radeonsi/gfx10: fix the wave size for compute-based culling

Acked-by: Pierre-Eric Pelloux-Prayer <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4269>

5 files changed, 19 insertions, 6 deletions

diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 41c5bd45036..400f2152243 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -1967,13 +1967,14 @@ 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 es)
+					bool ngg, bool es, bool prim_discard_cs)
 {
 	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_VERTEX && es && !ngg) ||
+	else if ((shader_type == PIPE_SHADER_VERTEX && prim_discard_cs) || /* only Wave64 implemented */
+		 (shader_type == PIPE_SHADER_VERTEX && es && !ngg) ||
 		 (shader_type == PIPE_SHADER_TESS_EVAL && es && !ngg) ||
 		 (shader_type == PIPE_SHADER_GEOMETRY && !ngg)) /* legacy GS only supports Wave64 */
 		return 64;
@@ -1984,7 +1985,8 @@ 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_es);
+				shader->key.as_ngg, shader->key.as_es,
+				shader->key.opt.vs_as_prim_discard_cs);
 }
 
 #define PRINT_ERR(fmt, args...) \
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 8c8f90b80e2..f0e60087dbf 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1636,6 +1636,7 @@ static void si_get_vs_prolog_key(const struct si_shader_info *info,
 	key->vs_prolog.as_ls = shader_out->key.as_ls;
 	key->vs_prolog.as_es = shader_out->key.as_es;
 	key->vs_prolog.as_ngg = shader_out->key.as_ngg;
+	key->vs_prolog.as_prim_discard_cs = shader_out->key.opt.vs_as_prim_discard_cs;
 
 	if (ngg_cull_shader) {
 		key->vs_prolog.gs_fast_launch_tri_list = !!(shader_out->key.opt.ngg_culling &
@@ -1656,6 +1657,12 @@ static void si_get_vs_prolog_key(const struct si_shader_info *info,
 		key->vs_prolog.num_merged_next_stage_vgprs = 5;
 	}
 
+	/* Only one of these combinations can be set. as_ngg can be set with as_es. */
+	assert(key->vs_prolog.as_ls +
+	       key->vs_prolog.as_ngg +
+	       (key->vs_prolog.as_es && !key->vs_prolog.as_ngg) +
+	       key->vs_prolog.as_prim_discard_cs <= 1);
+
 	/* Enable loading the InstanceID VGPR. */
 	uint16_t input_mask = u_bit_consecutive(0, info->num_inputs);
 
@@ -2088,6 +2095,7 @@ si_get_shader_part(struct si_screen *sscreen,
 		shader.key.as_ls = key->vs_prolog.as_ls;
 		shader.key.as_es = key->vs_prolog.as_es;
 		shader.key.as_ngg = key->vs_prolog.as_ngg;
+		shader.key.opt.vs_as_prim_discard_cs = key->vs_prolog.as_prim_discard_cs;
 		break;
 	case PIPE_SHADER_TESS_CTRL:
 		assert(!prolog);
@@ -2110,7 +2118,8 @@ si_get_shader_part(struct si_screen *sscreen,
 	struct si_shader_context ctx;
 	si_llvm_context_init(&ctx, sscreen, compiler,
 			     si_get_wave_size(sscreen, type, shader.key.as_ngg,
-					      shader.key.as_es));
+					      shader.key.as_es,
+					      shader.key.opt.vs_as_prim_discard_cs));
 	ctx.shader = &shader;
 	ctx.type = type;
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 1ee9aba9faf..ef571a5d684 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -566,6 +566,7 @@ union si_shader_part_key {
 		unsigned	as_ls:1;
 		unsigned	as_es:1;
 		unsigned	as_ngg:1;
+		unsigned	as_prim_discard_cs:1;
 		unsigned	has_ngg_cull_inputs:1; /* from the NGG cull shader */
 		unsigned	gs_fast_launch_tri_list:1; /* for NGG culling */
 		unsigned	gs_fast_launch_tri_strip:1; /* for NGG culling */
diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
index ac734d0924b..99ffdd2e980 100644
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
@@ -523,7 +523,8 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
 	shader->is_gs_copy_shader = true;
 
 	si_llvm_context_init(&ctx, sscreen, compiler,
-			     si_get_wave_size(sscreen, PIPE_SHADER_VERTEX, false, false));
+			     si_get_wave_size(sscreen, PIPE_SHADER_VERTEX,
+					      false, false, false));
 	ctx.shader = shader;
 	ctx.type = PIPE_SHADER_VERTEX;
 
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index ce7e791be3d..be7cda1d332 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -71,7 +71,7 @@ void si_get_ir_cache_key(struct si_shader_selector *sel, bool ngg, bool es,
 		shader_variant_flags |= 1 << 0;
 	if (sel->nir)
 		shader_variant_flags |= 1 << 1;
-	if (si_get_wave_size(sel->screen, sel->type, ngg, es) == 32)
+	if (si_get_wave_size(sel->screen, sel->type, ngg, es, false) == 32)
 		shader_variant_flags |= 1 << 2;
 	if (sel->type == PIPE_SHADER_FRAGMENT &&
 	    sel->info.uses_derivatives &&