[libc] Fix lane-id utility function not using built-in (#84902)

jhuber6 · web-flow · commit c167a2588737 · 2024-03-12T10:40:35.000-05:00
Summary:
Previously we got the lane-id from taking the global thread ID and
taking off the bottom 5 bits. This works but is inefficient compared to
the NVPTX intrinsic simply dedicated to get this value.
diff --git a/libc/src/__support/GPU/nvptx/utils.h b/libc/src/__support/GPU/nvptx/utils.h
@@ -97,7 +97,7 @@ LIBC_INLINE uint32_t get_lane_size() { return 32; }
 
 /// Returns the id of the thread inside of a CUDA warp executing together.
 [[clang::convergent]] LIBC_INLINE uint32_t get_lane_id() {
-  return get_thread_id() & (get_lane_size() - 1);
+  return __nvvm_read_ptx_sreg_laneid();
 }
 
 /// Returns the bit-mask of active threads in the current warp.

Original file line number	Diff line number	Diff line change
`@@ -97,7 +97,7 @@ LIBC_INLINE uint32_t get_lane_size() { return 32; }`
`97`	`97`
`98`	`98`	`/// Returns the id of the thread inside of a CUDA warp executing together.`
`99`	`99`	`[[clang::convergent]] LIBC_INLINE uint32_t get_lane_id() {`
`100`		`- return get_thread_id() & (get_lane_size() - 1);`
	`100`	`+ return __nvvm_read_ptx_sreg_laneid();`
`101`	`101`	`}`
`102`	`102`
`103`	`103`	`/// Returns the bit-mask of active threads in the current warp.`