From 08955de1ed724bc3c6740c221092d6893f8831d7 Mon Sep 17 00:00:00 2001 From: Alexander Droste Date: Fri, 23 Oct 2026 12:18:15 +0100 Subject: [PATCH] perf(cuda): disable timing on pinned transfer events Replace BLOCKING_SYNC with DISABLE_TIMING on pinned-buffer events. This also changes host event waits from blocking to busy-waiting; buffer ownership and completion-event reuse are unchanged. Original-commit: 231277f878 Signed-off-by: Alexander Droste --- vortex-cuda/src/pinned.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vortex-cuda/src/pinned.rs b/vortex-cuda/src/pinned.rs index 40371a042af..afc3279591c 100644 --- a/vortex-cuda/src/pinned.rs +++ b/vortex-cuda/src/pinned.rs @@ -60,7 +60,7 @@ impl PinnedByteBuffer { "pinned host buffer length {logical_len} exceeds capacity {capacity}" ); let event = ctx - .new_event(Some(CUevent_flags::CU_EVENT_BLOCKING_SYNC)) + .new_event(Some(CUevent_flags::CU_EVENT_DISABLE_TIMING)) .map_err(|e| vortex_err!("failed to create pinned host buffer event: {e}"))?; // Keep file-I/O staging cacheable: cudarc's allocator uses write-combined memory, // which makes CPU reads expensive.