Mesa (master): vc4: Move some FS input lowering into NIR.

[Eric Anholt]

Module: Mesa
Branch: master
Commit: 5a8c57b52287ba2bb8faa4447e7d1cc46ef1a3d4
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=5a8c57b52287ba2bb8faa4447e7d1cc46ef1a3d4

Author: Eric Anholt <eric at anholt.net>
Date:   Wed Jul 29 17:27:54 2015 -0700

vc4: Move some FS input lowering into NIR.

---

 src/gallium/drivers/vc4/vc4_nir_lower_io.c |   47 ++++++++++++++++++++++++++++
 src/gallium/drivers/vc4/vc4_program.c      |   38 ++--------------------
 2 files changed, 50 insertions(+), 35 deletions(-)

diff --git a/src/gallium/drivers/vc4/vc4_nir_lower_io.c b/src/gallium/drivers/vc4/vc4_nir_lower_io.c
index 9882b6b..fa06c89 100644
--- a/src/gallium/drivers/vc4/vc4_nir_lower_io.c
+++ b/src/gallium/drivers/vc4/vc4_nir_lower_io.c
@@ -42,6 +42,17 @@ vc4_nir_lower_input(struct vc4_compile *c, nir_builder *b,
 
         nir_builder_insert_before_instr(b, &intr->instr);
 
+        nir_variable *input_var = NULL;
+        foreach_list_typed(nir_variable, var, node, &c->s->inputs) {
+                if (var->data.driver_location == intr->const_index[0]) {
+                        input_var = var;
+                        break;
+                }
+        }
+        assert(input_var);
+        int semantic_name = input_var->data.location;
+        int semantic_index = input_var->data.index;
+
         /* Generate scalar loads equivalent to the original VEC4. */
         nir_ssa_def *dests[4];
         for (unsigned i = 0; i < intr->num_components; i++) {
@@ -55,6 +66,42 @@ vc4_nir_lower_input(struct vc4_compile *c, nir_builder *b,
                 dests[i] = &intr_comp->dest.ssa;
         }
 
+        switch (c->stage) {
+        case QSTAGE_FRAG:
+                switch (semantic_name) {
+                case TGSI_SEMANTIC_FACE:
+                        dests[0] = nir_fsub(b,
+                                            nir_imm_float(b, 1.0),
+                                            nir_fmul(b,
+                                                     nir_i2f(b, dests[0]),
+                                                     nir_imm_float(b, 2.0)));
+                        dests[1] = nir_imm_float(b, 0.0);
+                        dests[2] = nir_imm_float(b, 0.0);
+                        dests[3] = nir_imm_float(b, 1.0);
+                        break;
+                case TGSI_SEMANTIC_GENERIC:
+                        if (c->fs_key->point_sprite_mask &
+                            (1 << semantic_index)) {
+                                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);
+                        }
+                        break;
+                }
+                break;
+        case QSTAGE_COORD:
+        case QSTAGE_VERT:
+                break;
+        }
+
         /* Batch things back together into a vec4.  This will get split by the
          * later ALU scalarization pass.
          */
diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
index b2efd68..ddc9970 100644
--- a/src/gallium/drivers/vc4/vc4_program.c
+++ b/src/gallium/drivers/vc4/vc4_program.c
@@ -707,26 +707,6 @@ emit_fragcoord_input(struct vc4_compile *c, int attr)
         c->inputs[attr * 4 + 3] = qir_RCP(c, qir_FRAG_W(c));
 }
 
-static void
-emit_point_coord_input(struct vc4_compile *c, int attr)
-{
-        if (c->point_x.file == QFILE_NULL) {
-                c->point_x = qir_uniform_f(c, 0.0);
-                c->point_y = qir_uniform_f(c, 0.0);
-        }
-
-        c->inputs[attr * 4 + 0] = c->point_x;
-        if (c->fs_key->point_coord_upper_left) {
-                c->inputs[attr * 4 + 1] = qir_FSUB(c,
-                                                   qir_uniform_f(c, 1.0),
-                                                   c->point_y);
-        } else {
-                c->inputs[attr * 4 + 1] = c->point_y;
-        }
-        c->inputs[attr * 4 + 2] = qir_uniform_f(c, 0.0);
-        c->inputs[attr * 4 + 3] = qir_uniform_f(c, 1.0);
-}
-
 static struct qreg
 emit_fragment_varying(struct vc4_compile *c, uint8_t semantic,
                       uint8_t index, uint8_t swizzle)
@@ -768,19 +748,6 @@ emit_fragment_input(struct vc4_compile *c, int attr,
 }
 
 static void
-emit_face_input(struct vc4_compile *c, int attr)
-{
-        c->inputs[attr * 4 + 0] = qir_FSUB(c,
-                                           qir_uniform_f(c, 1.0),
-                                           qir_FMUL(c,
-                                                    qir_ITOF(c, qir_FRAG_REV_FLAG(c)),
-                                                    qir_uniform_f(c, 2.0)));
-        c->inputs[attr * 4 + 1] = qir_uniform_f(c, 0.0);
-        c->inputs[attr * 4 + 2] = qir_uniform_f(c, 0.0);
-        c->inputs[attr * 4 + 3] = qir_uniform_f(c, 1.0);
-}
-
-static void
 add_output(struct vc4_compile *c,
            uint32_t decl_offset,
            uint8_t semantic_name,
@@ -1707,11 +1674,12 @@ ntq_setup_inputs(struct vc4_compile *c)
                         if (semantic_name == TGSI_SEMANTIC_POSITION) {
                                 emit_fragcoord_input(c, loc);
                         } else if (semantic_name == TGSI_SEMANTIC_FACE) {
-                                emit_face_input(c, loc);
+                                c->inputs[loc * 4 + 0] = qir_FRAG_REV_FLAG(c);
                         } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
                                    (c->fs_key->point_sprite_mask &
                                     (1 << semantic_index))) {
-                                emit_point_coord_input(c, loc);
+                                c->inputs[loc * 4 + 0] = c->point_x;
+                                c->inputs[loc * 4 + 1] = c->point_y;
                         } else {
                                 emit_fragment_input(c, loc,
                                                     semantic_name,