radeonsi: increase the number of compiler threads depending on the CPU

The compiler queue was limited to 3 threads, so shader-db running
on a 16-thread CPU would have a bottleneck on the 3-thread queue.

Reviewed-by: Timothy Arceri <[email protected]>
Tested-by: Benedikt Schemmer <ben at besd.de>
Reviewed-by: Nicolai Hähnle <[email protected]>

2 files changed, 29 insertions, 16 deletions

diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index c7f1e27f6ba..1ca38ed55cb 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -832,7 +832,7 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 					   const struct pipe_screen_config *config)
 {
 	struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
-	unsigned num_threads, num_compiler_threads, num_compiler_threads_lowprio, i;
+	unsigned hw_threads, num_comp_hi_threads, num_comp_lo_threads, i;
 
 	if (!sscreen) {
 		return NULL;
@@ -889,17 +889,30 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 
 	si_disk_cache_create(sscreen);
 
-	/* Only enable as many threads as we have target machines, but at most
-	 * the number of CPUs - 1 if there is more than one.
-	 */
-	num_threads = sysconf(_SC_NPROCESSORS_ONLN);
-	num_threads = MAX2(1, num_threads - 1);
-	num_compiler_threads = MIN2(num_threads, ARRAY_SIZE(sscreen->compiler));
-	num_compiler_threads_lowprio =
-		MIN2(num_threads, ARRAY_SIZE(sscreen->compiler_lowp));
+	/* Determine the number of shader compiler threads. */
+	hw_threads = sysconf(_SC_NPROCESSORS_ONLN);
+
+	if (hw_threads >= 12) {
+		num_comp_hi_threads = hw_threads * 3 / 4;
+		num_comp_lo_threads = hw_threads / 3;
+	} else if (hw_threads >= 6) {
+		num_comp_hi_threads = hw_threads - 2;
+		num_comp_lo_threads = hw_threads / 2;
+	} else if (hw_threads >= 2) {
+		num_comp_hi_threads = hw_threads - 1;
+		num_comp_lo_threads = hw_threads / 2;
+	} else {
+		num_comp_hi_threads = 1;
+		num_comp_lo_threads = 1;
+	}
+
+	num_comp_hi_threads = MIN2(num_comp_hi_threads,
+				   ARRAY_SIZE(sscreen->compiler));
+	num_comp_lo_threads = MIN2(num_comp_lo_threads,
+				   ARRAY_SIZE(sscreen->compiler_lowp));
 
 	if (!util_queue_init(&sscreen->shader_compiler_queue, "si_shader",
-			     32, num_compiler_threads,
+			     64, num_comp_hi_threads,
 			     UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
 		si_destroy_shader_cache(sscreen);
 		FREE(sscreen);
@@ -908,7 +921,7 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 
 	if (!util_queue_init(&sscreen->shader_compiler_queue_low_priority,
 			     "si_shader_low",
-			     32, num_compiler_threads_lowprio,
+			     64, num_comp_lo_threads,
 			     UTIL_QUEUE_INIT_RESIZE_IF_FULL |
 			     UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY)) {
 	       si_destroy_shader_cache(sscreen);
@@ -1059,9 +1072,9 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
 	if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
 		sscreen->debug_flags |= DBG_ALL_SHADERS;
 
-	for (i = 0; i < num_compiler_threads; i++)
+	for (i = 0; i < num_comp_hi_threads; i++)
 		si_init_compiler(sscreen, &sscreen->compiler[i]);
-	for (i = 0; i < num_compiler_threads_lowprio; i++)
+	for (i = 0; i < num_comp_lo_threads; i++)
 		si_init_compiler(sscreen, &sscreen->compiler_lowp[i]);
 
 	/* Create the auxiliary context. This must be done last. */
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index a67786c84d9..27efc5099f0 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -529,13 +529,13 @@ struct si_screen {
 	struct util_queue		shader_compiler_queue;
 	/* Use at most 3 normal compiler threads on quadcore and better.
 	 * Hyperthreaded CPUs report the number of threads, but we want
-	 * the number of cores. */
-	struct si_compiler		compiler[3]; /* used by the queue only */
+	 * the number of cores. We only need this many threads for shader-db. */
+	struct si_compiler		compiler[24]; /* used by the queue only */
 
 	struct util_queue		shader_compiler_queue_low_priority;
 	/* Use at most 2 low priority threads on quadcore and better.
 	 * We want to minimize the impact on multithreaded Mesa. */
-	struct si_compiler		compiler_lowp[2]; /* at most 2 threads */
+	struct si_compiler		compiler_lowp[10];
 };
 
 struct si_blend_color {