Commenting async copy for now in read/write buf functions

src/update_halo.jl

@@ -492,15 +492,16 @@ let
     function iwrite_sendbufs!(n::Integer, dim::Integer, A::ROCArray{T}, i::Integer) where T <: GGNumber
         if ol(dim,A) >= 2  # There is only a halo and thus a halo update if the overlap is at least 2...
             # DEBUG: the follow section needs perf testing
-            if dim == 1 || amdgpuaware_MPI(dim) # Use a custom copy kernel for the first dimension to obtain a good copy performance (the CUDA 3-D memcopy does not perform well for this extremely strided case).
+            # DEBUG 2: commenting read_h2d_async! for now
+            # if dim == 1 || amdgpuaware_MPI(dim) # Use a custom copy kernel for the first dimension to obtain a good copy performance (the CUDA 3-D memcopy does not perform well for this extremely strided case).
                 ranges = sendranges(n, dim, A);
                 nthreads = (dim==1) ? (1, 32, 1) : (32, 1, 1);
                 halosize = [r[end] - r[1] + 1 for r in ranges];
                 nblocks  = Tuple(ceil.(Int, halosize./nthreads));
                 @roc gridsize=nblocks groupsize=nthreads stream=rocstreams[n,i] write_d2x!(gpusendbuf(n,dim,i,A), A, ranges[1], ranges[2], ranges[3], dim);
-            else
-                write_d2h_async!(sendbuf_flat(n,dim,i,A), A, sendranges(n,dim,A), rocstreams[n,i]);
-            end
+            # else
+            #     write_d2h_async!(sendbuf_flat(n,dim,i,A), A, sendranges(n,dim,A), rocstreams[n,i]);
+            # end
         end
     end
 end
@@ -521,15 +522,16 @@ let
     function iread_recvbufs!(n::Integer, dim::Integer, A::ROCArray{T}, i::Integer) where T <: GGNumber
         if ol(dim,A) >= 2  # There is only a halo and thus a halo update if the overlap is at least 2...
             # DEBUG: the follow section needs perf testing
-            if dim == 1 || amdgpuaware_MPI(dim)  # Use a custom copy kernel for the first dimension to obtain a good copy performance (the CUDA 3-D memcopy does not perform well for this extremely strided case).
+            # DEBUG 2: commenting read_h2d_async! for now
+            # if dim == 1 || amdgpuaware_MPI(dim)  # Use a custom copy kernel for the first dimension to obtain a good copy performance (the CUDA 3-D memcopy does not perform well for this extremely strided case).
                 ranges = recvranges(n, dim, A);
                 nthreads = (dim==1) ? (1, 32, 1) : (32, 1, 1);
                 halosize = [r[end] - r[1] + 1 for r in ranges];
                 nblocks  = Tuple(ceil.(Int, halosize./nthreads));
                 @roc gridsize=nblocks groupsize=nthreads stream=rocstreams[n,i] read_x2d!(gpurecvbuf(n,dim,i,A), A, ranges[1], ranges[2], ranges[3], dim);
-            else
-                read_h2d_async!(recvbuf_flat(n,dim,i,A), A, recvranges(n,dim,A), rocstreams[n,i]);
-            end
+            # else
+            #     read_h2d_async!(recvbuf_flat(n,dim,i,A), A, recvranges(n,dim,A), rocstreams[n,i]);
+            # end
         end
     end