Clear extra attributes of activation op in OpMaker (#45772)

* clear extra attr of activation op in opmaker

* fix syntax bug

* fix mkldnn kernel

* fix merge conflict

* fix bug

paddle/fluid/framework/grad_op_desc_maker.h

@@ -213,8 +213,8 @@ class SingleGradOpMaker<OpDesc> : public GradOpDescMakerBase {
     std::vector<std::unique_ptr<OpDesc>> retv;
     retv.emplace_back(new OpDesc());
     try {
-      this->Apply(retv.front().get());
       retv.front()->SetRuntimeAttrMap(this->RuntimeAttrs());
+      this->Apply(retv.front().get());
     } catch (platform::EnforceNotMet& exception) {
       framework::AppendErrorOpHint(retv.front().get()->Type(), &exception);
       throw std::move(exception);

paddle/fluid/operators/activation_op.cc

@@ -50,15 +50,6 @@ static constexpr bool CanInplaceAct() {
       AddOutput("Out",                                              \
                 "Output of " #OP_NAME                               \
                 " operator, a Tensor with shape same as input.");   \
-      AddAttr<bool>("use_mkldnn",                                            \
-                    "(bool, default false) Only used in mkldnn kernel")      \
-          .SetDefault(false)                                                 \
-          .AsExtra();                                                        \
-      AddAttr<bool>("use_cudnn",                                             \
-                    "(bool, default false) Only used in cudnn kernel, need " \
-                    "install cudnn")                                         \
-          .SetDefault(false)                                                 \
-          .AsExtra();                                                        \
       AddComment(OP_COMMENT);                                       \
     }                                                               \
   }
@@ -107,8 +98,7 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
 //   }
 // #endif
 #ifdef PADDLE_WITH_MKLDNN
-  auto it = oper.Attrs().find("use_mkldnn");
-  if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() &&
+  if (library == framework::LibraryType::kPlain &&
       oper.CanMKLDNNBeUsed(ctx, data_type)) {
     library = framework::LibraryType::kMKLDNN;
     layout = framework::DataLayout::kMKLDNN;
@@ -458,10 +448,6 @@ class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker {
         "A LoDTensor or Tensor with the same type and size as that of x.");
     AddAttr<float>("alpha", "Slope of the activation function at x < 0.")
         .SetDefault(0.02f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 LeakyRelu Activation Operator.
 
@@ -483,35 +469,6 @@ class SoftplusOpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<float>("beta", "The value of beta for Softplus.").SetDefault(1.0f);
     AddAttr<float>("threshold", "The value of threshold for Softplus.")
         .SetDefault(20.0f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel.")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<bool>(
-        "use_cudnn",
-        "(bool, default false) Only used in cudnn kernel, need install cudnn.")
-        .SetDefault(false)
-        .AsExtra();
-    AddAttr<std::string>(
-        "fuse_activation_type",
-        "Fused activation type used in softplus OneDNN kernel.")
-        .SetDefault("")
-        .AsExtra();
-    AddAttr<float>(
-        "fuse_activation_alpha",
-        "Fused activation alpha parameter type used in softplus OneDNN kernel.")
-        .SetDefault(0.0f)
-        .AsExtra();
-    AddAttr<float>(
-        "fuse_activation_beta",
-        "Fused activation beta parameter type used in softplus OneDNN kernel.")
-        .SetDefault(0.0f)
-        .AsExtra();
-    AddAttr<float>(
-        "fuse_activation_scale",
-        "Fused activation scale parameter type used in softplus OneDNN kernel.")
-        .SetDefault(1.0f)
-        .AsExtra();
     AddComment(R"DOC(
 :strong:`Softplus Activation Operator`
 
@@ -613,10 +570,6 @@ class ELUOpMaker : public framework::OpProtoAndCheckerMaker {
               "The output is a multi-dimensional Tensor which has same "
               "dimension and data type as the ``x``.");
     AddAttr<float>("alpha", "The alpha value of ELU").SetDefault(1.0f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 ELU Activation Operator.
 
@@ -712,10 +665,6 @@ class Relu6OpMaker : public framework::OpProtoAndCheckerMaker {
     AddAttr<float>("threshold",
                    "The threshold value of Relu6. Default is 6.0. ")
         .SetDefault(6.0f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Relu6 Activation Operator.
 
@@ -817,10 +766,6 @@ class SwishOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "Input of Swish operator");
     AddOutput("Out", "Output of Swish operator");
     AddAttr<float>("beta", "Constant beta of swish operator").SetDefault(1.0f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Swish Activation Operator.
 
@@ -841,10 +786,6 @@ class MishOpMaker : public framework::OpProtoAndCheckerMaker {
         "of softplus will be used if absolute value of input is greater than "
         ":attr:`threshold`")
         .SetDefault(20.f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 Mish Activation Operator.
 
@@ -871,10 +812,6 @@ class HardSwishOpMaker : public framework::OpProtoAndCheckerMaker {
         .SetDefault(6.0f);
     AddAttr<float>("offset", "The offset parameter of HardSwish operator")
         .SetDefault(3.0f);
-    AddAttr<bool>("use_mkldnn",
-                  "(bool, default false) Only used in mkldnn kernel")
-        .SetDefault(false)
-        .AsExtra();
     AddComment(R"DOC(
 HardSwish Activation Operator.
 

paddle/phi/api/yaml/op_compat.yaml

@@ -3,6 +3,11 @@
   extra :
     attrs : [bool use_cudnn = false, bool use_mkldnn = false]
 
+- api : acosh
+  backward : acosh_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : addmm
   backward : addmm_grad
   extra :
@@ -18,12 +23,22 @@
   extra :
     attrs : [bool use_cudnn = false, bool use_mkldnn = false]
 
+- api : asinh
+  backward : asinh_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : atan2
   inputs :
     {x : X1, y : X2}
   outputs :
     out : Out
 
+- api : atanh
+  backward : atanh_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : batch_norm
   backward : batch_norm_grad
   extra :
@@ -45,6 +60,11 @@
   extra :
     attrs : [bool use_mkldnn = false]
 
+- api : ceil
+  backward : ceil_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : cholesky
   inputs :
     x : X
@@ -107,6 +127,16 @@
   extra :
     attrs : [bool use_cudnn = true, bool use_mkldnn = false, int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB()]
 
+- api : cos
+  backward : cos_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : cosh
+  backward : cosh_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : cross
   inputs :
     {x : X, y : Y}
@@ -179,6 +209,11 @@
   extra :
     attrs : [bool fix_seed = false, int seed = 0]
 
+- api : elu
+  backward : elu_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : erf
   inputs :
     x : X
@@ -191,6 +226,16 @@
   outputs :
     out : Out
 
+- api : exp
+  backward : exp_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : expm1
+  backward : expm1_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : fft_c2c
   inputs: {x: X}
   outputs: {out: Out}
@@ -203,6 +248,11 @@
   inputs: {x: X}
   outputs: {out: Out}
 
+- api : floor
+  backward : floor_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : frobenius_norm
   backward : frobenius_norm_grad
   extra :
@@ -223,6 +273,11 @@
   extra :
     attrs : [bool is_test = false]
 
+- api : hard_swish
+  backward : hard_swish_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : inplace_abn
   backward : inplace_abn_grad
   extra :
@@ -233,6 +288,11 @@
   extra :
     attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
 
+- api : leaky_relu
+  backward : leaky_relu_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : lgamma
   inputs :
     x : X
@@ -244,11 +304,36 @@
   extra :
     attrs : [bool use_mkldnn = false]
 
+- api : log
+  backward : log_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : log10
+  backward : log10_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : log1p
+  backward : log1p_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : log2
+  backward : log2_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : log_softmax
   backward : log_softmax_grad
   extra :
     attrs : [bool use_mkldnn = false]
 
+- api : logsigmoid
+  backward : logsigmoid_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : lrn
   backward : lrn_grad
   extra :
@@ -261,6 +346,11 @@
              str mkldnn_data_type = "float32", 'int[] fused_reshape_X = {}', 'int[] fused_reshape_Y = {}',
              'int[] fused_transpose_X = {}', 'int[] fused_transpose_Y = {}',]
 
+- api : mish
+  backward : mish_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : mv
   inputs :
     {x : X, vec : Vec}
@@ -293,6 +383,21 @@
   outputs :
     out : Out
 
+- api : prelu
+  backward : prelu_grad
+  extra :
+    attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
+
+- api : prelu
+  backward : prelu_grad
+  extra :
+    attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
+
+- api : reciprocal
+  backward : reciprocal_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : reduce_all
   extra :
     attrs : [bool use_mkldnn = false]
@@ -336,15 +441,30 @@
   extra :
     attrs : [bool use_mkldnn = false]
 
+- api : relu
+  backward : relu_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : relu6
+  backward : relu6_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : renorm
   backward : renorm_grad
   extra :
     attrs : [bool use_mkldnn = false, bool use_cudnn = false]
 
-- api : rnn
-  backward : rnn_grad
+- api : round
+  backward : round_grad
   extra :
-    attrs : [bool is_test = false]
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : rsqrt
+  backward : rsqrt_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
 
 - api : seed
   extra :
@@ -359,6 +479,26 @@
   extra :
     attrs : [bool use_mkldnn = false]
 
+- api : sigmoid
+  backward : sigmoid_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : silu
+  backward : silu_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : sin
+  backward : sin_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : sinh
+  backward : sinh_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : slice
   backward : slice_grad
   extra :
@@ -368,10 +508,21 @@
   backward : softmax_grad
   extra :
     attrs : [bool use_cudnn = false, bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
-- api : prelu
-  backward : prelu_grad
+
+- api : softplus
+  backward : softplus_grad
   extra :
-    attrs : [bool use_mkldnn = false, str mkldnn_data_type = "float32", bool is_test = false]
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false, str fuse_activation_type = "", float fuse_activation_alpha = 0.0f,
+             float fuse_activation_beta = 0.0f, float fuse_activation_scale = 1.0f]
+
+- api : softsign
+  backward : softsign_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+- api : rnn
+  backward : rnn_grad
+  extra :
+    attrs : [bool is_test = false]
 
 - api : solve
   inputs :
@@ -379,6 +530,16 @@
   outputs :
     out : Out
 
+- api : sqrt
+  backward : sqrt_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : square
+  backward : square_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : squeeze (squeeze2)
   backward : squeeze_grad (squeeze2_grad)
   extra :
@@ -389,11 +550,31 @@
   extra :
     attrs : [bool use_mkldnn = false]
 
+- api : swish
+  backward : swish_grad
+  extra :
+    attrs : [bool use_mkldnn = false]
+
 - api : sync_batch_norm
   backward : sync_batch_norm_grad
   extra :
     attrs : [bool use_mkldnn = false, bool fuse_with_relu = false]
 
+- api : tan
+  backward : tan_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : tanh
+  backward : tanh_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
+- api : tanh_shrink
+  backward : tanh_shrink_grad
+  extra :
+    attrs : [bool use_mkldnn = false, bool use_cudnn = false]
+
 - api : trace
   inputs :
     x : Input