[Mesa-dev] [PATCH 47/61] radeonsi/gfx9: always compile monolithic ES-GS (asynchronously)

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

In addition to the non-monolithic variant.
---
 src/gallium/drivers/radeonsi/si_shader.h        | 10 +++++++++-
 src/gallium/drivers/radeonsi/si_state_shaders.c | 19 +++++++++++++++++++
 2 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index e0227e4..4fb79e6 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -438,26 +438,34 @@ struct si_shader_key {
 	unsigned as_ls:1; /* local shader, which precedes TCS */
 
 	/* Flags for monolithic compilation only. */
 	struct {
 		/* One byte for every input: SI_FIX_FETCH_* enums. */
 		uint8_t		vs_fix_fetch[SI_MAX_ATTRIBS];
 		uint64_t	ff_tcs_inputs_to_copy; /* for fixed-func TCS */
 	} mono;
 
 	/* Optimization flags for asynchronous compilation only. */
-	union {
+	struct {
 		struct {
 			uint64_t	kill_outputs; /* "get_unique_index" bits */
 			uint32_t	kill_outputs2; /* "get_unique_index2" bits */
 			unsigned	clip_disable:1;
 		} hw_vs; /* HW VS (it can be VS, TES, GS) */
+
+		/* For shaders where monolithic variants have better code.
+		 *
+		 * This is a flag that has no effect on code generation,
+		 * but forces monolithic shaders to be used as soon as
+		 * possible, because it's in the "opt" group.
+		 */
+		unsigned	prefer_mono:1;
 	} opt;
 };
 
 struct si_shader_config {
 	unsigned			num_sgprs;
 	unsigned			num_vgprs;
 	unsigned			spilled_sgprs;
 	unsigned			spilled_vgprs;
 	unsigned			private_mem_vgprs;
 	unsigned			lds_size;
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 6bb3f50..22bf3cf 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -1271,20 +1271,39 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 		break;
 	case PIPE_SHADER_GEOMETRY:
 		if (sctx->b.chip_class >= GFX9) {
 			if (sctx->tes_shader.cso) {
 				key->part.gs.es = sctx->tes_shader.cso;
 			} else {
 				si_shader_selector_key_vs(sctx, sctx->vs_shader.cso,
 							  key, &key->part.gs.vs_prolog);
 				key->part.gs.es = sctx->vs_shader.cso;
 			}
+
+			/* Merged ES-GS can have unbalanced wave usage.
+			 *
+			 * ES threads are per-vertex, while GS threads are
+			 * per-primitive. So without any amplification, there
+			 * are fewer GS threads than ES threads, which can result
+			 * in empty (no-op) GS waves. With too much amplification,
+			 * there are more GS threads than ES threads, which
+			 * can result in empty (no-op) ES waves.
+			 *
+			 * Non-monolithic shaders are implemented by setting EXEC
+			 * at the beginning of shader parts, and don't jump to
+			 * the end if EXEC is 0.
+			 *
+			 * Monolithic shaders use conditional blocks, so they can
+			 * jump and skip empty waves of ES or GS. So set this to
+			 * always use optimized variants, which are monolithic.
+			 */
+			key->opt.prefer_mono = 1;
 		}
 		key->part.gs.prolog.tri_strip_adj_fix = sctx->gs_tri_strip_adj_fix;
 		break;
 	case PIPE_SHADER_FRAGMENT: {
 		struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
 		struct si_state_blend *blend = sctx->queued.named.blend;
 
 		if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
 		    sel->info.colors_written == 0x1)
 			key->part.ps.epilog.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
-- 
2.7.4