initial cuda build

[ROCm/composable_kernel commit: 2f2cf35bf4]

src/include/direct_convolution.cuh

@@ -0,0 +1,12 @@
+#pragma once
+#include "device_tensor.cuh"
+
+template <class TFloat, int NBlockDim>
+__global__ void direct_convolution(DeviceTensorDescriptor in_desc,
+                                   TFloat* const in,
+                                   DeviceTensorDescriptor wei_desc,
+                                   TFloat* const wei,
+                                   DeviceTensorDescriptor out_desc,
+                                   TFloat* out)
+{
+}