Fea/infer for1.3 (#15760)

paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc

@@ -38,9 +38,13 @@ std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
                       ->assert_is_op("scale")
                       ->assert_op_attr<float>("scale", 1.)
                       ->assert_op_attr<float>("bias", 0.);
-  auto scale_out = detector.mutable_pattern()
-                       ->NewNode("scale_out")
-                       ->assert_is_op_output("scale");
+  auto scale_out =
+      detector.mutable_pattern()
+          ->NewNode("scale_out")
+          ->assert_is_op_output("scale")
+          // scale's output var should has only one consumer, or it can't be
+          // removed.
+          ->assert_more([](Node* x) { return x->outputs.size() == 1UL; });
 
   pre_op->LinksTo({scale_in});
   scale_op->LinksFrom({scale_in}).LinksTo({scale_out});

paddle/fluid/inference/api/paddle_api.h

@@ -16,6 +16,12 @@
 /*! \file paddle_api.h
  */
 
+/*! \mainpage Paddle Inference APIs
+ * \section intro_sec Introduction
+ * The Paddle inference library aims to offer an high performance inference SDK
+ * for Paddle users.
+ */
+
 #include <cassert>
 #include <memory>
 #include <string>
@@ -34,26 +40,49 @@ enum PaddleDType {
 };
 
 /**
- *\brief Memory menager for PaddleTensor.
+ * \brief Memory manager for `PaddleTensor`.
  *
- *The PaddleBuf holds a buffer for data input or output. The memory can be
- *allocated by user or by PaddleBuf itself, but in any case, the PaddleBuf
- *should be reused for better performance.
+ * The PaddleBuf holds a buffer for data input or output. The memory can be
+ * allocated by user or by PaddleBuf itself, but in any case, the PaddleBuf
+ * should be reused for better performance.
  *
- *For user allocated memory, the following API can be used:
- *- PaddleBuf(void* data, size_t length) to set an external memory by
- *specifying
- *  the memory address and length.
- *- Reset(void* data, size_t length) to reset the PaddleBuf with an external
+ * For user allocated memory, the following API can be used:
+ * - PaddleBuf(void* data, size_t length) to set an external memory by
+ * specifying the memory address and length.
+ * - Reset(void* data, size_t length) to reset the PaddleBuf with an external
  *memory.
- *ATTENTION, for user allocated memory, deallocation should be done by users
+ * ATTENTION, for user allocated memory, deallocation should be done by users
  *externally after the program finished. The PaddleBuf won't do any allocation
  *or deallocation.
  *
- *To have the PaddleBuf allocate and manage the memory:
- *- PaddleBuf(size_t length) will allocate a memory of size `length`.
- *- Resize(size_t length) resize the memory to no less than `length`, ATTENTION
+ * To have the PaddleBuf allocate and manage the memory:
+ * - PaddleBuf(size_t length) will allocate a memory of size `length`.
+ * - Resize(size_t length) resize the memory to no less than `length`, ATTENTION
  *  if the allocated memory is larger than `length`, nothing will done.
+ *
+ * Usage:
+ *
+ * Let PaddleBuf manage the memory internally.
+ * \code{cpp}
+ * const int num_elements = 128;
+ * PaddleBuf buf(num_elements * sizeof(float));
+ * \endcode
+ *
+ * Or
+ * \code{cpp}
+ * PaddleBuf buf;
+ * buf.Resize(num_elements * sizeof(float));
+ * \endcode
+ * Works the exactly the same.
+ *
+ * One can also make the `PaddleBuf` use the external memory.
+ * \code{cpp}
+ * PaddleBuf buf;
+ * void* external_memory = new float[num_elements];
+ * buf.Reset(external_memory, num_elements*sizeof(float));
+ * ...
+ * delete[] external_memory; // manage the memory lifetime outside.
+ * \endcode
  */
 class PaddleBuf {
  public:
@@ -78,7 +107,7 @@ class PaddleBuf {
   /** Tell whether the buffer is empty.
    */
   bool empty() const { return length_ == 0; }
-  /** Get the memory address.
+  /** Get the data's memory address.
    */
   void* data() const { return data_; }
   /** Get the memory length.
@@ -110,7 +139,8 @@ struct PaddleTensor {
 };
 
 enum class PaddlePlace { kUNK = -1, kCPU, kGPU };
-/** Tensor without copy, currently only supports AnalysisPredictor.
+
+/** Tensor without copy, currently only supports `AnalysisPredictor`.
  */
 class ZeroCopyTensor {
  public:
@@ -269,9 +299,11 @@ struct NativeConfig : public PaddlePredictor::Config {
  *
  * Usage:
  *
+ * \code{.cpp}
  * NativeConfig config;
  * ... // change the configs.
  * auto native_predictor = CreatePaddlePredictor(config);
+ * \endcode
  *
  * FOR EXTENSION DEVELOPER:
  * Different predictors are designated by config type. Similar configs can be

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -66,8 +66,54 @@ void GpuPassStrategy::EnableMKLDNN() {
   LOG(ERROR) << "GPU not support MKLDNN yet";
 }
 
+GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
+  passes_.assign({
+    "infer_clean_graph_pass",                        //
+        "identity_scale_op_clean_pass",              //
+        "conv_affine_channel_fuse_pass",             //
+        "conv_eltwiseadd_affine_channel_fuse_pass",  //
+        "conv_bn_fuse_pass",                         //
+#if CUDNN_VERSION >= 7100  // To run conv_fusion, the version of cudnn must be
+                           // guaranteed at least v7
+        "conv_elementwise_add_act_fuse_pass",   //
+        "conv_elementwise_add2_act_fuse_pass",  //
+        "conv_elementwise_add_fuse_pass",       //
+#endif
+  });
+
+  for (int i = 6; i >= 3; i--) {
+    passes_.push_back("transpose_flatten" + std::to_string(i) +
+                      "_concat_fuse_pass");
+  }
+  use_gpu_ = true;
+}
+
 void PaddlePassBuilder::AppendAnalysisPass(const std::string &pass) {
   analysis_passes_.push_back(pass);
 }
 
+CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
+  // NOTE the large fusions should be located in the front, so that they will
+  // not be damaged by smaller ones.
+  passes_.assign({
+      "infer_clean_graph_pass",         //
+      "attention_lstm_fuse_pass",       //
+      "seqpool_concat_fuse_pass",       //
+      "seqconv_eltadd_relu_fuse_pass",  //
+      // "embedding_fc_lstm_fuse_pass", //
+      "fc_lstm_fuse_pass",             //
+      "mul_lstm_fuse_pass",            //
+      "fc_gru_fuse_pass",              //
+      "mul_gru_fuse_pass",             //
+      "seq_concat_fc_fuse_pass",       //
+      "fc_fuse_pass",                  //
+      "repeated_fc_relu_fuse_pass",    //
+      "squared_mat_sub_fuse_pass",     //
+      "conv_bn_fuse_pass",             //
+      "conv_eltwiseadd_bn_fuse_pass",  //
+      "is_test_pass",                  //
+      "identity_scale_op_clean_pass",  //
+  });
+  use_gpu_ = false;
+}
 }  // namespace paddle

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -97,30 +97,7 @@ class PassStrategy : public PaddlePassBuilder {
  */
 class CpuPassStrategy : public PassStrategy {
  public:
-  CpuPassStrategy() : PassStrategy({}) {
-    // NOTE the large fusions should be located in the front, so that they will
-    // not be damaged by smaller ones.
-    passes_.assign({
-        "infer_clean_graph_pass",         //
-        "attention_lstm_fuse_pass",       //
-        "seqpool_concat_fuse_pass",       //
-        "seqconv_eltadd_relu_fuse_pass",  //
-        // "embedding_fc_lstm_fuse_pass", //
-        "fc_lstm_fuse_pass",             //
-        "mul_lstm_fuse_pass",            //
-        "fc_gru_fuse_pass",              //
-        "mul_gru_fuse_pass",             //
-        "seq_concat_fc_fuse_pass",       //
-        "fc_fuse_pass",                  //
-        "repeated_fc_relu_fuse_pass",    //
-        "squared_mat_sub_fuse_pass",     //
-        "conv_bn_fuse_pass",             //
-        "conv_eltwiseadd_bn_fuse_pass",  //
-        "is_test_pass",                  //
-        "identity_scale_op_clean_pass",  //
-    });
-    use_gpu_ = false;
-  }
+  CpuPassStrategy();
 
   explicit CpuPassStrategy(const CpuPassStrategy &other)
       : PassStrategy(other.AllPasses()) {}
@@ -153,27 +130,7 @@ class CpuPassStrategy : public PassStrategy {
  */
 class GpuPassStrategy : public PassStrategy {
  public:
-  GpuPassStrategy() : PassStrategy({}) {
-    passes_.assign({
-      "infer_clean_graph_pass",                        //
-          "identity_scale_op_clean_pass",              //
-          "conv_affine_channel_fuse_pass",             //
-          "conv_eltwiseadd_affine_channel_fuse_pass",  //
-          "conv_bn_fuse_pass",                         //
-#if CUDNN_VERSION >= 7100  // To run conv_fusion, the version of cudnn must be
-                           // guaranteed at least v7
-          "conv_elementwise_add_act_fuse_pass",   //
-          "conv_elementwise_add2_act_fuse_pass",  //
-          "conv_elementwise_add_fuse_pass",       //
-#endif
-    });
-
-    for (int i = 6; i >= 3; i--) {
-      passes_.push_back("transpose_flatten" + std::to_string(i) +
-                        "_concat_fuse_pass");
-    }
-    use_gpu_ = true;
-  }
+  GpuPassStrategy();
 
   explicit GpuPassStrategy(const GpuPassStrategy &other)
       : PassStrategy(other.AllPasses()) {