radeonsi: initialize shader compilers in threads on demand

It takes a noticable amount of time with piglit.

Reviewed-by: Timothy Arceri <[email protected]>

4 files changed, 14 insertions, 8 deletions

diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index 19d4cca0dba..6d9b5d0ad6e 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -121,6 +121,9 @@ static void si_create_compute_state_async(void *job, int thread_index)
 	assert(thread_index < ARRAY_SIZE(sscreen->compiler));
 	compiler = &sscreen->compiler[thread_index];
 
+	if (!compiler->passes)
+		si_init_compiler(sscreen, compiler);
+
 	if (program->ir_type == PIPE_SHADER_IR_TGSI) {
 		tgsi_scan_shader(sel->tokens, &sel->info);
 	} else {
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 1b7fdf23c60..be7424e9b7b 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -127,8 +127,7 @@ static const struct debug_named_value debug_options[] = {
 	DEBUG_NAMED_VALUE_END /* must be last */
 };
 
-static void si_init_compiler(struct si_screen *sscreen,
-			     struct ac_llvm_compiler *compiler)
+void si_init_compiler(struct si_screen *sscreen, struct ac_llvm_compiler *compiler)
 {
 	/* Only create the less-optimizing version of the compiler on APUs
 	 * predating Ryzen (Raven). */
@@ -938,7 +937,7 @@ radeonsi_screen_create_impl(struct radeon_winsys *ws,
 			    const struct pipe_screen_config *config)
 {
 	struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
-	unsigned hw_threads, num_comp_hi_threads, num_comp_lo_threads, i;
+	unsigned hw_threads, num_comp_hi_threads, num_comp_lo_threads;
 
 	if (!sscreen) {
 		return NULL;
@@ -1224,11 +1223,6 @@ radeonsi_screen_create_impl(struct radeon_winsys *ws,
 		}
 	}
 
-	for (i = 0; i < num_comp_hi_threads; i++)
-		si_init_compiler(sscreen, &sscreen->compiler[i]);
-	for (i = 0; i < num_comp_lo_threads; i++)
-		si_init_compiler(sscreen, &sscreen->compiler_lowp[i]);
-
 	sscreen->ge_wave_size = 64;
 	sscreen->ps_wave_size = 64;
 	sscreen->compute_wave_size = 64;
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 1a65c309e25..62183834416 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -1426,6 +1426,9 @@ void si_dispatch_prim_discard_cs_and_draw(struct si_context *sctx,
 					  unsigned input_indexbuf_max_elements);
 void si_initialize_prim_discard_tunables(struct si_context *sctx);
 
+/* si_pipe.c */
+void si_init_compiler(struct si_screen *sscreen, struct ac_llvm_compiler *compiler);
+
 /* si_perfcounters.c */
 void si_init_perfcounters(struct si_screen *screen);
 void si_destroy_perfcounters(struct si_screen *screen);
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index e5e7d523cd4..a6307950400 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -2105,6 +2105,9 @@ static void si_build_shader_variant(struct si_shader *shader,
 		compiler = shader->compiler_ctx_state.compiler;
 	}
 
+	if (!compiler->passes)
+		si_init_compiler(sscreen, compiler);
+
 	if (unlikely(!si_shader_create(sscreen, compiler, shader, debug))) {
 		PRINT_ERR("Failed to build shader variant (type=%u)\n",
 			  sel->type);
@@ -2472,6 +2475,9 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 	assert(thread_index < ARRAY_SIZE(sscreen->compiler));
 	compiler = &sscreen->compiler[thread_index];
 
+	if (!compiler->passes)
+		si_init_compiler(sscreen, compiler);
+
 	/* Compile the main shader part for use with a prolog and/or epilog.
 	 * If this fails, the driver will try to compile a monolithic shader
 	 * on demand.