Change CUDA DOT thread-blocks to 1024

This improves the performance on Ampere (A100) GPUs. Fixes #137.
UoB-HPC · Jun 12, 2023 · 092ee67 · 092ee67
1 parent 893af9f
commit 092ee67
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/src/cuda/CUDAStream.h b/src/cuda/CUDAStream.h
@@ -22,7 +22,7 @@
 #endif
 
 #define TBSIZE 1024
-#define DOT_NUM_BLOCKS 256
+#define DOT_NUM_BLOCKS 1024
 
 template <class T>
 class CUDAStream : public Stream<T>