multiclass_nms")
+                               ->assert_op_has_n_inputs("multiclass_nms", 2);
+
+  auto multiclass_nms_out = pattern->NewNode(GetNodeName("multiclass_nms_out"))
+                                ->assert_is_op_output("multiclass_nms")
+                                ->AsOutput();
+
+  std::vector<PDNode *> reshape1_outs;
+  std::vector<PDNode *> reshape2_outs;
+
+  for (int i = 0; i < times; i++) {
+    conv_in[i]->AsInput();
+    // prior_box
+    nodes[i * kNumFields]->LinksFrom({conv_in[i]});
+    // prior_box box out
+    nodes[i * kNumFields + kPriorBoxLocOffset]->LinksFrom(
+        {nodes[i * kNumFields]});
+    // reshape
+    nodes[i * kNumFields + kReshape1Offset]->LinksFrom(
+        {nodes[i * kNumFields + kPriorBoxLocOffset]});
+    // reshape_out
+    nodes[i * kNumFields + kReshape1OutOffset]->LinksFrom(
+        {nodes[i * kNumFields + kReshape1Offset]});
+
+    nodes[i * kNumFields + kPriorBoxVarOffset]->LinksFrom(
+        {nodes[i * kNumFields]});
+    // reshape
+    nodes[i * kNumFields + kReshape2Offset]->LinksFrom(
+        {nodes[i * kNumFields + kPriorBoxVarOffset]});
+    // reshape_out
+    nodes[i * kNumFields + kReshape2OutOffset]->LinksFrom(
+        {nodes[i * kNumFields + kReshape2Offset]});
+
+    reshape1_outs.push_back(nodes[i * kNumFields + kReshape1OutOffset]);
+    reshape2_outs.push_back(nodes[i * kNumFields + kReshape2OutOffset]);
+  }
+
+  concat_op1->LinksFrom(reshape1_outs);
+  concat_op2->LinksFrom(reshape2_outs);
+  concat_out1->LinksFrom({concat_op1});
+  concat_out2->LinksFrom({concat_op2});
+
+  conv_in[kBoxCoderThirdInputOffset]->AsInput();
+  conv_in[kMultiClassSecondInputNmsOffset]->AsInput();
+
+  box_coder_op->LinksFrom(
+      {concat_out1, concat_out2, conv_in[kBoxCoderThirdInputOffset]});
+  box_coder_out->LinksFrom({box_coder_op});
+
+  transpose_before_nms->LinksFrom({conv_in[kMultiClassSecondInputNmsOffset]});
+  transpose_before_nms_out->LinksFrom({transpose_before_nms});
+
+  multiclass_nms_op->LinksFrom({box_coder_out, transpose_before_nms_out})
+      .LinksTo({multiclass_nms_out});
+
+  return multiclass_nms_out;
+}
+
+PDNode *patterns::AnakinFillConstantElementWiseMulFuse::operator()(
+    PDNode *elementwise_op_input) {
+  auto fill_constant =
+      pattern->NewNode(fill_constant_repr())->assert_is_op("fill_constant");
+
+  auto fill_constant_out = pattern->NewNode(fill_constant_out_repr())
+                               ->assert_is_op_output("fill_constant")
+                               ->assert_is_op_input("elementwise_mul", "Y")
+                               ->AsIntermediate();
+
+  auto elementwise_mul_op =
+      pattern->NewNode(elementwise_mul_repr())->assert_is_op("elementwise_mul");
+
+  auto elementwise_mul_out = pattern->NewNode(elementwise_mul_out_repr())
+                                 ->assert_is_op_output("elementwise_mul")
+                                 ->AsOutput();
+
+  fill_constant_out->LinksFrom({fill_constant});
+  elementwise_mul_op->LinksFrom({elementwise_op_input, fill_constant_out});
+  elementwise_mul_out->LinksFrom({elementwise_mul_op});
+  return elementwise_mul_out;
+}
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle