Cherry pick deconv & jetson single arch (#33387)

* fix conv2d_transpose trt bugs (#33242)

* fix jetson arch when compiling with single arch (#33269)

cmake/cuda.cmake

@@ -95,11 +95,23 @@ function(select_nvcc_arch_flags out_variable)
   if(${CUDA_ARCH_NAME} STREQUAL "Kepler")
     set(cuda_arch_bin "30 35")
   elseif(${CUDA_ARCH_NAME} STREQUAL "Maxwell")
-    set(cuda_arch_bin "50")
+    if (WITH_NV_JETSON)
+      set(cuda_arch_bin "53")
+    else()
+      set(cuda_arch_bin "50")
+    endif()
   elseif(${CUDA_ARCH_NAME} STREQUAL "Pascal")
-    set(cuda_arch_bin "60 61")
+    if (WITH_NV_JETSON)
+      set(cuda_arch_bin "62")
+    else()
+      set(cuda_arch_bin "60 61")
+    endif()
   elseif(${CUDA_ARCH_NAME} STREQUAL "Volta")
-    set(cuda_arch_bin "70")
+    if (WITH_NV_JETSON)
+      set(cuda_arch_bin "72")
+    else()
+      set(cuda_arch_bin "70")
+    endif()
   elseif(${CUDA_ARCH_NAME} STREQUAL "Turing")
     set(cuda_arch_bin "75")
   elseif(${CUDA_ARCH_NAME} STREQUAL "Ampere")

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

@@ -103,11 +103,18 @@ void ConvertConv2d(TensorRTEngine* engine, const framework::proto::OpDesc& op,
 
   TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT,
                               static_cast<void*>(bias_data), bias_size};
-  auto* layer = fadd_layer(const_cast<nvinfer1::ITensor*>(X), n_output, n_input,
-                           nv_ksize, weight, bias);
-  PADDLE_ENFORCE_NOT_NULL(layer,
-                          platform::errors::Fatal("TensorRT create conv2d"
-                                                  " layer error."));
+  // In conv2d_transpose and depthwise_conv2d_transpose,
+  // output channels = filter_dims[1] * groups
+  auto* layer = (op_desc.Type() == "conv2d_transpose" ||
+                 op_desc.Type() == "depthwise_conv2d_transpose")
+                    ? fadd_layer(const_cast<nvinfer1::ITensor*>(X),
+                                 n_input * groups, nv_ksize, weight, bias)
+                    : fadd_layer(const_cast<nvinfer1::ITensor*>(X), n_output,
+                                 nv_ksize, weight, bias);
+
+  PADDLE_ENFORCE_NOT_NULL(
+      layer, platform::errors::Fatal("TensorRT create conv2d/conv2d_transpose"
+                                     " layer failed."));
   layer->setStride(nv_strides);
   layer->setPadding(nv_paddings);
   layer->setNbGroups(groups);
@@ -134,7 +141,6 @@ class Conv2dOpConverter : public OpConverter {
     ConvertConv2d(
         engine_, op, scope, test_mode,
         [&](nvinfer1::ITensor* inputs, int n_output, /* Conv output maps */
-            int n_input,                             /* Conv input maps */
             nvinfer1::DimsHW& ksize, TensorRTEngine::Weight& weight,
             TensorRTEngine::Weight& bias) -> nvinfer1::IConvolutionLayer* {
           auto* layer =
@@ -156,7 +162,6 @@ class Deconv2dOpConverter : public OpConverter {
     ConvertConv2d(
         engine_, op, scope, test_mode,
         [&](nvinfer1::ITensor* inputs, int n_output, /* Deconv input maps */
-            int n_input,                             /* Deconv output maps */
             nvinfer1::DimsHW& ksize, TensorRTEngine::Weight& weight,
             TensorRTEngine::Weight& bias) -> nvinfer1::IDeconvolutionLayer* {
           auto* layer =

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

@@ -36,6 +36,7 @@ class TensorRTSubgraphPassConvTest(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"),
@@ -50,6 +51,7 @@ class TensorRTSubgraphPassConvTest(InferencePassTest):
         self.conv_filter_size = 6
         self.conv_groups = 3
         self.conv_padding = [1, 1]
+        self.use_cudnn = True
 
     def test_check_output(self):
         if core.is_compiled_with_cuda():
@@ -65,6 +67,7 @@ class TensorRTSubgraphPassConvValidPaddingTest(TensorRTSubgraphPassConvTest):
         self.conv_filter_size = 6
         self.conv_groups = 3
         self.conv_padding = 'VALID'
+        self.use_cudnn = True
 
 
 class TensorRTSubgraphPassConvSamePaddingTest(InferencePassTest):
@@ -73,6 +76,7 @@ class TensorRTSubgraphPassConvSamePaddingTest(InferencePassTest):
         self.conv_filter_size = 6
         self.conv_groups = 3
         self.conv_padding = 'SAME'
+        self.use_cudnn = True
 
 
 class TensorRTSubgraphPassDepthwiseConvTest(TensorRTSubgraphPassConvTest):
@@ -81,6 +85,16 @@ class TensorRTSubgraphPassDepthwiseConvTest(TensorRTSubgraphPassConvTest):
         self.conv_filter_size = 6
         self.conv_groups = 6
         self.conv_padding = [1, 1]
+        self.use_cudnn = False
+
+
+class TensorRTSubgraphPassDepthwiseConv2Test(TensorRTSubgraphPassConvTest):
+    def set_params(self):
+        self.conv_num_filters = 12
+        self.conv_filter_size = 6
+        self.conv_groups = 6
+        self.conv_padding = [1, 1]
+        self.use_cudnn = False
 
 
 class TensorRTSubgraphPassConvTransposeTest(InferencePassTest):
@@ -151,6 +165,16 @@ class TensorRTSubgraphPassConvTransposeMultiGroupTest(
         self.use_cudnn = True
 
 
+class TensorRTSubgraphPassConvTranspose2Test(
+        TensorRTSubgraphPassConvTransposeTest):
+    def set_params(self):
+        self.conv_num_filters = 12
+        self.conv_filter_size = 4
+        self.conv_groups = 6
+        self.conv_padding = [1, 1]
+        self.use_cudnn = False
+
+
 class TensorRTSubgraphPassDepthwiseConvTransposeTest(
         TensorRTSubgraphPassConvTransposeTest):
     def set_params(self):