document GPU temporary copy

eth-cscs · Nov 4, 2021 · a0f9199 · a0f9199
1 parent a94417a
commit a0f9199
Showing 1 changed file with 9 additions and 0 deletions.
diff --git a/include/dlaf/communication/kernels/broadcast.h b/include/dlaf/communication/kernels/broadcast.h
@@ -89,6 +89,15 @@ struct ScheduleRecvBcast {
     using matrix::duplicateIfNeeded;
     using matrix::copy_o;
 
+    // Note:
+    //
+    // TILE_GPU -+-> duplicateIfNeeded<CPU> ---> TILE_CPU ---> recvBcast ---> TILE_CPU -+-> copy
+    //           |                                                                      |
+    //           +----------------------------------------------------------------------+
+    //
+    // Actually `duplicateIfNeeded` always makes a copy, because it is always needed since this
+    // is the specialization for GPU input and MPI withuot CUDA_RDMA requires CPU memory.
+
     auto tile_gpu = tile.share();
     auto tile_cpu = duplicateIfNeeded<Device::CPU>(tile_gpu);