fix(//core/conversion/converters): Fix plugin implementation for TRT 7

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>

Author: narendasan

Diff for: core/conversion/converters/impl/conv_deconv.cpp

@@ -45,9 +45,10 @@ auto conv_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
 
                 deconv->setStrideNd(stride);
                 deconv->setPaddingNd(padding);
+#if NV_TENSORRT_MAJOR > 7 || (NV_TENSORRT_MAJOR == 7 && NV_TENSORRT_MINOR == 1)
                 deconv->setDilationNd(dilation);
                 deconv->setNbGroups(groups);
-
+#endif
                 new_layer = deconv;
             } else {
                 nvinfer1::IConvolutionLayer* conv;

Diff for: core/conversion/converters/impl/interpolate.cpp

@@ -153,7 +153,7 @@ auto interpolate_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
                     // align_corners not supported in TensorRT, create plugin and run layer through PyTorch
                     create_plugin(ctx, n, in, "linear1d", in_shape, out_shape, out_size, std::string("linear"));
                 } else {
-                    resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR. true);
+                    resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR, true);
                 }
 #else
                 resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR, align_corners);
@@ -185,7 +185,7 @@ auto interpolate_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
                     // align_corners not supported in TensorRT, create plugin and run layer through PyTorch
                     create_plugin(ctx, n, in, "bilinear2d", in_shape, out_shape, out_size, std::string("bilinear"));
                 } else {
-                    resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR. true);
+                    resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR, true);
                 }
 #else
                 resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR, align_corners);
@@ -217,7 +217,7 @@ auto interpolate_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
                     // align_corners not supported in TensorRT, create plugin and run layer through PyTorch
                     create_plugin(ctx, n, in, "trilinear3d", in_shape, out_shape, out_size, std::string("trilinear"));
                 } else {
-                    resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR. true);
+                    resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR, true);
                 }
 #else
                 resize_layer_size(ctx, n, in, out_shape, nvinfer1::ResizeMode::kLINEAR, align_corners);

Diff for: core/conversion/converters/impl/plugins/interpolate_plugin.cpp

@@ -178,13 +178,13 @@ int InterpolatePlugin::enqueue(const nvinfer1::PluginTensorDesc* inputDesc, cons
 
     cudaStreamWaitEvent(torch_stream.stream(), event, 0);
 
-    if (mode == "linear") {
+    if (mode_ == "linear") {
         at::upsample_linear1d_out(output, input, {size_[0]}, align_corners_);
-    } else if (mode == "bilinear") {
+    } else if (mode_ == "bilinear") {
         at::upsample_bilinear2d_out(output, input, {size_[0], size_[1]}, align_corners_);
-    } else if (mode == "trilinear") {
+    } else if (mode_ == "trilinear") {
         at::upsample_trilinear3d_out(output, input, {size_[0], size_[1], size_[2]}, align_corners_);
-    } else if (mode == "adaptive_pool2d") {
+    } else if (mode_ == "adaptive_pool2d") {
         at::adaptive_avg_pool2d_out(output, input, {size_[0], size_[1]});
     }
 
@@ -212,11 +212,6 @@ int InterpolatePlugin::enqueue(const nvinfer1::PluginTensorDesc* inputDesc, cons
         output = at::adaptive_avg_pool2d(input, {size_[0], size_[1]});
     }
 
-    output = output.contiguous();
-    for (int i = 0; i < util::volume(outputDesc->dims); i++) {
-        std::cout << ((float*)output.data_ptr())[i] << std::endl;
-    }
-
     cudaMemcpyAsync(outputs[0], output.data_ptr(), util::volume(outputDesc->dims) * sizeof(float), cudaMemcpyHostToDevice, stream);
     cudaStreamSynchronize(stream);