add TransferCastOpPass, DeleteScaleOpPass (#38985)

Co-authored-by: NXiaobing Wang <xiaobingw@graphcore.ai>
Co-authored-by: NAllen Guo <alleng@graphcore.ai>
Co-authored-by: NZhixin Yao <zhixiny@graphcore.ai>
Co-authored-by: NHaicheng Jiang <haichengj@graphcore.ai>
Co-authored-by: NHan Zhao <hanzhao@graphcore.ai>
Co-authored-by: NXiaobing Wang <xiaobingw@graphcore.ai>
Co-authored-by: NZhixin Yao <zhixiny@graphcore.ai>
Co-authored-by: NHaicheng Jiang <haichengj@graphcore.ai>
Co-authored-by: NHan Zhao <hanzhao@graphcore.ai>

paddle/fluid/framework/ir/ipu/delete_scale_op_pass.cc

+// Copyright (c) 2022 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 "paddle/fluid/framework/ir/ipu/delete_scale_op_pass.h"
+
+#include "paddle/fluid/framework/ddim.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+#include "paddle/fluid/framework/ir/pass_tester_helper.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/platform/device/ipu/popart_canonicalization/canonicalization_utils.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+// this pass is used to optimize scale operator whose scale = 1 and bias = 0.
+// scale will not be optimized if it is the only one operator in the graph.
+void DeleteScaleOpPass::ApplyImpl(ir::Graph* graph) const {
+  VLOG(10) << "enter DeleteScaleOpPass::ApplyImpl";
+  VLOG(10) << "Raw Graph: ";
+  VLOG(10) << DebugString(graph);
+
+  auto nodes = ir::TopologySortOperations(*graph);
+
+  // delete op
+  for (auto node : nodes) {
+    if (!node->Op()) {
+      continue;
+    }
+    auto op = node->Op();
+    if (op->Type() == "scale") {
+      auto input_var_node = node->inputs[0];
+      auto output_var_node = node->outputs[0];
+      // only optimize scale *1 + 0
+      auto scale = BOOST_GET_CONST(float, op->GetAttr("scale"));
+      auto bias = BOOST_GET_CONST(float, op->GetAttr("bias"));
+      if (scale != 1 || bias != 0) {
+        return;
+      }
+      // only one op and it is scale , do not optimize
+      if (input_var_node->inputs.size() == 0 &&
+          output_var_node->outputs.size() == 0) {
+        return;
+      }
+      VLOG(10) << "scale is to be optimized "
+               << " scale: " << scale << " bias: " << bias;
+      // build link in nodes
+      ir::Node* next_op_node = nullptr;
+      ir::Node* pre_op_node = nullptr;
+      // scale is not the last one
+      if (node->outputs[0]->outputs.size() > 0) {
+        next_op_node = node->outputs[0]->outputs[0];
+        input_var_node->outputs.push_back(next_op_node);
+        next_op_node->inputs.push_back(input_var_node);
+        platform::ipu::DisConnectNodes(output_var_node, node);
+        platform::ipu::DisConnectNodes(input_var_node, node);
+        auto var_map = next_op_node->Op()->Inputs();
+        for (auto& name_m : var_map) {
+          if (std::find(name_m.second.begin(), name_m.second.end(),
+                        output_var_node->Name()) != name_m.second.end()) {
+            std::vector<std::string> new_inputs;
+            for (auto& i_n : name_m.second) {
+              if (i_n != output_var_node->Name()) {
+                new_inputs.push_back(i_n);
+              }
+            }
+            new_inputs.push_back(input_var_node->Name());
+            next_op_node->Op()->SetInput(name_m.first, new_inputs);
+            next_op_node->Op()->Flush();
+          }
+        }
+        GraphSafeRemoveNodes(graph, {node, output_var_node});
+      } else {  // scale is not the first one
+        pre_op_node = node->inputs[0]->inputs[0];
+        output_var_node->inputs.push_back(pre_op_node);
+        pre_op_node->outputs.push_back(output_var_node);
+        platform::ipu::DisConnectNodes(input_var_node, node);
+        platform::ipu::DisConnectNodes(output_var_node, node);
+
+        auto var_map = pre_op_node->Op()->Inputs();
+        std::vector<std::string> new_outputs;
+        for (auto& name_m : var_map) {
+          if (std::find(name_m.second.begin(), name_m.second.end(),
+                        input_var_node->Name()) != name_m.second.end()) {
+            for (auto& i_n : name_m.second) {
+              if (i_n != input_var_node->Name()) {
+                new_outputs.push_back(i_n);
+              }
+            }
+            new_outputs.push_back(output_var_node->Name());
+            pre_op_node->Op()->SetOutput(name_m.first, new_outputs);
+            pre_op_node->Op()->Flush();
+          }
+        }
+        GraphSafeRemoveNodes(graph, {node, input_var_node});
+      }
+    }
+  }
+  VLOG(10) << "Post Graph: ";
+  VLOG(10) << DebugString(graph);
+  VLOG(10) << "leave DeleteScaleOpPass::ApplyImpl";
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(delete_scale_op_pass, paddle::framework::ir::DeleteScaleOpPass);

paddle/fluid/framework/ir/ipu/ipu_pass_base.h → paddle/fluid/framework/ir/ipu/delete_scale_op_pass.h

-// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2022 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.
@@ -14,22 +14,15 @@
 
 #pragma once
 
-#include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/pass.h"
-#include "paddle/fluid/framework/scope.h"
-
+#include "paddle/fluid/platform/device/ipu/ipu_utils.h"
 namespace paddle {
 namespace framework {
 namespace ir {
 
-class IPUPassBase : public Pass {
- public:
-  void Init(const std::string& repr, Graph* graph) const;
-  virtual ~IPUPassBase() {}
-
+class DeleteScaleOpPass : public Pass {
  protected:
-  mutable Graph* graph_;
-  mutable std::string repr_;
+  void ApplyImpl(ir::Graph* graph) const override;
 };
 
 }  // namespace ir

paddle/fluid/framework/ir/ipu/transfer_cast_op_pass.cc

+// Copyright (c) 2022 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 "paddle/fluid/framework/ir/ipu/transfer_cast_op_pass.h"
+
+#include "paddle/fluid/framework/ir/pass_tester_helper.h"
+#include "paddle/fluid/platform/device/ipu/ipu_backend.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+// Transfer the target dtype of Cast Op to FP16 if the original target is FP32
+// and enable FP16 mode.
+void TransferCastOpPass::ApplyImpl(ir::Graph* graph) const {
+  VLOG(10) << "enter TransferCastOpPass::ApplyImpl";
+  VLOG(10) << "Raw Graph: ";
+  VLOG(10) << DebugString(graph);
+
+  auto ipu_backend = platform::ipu::IpuBackend::GetInstance();
+  auto enable_fp16 = ipu_backend->GetIpuStrategy()->enable_fp16;
+  if (enable_fp16) {
+    for (auto* node : graph->Nodes()) {
+      if (node->IsOp() && node->Op()->Type() == "popart_cast") {
+        if (BOOST_GET_CONST(std::string, node->Op()->GetAttr("to")) ==
+            "FLOAT") {
+          node->Op()->SetAttr("to", std::string("FLOAT16"));
+        }
+      }
+    }
+  }
+
+  VLOG(10) << "Post Graph: ";
+  VLOG(10) << DebugString(graph);
+  VLOG(10) << "leave TransferCastOpPass::ApplyImpl";
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(transfer_cast_op_pass, paddle::framework::ir::TransferCastOpPass);

paddle/fluid/framework/ir/ipu/ipu_pass_base.cc → paddle/fluid/framework/ir/ipu/transfer_cast_op_pass.h

-// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2022 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.
@@ -12,16 +12,18 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/ipu/ipu_pass_base.h"
+#pragma once
+
+#include "paddle/fluid/framework/ir/pass.h"
 
 namespace paddle {
 namespace framework {
 namespace ir {
 
-void IPUPassBase::Init(const std::string& repr, Graph* graph) const {
-  repr_ = repr;
-  graph_ = graph;
-}
+class TransferCastOpPass : public Pass {
+ protected:
+  void ApplyImpl(ir::Graph* graph) const override;
+};
 
 }  // namespace ir
 }  // namespace framework