fix(cuda_std): use correct PTX scope suffix in block acqrel fence

Snehal-Reddy · Snehal-Reddy · commit deb2e1022eb1 · 2026-04-19T14:22:16.000+01:00
This updates the inline assembly to use the correct `.cta` (cooperative thread array) scope suffix, ensuring that block-level fences don't incur the unnecessary performance overhead of a system-wide synchronization.
diff --git a/crates/cuda_std/src/atomic/intrinsics.rs b/crates/cuda_std/src/atomic/intrinsics.rs
@@ -42,7 +42,7 @@ pub unsafe fn fence_acqrel_device() {
 
 #[gpu_only]
 pub unsafe fn fence_acqrel_block() {
-    asm!("fence.acq_rel.sys;");
+    asm!("fence.acq_rel.cta;");
 }
 
 #[gpu_only]

Original file line number	Diff line number	Diff line change
`@@ -42,7 +42,7 @@ pub unsafe fn fence_acqrel_device() {`
`42`	`42`
`43`	`43`	`#[gpu_only]`
`44`	`44`	`pub unsafe fn fence_acqrel_block() {`
`45`		`- asm!("fence.acq_rel.sys;");`
	`45`	`+ asm!("fence.acq_rel.cta;");`
`46`	`46`	`}`
`47`	`47`
`48`	`48`	`#[gpu_only]`