Mesa (master): radeonsi: only support at most 1024 threads per block

GitLab Mirror gitlab-mirror at kemper.freedesktop.org
Tue Sep 10 03:44:11 UTC 2019


Module: Mesa
Branch: master
Commit: 42ea0b7b52d78fc923e50a0825859fe079fd8c35
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=42ea0b7b52d78fc923e50a0825859fe079fd8c35

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Tue Aug 27 17:39:02 2019 -0400

radeonsi: only support at most 1024 threads per block

LLVM 10 won't support 2048.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer at amd.com>

---

 src/gallium/drivers/radeonsi/si_get.c | 10 ++--------
 1 file changed, 2 insertions(+), 8 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_get.c b/src/gallium/drivers/radeonsi/si_get.c
index 118301fbf4b..111279674e6 100644
--- a/src/gallium/drivers/radeonsi/si_get.c
+++ b/src/gallium/drivers/radeonsi/si_get.c
@@ -752,14 +752,8 @@ static unsigned get_max_threads_per_block(struct si_screen *screen,
 	if (ir_type == PIPE_SHADER_IR_NATIVE)
 		return 256;
 
-	/* Only 16 waves per thread-group on gfx9. */
-	if (screen->info.chip_class >= GFX9)
-		return 1024;
-
-	/* Up to 40 waves per thread-group on GCN < gfx9. Expose a nice
-	 * round number.
-	 */
-	return 2048;
+        /* LLVM 10 only supports 1024 threads per block. */
+	return 1024;
 }
 
 static int si_get_compute_param(struct pipe_screen *screen,




More information about the mesa-commit mailing list