vc4: Convert to using nir_lower_io_scalar for FS inputs.

[mesa.git] / src / gallium / drivers / vc4 / vc4_nir_lower_io.c

diff --git a/src/gallium/drivers/vc4/vc4_nir_lower_io.c b/src/gallium/drivers/vc4/vc4_nir_lower_io.c
index 85f31f34131da45d2425f3ca5c7e2216449bdf4a..d2b27bd7c8a7d169461b6bc3cb76ef9617548ae0 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_nir_lower_io.c
+++ b/src/gallium/drivers/vc4/vc4_nir_lower_io.c
@@ -29,9 +29,10 @@
  * Walks the NIR generated by TGSI-to-NIR to lower its io intrinsics into
  * something amenable to the VC4 architecture.
  *
- * Currently, it split inputs, outputs, and uniforms into scalars, drops any
- * non-position outputs in coordinate shaders, and fixes up the addressing on
- * indirect uniform loads.
+ * Currently, it splits outputs, VS inputs, and uniforms into scalars, drops
+ * any non-position outputs in coordinate shaders, and fixes up the addressing
+ * on indirect uniform loads.  FS input scalarization is handled by
+ * nir_lower_io_to_scalar().
  */
 
 static void
@@ -228,11 +229,22 @@ vc4_nir_lower_vertex_attr(struct vc4_compile *c, nir_builder *b,
         replace_intrinsic_with_vec4(b, intr, dests);
 }
 
+static bool
+is_point_sprite(struct vc4_compile *c, nir_variable *var)
+{
+        if (var->data.location < VARYING_SLOT_VAR0 ||
+            var->data.location > VARYING_SLOT_VAR31)
+                return false;
+
+        return (c->fs_key->point_sprite_mask &
+                (1 << (var->data.location - VARYING_SLOT_VAR0)));
+}
+
 static void
 vc4_nir_lower_fs_input(struct vc4_compile *c, nir_builder *b,
                        nir_intrinsic_instr *intr)
 {
-        b->cursor = nir_before_instr(&intr->instr);
+        b->cursor = nir_after_instr(&intr->instr);
 
         if (nir_intrinsic_base(intr) >= VC4_NIR_TLB_COLOR_READ_INPUT &&
             nir_intrinsic_base(intr) < (VC4_NIR_TLB_COLOR_READ_INPUT +
@@ -250,50 +262,42 @@ vc4_nir_lower_fs_input(struct vc4_compile *c, nir_builder *b,
         }
         assert(input_var);
 
-        /* All TGSI-to-NIR inputs are vec4. */
-        assert(intr->num_components == 4);
+        int comp = nir_intrinsic_component(intr);
 
-        /* We only accept direct inputs and TGSI only ever gives them to us
-         * with an offset value of 0.
-         */
-        assert(nir_src_as_const_value(intr->src[0]) &&
-               nir_src_as_const_value(intr->src[0])->u32[0] == 0);
+        /* Lower away point coordinates, and fix up PNTC. */
+        if (is_point_sprite(c, input_var) ||
+            input_var->data.location == VARYING_SLOT_PNTC) {
+                assert(intr->num_components == 1);
 
-        /* Generate scalar loads equivalent to the original VEC4. */
-        nir_ssa_def *dests[4];
-        for (unsigned i = 0; i < intr->num_components; i++) {
-                nir_intrinsic_instr *intr_comp =
-                        nir_intrinsic_instr_create(c->s, nir_intrinsic_load_input);
-                intr_comp->num_components = 1;
-                nir_intrinsic_set_base(intr_comp,
-                                       nir_intrinsic_base(intr) * 4 + i);
-                intr_comp->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
+                nir_ssa_def *result = &intr->dest.ssa;
 
-                nir_ssa_dest_init(&intr_comp->instr, &intr_comp->dest, 1, 32, NULL);
-                nir_builder_instr_insert(b, &intr_comp->instr);
+                switch (comp) {
+                case 0:
+                case 1:
+                        /* If we're not rendering points, we need to set a
+                         * defined value for the input that would come from
+                         * PNTC.
+                         */
+                        if (!c->fs_key->is_points)
+                                result = nir_imm_float(b, 0.0);
+                        break;
+                case 2:
+                        result = nir_imm_float(b, 0.0);
+                        break;
+                case 3:
+                        result = nir_imm_float(b, 1.0);
+                        break;
+                }
 
-                dests[i] = &intr_comp->dest.ssa;
-        }
+                if (c->fs_key->point_coord_upper_left && comp == 1)
+                        result = nir_fsub(b, nir_imm_float(b, 1.0), result);
 
-        if (input_var->data.location >= VARYING_SLOT_VAR0) {
-                if (c->fs_key->point_sprite_mask &
-                    (1 << (input_var->data.location -
-                           VARYING_SLOT_VAR0))) {
-                        if (!c->fs_key->is_points) {
-                                dests[0] = nir_imm_float(b, 0.0);
-                                dests[1] = nir_imm_float(b, 0.0);
-                        }
-                        if (c->fs_key->point_coord_upper_left) {
-                                dests[1] = nir_fsub(b,
-                                                    nir_imm_float(b, 1.0),
-                                                    dests[1]);
-                        }
-                        dests[2] = nir_imm_float(b, 0.0);
-                        dests[3] = nir_imm_float(b, 1.0);
+                if (result != &intr->dest.ssa) {
+                        nir_ssa_def_rewrite_uses_after(&intr->dest.ssa,
+                                                       nir_src_for_ssa(result),
+                                                       result->parent_instr);
                 }
         }
-
-        replace_intrinsic_with_vec4(b, intr, dests);
 }
 
 static void