add conv_gelu_mkldnn_fuse_pass (#38107)

* add conv_gelu_mkldnn_fuse_pass

* add post ops

paddle/fluid/framework/ir/mkldnn/conv_activation_mkldnn_fuse_pass.cc

@@ -69,7 +69,15 @@ void ConvActivationFusePass::ApplyImpl(ir::Graph* graph) const {
     desc->SetOutput("Output",
                     std::vector<std::string>({activation_out->Name()}));
 
+    if (activation_type() == "gelu" &&
+        activation->Op()->HasAttr("approximate")) {
+      bool approximate =
+          BOOST_GET_CONST(bool, activation->Op()->GetAttr("approximate"));
+      std::string type = approximate ? "_tanh" : "_erf";
+      desc->SetAttr("fuse_activation", "gelu" + type);
+    } else {
       desc->SetAttr("fuse_activation", activation_type());
+    }
 
     // MKLDNN ops use alpha and beta as activation parameters but paddle ops are
     // not generalized
@@ -240,6 +248,19 @@ Conv2DHardSigmoidFusePass::Conv2DHardSigmoidFusePass() {
       .End();
 }
 
+Conv2DGeluFusePass::Conv2DGeluFusePass() {
+  AddOpCompat(OpCompat("gelu"))
+      .AddInput("X")
+      .IsTensor()
+      .End()
+      .AddOutput("Out")
+      .IsTensor()
+      .End()
+      .AddAttr("approximate")
+      .IsType<bool>()
+      .End();
+}
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle
@@ -294,3 +315,11 @@ REGISTER_PASS_CAPABILITY(conv_hard_sigmoid_mkldnn_fuse_pass)
         paddle::framework::compatible::OpVersionComparatorCombination()
             .LE("conv2d", 1)
             .EQ("hard_sigmoid", 0));
+
+REGISTER_PASS(conv_gelu_mkldnn_fuse_pass,
+              paddle::framework::ir::Conv2DGeluFusePass);
+REGISTER_PASS_CAPABILITY(conv_gelu_mkldnn_fuse_pass)
+    .AddCombination(
+        paddle::framework::compatible::OpVersionComparatorCombination()
+            .LE("conv2d", 1)
+            .EQ("gelu", 0));

paddle/fluid/framework/ir/mkldnn/conv_activation_mkldnn_fuse_pass.h

@@ -81,6 +81,15 @@ class Conv2DHardSigmoidFusePass : public ConvActivationFusePass {
   std::string activation_type() const { return "hard_sigmoid"; }
 };
 
+/*
+ * Fuse Conv and Gelu class
+ */
+class Conv2DGeluFusePass : public ConvActivationFusePass {
+ public:
+  Conv2DGeluFusePass();
+  std::string activation_type() const { return "gelu"; }
+};
+
 }  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -250,6 +250,7 @@ void CpuPassStrategy::EnableMKLDNN() {
              "conv_swish_mkldnn_fuse_pass",         //
              "conv_hard_swish_mkldnn_fuse_pass",    //
              "conv_hard_sigmoid_mkldnn_fuse_pass",  //
+             "conv_gelu_mkldnn_fuse_pass",
              "scale_matmul_fuse_pass",                        //
              "reshape_transpose_matmul_mkldnn_fuse_pass",     //
              "reshape_transpose_matmul_v2_mkldnn_fuse_pass",  //

paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc

@@ -510,6 +510,12 @@ class ConvMKLDNNHandlerT
                                      fuse_alpha, fuse_beta);
       post_operations.append_eltwise(scale, dnnl::algorithm::eltwise_clip, 0.0f,
                                      1.0f);
+    } else if (fuse_activation == "gelu_tanh") {
+      post_operations.append_eltwise(scale, dnnl::algorithm::eltwise_gelu_tanh,
+                                     0.0f, 0.0f);
+    } else if (fuse_activation == "gelu_erf") {
+      post_operations.append_eltwise(scale, dnnl::algorithm::eltwise_gelu_erf,
+                                     0.0f, 0.0f);
     }
     conv_attr.set_post_ops(post_operations);
     return conv_attr;

python/paddle/fluid/tests/unittests/ir/inference/test_mkldnn_conv_gelu_fuse_pass.py

+# Copyright (c) 2021 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.
+
+from auto_scan_test import PassAutoScanTest, SkipReasons
+from program_config import TensorConfig, ProgramConfig
+import numpy as np
+import paddle.inference as paddle_infer
+from functools import partial
+from typing import Optional, List, Callable, Dict, Any, Set
+import unittest
+
+import hypothesis
+from hypothesis import given, settings, seed, example, assume
+import hypothesis.strategies as st
+
+
+class TestConvGeluMkldnnFusePass(PassAutoScanTest):
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        attrs = [
+            program_config.ops[i].attrs
+            for i in range(len(program_config.ops))
+        ]
+        # If the problem has been fixed, the judgment 
+        # needs to be deleted!!!
+        if attrs[0]['data_format'] == "NHWC":
+            return False
+
+        return True
+
+    def sample_program_config(self, draw):
+        data_format = draw(st.sampled_from(["NCHW", "NHWC"]))
+        dilations = draw(st.sampled_from([[1, 1], [2, 2], [1, 2]]))
+        padding_algorithm = draw(st.sampled_from(["EXPLICIT", "SAME", "VALID"]))
+        groups = draw(st.sampled_from([1, 2, 4]))
+        paddings = draw(st.sampled_from([[0, 3], [1, 2, 3, 4]]))
+        strides = draw(st.sampled_from([[1, 1], [2, 2], [1, 2]]))
+        approximate = draw(st.booleans())
+        batch_size = draw(st.integers(min_value=1, max_value=4))
+
+        def generate_input():
+            if data_format == "NCHW":
+                return np.random.random(
+                    [batch_size, 48, 64, 64]).astype(np.float32)
+            else:
+                return np.random.random(
+                    [batch_size, 64, 64, 48]).astype(np.float32)
+
+        def generate_weight():
+            return np.random.random(
+                [16, int(48 / groups), 3, 3]).astype(np.float32)
+
+        ops_config = [{
+            "op_type": "conv2d",
+            "op_inputs": {
+                "Input": ["input_data"],
+                "Filter": ["input_weight"]
+            },
+            "op_outputs": {
+                "Output": ["conv_output"]
+            },
+            "op_attrs": {
+                "data_format": data_format,
+                "dilations": dilations,
+                "padding_algorithm": padding_algorithm,
+                "groups": groups,
+                "paddings": paddings,
+                "strides": strides
+            }
+        }, {
+            "op_type": "gelu",
+            "op_inputs": {
+                "X": ["conv_output"]
+            },
+            "op_outputs": {
+                "Out": ["gelu_output"]
+            },
+            "op_attrs": {
+                "approximate": approximate,
+            },
+        }]
+
+        ops = self.generate_op_config(ops_config)
+
+        program_config = ProgramConfig(
+            ops=ops,
+            weights={
+                "input_weight": TensorConfig(data_gen=partial(generate_weight))
+            },
+            inputs={
+                "input_data": TensorConfig(data_gen=partial(generate_input)),
+            },
+            outputs=["gelu_output"])
+
+        return program_config
+
+    def sample_predictor_configs(self, program_config):
+        config = self.create_inference_config(use_mkldnn=True)
+        yield config, ["conv2d"], (1e-5, 1e-5)
+
+    # If the problem has been fixed, the judgment 
+    # needs to be deleted!!!
+    def add_ignore_pass_case(self):
+        def teller1(program_config, predictor_config):
+            if program_config.ops[0].attrs['data_format'] == "NHWC":
+                return True
+            return False
+
+        self.add_ignore_check_case(
+            teller1, SkipReasons.PASS_ACCURACY_ERROR,
+            "The output format of conv2d is wrong when data_format attribute is NHWC"
+        )
+
+    def test(self):
+        self.run_and_statis(quant=False, passes=["conv_gelu_mkldnn_fuse_pass"])
+
+
+if __name__ == "__main__":
+    unittest.main()