From 85e697d74933d2251d25192a2bcf381adff7d433 Mon Sep 17 00:00:00 2001
From: Pei Yang <peiyang@baidu.com>
Date: Tue, 27 Apr 2021 12:32:36 +0800
Subject: [PATCH] support depthwise_conv2d_transpose (#32593)

---
 .../inference/tensorrt/convert/conv2d_op.cc   |  2 +-
 .../inference/tensorrt/convert/op_converter.h |  6 ++++++
 paddle/fluid/inference/tensorrt/op_teller.cc  |  7 +++++--
 .../ir/inference/test_trt_conv_pass.py        | 19 +++++++++++++++++--
 4 files changed, 29 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc
index ba47358b147..61199724bcf 100644
--- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc
@@ -160,7 +160,7 @@ class Deconv2dOpConverter : public OpConverter {
             nvinfer1::DimsHW& ksize, TensorRTEngine::Weight& weight,
             TensorRTEngine::Weight& bias) -> nvinfer1::IDeconvolutionLayer* {
           auto* layer =
-              TRT_ENGINE_ADD_LAYER(engine_, Deconvolution, *inputs, n_input,
+              TRT_ENGINE_ADD_LAYER(engine_, Deconvolution, *inputs, n_output,
                                    ksize, weight.get(), bias.get());
           return layer;
         },
diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h
index 8de16df0a2f..f72ae2c3ec2 100644
--- a/paddle/fluid/inference/tensorrt/convert/op_converter.h
+++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h
@@ -109,6 +109,12 @@ class OpConverter {
           it, platform::errors::Unimplemented("no OpConverter for optype [%s]",
                                               op_desc.Type()));
     }
+    if (op_desc.Type() == "depthwise_conv2d_transpose") {
+      it = Registry<OpConverter>::Global().Lookup("conv2d_transpose");
+      PADDLE_ENFORCE_NOT_NULL(
+          it, platform::errors::Unimplemented("no OpConverter for optype [%s]",
+                                              op_desc.Type()));
+    }
     if (op_desc.Type() == "transpose2") {
       it = Registry<OpConverter>::Global().Lookup("transpose");
       PADDLE_ENFORCE_NOT_NULL(
diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc
index c8dfc169535..48c7b7fdd0d 100644
--- a/paddle/fluid/inference/tensorrt/op_teller.cc
+++ b/paddle/fluid/inference/tensorrt/op_teller.cc
@@ -102,6 +102,7 @@ struct SimpleOpTypeSetTeller : public Teller {
       "dropout",
       "prelu",
       "conv2d_transpose",
+      "depthwise_conv2d_transpose",
       "leaky_relu",
       "fc",
       "shuffle_channel",
@@ -172,7 +173,8 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
     }
 
     if (op_type == "conv2d" || op_type == "conv2d_transpose" ||
-        op_type == "conv2d_fusion") {
+        op_type == "conv2d_fusion" || op_type == "depthwise_conv2d" ||
+        op_type == "depthwise_conv2d_transpose") {
       std::vector<int> paddings =
           BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
 
@@ -202,7 +204,8 @@ bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
         }
       }
 
-      if (op_type == "conv2d_transpose") {
+      if (op_type == "conv2d_transpose" ||
+          op_type == "depthwise_conv2d_transpose") {
         if (!desc.HasAttr("dilations")) {
           return false;
         } else {
diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py
index 0821b390e5e..ec3955a9ae1 100644
--- a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py
+++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_conv_pass.py
@@ -96,6 +96,7 @@ class TensorRTSubgraphPassConvTransposeTest(InferencePassTest):
                 groups=self.conv_groups,
                 padding=self.conv_padding,
                 bias_attr=False,
+                use_cudnn=self.use_cudnn,
                 act=None)
         self.feeds = {
             "data": np.random.random([1, 6, 64, 64]).astype("float32"),
@@ -110,6 +111,7 @@ class TensorRTSubgraphPassConvTransposeTest(InferencePassTest):
         self.conv_filter_size = 6
         self.conv_groups = 1
         self.conv_padding = [1, 1]
+        self.use_cudnn = True
 
     def test_check_output(self):
         if core.is_compiled_with_cuda():
@@ -126,6 +128,7 @@ class TensorRTSubgraphPassConvTransposeValidPaddingTest(
         self.conv_filter_size = 6
         self.conv_groups = 1
         self.conv_padding = 'VALID'
+        self.use_cudnn = True
 
 
 class TensorRTSubgraphPassConvTransposeSamePaddingTest(
@@ -135,15 +138,27 @@ class TensorRTSubgraphPassConvTransposeSamePaddingTest(
         self.conv_filter_size = 6
         self.conv_groups = 1
         self.conv_padding = 'SAME'
+        self.use_cudnn = True
 
 
-class TensorRTSubgraphPassDepthwiseConvTransposeTest(
+class TensorRTSubgraphPassConvTransposeMultiGroupTest(
         TensorRTSubgraphPassConvTransposeTest):
     def set_params(self):
         self.conv_num_filters = 6
         self.conv_filter_size = 6
-        self.conv_groups = 1
+        self.conv_groups = 2
+        self.conv_padding = [1, 1]
+        self.use_cudnn = True
+
+
+class TensorRTSubgraphPassDepthwiseConvTransposeTest(
+        TensorRTSubgraphPassConvTransposeTest):
+    def set_params(self):
+        self.conv_num_filters = 6
+        self.conv_filter_size = 4
+        self.conv_groups = 6
         self.conv_padding = [1, 1]
+        self.use_cudnn = False
 
 
 if __name__ == "__main__":
-- 
GitLab