st/mesa: add environment variable pin_app_thread for faster glthread on AMD Zen

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4369>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4369>

diff --git a/src/gallium/auxiliary/util/u_helpers.c b/src/gallium/auxiliary/util/u_helpers.c
index 5e5eeedbf45a92c44600ebc585e5c78469e8151a..31b7d533952d158ebef68ce6c28c478d4d384d0c 100644 (file)
--- a/src/gallium/auxiliary/util/u_helpers.c
+++ b/src/gallium/auxiliary/util/u_helpers.c
@@ -192,6 +192,16 @@ util_pin_driver_threads_to_random_L3(struct pipe_context *ctx,
    /* Do the same for the upper level thread if there is any (e.g. glthread) */
    if (upper_thread)
       util_pin_thread_to_L3(*upper_thread, cache, util_cpu_caps.cores_per_L3);
+
+   /* Optionally pin the application thread to the same L3 to get maximum
+    * performance with glthread on AMD Zen. (this function is only called
+    * with glthread) This is used to estimate and remove the overhead of
+    * Infinity Fabric between L3 caches.
+    */
+#if defined(HAVE_PTHREAD)
+   if (debug_get_bool_option("pin_app_thread", false))
+      util_pin_thread_to_L3(pthread_self(), cache, util_cpu_caps.cores_per_L3);
+#endif
 }
 
 /* This is a helper for hardware bring-up. Don't remove. */