v3d: clamp gl_PointSize to a minimum of 1.0

[mesa.git] / src / broadcom / compiler / vir.c

diff --git a/src/broadcom/compiler/vir.c b/src/broadcom/compiler/vir.c
index b785b53c62c9cf72a7f1cb07467fc0b40ba6b04c..78362a2949ce8382bc9c3418d748a03c00aa4b57 100644 (file)
--- a/src/broadcom/compiler/vir.c
+++ b/src/broadcom/compiler/vir.c
@@ -75,6 +75,8 @@ vir_has_side_effects(struct v3d_compile *c, struct qinst *inst)
 
         if (inst->qpu.sig.ldtmu ||
             inst->qpu.sig.ldvary ||
+            inst->qpu.sig.ldtlbu ||
+            inst->qpu.sig.ldtlb ||
             inst->qpu.sig.wrtmuc ||
             inst->qpu.sig.thrsw) {
                 return true;
@@ -562,8 +564,29 @@ v3d_lower_nir(struct v3d_compile *c)
                 }
         }
 
+        /* CS textures may not have return_size reflecting the shadow state. */
+        nir_foreach_variable(var, &c->s->uniforms) {
+                const struct glsl_type *type = glsl_without_array(var->type);
+                unsigned array_len = MAX2(glsl_get_length(var->type), 1);
+
+                if (!glsl_type_is_sampler(type) ||
+                    !glsl_sampler_type_is_shadow(type))
+                        continue;
+
+                for (int i = 0; i < array_len; i++) {
+                        tex_options.lower_tex_packing[var->data.binding + i] =
+                                nir_lower_tex_packing_16;
+                }
+        }
+
         NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
         NIR_PASS_V(c->s, nir_lower_system_values);
+
+        NIR_PASS_V(c->s, nir_lower_vars_to_scratch,
+                   nir_var_function_temp,
+                   0,
+                   glsl_get_natural_size_align_bytes);
+        NIR_PASS_V(c->s, v3d_nir_lower_scratch);
 }
 
 static void
@@ -667,6 +690,17 @@ v3d_fs_set_prog_data(struct v3d_compile *c,
         prog_data->writes_z = c->writes_z;
         prog_data->disable_ez = !c->s->info.fs.early_fragment_tests;
         prog_data->uses_center_w = c->uses_center_w;
+        prog_data->uses_implicit_point_line_varyings =
+                c->uses_implicit_point_line_varyings;
+        prog_data->lock_scoreboard_on_first_thrsw =
+                c->lock_scoreboard_on_first_thrsw;
+}
+
+static void
+v3d_cs_set_prog_data(struct v3d_compile *c,
+                     struct v3d_compute_prog_data *prog_data)
+{
+        prog_data->shared_size = c->s->info.cs.shared_size;
 }
 
 static void
@@ -679,7 +713,9 @@ v3d_set_prog_data(struct v3d_compile *c,
 
         v3d_set_prog_data_uniforms(c, prog_data);
 
-        if (c->s->info.stage == MESA_SHADER_VERTEX) {
+        if (c->s->info.stage == MESA_SHADER_COMPUTE) {
+                v3d_cs_set_prog_data(c, (struct v3d_compute_prog_data *)prog_data);
+        } else if (c->s->info.stage == MESA_SHADER_VERTEX) {
                 v3d_vs_set_prog_data(c, (struct v3d_vs_prog_data *)prog_data);
         } else {
                 assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
@@ -722,9 +758,16 @@ v3d_nir_lower_vs_early(struct v3d_compile *c)
         NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
         v3d_optimize_nir(c->s);
         NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in);
+
+        /* This must go before nir_lower_io */
+        if (c->vs_key->per_vertex_point_size)
+                NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
+
         NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
                    type_size_vec4,
                    (nir_lower_io_options)0);
+        /* clean up nir_lower_io's deref_var remains */
+        NIR_PASS_V(c->s, nir_opt_dce);
 }
 
 static void
@@ -763,6 +806,8 @@ v3d_nir_lower_fs_early(struct v3d_compile *c)
         if (c->fs_key->int_color_rb || c->fs_key->uint_color_rb)
                 v3d_fixup_fs_output_types(c);
 
+        NIR_PASS_V(c->s, v3d_nir_lower_logic_ops, c);
+
         /* If the shader has no non-TLB side effects, we can promote it to
          * enabling early_fragment_tests even if the user didn't.
          */
@@ -865,13 +910,17 @@ uint64_t *v3d_compile(const struct v3d_compiler *compiler,
                 c->fs_key = (struct v3d_fs_key *)key;
                 prog_data = rzalloc_size(NULL, sizeof(struct v3d_fs_prog_data));
                 break;
+        case MESA_SHADER_COMPUTE:
+                prog_data = rzalloc_size(NULL,
+                                         sizeof(struct v3d_compute_prog_data));
+                break;
         default:
                 unreachable("unsupported shader stage");
         }
 
         if (c->s->info.stage == MESA_SHADER_VERTEX) {
                 v3d_nir_lower_vs_early(c);
-        } else {
+        } else if (c->s->info.stage != MESA_SHADER_COMPUTE) {
                 assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
                 v3d_nir_lower_fs_early(c);
         }
@@ -880,7 +929,7 @@ uint64_t *v3d_compile(const struct v3d_compiler *compiler,
 
         if (c->s->info.stage == MESA_SHADER_VERTEX) {
                 v3d_nir_lower_vs_late(c);
-        } else {
+        } else if (c->s->info.stage != MESA_SHADER_COMPUTE)  {
                 assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
                 v3d_nir_lower_fs_late(c);
         }
@@ -903,7 +952,8 @@ uint64_t *v3d_compile(const struct v3d_compiler *compiler,
         char *shaderdb;
         int ret = asprintf(&shaderdb,
                            "%s shader: %d inst, %d threads, %d loops, "
-                           "%d uniforms, %d max-temps, %d:%d spills:fills",
+                           "%d uniforms, %d max-temps, %d:%d spills:fills, "
+                           "%d sfu-stalls, %d inst-and-stalls",
                            vir_get_stage_name(c),
                            c->qpu_inst_count,
                            c->threads,
@@ -911,7 +961,9 @@ uint64_t *v3d_compile(const struct v3d_compiler *compiler,
                            c->num_uniforms,
                            vir_get_max_temps(c),
                            c->spills,
-                           c->fills);
+                           c->fills,
+                           c->qpu_inst_stalled_count,
+                           c->qpu_inst_count + c->qpu_inst_stalled_count);
         if (ret >= 0) {
                 if (V3D_DEBUG & V3D_DEBUG_SHADERDB)
                         fprintf(stderr, "SHADER-DB: %s\n", shaderdb);