st/mesa: pin driver threads to a fixed CCX when glthread is enabled

radeonsi has 3 driver threads (glthread, gallium, winsys), other drivers
may have 2 (glthread, gallium), so it makes sense to pin them to a random
CCX and keep that irrespective of the app thread.

Reviewed-by: Dave Airlie <airlied@redhat.com>

diff --git a/src/gallium/auxiliary/util/u_helpers.c b/src/gallium/auxiliary/util/u_helpers.c
index 4c70c004178b36bbc8973ae0fad882b1b6611186..821242c01818ed7aa95c84163027a70aaaac2168 100644 (file)
--- a/src/gallium/auxiliary/util/u_helpers.c
+++ b/src/gallium/auxiliary/util/u_helpers.c
@@ -121,43 +121,6 @@ util_upload_index_buffer(struct pipe_context *pipe,
    return *out_buffer != NULL;
 }
 
-#ifdef HAVE_PTHREAD_SETAFFINITY
-
-static unsigned L3_cache_number;
-static once_flag thread_pinning_once_flag = ONCE_FLAG_INIT;
-
-static void
-util_set_full_cpu_affinity(void)
-{
-   cpu_set_t cpuset;
-
-   CPU_ZERO(&cpuset);
-   for (unsigned i = 0; i < CPU_SETSIZE; i++)
-      CPU_SET(i, &cpuset);
-
-   pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset);
-}
-
-static void
-util_init_thread_pinning(void)
-{
-   /* Get a semi-random number. */
-   int64_t t = os_time_get_nano();
-   L3_cache_number = (t ^ (t >> 8) ^ (t >> 16));
-
-   /* Reset thread affinity for all child processes to prevent them from
-    * inheriting the current thread's affinity.
-    *
-    * XXX: If the driver is unloaded after this, and the app later calls
-    * fork(), the child process will likely crash before fork() returns,
-    * because the address where util_set_full_cpu_affinity was located
-    * will either be unmapped or point to random other contents.
-    */
-   pthread_atfork(NULL, NULL, util_set_full_cpu_affinity);
-}
-
-#endif
-
 /**
  * Called by MakeCurrent. Used to notify the driver that the application
  * thread may have been changed.
@@ -170,30 +133,21 @@ util_init_thread_pinning(void)
  *                      pinned.
  */
 void
-util_context_thread_changed(struct pipe_context *ctx, thrd_t *upper_thread)
+util_pin_driver_threads_to_random_L3(struct pipe_context *ctx,
+                                     thrd_t *upper_thread)
 {
-#ifdef HAVE_PTHREAD_SETAFFINITY
    /* If pinning has no effect, don't do anything. */
    if (util_cpu_caps.nr_cpus == util_cpu_caps.cores_per_L3)
       return;
 
-   thrd_t current = thrd_current();
-   int cache = util_get_L3_for_pinned_thread(current,
-                                             util_cpu_caps.cores_per_L3);
-
-   call_once(&thread_pinning_once_flag, util_init_thread_pinning);
+   unsigned num_L3_caches = util_cpu_caps.nr_cpus /
+                            util_cpu_caps.cores_per_L3;
 
-   /* If the main thread is not pinned, choose the L3 cache. */
-   if (cache == -1) {
-      unsigned num_L3_caches = util_cpu_caps.nr_cpus /
-                               util_cpu_caps.cores_per_L3;
-
-      /* Choose a different L3 cache for each subsequent MakeCurrent. */
-      cache = p_atomic_inc_return(&L3_cache_number) % num_L3_caches;
-      util_pin_thread_to_L3(current, cache, util_cpu_caps.cores_per_L3);
-   }
+   /* Get a semi-random number. */
+   int64_t t = os_time_get_nano();
+   unsigned cache = (t ^ (t >> 8) ^ (t >> 16)) % num_L3_caches;
 
-   /* Tell the driver to pin its threads to the same L3 cache. */
+   /* Tell the driver to pin its threads to the selected L3 cache. */
    if (ctx->set_context_param) {
       ctx->set_context_param(ctx, PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
                              cache);
@@ -202,7 +156,6 @@ util_context_thread_changed(struct pipe_context *ctx, thrd_t *upper_thread)
    /* 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);
-#endif
 }
 
 /* This is a helper for hardware bring-up. Don't remove. */

diff --git a/src/gallium/auxiliary/util/u_helpers.h b/src/gallium/auxiliary/util/u_helpers.h
index 38c47c1cc98e6cc0fb644816fe434367a4450951..ed8467291ba6e23e8228ca05eafdfbe5b90348b3 100644 (file)
--- a/src/gallium/auxiliary/util/u_helpers.h
+++ b/src/gallium/auxiliary/util/u_helpers.h
@@ -52,7 +52,8 @@ bool util_upload_index_buffer(struct pipe_context *pipe,
                               unsigned *out_offset);
 
 void
-util_context_thread_changed(struct pipe_context *ctx, thrd_t *upper_thread);
+util_pin_driver_threads_to_random_L3(struct pipe_context *ctx,
+                                     thrd_t *upper_thread);
 
 struct pipe_query *
 util_begin_pipestat_query(struct pipe_context *ctx);

diff --git a/src/mesa/state_tracker/st_manager.c b/src/mesa/state_tracker/st_manager.c
index 076ad42646d0a174828d1f3b27d2084213ce2746..73729d745455916352ab0fa7f5c4d45c264b2198 100644 (file)
--- a/src/mesa/state_tracker/st_manager.c
+++ b/src/mesa/state_tracker/st_manager.c
@@ -800,6 +800,17 @@ st_start_thread(struct st_context_iface *stctxi)
    struct st_context *st = (struct st_context *) stctxi;
 
    _mesa_glthread_init(st->ctx);
+
+   /* Pin all driver threads to one L3 cache for optimal performance
+    * on AMD Zen. This is only done if glthread is enabled.
+    *
+    * If glthread is disabled, st_draw.c re-pins driver threads regularly
+    * based on the location of the app thread.
+    */
+   struct glthread_state *glthread = st->ctx->GLThread;
+   if (glthread && st->pipe->set_context_param) {
+      util_pin_driver_threads_to_random_L3(st->pipe, &glthread->queue.threads[0]);
+   }
 }