intel: Move unlit centroid workaround into the elk compiler

This was only needed on Sandybridge.  We can delete the brw code,
and replace the generic devinfo bit with a helper inside the elk
compiler itself.

Thanks to Iván Briano for noticing we still had dead brw code for this.

Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33764>

src/intel/compiler/brw_compile_fs.cpp

@@ -504,8 +504,6 @@ brw_emit_interpolation_setup(brw_shader &s)
    }
 
    if (wm_key->persample_interp == INTEL_SOMETIMES) {
-      assert(!devinfo->needs_unlit_centroid_workaround);
-
       const brw_builder ubld = bld.exec_all().group(16, 0);
       bool loaded_flag = false;
 
@@ -556,42 +554,6 @@ brw_emit_interpolation_setup(brw_shader &s)
       s.delta_xy[i] = brw_fetch_barycentric_reg(
          bld, payload.barycentric_coord_reg[i]);
    }
-
-   uint32_t centroid_modes = wm_prog_data->barycentric_interp_modes &
-      (1 << INTEL_BARYCENTRIC_PERSPECTIVE_CENTROID |
-       1 << INTEL_BARYCENTRIC_NONPERSPECTIVE_CENTROID);
-
-   if (devinfo->needs_unlit_centroid_workaround && centroid_modes) {
-      /* Get the pixel/sample mask into f0 so that we know which
-       * pixels are lit.  Then, for each channel that is unlit,
-       * replace the centroid data with non-centroid data.
-       */
-      for (unsigned i = 0; i < DIV_ROUND_UP(s.dispatch_width, 16); i++) {
-         bld.exec_all().group(1, 0)
-            .MOV(retype(brw_flag_reg(0, i), BRW_TYPE_UW),
-                 retype(brw_vec1_grf(1 + i, 7), BRW_TYPE_UW));
-      }
-
-      for (int i = 0; i < INTEL_BARYCENTRIC_MODE_COUNT; ++i) {
-         if (!(centroid_modes & (1 << i)))
-            continue;
-
-         const brw_reg centroid_delta_xy = s.delta_xy[i];
-         const brw_reg &pixel_delta_xy = s.delta_xy[i - 1];
-
-         s.delta_xy[i] = bld.vgrf(BRW_TYPE_F, 2);
-
-         for (unsigned c = 0; c < 2; c++) {
-            for (unsigned q = 0; q < s.dispatch_width / 8; q++) {
-               set_predicate(BRW_PREDICATE_NORMAL,
-                  bld.quarter(q).SEL(
-                     quarter(offset(s.delta_xy[i], bld, c), q),
-                     quarter(offset(centroid_delta_xy, bld, c), q),
-                     quarter(offset(pixel_delta_xy, bld, c), q)));
-            }
-         }
-      }
-   }
 }
 
 
@@ -664,17 +626,6 @@ brw_emit_repclear_shader(brw_shader &s)
    brw_lower_scoreboard(s);
 }
 
-/**
- * Turn one of the two CENTROID barycentric modes into PIXEL mode.
- */
-static enum intel_barycentric_mode
-centroid_to_pixel(enum intel_barycentric_mode bary)
-{
-   assert(bary == INTEL_BARYCENTRIC_PERSPECTIVE_CENTROID ||
-          bary == INTEL_BARYCENTRIC_NONPERSPECTIVE_CENTROID);
-   return (enum intel_barycentric_mode) ((unsigned) bary - 1);
-}
-
 static void
 calculate_urb_setup(const struct intel_device_info *devinfo,
                     const struct brw_wm_prog_key *key,
@@ -976,15 +927,10 @@ brw_compute_barycentric_interp_modes(const struct intel_device_info *devinfo,
             if (!is_used_in_not_interp_frag_coord(&intrin->def))
                continue;
 
-            nir_intrinsic_op bary_op = intrin->intrinsic;
             enum intel_barycentric_mode bary =
                brw_barycentric_mode(key, intrin);
 
             barycentric_interp_modes |= 1 << bary;
-
-            if (devinfo->needs_unlit_centroid_workaround &&
-                bary_op == nir_intrinsic_load_barycentric_centroid)
-               barycentric_interp_modes |= 1 << centroid_to_pixel(bary);
          }
       }
    }

src/intel/compiler/elk/elk_fs.cpp

@@ -6415,7 +6415,7 @@ elk_compute_barycentric_interp_modes(const struct intel_device_info *devinfo,
 
             barycentric_interp_modes |= 1 << bary;
 
-            if (devinfo->needs_unlit_centroid_workaround &&
+            if (elk_needs_unlit_centroid_workaround(devinfo) &&
                 bary_op == nir_intrinsic_load_barycentric_centroid)
                barycentric_interp_modes |= 1 << centroid_to_pixel(bary);
          }

src/intel/compiler/elk/elk_fs_visitor.cpp

@@ -31,6 +31,7 @@
 #include "elk_fs.h"
 #include "elk_fs_builder.h"
 #include "elk_nir.h"
+#include "elk_private.h"
 #include "compiler/glsl_types.h"
 
 using namespace elk;
@@ -289,7 +290,7 @@ elk_fs_visitor::emit_interpolation_setup_gfx6()
    }
 
    if (wm_key->persample_interp == ELK_SOMETIMES) {
-      assert(!devinfo->needs_unlit_centroid_workaround);
+      assert(!elk_needs_unlit_centroid_workaround(devinfo));
 
       const fs_builder ubld = bld.exec_all().group(16, 0);
       bool loaded_flag = false;
@@ -346,7 +347,7 @@ elk_fs_visitor::emit_interpolation_setup_gfx6()
       (1 << ELK_BARYCENTRIC_PERSPECTIVE_CENTROID |
        1 << ELK_BARYCENTRIC_NONPERSPECTIVE_CENTROID);
 
-   if (devinfo->needs_unlit_centroid_workaround && centroid_modes) {
+   if (elk_needs_unlit_centroid_workaround(devinfo) && centroid_modes) {
       /* Get the pixel/sample mask into f0 so that we know which
        * pixels are lit.  Then, for each channel that is unlit,
        * replace the centroid data with non-centroid data.

src/intel/compiler/elk/elk_private.h

@@ -60,6 +60,18 @@ inline bool elk_simd_any_compiled(const elk_simd_selection_state &state)
    return elk_simd_first_compiled(state) >= 0;
 }
 
+inline bool
+elk_needs_unlit_centroid_workaround(const struct intel_device_info *devinfo)
+{
+   /* Sandybridge doesn't do centroid interpolation correctly on unlit pixels,
+    * causing incorrect values for derivatives near triangle edges.  Enabling
+    * this flag causes the fragment shader to use non-centroid interpolation
+    * for unlit pixels, at the expense of two extra fragment shader
+    * instructions."
+    */
+   return devinfo->ver == 6;
+}
+
 bool elk_simd_should_compile(elk_simd_selection_state &state, unsigned simd);
 
 void elk_simd_mark_compiled(elk_simd_selection_state &state, unsigned simd, bool spilled);

src/intel/dev/intel_device_info.c

@@ -172,7 +172,6 @@ static const struct intel_device_info intel_device_info_snb_gt1 = {
    .has_hiz_and_separate_stencil = true,
    .has_llc = true,
    .has_pln = true,
-   .needs_unlit_centroid_workaround = true,
    .num_slices = 1,
    .num_subslices = { 1, },
    .max_eus_per_subslice = 6,
@@ -202,7 +201,6 @@ static const struct intel_device_info intel_device_info_snb_gt2 = {
    .has_hiz_and_separate_stencil = true,
    .has_llc = true,
    .has_pln = true,
-   .needs_unlit_centroid_workaround = true,
    .num_slices = 1,
    .num_subslices = { 1, },
    .max_eus_per_subslice = 12,

src/intel/dev/intel_device_info.py

@@ -310,14 +310,6 @@ Struct("intel_device_info",
 
         Member("bool", "has_compute_engine", comment="Whether this platform has compute engine"),
 
-        Member("bool", "needs_unlit_centroid_workaround", compiler_field=True,
-               comment=dedent("""\
-               Some versions of Gen hardware don't do centroid interpolation correctly
-               on unlit pixels, causing incorrect values for derivatives near triangle
-               edges.  Enabling this flag causes the fragment shader to use
-               non-centroid interpolation for unlit pixels, at the expense of two extra
-               fragment shader instructions.""")),
-
         Member("bool", "needs_null_push_constant_tbimr_workaround",
                comment=dedent("""\
                Whether the platform needs an undocumented workaround for a hardware bug