Enable the detection of subgraph composed of grad ops (PaddlePaddle#2…

…1223)

* Add the first implememtation of fusion_group op PaddlePaddle#19621 (#3)

* Add the dynamic load of nvrtc, and support runtime compiling of CUDA kernel using nvrtc.
test=develop

* Call CUDA driver api to launch the kernel compiled by nvrtc.
test=develop

* Disable for mac and windows.
test=develop

* Refine the codes to support manually specified num_threads and workload_per_thread.
test=develop

* Refine the CUDA kernel to support large dims.
test=develop

* Add DeviceCodePool to manage all device codes.

* Add the first implementation fusion_group op.

* Add unit-test for fusion_group op.

* Add the check of result.

* Add the check of nvrtc in unit-test.
test=develop

* Add comment to explain the inputs, outputs and features of fusion_group op.
test=develop

* Disable fusion_group op for mac and windows.
test=develop

* Make the compiling of device code return status instead of hanging up.
test=develop

* Add the check of whether there is CUDA driver library, and do not core dump when failing to call the CUDA driver API.

* Unify fusion_group_op's input and output names.
test=develop

* Add the check of CUDA driver library in unittest.
test=develop

* Enable generating code for a given subgraph. PaddlePaddle#21126 (#4)

* Enable generating code for a given subgraph.

* Support sorting the subgraph.

* Remove the rearange of expressions because we use the sorted subgraph directly.

* Enable generating code for a subgraph which is composed of grad ops.

* Use expression information to check the accuracy in unittest.

* Separate load and store from computation expressions.
test=develop

* Improve the loading statements in generated codes.
test=develop

* Remove unused arguments from formal list.
test=develop

* Enable the detection of subgraph of grad ops.

* Generate code for detected subgraph in fusion_group_pass.

* Add an option in BuildStrategy to enable fusion_group_pass and add unittest.
test=develop

* Fix a bug when checking whether the shape of all inputs are the same.

* Add debug information.

* Remove subgraph_detector from inference/analysis to the common framework/ir directory. (#5)

test=develop

* Call subgraph_detector in fusion_group pass.
test=develop

* Disable fusion_group when WITH_GPU is OFF.
test=develop

* Refine all PADDLE_ENFORCE message.
test=develop

* Fix the case that some inputs are not defined in grad ops, and set op_role for fused op.
test=develop

* Follow review comments.
test=develop

paddle/fluid/framework/details/CMakeLists.txt

@@ -64,7 +64,14 @@ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope d
 
 cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
 
-set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass buffer_shared_inplace_op_pass buffer_shared_cross_op_memory_reuse_pass)
+set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto
+    sequential_execution_pass
+    modify_op_lock_and_record_event_pass
+    all_reduce_deps_pass
+    reference_count_pass
+    eager_deletion_pass
+    buffer_shared_inplace_op_pass
+    buffer_shared_cross_op_memory_reuse_pass)
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
 
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
@@ -91,23 +98,22 @@ cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executo
         DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context)
 cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle)
 
+set(IR_PASS_DEPS graph_viz_pass multi_devices_graph_pass
+    multi_devices_graph_print_pass multi_devices_graph_check_pass
+    fuse_elewise_add_act_pass fuse_bn_act_pass 
+    multi_batch_merge_pass 
+    fuse_relu_depthwise_conv_pass
+    lock_free_optimize_pass
+    coalesce_grad_tensor_pass fuse_all_reduce_op_pass backward_optimizer_op_deps_pass
+    fuse_adam_op_pass fuse_sgd_op_pass fuse_momentum_op_pass
+    sync_batch_norm_pass runtime_context_cache_pass)
+if(WITH_GPU)
+  set(IR_PASS_DEPS ${IR_PASS_DEPS} fusion_group_pass)
+endif()
 if(WITH_NGRAPH) 
-  set(NGRAPH_BS_DEPS ngraph)
-else()
-  set(NGRAPH_BS_DEPS)
+  set(IR_PASS_DEPS ${IR_PASS_DEPS} ngraph)
 endif()
-
-cc_library(build_strategy SRCS build_strategy.cc DEPS
-        graph_viz_pass multi_devices_graph_pass
-        multi_devices_graph_print_pass multi_devices_graph_check_pass
-        fuse_elewise_add_act_pass fuse_bn_act_pass multi_batch_merge_pass
-        fuse_relu_depthwise_conv_pass
-        lock_free_optimize_pass
-        coalesce_grad_tensor_pass fuse_all_reduce_op_pass backward_optimizer_op_deps_pass
-        fuse_adam_op_pass fuse_sgd_op_pass fuse_momentum_op_pass
-        sync_batch_norm_pass runtime_context_cache_pass
-        pass_builder
-        ${NGRAPH_BS_DEPS})
+cc_library(build_strategy SRCS build_strategy.cc DEPS pass_builder ${IR_PASS_DEPS})
 
 if (WITH_MKLDNN)
   target_link_libraries(build_strategy mkldnn_placement_pass)

paddle/fluid/framework/details/build_strategy.cc

@@ -165,9 +165,12 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   void AppendOpFusePasses() {
     AppendPassWithCheck(strategy_.fuse_relu_depthwise_conv_,
                         "fuse_relu_depthwise_conv_pass");
+    AppendPassWithCheck(strategy_.fuse_bn_act_ops_, "fuse_bn_act_pass");
+#ifdef PADDLE_WITH_CUDA
+    AppendPassWithCheck(strategy_.enable_auto_fusion_, "fusion_group_pass");
+#endif
     AppendPassWithCheck(strategy_.fuse_elewise_add_act_ops_,
                         "fuse_elewise_add_act_pass");
-    AppendPassWithCheck(strategy_.fuse_bn_act_ops_, "fuse_bn_act_pass");
     // for single card training, fuse_all_reduce_ops is unnecessary.
     // coalesce_grad_tensor_pass should be before of MultiDevPass.
     AppendPassWithCheck(strategy_.fuse_all_reduce_ops_,
@@ -370,6 +373,12 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
                         "GPU, skipped.";
         continue;
       }
+    } else if (pass->Type() == "fusion_group_pass") {
+      pass->Set<bool>("use_gpu", new bool(use_cuda));
+      if (!use_cuda) {
+        LOG(WARNING) << "fusion_group_pass is only supported on GPU, skipped.";
+        continue;
+      }
     } else if (pass->Type() == "fuse_bn_act_pass") {
       if (!use_cuda) {
         LOG(WARNING) << "fuse_bn_act_pass is only supported on "
@@ -427,3 +436,6 @@ USE_PASS(mkldnn_placement_pass);
 #ifdef PADDLE_WITH_NGRAPH
 USE_PASS(ngraph_subgraph_pass);
 #endif
+#ifdef PADDLE_WITH_CUDA
+USE_PASS(fusion_group_pass);
+#endif

paddle/fluid/framework/details/build_strategy.h

@@ -86,8 +86,9 @@ struct BuildStrategy {
   // Operator fusion
   // TODO(dev-paddle): fuse_elewise_add_act_ops may cause some models have
   // cycle.
-  bool fuse_elewise_add_act_ops_{false};
   bool fuse_bn_act_ops_{false};
+  bool fuse_elewise_add_act_ops_{false};
+  bool enable_auto_fusion_{false};
   // Fuse_all_optimizer_ops and fuse_all_reduce_ops require that gradients
   // should not be sparse types
   boost::optional<bool> fuse_all_optimizer_ops_{false};

paddle/fluid/framework/ir/CMakeLists.txt

@@ -6,7 +6,7 @@ file(APPEND ${pass_file} "\#include \"paddle/fluid/framework/ir/pass.h\"\n")
 add_subdirectory(fuse_optimizer_ops_pass)
 add_subdirectory(memory_optimize_pass)
 add_subdirectory(multi_devices_graph_pass)
-if(NOT APPLE AND NOT WIN32)
+if(NOT APPLE AND NOT WIN32 AND WITH_GPU)
     add_subdirectory(fusion_group)
 endif()
 

paddle/fluid/framework/ir/fusion_group/CMakeLists.txt

@@ -1,9 +1,11 @@
-cc_library(code_generator SRCS operation.cc code_generator.cc code_generator_helper.cc DEPS graph)
+cc_library(code_generator
+    SRCS operation.cc code_generator.cc code_generator_helper.cc
+    DEPS graph subgraph_detector)
 if(WITH_GPU)
     cc_test(test_code_generator SRCS code_generator_tester.cc DEPS code_generator device_code lod_tensor graph_viz_pass)
 endif()
 
 cc_library(fusion_group_pass
     SRCS fusion_group_pass.cc elementwise_group_detector.cc
-    DEPS graph_pattern_detector pass code_generator)
+    DEPS subgraph_detector fuse_pass_base code_generator device_code)
 cc_test(test_fusion_group_pass SRCS fusion_group_pass_tester.cc DEPS fusion_group_pass graph_viz_pass)

paddle/fluid/framework/ir/fusion_group/code_generator.cc

@@ -33,7 +33,7 @@ CodeGenerator::CodeGenerator() {
 
 std::string CodeGenerator::Generate(SubGraph* subgraph) {
   std::vector<OperationExpression> expressions = ConvertToExpressions(subgraph);
-  return Generate(subgraph->func_name, expressions);
+  return Generate(subgraph->GetFuncName(), expressions);
 }
 
 static bool HasInput(Node* n, std::string name) {

paddle/fluid/framework/ir/fusion_group/code_generator_tester.cc

@@ -227,7 +227,7 @@ std::vector<fusion_group::OperationExpression> TestMain(
   std::string code_str = code_generator.Generate(subgraph);
   VLOG(3) << code_str;
 
-  TestMainImpl(subgraph->func_name, code_str, cpu_tensors, n, input_ids,
+  TestMainImpl(subgraph->GetFuncName(), code_str, cpu_tensors, n, input_ids,
                output_ids);
 
   // Need to check the accuracy according to expressions.

paddle/fluid/framework/ir/fusion_group/elementwise_group_detector.cc

@@ -13,8 +13,11 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/ir/fusion_group/elementwise_group_detector.h"
+#include <string>
+#include <unordered_set>
+#include <vector>
 #include "paddle/fluid/framework/ir/fusion_group/operation.h"
-#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+#include "paddle/fluid/framework/ir/subgraph_detector.h"
 
 namespace paddle {
 namespace framework {
@@ -26,20 +29,22 @@ static std::unordered_set<std::string> unary_op_types;
 
 static std::unordered_set<std::string>& GetBinaryOpTypes() {
   if (binary_op_types.empty()) {
-    binary_op_types = OperationMap::Instance().Find(0, 2);
+    binary_op_types =
+        OperationMap::Instance().Find(/* type= */ 0, /* num_operands= */ 2);
   }
   return binary_op_types;
 }
 
 static std::unordered_set<std::string>& GetUnaryOpTypes() {
   if (unary_op_types.empty()) {
-    unary_op_types = OperationMap::Instance().Find(0, 1);
+    unary_op_types =
+        OperationMap::Instance().Find(/* type= */ 0, /* num_operands= */ 1);
   }
   return unary_op_types;
 }
 
 static bool IsSpecifiedOp(const std::unordered_set<std::string>& op_types,
-                          Node* n) {
+                          const Node* n) {
   if (n && n->IsOp() && n->Op() && n->outputs.size() > 0U) {
     auto iter = op_types.find(n->Op()->Type());
     if (iter != op_types.end()) {
@@ -49,114 +54,63 @@ static bool IsSpecifiedOp(const std::unordered_set<std::string>& op_types,
   return false;
 }
 
-static bool IsBinaryOp(Node* n) {
-  if (IsSpecifiedOp(GetBinaryOpTypes(), n) && n->inputs.size() == 2U) {
-    auto* x = n->inputs[0];
-    auto* y = n->inputs[1];
+static bool IsGradOp(const Node* n) {
+  PADDLE_ENFORCE_EQ(n && n->IsOp() && n->Op(), true,
+                    platform::errors::InvalidArgument(
+                        "Expected node %p to be an operator node.", n));
+  std::string suffix = "_grad";
+  std::string op_type = n->Op()->Type();
+  size_t pos = op_type.rfind(suffix);
+  return pos != std::string::npos &&
+         pos == (op_type.length() - suffix.length());
+}
 
-    std::vector<int64_t> x_shape;
-    std::vector<int64_t> y_shape;
-    if (x && x->IsVar() && x->Var()) {
-      x_shape = x->Var()->GetShape();
-    }
-    if (y && y->IsVar() && y->Var()) {
-      y_shape = y->Var()->GetShape();
-    }
-    if (x_shape.size() == 0U || x_shape.size() != y_shape.size()) {
+static bool IsEqualAndNotEmpty(const std::vector<int64_t>& l,
+                               const std::vector<int64_t>& r) {
+  return l.size() != 0U && r.size() != 0U && l == r;
+}
+
+static bool IsBinaryOp(const Node* n) {
+  if (IsSpecifiedOp(GetBinaryOpTypes(), n)) {
+    if ((!IsGradOp(n) && n->inputs.size() != 2U) || n->inputs.size() == 0U) {
       return false;
     }
-    for (size_t i = 0; i < x_shape.size(); ++i) {
-      if (x_shape[i] != y_shape[i]) {
+
+    // The shape of all inputs should be the same.
+    std::vector<int64_t> shape_0;
+    for (size_t i = 0; i < n->inputs.size(); ++i) {
+      auto* in_i = n->inputs[i];
+      if (!(in_i && in_i->IsVar() && in_i->Var())) {
         return false;
       }
-    }
-    return true;
-  }
-  return false;
-}
-
-static bool IsUnaryOp(Node* n) { return IsSpecifiedOp(GetUnaryOpTypes(), n); }
-
-bool ElementwiseGroupDetector::IsElementwiseOp(Node* n) {
-  return IsBinaryOp(n) || IsUnaryOp(n);
-}
 
-bool ElementwiseGroupDetector::IsInputOfElementwiseOp(Node* n,
-                                                      std::string name) {
-  if (n && n->IsVar() && n->Var()) {
-    for (auto* op : n->outputs) {
-      if (IsElementwiseOp(op)) {
-        if (name.empty()) {
-          return true;
-        } else if (IsNthInput(n, op, name, 0)) {
-          return true;
+      std::vector<int64_t> shape_i = in_i->Var()->GetShape();
+      if (i == 0U) {
+        shape_0 = shape_i;
+      } else {
+        if (!IsEqualAndNotEmpty(shape_0, shape_i)) {
+          return false;
         }
       }
     }
+    return true;
   }
   return false;
 }
 
-bool ElementwiseGroupDetector::IsOutputOfElementwiseOp(Node* n) {
-  if (n && n->IsVar() && n->Var()) {
-    for (auto* op : n->inputs) {
-      if (IsElementwiseOp(op)) {
-        return true;
-      }
-    }
-  }
-  return false;
+static bool IsUnaryOp(const Node* n) {
+  return IsSpecifiedOp(GetUnaryOpTypes(), n);
 }
 
-int ElementwiseGroupDetector::Search(Node* n, std::vector<Node*> except_nodes) {
-  std::unordered_set<Node*> except_nodes_set;
-  for (size_t i = 0; i < except_nodes.size(); ++i) {
-    except_nodes_set.insert(except_nodes[i]);
-  }
-
-  int num_operations = 0;
-  if (IsElementwiseOp(n)) {
-    subgraph_.Insert(n);
-    num_operations += 1;
-    for (auto* var : n->inputs) {
-      subgraph_.Insert(var);
-      if (except_nodes_set.find(var) == except_nodes_set.end()) {
-        num_operations += Search(var, {n});
-      }
-    }
-    for (auto* var : n->outputs) {
-      subgraph_.Insert(var);
-      if (except_nodes_set.find(var) == except_nodes_set.end()) {
-        num_operations += Search(var, {n});
-      }
-    }
-  } else if (n && n->IsVar() && n->Var()) {
-    for (auto* op : n->inputs) {
-      if (IsElementwiseOp(op) &&
-          except_nodes_set.find(op) == except_nodes_set.end()) {
-        num_operations += Search(op, {n});
-      }
-    }
-    for (auto* op : n->outputs) {
-      if (IsElementwiseOp(op) &&
-          except_nodes_set.find(op) == except_nodes_set.end()) {
-        num_operations += Search(op, {n});
-      }
-    }
-  }
-  return num_operations;
+bool ElementwiseGroupDetector::IsElementwiseOp(const Node* n) {
+  return IsBinaryOp(n) || IsUnaryOp(n);
 }
 
-int ElementwiseGroupDetector::operator()(Node* n) {
-  if (!IsOutputOfElementwiseOp(n) && IsInputOfElementwiseOp(n, "X")) {
-    name_ = n->Name();
-    subgraph_.Insert(n);
-    num_operations_ = Search(n, n->inputs);
-    VLOG(4) << "Detect elementwise subgraph begin with " << name_ << ", "
-            << num_operations_ << " operations, " << GetSubgraph().GetNumNodes()
-            << " nodes";
-  }
-  return num_operations_;
+std::vector<std::vector<Node*>> ElementwiseGroupDetector::operator()(
+    Graph* graph) {
+  auto teller = [&](const Node* n) -> bool { return IsElementwiseOp(n); };
+
+  return SubgraphDetector(graph, teller)();
 }
 
 }  // namespace fusion_group

paddle/fluid/framework/ir/fusion_group/elementwise_group_detector.h

@@ -14,10 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#include <string>
-#include <unordered_set>
 #include <vector>
-#include "paddle/fluid/framework/ir/fusion_group/subgraph.h"
+#include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/node.h"
 
 namespace paddle {
@@ -27,21 +25,10 @@ namespace fusion_group {
 
 class ElementwiseGroupDetector {
  public:
-  int operator()(Node* n);
-
-  SubGraph GetSubgraph() const { return subgraph_; }
-
- private:
-  bool IsElementwiseOp(Node* n);
-  bool IsInputOfElementwiseOp(Node* n, std::string name = "");
-  bool IsOutputOfElementwiseOp(Node* n);
-
-  int Search(Node* n, std::vector<Node*> except_nodes = {});
+  std::vector<std::vector<Node*>> operator()(Graph* graph);
 
  private:
-  std::string name_;
-  int num_operations_{0};
-  SubGraph subgraph_;
+  bool IsElementwiseOp(const Node* n);
 };
 
 }  // namespace fusion_group