[Mesa-dev] [PATCH 8/8] radeonsi: add radeonsi_sync_compile option

[Nicolai Hähnle]

From: Nicolai Hähnle <nicolai.haehnle at amd.com>

Force the driver thread to sync immediately with a compiler thread (but
compilation still happens in a separate thread).

This can be useful to simplify debugging compiler issues.
---
 src/gallium/drivers/radeonsi/si_debug_options.inc |  1 +
 src/gallium/drivers/radeonsi/si_state_shaders.c   | 13 ++++++++++---
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_debug_options.inc b/src/gallium/drivers/radeonsi/si_debug_options.inc
index f4c3e19ed95..019256ca1d1 100644
--- a/src/gallium/drivers/radeonsi/si_debug_options.inc
+++ b/src/gallium/drivers/radeonsi/si_debug_options.inc
@@ -1,5 +1,6 @@
 OPT_BOOL(clear_db_cache_before_clear, false, "Clear DB cache before fast depth clear")
 OPT_BOOL(enable_nir, false, "Enable NIR")
 OPT_BOOL(aux_debug, false, "Generate ddebug_dumps for the auxiliary context")
+OPT_BOOL(sync_compile, false, "Always compile synchronously (will cause stalls)")
 
 #undef OPT_BOOL
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 5bdfd4f6ac1..f57e7730905 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -1945,20 +1945,24 @@ current_not_ready:
 			sel->first_variant = shader;
 			sel->last_variant = shader;
 		} else {
 			sel->last_variant->next_variant = shader;
 			sel->last_variant = shader;
 		}
 
 		/* Use the default (unoptimized) shader for now. */
 		memset(&key->opt, 0, sizeof(key->opt));
 		mtx_unlock(&sel->mutex);
+
+		if (sscreen->options.sync_compile)
+			util_queue_fence_wait(&shader->ready);
+
 		goto again;
 	}
 
 	/* Reset the fence before adding to the variant list. */
 	util_queue_fence_reset(&shader->ready);
 
 	if (!sel->last_variant) {
 		sel->first_variant = shader;
 		sel->last_variant = shader;
 	} else {
@@ -2157,38 +2161,41 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 }
 
 void si_schedule_initial_compile(struct si_context *sctx, unsigned processor,
 				 struct util_queue_fence *ready_fence,
 				 struct si_compiler_ctx_state *compiler_ctx_state,
 				 void *job, util_queue_execute_func execute)
 {
 	util_queue_fence_init(ready_fence);
 
 	struct util_async_debug_callback async_debug;
-	bool wait =
+	bool debug =
 		(sctx->debug.debug_message && !sctx->debug.async) ||
 		sctx->is_debug ||
 		si_can_dump_shader(sctx->screen, processor);
 
-	if (wait) {
+	if (debug) {
 		u_async_debug_init(&async_debug);
 		compiler_ctx_state->debug = async_debug.base;
 	}
 
 	util_queue_add_job(&sctx->screen->shader_compiler_queue, job,
 			   ready_fence, execute, NULL);
 
-	if (wait) {
+	if (debug) {
 		util_queue_fence_wait(ready_fence);
 		u_async_debug_drain(&async_debug, &sctx->debug);
 		u_async_debug_cleanup(&async_debug);
 	}
+
+	if (sctx->screen->options.sync_compile)
+		util_queue_fence_wait(ready_fence);
 }
 
 /* Return descriptor slot usage masks from the given shader info. */
 void si_get_active_slot_masks(const struct tgsi_shader_info *info,
 			      uint32_t *const_and_shader_buffers,
 			      uint64_t *samplers_and_images)
 {
 	unsigned start, num_shaderbufs, num_constbufs, num_images, num_samplers;
 
 	num_shaderbufs = util_last_bit(info->shader_buffers_declared);
-- 
2.20.1