I've committed this single-liner (and added a comment) to OG8. This allows detection of Compute Capability 7.0 (Volta) and lets libgomp allocate a more reasonable default gang number for Volta GPUs.
Tested without regressions on a powerpc64le-linux system. Note that mainline has different code for doing this task, so this patch doesn't apply there (not needed there). Chung-Lin [og] Handle Compute Capability 7.0 (Volta) libgomp/ * plugin/plugin-nvptx.c (GOMP_OFFLOAD_load_image): Handle up to Compute Capability 7.0.
diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c index c2d3b71..706a36f 100644 --- a/libgomp/plugin/plugin-nvptx.c +++ b/libgomp/plugin/plugin-nvptx.c @@ -1273,8 +1273,9 @@ GOMP_OFFLOAD_load_image (int ord, unsigned version, const void *target_data, dev->register_allocation_unit_size = 256; dev->register_allocation_granularity = 2; } - else if (dev->binary_version <= 62) + else if (dev->binary_version <= 70) { + /* Compute Capability 6.1, 6.2, 7.0 share same parameters. */ dev->register_allocation_unit_size = 256; dev->register_allocation_granularity = 4; }