radeonsi/gfx10: add as_ngg variant for TES as ES to select Wave32/64

Legacy GS has to use Wave64, so TES before GS has to use Wave64 too.

Acked-by: Pierre-Eric Pelloux-Prayer <[email protected]>
Reviewed-by: Samuel Pitoiset <[email protected]>

Author: Marek Olšák

src/gallium/drivers/radeonsi/si_pipe.h

@@ -1895,13 +1895,14 @@ static inline bool si_compute_prim_discard_enabled(struct si_context *sctx)
 
 static inline unsigned si_get_wave_size(struct si_screen *sscreen,
 					enum pipe_shader_type shader_type,
-					bool ngg)
+					bool ngg, bool es)
 {
 	if (shader_type == PIPE_SHADER_COMPUTE)
 		return sscreen->compute_wave_size;
 	else if (shader_type == PIPE_SHADER_FRAGMENT)
 		return sscreen->ps_wave_size;
-	else if (shader_type == PIPE_SHADER_GEOMETRY && !ngg) /* legacy GS only supports Wave64 */
+	else if ((shader_type == PIPE_SHADER_TESS_EVAL && es && !ngg) ||
+		 (shader_type == PIPE_SHADER_GEOMETRY && !ngg)) /* legacy GS only supports Wave64 */
 		return 64;
 	else
 		return sscreen->ge_wave_size;
@@ -1910,7 +1911,7 @@ static inline unsigned si_get_wave_size(struct si_screen *sscreen,
 static inline unsigned si_get_shader_wave_size(struct si_shader *shader)
 {
 	return si_get_wave_size(shader->selector->screen, shader->selector->type,
-				shader->key.as_ngg);
+				shader->key.as_ngg, shader->key.as_es);
 }
 
 #define PRINT_ERR(fmt, args...) \

src/gallium/drivers/radeonsi/si_shader.c

@@ -5727,7 +5727,7 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
 	shader->is_gs_copy_shader = true;
 
 	si_init_shader_ctx(&ctx, sscreen, compiler,
-			   si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false));
+			   si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false));
 	ctx.shader = shader;
 	ctx.type = PIPE_SHADER_VERTEX;
 
@@ -6172,7 +6172,8 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx)
 		}
 	}
 
-	if (shader->key.as_ngg && ctx->type != PIPE_SHADER_GEOMETRY) {
+	if (ctx->type != PIPE_SHADER_GEOMETRY &&
+	    (shader->key.as_ngg && !shader->key.as_es)) {
 		/* Unconditionally declare scratch space base for streamout and
 		 * vertex compaction. Whether space is actually allocated is
 		 * determined during linking / PM4 creation.
@@ -6219,13 +6220,13 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx)
 						ctx->param_merged_wave_info, 0);
 		} else if (ctx->type == PIPE_SHADER_TESS_CTRL ||
 			   ctx->type == PIPE_SHADER_GEOMETRY ||
-			   shader->key.as_ngg) {
+			   (shader->key.as_ngg && !shader->key.as_es)) {
 			LLVMValueRef num_threads;
 			bool nested_barrier;
 
 			if (!shader->is_monolithic ||
 			    (ctx->type == PIPE_SHADER_TESS_EVAL &&
-			     shader->key.as_ngg))
+			     (shader->key.as_ngg && !shader->key.as_es)))
 				ac_init_exec_full_mask(&ctx->ac);
 
 			if (ctx->type == PIPE_SHADER_TESS_CTRL ||
@@ -7048,6 +7049,7 @@ int si_compile_tgsi_shader(struct si_screen *sscreen,
 			struct si_shader shader_es = {};
 			shader_es.selector = es;
 			shader_es.key.as_es = 1;
+			shader_es.key.as_ngg = shader->key.as_ngg;
 			shader_es.key.mono = shader->key.mono;
 			shader_es.key.opt = shader->key.opt;
 			shader_es.is_monolithic = true;
@@ -7305,7 +7307,8 @@ si_get_shader_part(struct si_screen *sscreen,
 
 	struct si_shader_context ctx;
 	si_init_shader_ctx(&ctx, sscreen, compiler,
-			   si_get_wave_size(sscreen, type, shader.key.as_ngg));
+			   si_get_wave_size(sscreen, type, shader.key.as_ngg,
+					    shader.key.as_es));
 	ctx.shader = &shader;
 	ctx.type = type;
 
@@ -7703,10 +7706,15 @@ static bool si_shader_select_gs_parts(struct si_screen *sscreen,
 				      struct pipe_debug_callback *debug)
 {
 	if (sscreen->info.chip_class >= GFX9) {
-		struct si_shader *es_main_part =
-			shader->key.part.gs.es->main_shader_part_es;
+		struct si_shader *es_main_part;
+		enum pipe_shader_type es_type = shader->key.part.gs.es->type;
 
-		if (shader->key.part.gs.es->type == PIPE_SHADER_VERTEX &&
+		if (es_type == PIPE_SHADER_TESS_EVAL && shader->key.as_ngg)
+			es_main_part = shader->key.part.gs.es->main_shader_part_ngg_es;
+		else
+			es_main_part = shader->key.part.gs.es->main_shader_part_es;
+
+		if (es_type == PIPE_SHADER_VERTEX &&
 		    !si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
 				      &shader->key.part.gs.vs_prolog))
 			return false;

src/gallium/drivers/radeonsi/si_shader.h

@@ -337,6 +337,7 @@ struct si_shader_selector {
 	struct si_shader	*main_shader_part_ls; /* as_ls is set in the key */
 	struct si_shader	*main_shader_part_es; /* as_es is set in the key */
 	struct si_shader	*main_shader_part_ngg; /* as_ngg is set in the key */
+	struct si_shader	*main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
 
 	struct si_shader	*gs_copy_shader;
 
@@ -789,6 +790,8 @@ si_get_main_shader_part(struct si_shader_selector *sel,
 {
 	if (key->as_ls)
 		return &sel->main_shader_part_ls;
+	if (key->as_es && key->as_ngg)
+		return &sel->main_shader_part_ngg_es;
 	if (key->as_es)
 		return &sel->main_shader_part_es;
 	if (key->as_ngg)

src/gallium/drivers/radeonsi/si_state_shaders.c

@@ -1852,10 +1852,11 @@ static inline void si_shader_selector_key(struct pipe_context *ctx,
 			key->mono.u.ff_tcs_inputs_to_copy = sctx->vs_shader.cso->outputs_written;
 		break;
 	case PIPE_SHADER_TESS_EVAL:
+		key->as_ngg = stages_key.u.ngg;
+
 		if (sctx->gs_shader.cso)
 			key->as_es = 1;
 		else {
-			key->as_ngg = stages_key.u.ngg;
 			si_shader_selector_key_hw_vs(sctx, sel, key);
 
 			if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid)
@@ -2269,6 +2270,10 @@ int si_shader_select_with_key(struct si_screen *sscreen,
 			else
 				assert(0);
 
+			if (sel->type == PIPE_SHADER_GEOMETRY &&
+			    previous_stage_sel->type == PIPE_SHADER_TESS_EVAL)
+				shader1_key.as_ngg = key->as_ngg;
+
 			mtx_lock(&previous_stage_sel->mutex);
 			ok = si_check_missing_main_part(sscreen,
 							previous_stage_sel,
@@ -2429,7 +2434,7 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 	if (sel->nir) {
 		/* TODO: GS always sets wave size = default. Legacy GS will have
 		 * incorrect subgroup_size and ballot_bit_size. */
-		si_lower_nir(sel, si_get_wave_size(sscreen, sel->type, true));
+		si_lower_nir(sel, si_get_wave_size(sscreen, sel->type, true, false));
 	}
 
 	/* Compile the main shader part for use with a prolog and/or epilog.
@@ -2455,9 +2460,9 @@ static void si_init_shader_selector_async(void *job, int thread_index)
 					      sel->so.num_outputs != 0,
 					      &shader->key);
 		if (sscreen->info.chip_class >= GFX10 &&
-		    (((sel->type == PIPE_SHADER_VERTEX ||
-		       sel->type == PIPE_SHADER_TESS_EVAL) &&
+		    ((sel->type == PIPE_SHADER_VERTEX &&
 		      !shader->key.as_ls && !shader->key.as_es) ||
+		     sel->type == PIPE_SHADER_TESS_EVAL ||
 		     sel->type == PIPE_SHADER_GEOMETRY))
 			shader->key.as_ngg = 1;