Optimize attribute selected performence (#42294) (#42368)

* opt attr eaque perf * opt attr select code * fix one hot infermeta * polish get attr impl * fix tests failed * add testcases
PaddlePaddle · Apr 28, 2022 · e0e534a · e0e534a
1 parent d04a68d
commit e0e534a
Show file tree

Hide file tree

Showing 10 changed files with 660 additions and 531 deletions.
diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h
@@ -242,7 +242,7 @@ class AttrReader {
     return *attr_value;
   }
 
-  inline const Attribute& GetAttr(const std::string& name) const {
+  const Attribute* GetAttr(const std::string& name) const {
     auto it = attrs_.find(name);
     bool found = it != attrs_.end();
     if (!found) {
@@ -251,11 +251,10 @@ class AttrReader {
         found = it != default_attrs_->end();
       }
     }
-    PADDLE_ENFORCE_EQ(found, true,
-                      platform::errors::NotFound(
-                          "Attribute (%s) should be in AttributeMap.", name));
-
-    return it->second;
+    if (found) {
+      return &it->second;
+    }
+    return nullptr;
   }
 
  private:

diff --git a/paddle/fluid/framework/infershape_utils.cc b/paddle/fluid/framework/infershape_utils.cc
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
diff --git a/paddle/fluid/imperative/prepared_operator.h b/paddle/fluid/imperative/prepared_operator.h
diff --git a/paddle/phi/infermeta/unary.cc b/paddle/phi/infermeta/unary.cc
@@ -2945,7 +2945,7 @@ void UnStackInferMeta(const MetaTensor& x,
 }
 
 void OneHotRawInferMeta(const MetaTensor& x,
-                        int32_t depth,
+                        const Scalar& depth,
                         DataType dtype,
                         bool allow_out_of_range,
                         MetaTensor* out) {
@@ -2955,7 +2955,7 @@ void OneHotRawInferMeta(const MetaTensor& x,
       1,
       phi::errors::InvalidArgument("Rank of Input(X) should be at least 1."));
   auto out_dims_vec = phi::vectorize(x_dims);
-  out_dims_vec.push_back(depth);
+  out_dims_vec.push_back(depth.to<int>());
   auto out_dims = phi::make_ddim(out_dims_vec);
   out->set_dims(out_dims);
   out->share_lod(x);

diff --git a/paddle/phi/infermeta/unary.h b/paddle/phi/infermeta/unary.h
@@ -426,7 +426,7 @@ void UnStackInferMeta(const MetaTensor& x,
                       std::vector<MetaTensor*> outs);
 
 void OneHotRawInferMeta(const MetaTensor& x,
-                        int32_t depth,
+                        const Scalar& depth,
                         DataType dtype,
                         bool allow_out_of_range,
                         MetaTensor* out);

diff --git a/paddle/phi/kernels/cpu/one_hot_kernel.cc b/paddle/phi/kernels/cpu/one_hot_kernel.cc
@@ -64,18 +64,19 @@ struct OneHotV2OpFunctor {
 template <typename T, typename Context>
 void OneHotRawKernel(const Context& dev_ctx,
                      const DenseTensor& x,
-                     int32_t depth,
+                     const Scalar& depth,
                      DataType dtype,
                      bool allow_out_of_range,
                      DenseTensor* out) {
+  auto depth_v = depth.to<int>();
   auto out_dims = out->dims();
   if (out_dims[out_dims.size() - 1] == -1) {
-    out_dims[out_dims.size() - 1] = depth;
+    out_dims[out_dims.size() - 1] = depth_v;
     out->Resize(out_dims);
   }
 
   phi::VisitDataType(dtype,
-                     OneHotV2OpFunctor<Context, T>(&x, out, depth, dev_ctx));
+                     OneHotV2OpFunctor<Context, T>(&x, out, depth_v, dev_ctx));
 }
 
 }  // namespace phi

diff --git a/paddle/phi/kernels/gpu/one_hot_kernel.cu b/paddle/phi/kernels/gpu/one_hot_kernel.cu
@@ -73,18 +73,19 @@ struct OneHotV2OpCUDAFunctor {
 template <typename T, typename Context>
 void OneHotRawKernel(const Context& dev_ctx,
                      const DenseTensor& x,
-                     int32_t depth,
+                     const Scalar& depth,
                      DataType dtype,
                      bool allow_out_of_range,
                      DenseTensor* out) {
+  auto depth_v = depth.to<int>();
   auto out_dims = out->dims();
   if (out_dims[out_dims.size() - 1] == -1) {
-    out_dims[out_dims.size() - 1] = depth;
+    out_dims[out_dims.size() - 1] = depth_v;
     out->Resize(out_dims);
   }
 
   phi::VisitDataType(
-      dtype, OneHotV2OpCUDAFunctor<Context, T>(&x, out, depth, dev_ctx));
+      dtype, OneHotV2OpCUDAFunctor<Context, T>(&x, out, depth_v, dev_ctx));
 }
 
 }  // namespace phi

diff --git a/paddle/phi/kernels/one_hot_kernel.cc b/paddle/phi/kernels/one_hot_kernel.cc
@@ -24,9 +24,8 @@ void OneHotKernel(const Context& dev_ctx,
                   const DenseTensor& x,
                   const Scalar& num_classes_s,
                   DenseTensor* out) {
-  int num_classes = num_classes_s.to<int>();
   OneHotRawKernel<T>(
-      dev_ctx, x, num_classes, phi::DataType::FLOAT32, false, out);
+      dev_ctx, x, num_classes_s, phi::DataType::FLOAT32, false, out);
 }
 
 }  // namespace phi

diff --git a/paddle/phi/kernels/one_hot_kernel.h b/paddle/phi/kernels/one_hot_kernel.h
@@ -28,7 +28,7 @@ void OneHotKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void OneHotRawKernel(const Context& dev_ctx,
                      const DenseTensor& x,
-                     int32_t depth,
+                     const Scalar& depth,
                      DataType dtype,
                      bool allow_out_of_range,
                      DenseTensor* out);