use blocking synchronize to reduce poll waiting

CliMA · Oct 19, 2023 · 6a59fc8 · 6a59fc8
1 parent 9c22047
commit 6a59fc8
Showing 1 changed file with 4 additions and 4 deletions.
diff --git a/src/Spaces/dss_cuda.jl b/src/Spaces/dss_cuda.jl
@@ -427,13 +427,13 @@ function fill_send_buffer!(::ClimaComms.CUDADevice, dss_buffer::DSSBuffer)
     if nsend > 0
         nitems = nsend * nlevels * nfid
         nthreads, nblocks = _configure_threadblock(nitems)
-        CUDA.synchronize() # CUDA MPI uses a separate stream. This will synchronize across streams
+        CUDA.synchronize(;blocking=true) # CUDA MPI uses a separate stream. This will synchronize across streams
         @cuda threads = (nthreads) blocks = (nblocks) fill_send_buffer_kernel!(
             send_data,
             send_buf_idx,
             pperimeter_data,
         )
-        CUDA.synchronize() # CUDA MPI uses a separate stream. This will synchronize across streams
+        CUDA.synchronize(;blocking=true) # CUDA MPI uses a separate stream. This will synchronize across streams
     end
     return nothing
 end
@@ -468,13 +468,13 @@ function load_from_recv_buffer!(::ClimaComms.CUDADevice, dss_buffer::DSSBuffer)
     if nrecv > 0
         nitems = nrecv * nlevels * nfid
         nthreads, nblocks = _configure_threadblock(nitems)
-        CUDA.synchronize()
+        CUDA.synchronize(;blocking=true)
         @cuda threads = (nthreads) blocks = (nblocks) load_from_recv_buffer_kernel!(
             pperimeter_data,
             recv_data,
             recv_buf_idx,
         )
-        CUDA.synchronize()
+        CUDA.synchronize(;blocking=true)
     end
     return nothing
 end