CUDA kernel compilation fixes

src/kernels/level2/level2.opencl

@@ -34,7 +34,7 @@ R"(
 
 // Returns an element from a vector
 INLINE_FUNC real LoadVector(const int id, const int max,
-                            __global real* gm, const int offset, const int inc,
+                            const __global real* gm, const int offset, const int inc,
                             const int do_conjugate) {
   if (id < max) {
     real result = gm[id*inc + offset];

src/kernels/opencl_to_cuda.h

@@ -31,6 +31,11 @@ __device__ int get_group_id(const int x) {
   if (x == 1) { return blockIdx.y; }
   return blockIdx.z;
 }
+__device__ int get_local_size(const int x) {
+  if (x == 0) { return blockDim.x; }
+  if (x == 1) { return blockDim.y; }
+  return blockDim.z;
+}
 __device__ int get_num_groups(const int x) {
   if (x == 0) { return gridDim.x; }
   if (x == 1) { return gridDim.y; }