[XPU] add dropout bridge and unit test (#2650)

test=develop

lite/kernels/xpu/bridges/CMakeLists.txt

@@ -17,6 +17,7 @@ lite_cc_library(subgraph_bridge_batch_norm_op_xpu SRCS batch_norm_op.cc DEPS ${x
 lite_cc_library(subgraph_bridge_transpose_op_xpu SRCS transpose_op.cc DEPS ${xpu_subgraph_bridge_deps})
 lite_cc_library(subgraph_bridge_reshape_op_xpu SRCS reshape_op.cc DEPS ${xpu_subgraph_bridge_deps})
 lite_cc_library(subgraph_bridge_layer_norm_op_xpu SRCS layer_norm_op.cc DEPS ${xpu_subgraph_bridge_deps})
+lite_cc_library(subgraph_bridge_dropout_op_xpu SRCS dropout_op.cc DEPS ${xpu_subgraph_bridge_deps})
 
 set(xpu_subgraph_bridges
         subgraph_bridge_registry
@@ -32,6 +33,7 @@ set(xpu_subgraph_bridges
         subgraph_bridge_transpose_op_xpu
         subgraph_bridge_reshape_op_xpu
         subgraph_bridge_layer_norm_op_xpu
+        subgraph_bridge_dropout_op_xpu
         CACHE INTERNAL "xpu_subgraph_bridges")
 
 message(STATUS "+++++ xpu_subgraph_bridges: ${xpu_subgraph_bridges}")

lite/kernels/xpu/bridges/dropout_op.cc

+// Copyright (c) 2019 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 "lite/kernels/npu/bridges/registry.h"
+#include "lite/kernels/xpu/bridges/graph.h"
+#include "lite/kernels/xpu/bridges/utility.h"
+
+namespace paddle {
+namespace lite {
+namespace subgraph {
+namespace xpu {
+
+int DropoutConverter(void* ctx, OpLite* op) {
+  CHECK(ctx != nullptr);
+  CHECK(op != nullptr);
+  auto graph = static_cast<Graph*>(ctx);
+  auto op_info = op->op_info();
+  auto op_type = op_info->Type();
+  VLOG(3) << "[XPU] Converting " + op_type + "...";
+
+  // Create node and set params from op
+  auto x_var_name = op_info->Input("X").front();
+  auto out_var_name = op_info->Output("Out").front();
+  auto dropout_prob = op_info->GetAttr<float>("dropout_prob");
+  auto dropout_implementation =
+      op_info->GetAttr<std::string>("dropout_implementation");
+  double rate;
+  if (dropout_implementation == "downgrade_in_infer") {
+    rate = 1. - dropout_prob;
+  } else if (dropout_implementation == "upscale_in_train") {
+    rate = 1.;
+  } else {
+    LOG(FATAL) << "unsupported dropout_implementation == "
+               << dropout_implementation << " for dropout";
+  }
+  CHECK(graph->HasNode(x_var_name));
+  graph->AddNode(
+      out_var_name,
+      graph->builder_.CreateDropout(*graph->GetNode(x_var_name), rate));
+
+  return SUCCESS;
+}
+
+}  // namespace xpu
+}  // namespace subgraph
+}  // namespace lite
+}  // namespace paddle
+
+REGISTER_SUBGRAPH_BRIDGE(XPU,
+                         dropout,
+                         paddle::lite::subgraph::xpu::DropoutConverter);

lite/kernels/xpu/bridges/paddle_use_bridges.h

@@ -26,3 +26,4 @@ USE_SUBGRAPH_BRIDGE(XPU, transpose);
 USE_SUBGRAPH_BRIDGE(XPU, transpose2);
 USE_SUBGRAPH_BRIDGE(XPU, reshape);
 USE_SUBGRAPH_BRIDGE(XPU, reshape2);
+USE_SUBGRAPH_BRIDGE(XPU, dropout);

lite/tests/kernels/CMakeLists.txt

@@ -28,6 +28,7 @@ if((NOT LITE_WITH_OPENCL AND NOT LITE_WITH_FPGA) AND (LITE_WITH_X86 OR LITE_WITH
     lite_cc_test(test_kernel_transpose_compute SRCS transpose_compute_test.cc DEPS arena_framework ${xpu_kernels} ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
     lite_cc_test(test_kernel_reshape_compute SRCS reshape_compute_test.cc DEPS arena_framework ${xpu_kernels} ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
     lite_cc_test(test_kernel_layer_norm_compute SRCS layer_norm_compute_test.cc DEPS arena_framework ${xpu_kernels} ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
+    lite_cc_test(test_kernel_dropout_compute SRCS dropout_compute_test.cc DEPS arena_framework ${xpu_kernels} ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
     lite_cc_test(test_kernel_softmax_compute SRCS softmax_compute_test.cc DEPS arena_framework ${xpu_kernels} ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
 
 if(LITE_BUILD_EXTRA)

lite/tests/kernels/dropout_compute_test.cc

+// Copyright (c) 2019 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 <gtest/gtest.h>
+#include <cmath>
+#include <string>
+#include "lite/api/paddle_use_kernels.h"
+#include "lite/api/paddle_use_ops.h"
+#include "lite/core/arena/framework.h"
+
+namespace paddle {
+namespace lite {
+
+class DropoutComputeTester : public arena::TestCase {
+ protected:
+  // common attributes for this op.
+  std::string type_ = "dropout";
+  std::string input_ = "x";
+  std::string output_ = "out";
+  std::string mask_ = "mask";
+  DDim dims_{{1}};
+  float dropout_prob_ = 0.5;
+  bool fix_seed_ = true;
+  int seed_ = 1;
+  std::string dropout_implementation_ = "downgrade_in_infer";
+
+ public:
+  DropoutComputeTester(const Place& place,
+                       const std::string& alias,
+                       DDim dims,
+                       float dropout_prob,
+                       bool fix_seed,
+                       int seed,
+                       std::string dropout_implementation)
+      : TestCase(place, alias),
+        dims_(dims),
+        dropout_prob_(dropout_prob),
+        fix_seed_(fix_seed),
+        seed_(seed),
+        dropout_implementation_(dropout_implementation) {}
+
+  void RunBaseline(Scope* scope) override {
+    auto* out = scope->NewTensor(output_);
+    CHECK(out);
+    out->Resize(dims_);
+    auto* output_data = out->mutable_data<float>();
+
+    auto* x = scope->FindTensor(input_);
+    const auto* x_data = x->data<float>();
+
+    if (dropout_implementation_ == "downgrade_in_infer") {
+      float rate = 1 - dropout_prob_;
+      for (int64_t i = 0; i < dims_.production(); i++) {
+        output_data[i] = x_data[i] * rate;
+      }
+    } else if (dropout_implementation_ == "upscale_in_train") {
+      memcpy(output_data, x_data, sizeof(float) * dims_.production());
+    } else {
+      LOG(FATAL) << "unsupported dropout_implementation: "
+                 << dropout_implementation_;
+    }
+  }
+
+  void PrepareOpDesc(cpp::OpDesc* op_desc) {
+    op_desc->SetType(type_);
+    op_desc->SetInput("X", {input_});
+    op_desc->SetOutput("Out", {output_});
+    op_desc->SetOutput("Mask", {mask_});
+    op_desc->SetAttr("dropout_prob", dropout_prob_);
+    op_desc->SetAttr("fix_seed", fix_seed_);
+    op_desc->SetAttr("seed", seed_);
+    op_desc->SetAttr("dropout_implementation", dropout_implementation_);
+  }
+
+  void PrepareData() override {
+    std::vector<float> input_data(dims_.production());
+    for (int i = 0; i < dims_.production(); i++) {
+#if 0
+      float sign = i % 3 == 0 ? -1.0f : 1.0f;
+      input_data[i] = sign * static_cast<float>(i % 128) * 0.013f + 0.001;
+#else
+      input_data[i] = 1;
+#endif
+    }
+    SetCommonTensor(input_, dims_, input_data.data());
+  }
+};
+
+TEST(Dropout, precision) {
+  LOG(INFO) << "test dropout op";
+  float abs_error = 2e-5;
+  Place place;
+#if defined(LITE_WITH_XPU)
+  place = TARGET(kXPU);
+#else
+  return;
+#endif
+
+  std::vector<std::vector<int64_t>> dims{
+      /*{3} ,*/ {3, 4} /*, {3, 4, 5}, {1, 2, 3, 4}, {2, 3, 4, 5}*/};
+  for (auto dim : dims) {
+    for (auto dropout_prob : {/*0.,*/ 0.5 /*, 1.*/}) {
+      for (auto dropout_implementation :
+           {"downgrade_in_infer", "upscale_in_train"}) {
+        std::unique_ptr<arena::TestCase> tester(
+            new DropoutComputeTester(place,
+                                     "def",
+                                     DDim(dim),
+                                     dropout_prob,
+                                     true,
+                                     1,
+                                     dropout_implementation));
+        arena::Arena arena(std::move(tester), place, abs_error);
+        arena.TestPrecision({"mask"});
+      }
+    }
+  }
+}
+
+}  // namespace lite
+}  // namespace paddle