From 16e3e0d93bc8f21a9e3aac6ea941b8dc6e625136 Mon Sep 17 00:00:00 2001 From: Mel Henning Date: Mon, 24 Mar 2025 12:39:56 -0400 Subject: [PATCH] nvk: Support blackwell in max_warps_per_mp_for_sm Part-of: --- src/nouveau/winsys/nouveau_device.c | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/src/nouveau/winsys/nouveau_device.c b/src/nouveau/winsys/nouveau_device.c index 925df476be9..f70839c0659 100644 --- a/src/nouveau/winsys/nouveau_device.c +++ b/src/nouveau/winsys/nouveau_device.c @@ -111,6 +111,9 @@ sm_for_chipset(uint16_t chipset) static uint8_t max_warps_per_mp_for_sm(uint8_t sm) { + /* These are documented in each architecture's tuning guide. Eg. see: + * https://docs.nvidia.com/cuda/blackwell-tuning-guide/index.html#occupancy + */ switch (sm) { case 10: case 11: @@ -124,6 +127,7 @@ max_warps_per_mp_for_sm(uint8_t sm) case 86: case 87: case 89: + case 120: return 48; case 30: case 32: @@ -139,6 +143,7 @@ max_warps_per_mp_for_sm(uint8_t sm) case 72: case 80: case 90: + case 100: return 64; default: assert(!"unkown SM version");