asahi,agx: use hw clip distance

rather than lowering in the fragment shader. this is what we're "supposed" to
do, what apple does, and it saves a byte of the fs key

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>

src/asahi/compiler/agx_compile.c

@@ -449,14 +449,18 @@ agx_emit_store_vary(agx_builder *b, nir_intrinsic_instr *instr)
 
    unsigned imm_index = b->shader->out->varyings.vs.slots[sem.location];
 
-   if (sem.location == VARYING_SLOT_LAYER) {
+   if (sem.location == VARYING_SLOT_LAYER ||
+       sem.location == VARYING_SLOT_CLIP_DIST0) {
       /* Separate slots used for the sysval vs the varying. The default slot
        * above is for the varying. Change for the sysval.
        */
       assert(sem.no_sysval_output || sem.no_varying);
 
-      if (sem.no_varying)
-         imm_index = b->shader->out->varyings.vs.layer_viewport_slot;
+      if (sem.no_varying) {
+         imm_index = sem.location == VARYING_SLOT_LAYER
+                        ? b->shader->out->varyings.vs.layer_viewport_slot
+                        : b->shader->out->varyings.vs.clip_dist_slot;
+      }
    }
 
    assert(imm_index < ~0);
@@ -2589,6 +2593,12 @@ agx_remap_varyings_vs(nir_shader *nir, struct agx_varyings_vs *varyings,
       base += 1;
    }
 
+   if (nir->info.outputs_written & VARYING_BIT_CLIP_DIST0) {
+      varyings->clip_dist_slot = base;
+      varyings->nr_clip_dists = nir->info.clip_distance_array_size;
+      base += varyings->nr_clip_dists;
+   }
+
    /* All varyings linked now */
    varyings->nr_index = base;
 }
@@ -3071,6 +3081,10 @@ agx_compile_shader_nir(nir_shader *nir, struct agx_shader_key *key,
    if (nir->info.stage == MESA_SHADER_FRAGMENT)
       out->tag_write_disable = !nir->info.writes_memory;
 
+   if (nir->info.stage == MESA_SHADER_VERTEX &&
+       (nir->info.outputs_written & VARYING_BIT_CLIP_DIST0))
+      NIR_PASS(_, nir, agx_nir_lower_clip_distance);
+
    bool needs_libagx = nir->info.stage == MESA_SHADER_GEOMETRY;
 
    /* Late tilebuffer lowering creates multisampled image stores */

src/asahi/compiler/agx_compile.h

@@ -46,6 +46,12 @@ struct agx_varyings_vs {
     * is written. What's at slots[VARYING_SLOT_LAYER] is the varying output.
     */
    unsigned layer_viewport_slot;
+
+   /* Base slot for the clip distance sysval outputs, or ~0 if none is written.
+    * What's at slots[VARYING_SLOT_CLIP_DIST0] is the varying output.
+    */
+   unsigned clip_dist_slot;
+   unsigned nr_clip_dists;
 };
 
 /* Conservative bound, * 4 due to offsets (TODO: maybe worth eliminating

src/asahi/compiler/agx_nir.h

@@ -16,6 +16,7 @@ bool agx_nir_fuse_algebraic_late(struct nir_shader *shader);
 bool agx_nir_fence_images(struct nir_shader *shader);
 bool agx_nir_lower_multisampled_image_store(struct nir_shader *s);
 bool agx_nir_lower_layer(struct nir_shader *s);
+bool agx_nir_lower_clip_distance(struct nir_shader *s);
 bool agx_nir_lower_cull_distance_vs(struct nir_shader *s);
 bool agx_nir_lower_subgroups(struct nir_shader *s);
 

src/asahi/compiler/agx_nir_lower_clip_distance.c

@@ -0,0 +1,42 @@
+/*
+ * Copyright 2023 Valve Corporation
+ * SPDX-License-Identifier: MIT
+ */
+
+#include "compiler/nir/nir.h"
+#include "compiler/nir/nir_builder.h"
+#include "agx_nir.h"
+
+static bool
+lower(nir_builder *b, nir_intrinsic_instr *intr, UNUSED void *data)
+{
+   if (intr->intrinsic != nir_intrinsic_store_output)
+      return false;
+
+   nir_io_semantics sem = nir_intrinsic_io_semantics(intr);
+   if (sem.location != VARYING_SLOT_CLIP_DIST0)
+      return false;
+
+   nir_instr *clone = nir_instr_clone(b->shader, &intr->instr);
+   nir_intrinsic_instr *lowered = nir_instr_as_intrinsic(clone);
+
+   b->cursor = nir_after_instr(&intr->instr);
+   nir_builder_instr_insert(b, clone);
+
+   nir_io_semantics new_sem = sem;
+   new_sem.no_varying = true;
+   nir_intrinsic_set_io_semantics(lowered, new_sem);
+
+   sem.no_sysval_output = true;
+   nir_intrinsic_set_io_semantics(intr, sem);
+   return true;
+}
+
+bool
+agx_nir_lower_clip_distance(nir_shader *s)
+{
+   assert(s->info.outputs_written & VARYING_BIT_CLIP_DIST0);
+
+   return nir_shader_intrinsics_pass(
+      s, lower, nir_metadata_block_index | nir_metadata_dominance, NULL);
+}

src/asahi/compiler/meson.build

@@ -9,6 +9,7 @@ libasahi_agx_files = files(
   'agx_insert_waits.c',
   'agx_nir_fence_images.c',
   'agx_nir_lower_address.c',
+  'agx_nir_lower_clip_distance.c',
   'agx_nir_lower_cull_distance.c',
   'agx_nir_lower_frag_sidefx.c',
   'agx_nir_lower_sample_mask.c',

src/gallium/drivers/asahi/agx_state.c

@@ -2016,15 +2016,6 @@ agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
          NIR_PASS(_, nir, agx_nir_lower_stats_fs);
       }
 
-      /* Clip plane lowering creates discard instructions, so run that before
-       * lowering discards. Note: this introduces extra loads from the clip
-       * plane outputs, but they use smooth interpolation so it does not affect
-       * the flat/linear masks that get propagated back to the VS.
-       */
-      if (key->clip_plane_enable) {
-         NIR_PASS(_, nir, nir_lower_clip_fs, key->clip_plane_enable, false);
-      }
-
       /* Similarly for cull distancing lowering */
       if (key->cull_distance_size) {
          NIR_PASS(_, nir, agx_nir_lower_cull_distance_fs,
@@ -2587,7 +2578,6 @@ agx_update_fs(struct agx_batch *batch)
 
       .cull_distance_size =
          ctx->stage[MESA_SHADER_VERTEX].shader->info.cull_distance_size,
-      .clip_plane_enable = ctx->rast->base.clip_plane_enable,
 
       .polygon_stipple =
          ctx->rast->base.poly_stipple_enable &&
@@ -3772,6 +3762,14 @@ agx_encode_state(struct agx_batch *batch, uint8_t *out, bool is_lines,
          cfg.viewport_target = vs->info.writes_layer_viewport;
          cfg.render_target = vs->info.writes_layer_viewport;
          cfg.frag_coord_z = fs->info.varyings.fs.reads_z;
+         cfg.clip_distance_plane_0 = vs->info.varyings.vs.nr_clip_dists > 0;
+         cfg.clip_distance_plane_1 = vs->info.varyings.vs.nr_clip_dists > 1;
+         cfg.clip_distance_plane_2 = vs->info.varyings.vs.nr_clip_dists > 2;
+         cfg.clip_distance_plane_3 = vs->info.varyings.vs.nr_clip_dists > 3;
+         cfg.clip_distance_plane_4 = vs->info.varyings.vs.nr_clip_dists > 4;
+         cfg.clip_distance_plane_5 = vs->info.varyings.vs.nr_clip_dists > 5;
+         cfg.clip_distance_plane_6 = vs->info.varyings.vs.nr_clip_dists > 6;
+         cfg.clip_distance_plane_7 = vs->info.varyings.vs.nr_clip_dists > 7;
 
          assert(cfg.point_size || !is_points);
       }

src/gallium/drivers/asahi/agx_state.h

@@ -460,7 +460,6 @@ struct asahi_fs_shader_key {
    bool polygon_stipple;
 
    uint8_t cull_distance_size;
-   uint8_t clip_plane_enable;
    uint8_t nr_samples;
    enum pipe_format rt_formats[PIPE_MAX_COLOR_BUFS];
 };