Make sure torch inputs contiguous before passing data pointer to TRT execution

Author: bddppq

core/execution/register_trt_op.cpp

@@ -11,13 +11,15 @@ namespace execution {
 
 std::vector<at::Tensor> RunCudaEngine(nvinfer1::IExecutionContext* ctx, std::pair<uint64_t, uint64_t> io, std::vector<at::Tensor>& inputs) {
     std::vector<void*> gpu_handles;
+
+    std::vector<at::Tensor> contig_inputs{};
+    contig_inputs.reserve(inputs.size());
     for (size_t i = 0; i < inputs.size(); i++) {
-        auto in_gpu = inputs[i].to(at::kCUDA);
-        auto shape = core::util::toDimsPad(in_gpu.sizes(), 4);
-        //in_gpu = inputs[i].reshape(core::util::toVec(shape));
-        //LOG_DEBUG("In shape:" << in_gpu.sizes() );
+        contig_inputs.push_back(inputs[i].to(at::kCUDA).contiguous());
+        auto shape = core::util::toDimsPad(contig_inputs.back().sizes(), 4);
+        //LOG_DEBUG("In shape:" << contig_inputs.back().sizes() );
         ctx->setBindingDimensions(i, shape);
-        gpu_handles.push_back(in_gpu.data_ptr());
+        gpu_handles.push_back(contig_inputs.back().data_ptr());
     }
 
     TRTORCH_CHECK(ctx->allInputDimensionsSpecified(), "Not enough inputs provided (execution.RunCudaEngine)");