Reduce chunk_size used in CUDA kernels from 8 to 1

diff --git a/doc/news/changes/minor/20190423BrunoTurcksin b/doc/news/changes/minor/20190423BrunoTurcksin
new file mode 100644 (file)
index 0000000..3a3c21d
--- /dev/null
+++ b/doc/news/changes/minor/20190423BrunoTurcksin
@@ -0,0 +1,4 @@
+Changed: The chunk_size used in CUDA kernels has been reduced from 8 to 1 to
+improve performance on newer architectures.
+<br>
+(Bruno Turcksin, 2019/04/23)

diff --git a/include/deal.II/base/cuda_size.h b/include/deal.II/base/cuda_size.h
index 26e0a3c4d99008259e6f3938d28cfc17c0fd6ad7..274b418f9a75c2b40909a33f0ae155e0ae4041e5 100644 (file)
--- a/include/deal.II/base/cuda_size.h
+++ b/include/deal.II/base/cuda_size.h
@@ -23,14 +23,16 @@ DEAL_II_NAMESPACE_OPEN
 namespace CUDAWrappers
 {
   /**
-   * Define the size of a block when launching a CUDA kernel.
+   * Define the size of a block when launching a CUDA kernel. This number can be
+   * changed depending on the architecture the code is running on.
    */
   constexpr int block_size = 512;
 
   /**
-   * Define the size of chunk of data worked on by a thread.
+   * Define the size of chunk of data worked on by a thread. This number can be
+   * changed depending on the architecture the code is running on.
    */
-  constexpr int chunk_size = 8;
+  constexpr int chunk_size = 1;
 } // namespace CUDAWrappers
 
 DEAL_II_NAMESPACE_CLOSE