Made copy_tensor() use cudaMemcpyAsync() rather than cudaMemcpy().

dlib/dnn/cuda_dlib.cu

@@ -1401,7 +1401,7 @@ namespace dlib
 
             for (long i = 0; i < src.num_samples(); ++i)
             {
-                CHECK_CUDA(cudaMemcpy(dest_p, src_p, block_size * sizeof(float), cudaMemcpyDeviceToDevice));
+                CHECK_CUDA(cudaMemcpyAsync(dest_p, src_p, block_size * sizeof(float), cudaMemcpyDeviceToDevice));
 
                 dest_p += dest_sample_size;
                 src_p  += src_sample_size;

dlib/dnn/tensor_tools.h

@@ -1558,6 +1558,7 @@ namespace dlib { namespace tt
             - dest.k() - dest_k_offset >= count_k
             - src.k() - src_k_offset >= count_k
             - is_same_object(dest,src) == false
+            - The memory areas of src and dest do not overlap.
         ensures
             - performs: dest[i, k + dest_k_offset, r, c] = src[i, k + src_k_offset, r, c], where k in [0..count_k]
               Copies content of each sample from src in to corresponding place of sample at dest.