freedreno/ir3: add support for a650 tess shared storage

A650 uses LDL/STL, and the "local_primitive_id" in tess ctrl shader comes
from bits 16-21 in the header instead of 0-5.

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5764>

src/freedreno/ir3/ir3_compiler.c

@@ -93,6 +93,9 @@ ir3_compiler_create(struct fd_device *dev, uint32_t gpu_id)
 		 * TODO: is this true on earlier gen's?
 		 */
 		compiler->max_const_compute = 256;
+
+		if (compiler->gpu_id == 650)
+			compiler->tess_use_shared = true;
 	} else {
 		compiler->max_const_pipeline = 512;
 		compiler->max_const_geom = 512;

src/freedreno/ir3/ir3_compiler.h

@@ -74,6 +74,9 @@ struct ir3_compiler {
 	 */
 	bool samgq_workaround;
 
+	/* on a650, vertex shader <-> tess control io uses LDL/STL */
+	bool tess_use_shared;
+
 	/* The maximum number of constants, in vec4's, across the entire graphics
 	 * pipeline.
 	 */

src/freedreno/ir3/ir3_compiler_nir.c

@@ -928,6 +928,10 @@ emit_intrinsic_load_shared_ir3(struct ir3_context *ctx, nir_intrinsic_instr *int
 			create_immed(b, intr->num_components), 0,
 			create_immed(b, base), 0);
 
+	/* for a650, use LDL for tess ctrl inputs: */
+	if (ctx->so->type == MESA_SHADER_TESS_CTRL && ctx->compiler->tess_use_shared)
+		load->opc = OPC_LDL;
+
 	load->cat6.type = utype_dst(intr->dest);
 	load->regs[0]->wrmask = MASK(intr->num_components);
 
@@ -952,6 +956,11 @@ emit_intrinsic_store_shared_ir3(struct ir3_context *ctx, nir_intrinsic_instr *in
 		ir3_create_collect(ctx, value, intr->num_components), 0,
 		create_immed(b, intr->num_components), 0);
 
+	/* for a650, use STL for vertex outputs used by tess ctrl shader: */
+	if (ctx->so->type == MESA_SHADER_VERTEX && ctx->so->key.tessellation &&
+		ctx->compiler->tess_use_shared)
+		store->opc = OPC_STL;
+
 	store->cat6.dst_offset = nir_intrinsic_base(intr);
 	store->cat6.type = utype_src(intr->src[0]);
 	store->barrier_class = IR3_BARRIER_SHARED_W;

src/freedreno/ir3/ir3_nir.c

@@ -379,7 +379,7 @@ ir3_nir_lower_variant(struct ir3_shader_variant *so, nir_shader *s)
 			break;
 		case MESA_SHADER_TESS_CTRL:
 			NIR_PASS_V(s, ir3_nir_lower_tess_ctrl, so, so->key.tessellation);
-			NIR_PASS_V(s, ir3_nir_lower_to_explicit_input);
+			NIR_PASS_V(s, ir3_nir_lower_to_explicit_input, so->shader->compiler);
 			progress = true;
 			break;
 		case MESA_SHADER_TESS_EVAL:
@@ -389,7 +389,7 @@ ir3_nir_lower_variant(struct ir3_shader_variant *so, nir_shader *s)
 			progress = true;
 			break;
 		case MESA_SHADER_GEOMETRY:
-			NIR_PASS_V(s, ir3_nir_lower_to_explicit_input);
+			NIR_PASS_V(s, ir3_nir_lower_to_explicit_input, so->shader->compiler);
 			progress = true;
 			break;
 		default:

src/freedreno/ir3/ir3_nir.h

@@ -46,7 +46,7 @@ bool ir3_nir_lower_tex_prefetch(nir_shader *shader);
 
 void ir3_nir_lower_to_explicit_output(nir_shader *shader,
 		struct ir3_shader_variant *v, unsigned topology);
-void ir3_nir_lower_to_explicit_input(nir_shader *shader);
+void ir3_nir_lower_to_explicit_input(nir_shader *shader, struct ir3_compiler *compiler);
 void ir3_nir_lower_tess_ctrl(nir_shader *shader, struct ir3_shader_variant *v, unsigned topology);
 void ir3_nir_lower_tess_eval(nir_shader *shader, unsigned topology);
 void ir3_nir_lower_gs(nir_shader *shader);

src/freedreno/ir3/ir3_nir_lower_tess.c

@@ -42,6 +42,9 @@ struct state {
 
 	struct exec_list old_outputs;
 	struct exec_list emit_outputs;
+
+	/* tess ctrl shader on a650 gets the local primitive id at different bits: */
+	bool local_primitive_id_start;
 };
 
 static nir_ssa_def *
@@ -66,7 +69,7 @@ build_vertex_id(nir_builder *b, struct state *state)
 static nir_ssa_def *
 build_local_primitive_id(nir_builder *b, struct state *state)
 {
-	return bitfield_extract(b, state->header, 0, 63);
+	return bitfield_extract(b, state->header, state->local_primitive_id_start, 63);
 }
 
 static nir_variable *
@@ -301,10 +304,16 @@ lower_block_to_explicit_input(nir_block *block, nir_builder *b, struct state *st
 }
 
 void
-ir3_nir_lower_to_explicit_input(nir_shader *shader)
+ir3_nir_lower_to_explicit_input(nir_shader *shader, struct ir3_compiler *compiler)
 {
  	struct state state = { };
 
+	/* when using stl/ldl (instead of stlw/ldlw) for linking VS and HS,
+	 * HS uses a different primitive id, which starts at bit 16 in the header
+	 */
+	if (shader->info.stage == MESA_SHADER_TESS_CTRL && compiler->tess_use_shared)
+		state.local_primitive_id_start = 16;
+
 	nir_function_impl *impl = nir_shader_get_entrypoint(shader);
 	assert(impl);