Clear extra attributes of some Op in OpMaker (Part2) (#45684)

* remove extra attr of opmaker

* remove extra attr of gru

* revert scale

* fix bug of create_op

* add extra attr checker in infer_shape

* fix gelu

paddle/fluid/framework/op_desc.cc

@@ -1018,6 +1018,13 @@ void OpDesc::CheckAttrs() {
   }
   VLOG(10) << "begin to check attribute of " << Type();
   checker->Check(&attrs_);
+  const auto &extra_attr_checkers =
+      operators::ExtraInfoUtils::Instance().GetExtraAttrsChecker(Type());
+  if (!extra_attr_checkers.empty()) {
+    for (const auto &extra_checker : extra_attr_checkers) {
+      extra_checker(&runtime_attrs_, false);
+    }
+  }
 }
 
 void OpDesc::InferShape(const BlockDesc &block) {

paddle/fluid/framework/op_registry.cc

@@ -38,8 +38,17 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
     }
   }
   auto& info = OpInfoMap::Instance().Get(type);
-  if (attr_check && info.Checker() != nullptr) {
-    info.Checker()->Check(&standard_attrs);
+  if (attr_check) {
+    if (info.Checker() != nullptr) {
+      info.Checker()->Check(&standard_attrs);
+    }
+    const auto& extra_attr_checkers =
+        operators::ExtraInfoUtils::Instance().GetExtraAttrsChecker(type);
+    if (!extra_attr_checkers.empty()) {
+      for (const auto& checker : extra_attr_checkers) {
+        checker(&runtime_attrs, false);
+      }
+    }
   }
   auto op_base = std::unique_ptr<OperatorBase>(
       info.Creator()(type, inputs, outputs, standard_attrs));

paddle/fluid/operators/data_norm_op.cc

@@ -251,10 +251,6 @@ class DataNormOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<std::string>("data_layout", "").SetDefault("NCHW");
     AddAttr<bool>("sync_stats", "(bool, default false) only used in multi-GPU")
         .SetDefault(false);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddInput("X", "The input tensor");
     AddInput("BatchSize",
              "BatchSize is a 1-dimensional tensor of size C "

paddle/fluid/operators/dropout_op.cc

@@ -77,15 +77,6 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(bool, default false) Set to true for inference only, false "
                   "for training. Some layers may run faster when this is true.")
         .SetDefault(false);
-    AddAttr<bool>("fix_seed",
-                  "A flag indicating whether to use a fixed seed to generate "
-                  "random mask. NOTE: DO NOT set this flag to true in "
-                  "training. Setting this flag to true is only useful in "
-                  "unittest or for debug that always the same output units "
-                  "will be dropped.")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<int>("seed", "Dropout random seed.").SetDefault(0).AsExtra();
     AddAttr<std::string>(
         "dropout_implementation",
         "[\"downgrade_in_infer\"|\"upscale_in_train\"]"

paddle/fluid/operators/gelu_op.cc

@@ -39,9 +39,8 @@ class GeluOp : public framework::OperatorWithKernel {
     framework::DataLayout layout = framework::DataLayout::kAnyLayout;
     auto data_type = OperatorWithKernel::IndicateVarDataType(ctx, "X");
 #ifdef PADDLE_WITH_MKLDNN
-    auto it = this->Attrs().find("use_mkldnn");
     if (library == framework::LibraryType::kPlain &&
-        it != this->Attrs().end() && this->CanMKLDNNBeUsed(ctx, data_type)) {
+        this->CanMKLDNNBeUsed(ctx, data_type)) {
       library = framework::LibraryType::kMKLDNN;
       layout = framework::DataLayout::kMKLDNN;
     }
@@ -100,21 +99,6 @@ class GeluOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<bool>("approximate",
                   "(bool, default false) use approximation of gelu")
         .SetDefault(false);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<std::string>(
-        "mkldnn_data_type",
-        "(string, default \"float32\"). Data type of mkldnn kernel")
-        .SetDefault("float32")
-        .InEnum({"float32", "int8", "bfloat16"})
-        .AsExtra();
-    AddAttr<bool>("use_cudnn",
-                  "(bool, default false) Only used in cudnn kernel, need "
-                  "install cudnn")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Gelu Activation Operator. 
 

paddle/fluid/operators/grid_sampler_op.cc

@@ -67,12 +67,6 @@ class GridSampleOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Output",
               "(Tensor) Output tensor with shape [N, C, H, W] or shape [N,C, "
               "D, H ,W]");
-    AddAttr<bool>(
-        "use_cudnn",
-        "(bool, default true) Only used in cudnn kernel, need install cudnn")
-        .SetDefault(true)
-        .AsExtra();
-
     AddAttr<bool>(
         "align_corners",
         "(bool, default true) If align_corners is true, it will project"

paddle/fluid/operators/gru_op.cc

@@ -180,9 +180,6 @@ class GRUOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(bool, default: False) "
                   "whether to compute reversed GRU.")
         .SetDefault(false);
-    AddAttr<bool>("is_test", "True if in test phase.")
-        .SetDefault(false)
-        .AsExtra();
     AddAttr<bool>("origin_mode",
                   "bool"
                   "use origin mode in article https://arxiv.org/abs/1412.3555")

paddle/fluid/operators/interpolate_v2_op.cc

@@ -550,10 +550,6 @@ class InterpolateV2OpMaker : public framework::OpProtoAndCheckerMaker {
                  "can be \'0\' for src_idx = scale*(dst_indx+0.5)-0.5 , "
                  "can be \'1\' for src_idx = scale*dst_index .")
         .SetDefault(1);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
           This operator samples input X to given output shape by using specified
           interpolation method, the interpolation methods can be \"nearest\"

paddle/fluid/operators/layer_norm_op.cc

@@ -171,22 +171,6 @@ class LayerNormOpMaker : public framework::OpProtoAndCheckerMaker {
                                 "greater than zero. But received [%d].",
                                 begin_norm_axis));
         });
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<std::string>(
-        "mkldnn_data_type",
-        "(string, default \"float32\"). Data type of mkldnn kernel")
-        .SetDefault("float32")
-        .InEnum({"float32", "bfloat16"})
-        .AsExtra();
-    AddAttr<bool>("is_test",
-                  "(bool, default false) Set to true for inference only, false "
-                  "for training. Some layers may run faster when this is true.")
-        .SetDefault(false)
-        .AsExtra();
-
     AddComment(R"DOC(
 Assume feature vectors exist on dimensions
 :attr:`begin_norm_axis ... rank(input)` and calculate the moment statistics

paddle/fluid/operators/log_softmax_op.cc

@@ -57,10 +57,6 @@ class LogSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
                  "The dimension index of Input(x) to perform log_softmax,"
                  "default -1 for last dimension")
         .SetDefault(-1);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 LogSoftmax Operator.
 

paddle/fluid/operators/lrn_op.cc

@@ -302,10 +302,6 @@ class LRNOpMaker : public framework::OpProtoAndCheckerMaker {
                "beta is the power number.")
         .SetDefault(0.75)
         .GreaterThan(0.0);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddAttr<std::string>(
         "data_format",
         "(string, default NCHW) Only used in "
@@ -313,12 +309,6 @@ class LRNOpMaker : public framework::OpProtoAndCheckerMaker {
         "Defaults to \"NHWC\". Specify the data format of the output data, "
         "the input will be transformed automatically. ")
         .SetDefault("AnyLayout");
-    AddAttr<bool>("is_test",
-                  "(bool, default false) Set to true for inference only, false "
-                  "for training. Some layers may run faster when this is true.")
-        .SetDefault(false)
-        .AsExtra();
-
     AddComment(R"DOC(
 Local Response Normalization Operator.
 

paddle/fluid/operators/pad2d_op.cc

@@ -773,10 +773,6 @@ class Pad2dOpMaker : public framework::OpProtoAndCheckerMaker {
         "An optional string from: \"NHWC\", \"NCHW\". "
         "Defaults to \"NHWC\". Specify the data format of the input data.")
         .SetDefault("NCHW");
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Pad2d Operator.
 Pad 2-d images according to 'paddings' and 'mode'. 

paddle/fluid/operators/pad3d_op.cc

@@ -111,10 +111,6 @@ class Pad3dOpMaker : public framework::OpProtoAndCheckerMaker {
         "An optional string from: \"NDHWC\", \"NCDHW\". "
         "Defaults to \"NDHWC\". Specify the data format of the input data.")
         .SetDefault("NCDHW");
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Pad3d Operator.
 Pad 3-d images according to 'paddings' and 'mode'. 

paddle/fluid/operators/partial_sum_op.cc

@@ -155,11 +155,6 @@ class PartialSumOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "Input tensors of partial_sum operator.").AsDuplicable();
     AddOutput("Out", "Output tensor of partial_sum operator.");
-    AddAttr<bool>(
-        "use_mkldnn",
-        "(bool, default false) Indicates if MKL-DNN kernel will be used")
-        .SetDefault(false)
-        .AsExtra();
     AddAttr<int>("start_index", "The start index of tensor wanted to be added.")
         .SetDefault(0);
     AddAttr<int>("length", "The length of tensor wanted to be added.")

paddle/fluid/operators/prelu_op.cc

@@ -104,21 +104,6 @@ There are modes:
     AddAttr<std::string>("data_format",
                          "Data format that specifies the layout of input")
         .SetDefault("NCHW");
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<std::string>(
-        "mkldnn_data_type",
-        "(string, default \"float32\"). Data type of mkldnn kernel")
-        .SetDefault("float32")
-        .InEnum({"float32", "bfloat16"})
-        .AsExtra();
-    AddAttr<bool>("is_test",
-                  "(bool, default false) Set to true for inference only, false "
-                  "for training. Some layers may run faster when this is true.")
-        .SetDefault(false)
-        .AsExtra();
   }
 };
 

paddle/fluid/operators/renorm_op.cc

@@ -39,15 +39,6 @@ class RenormOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<int>("axis",
                  "int,the dimension to slice over to get the sub-tensors");
     AddAttr<float>("max_norm", "(float, the norm upper-bound");
-    AddAttr<bool>("use_cudnn",
-                  "(bool, default false) Only used in cudnn kernel, need "
-                  "install cudnn")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Renorm Operator.
 

paddle/phi/api/yaml/api_compat.yaml

@@ -35,6 +35,16 @@
   outputs :
     out : Out
 
+- api : bicubic_interp (bicubic_interp_v2)
+  backward : bicubic_interp_grad (bicubic_interp_v2_grad)
+  extra :
+    attrs : [bool use_mkldnn = false]
+
+- api : bilinear_interp (bilinear_interp_v2)
+  backward : bilinear_interp_grad (bilinear_interp_v2_grad)
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : cholesky
   inputs :
     x : X
@@ -105,6 +115,11 @@
   outputs :
     out : Out
 
+- api : data_norm
+  backward : data_norm_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : depthwise_conv2d
   backward : depthwise_conv2d_grad
   extra :
@@ -154,6 +169,16 @@
   outputs :
     out : Out
 
+- api : dropout
+  backward : dropout_grad
+  extra :
+    attrs : [bool fix_seed = false, int seed = 0]
+
+- api : dropout_nd
+  backward : dropout_nd_grad
+  extra :
+    attrs : [bool fix_seed = false, int seed = 0]
+
 - api : erf
   inputs :
     x : X
@@ -166,35 +191,117 @@
   outputs :
     out : Out
 
+- api : fft_c2c
+  inputs: {x: X}
+  outputs: {out: Out}
+
+- api : fft_c2r
+  inputs: {x: X}
+  outputs: {out: Out}
+
+- api : fft_r2c
+  inputs: {x: X}
+  outputs: {out: Out}
+
+- api : gelu
+  backward : gelu_grad
+  extra :
+    attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool use_cudnn = false]
+
+- api : grid_sampler
+  backward : grid_sampler_grad
+  extra :
+    attrs : [bool use_cudnn = true]
+
+- api : gru
+  backward : gru_grad
+  extra :
+    attrs : [bool is_test = false]
+
 - api : inplace_abn
   backward : inplace_abn_grad
   extra :
     attrs : [bool use_mkldnn = false, bool fuse_with_relu = false]
 
+- api : layer_norm
+  backward : layer_norm_grad
+  extra :
+    attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
+
 - api : lgamma
   inputs :
     x : X
   outputs :
     out : Out
 
+- api : linear_interp (linear_interp_v2)
+  backward : linear_interp_grad (linear_interp_v2_grad)
+  extra :
+    attrs : [bool use_mkldnn = false]
+
+- api : log_softmax
+  backward : log_softmax_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
+- api : lrn
+  backward : lrn_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool is_test = false]
+
 - api : mv
   inputs :
     {x : X, vec : Vec}
   outputs :
     out : Out
 
+- api : nearest_interp (nearest_interp_v2)
+  backward : nearest_interp_grad (nearest_interp_v2_grad)
+  extra :
+    attrs : [bool use_mkldnn = false]
+
+- api : pad2d
+  backward : pad2d_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
+- api : pad3d
+  backward : pad3d_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
+- api : partial_sum
+  backward : partial_sum_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : poisson
   inputs :
     x : X
   outputs :
     out : Out
 
+- api : prelu
+  backward : prelu_grad
+  extra :
+    attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
+
+- api : renorm
+  backward : renorm_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : solve
   inputs :
     {x : X, y : Y}
   outputs :
     out : Out
 
+- api : stack
+  backward : stack_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : sync_batch_norm
   backward : sync_batch_norm_grad
   extra :
@@ -206,20 +313,13 @@
   outputs :
     out : Out
 
+- api : trilinear_interp (trilinear_interp_v2)
+  backward : trilinear_interp_grad (trilinear_interp_v2_grad)
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : trunc
   inputs :
     x : X
   outputs :
     out : Out
-
-- api: fft_c2c
-  inputs: {x: X}
-  outputs: {out: Out}
-
-- api: fft_c2r
-  inputs: {x: X}
-  outputs: {out: Out}
-
-- api: fft_r2c
-  inputs: {x: X}
-  outputs: {out: Out}