asahi: lower poly stipple

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

src/compiler/nir/nir_intrinsics.py

@@ -1863,6 +1863,11 @@ intrinsic("sample_mask_agx", src_comp=[1, 1])
 # sample_mask_agx also triggers depth/stencil testing.
 intrinsic("discard_agx", src_comp=[1])
 
+# For a given row of the polygon stipple given as an integer source in [0, 31],
+# load the 32-bit stipple pattern for that row.
+intrinsic("load_polygon_stipple_agx", src_comp=[1], dest_comp=1, bit_sizes=[32],
+          flags=[CAN_ELIMINATE, CAN_ELIMINATE])
+
 # The fixed-function sample mask specified in the API (e.g. glSampleMask)
 system_value("api_sample_mask_agx", 1, bit_sizes=[16])
 

src/gallium/drivers/asahi/agx_nir_lower_sysvals.c

@@ -165,6 +165,14 @@ lower_intrinsic(nir_builder *b, nir_intrinsic_instr *intr,
       return load_sysval_root(b, 1, 32, &u->fixed_point_size);
    case nir_intrinsic_load_tex_sprite_mask_agx:
       return load_sysval_root(b, 1, 16, &u->sprite_mask);
+   case nir_intrinsic_load_polygon_stipple_agx: {
+      nir_def *base = load_sysval_root(b, 1, 64, &u->polygon_stipple);
+      nir_def *row = intr->src[0].ssa;
+      nir_def *addr = nir_iadd(b, base, nir_u2u64(b, nir_imul_imm(b, row, 4)));
+
+      return nir_load_global_constant(b, addr, 4, 1, 32);
+   }
+
    default:
       break;
    }

src/gallium/drivers/asahi/agx_state.c

@@ -918,9 +918,13 @@ agx_set_clip_state(struct pipe_context *ctx,
 }
 
 static void
-agx_set_polygon_stipple(struct pipe_context *ctx,
+agx_set_polygon_stipple(struct pipe_context *pctx,
                         const struct pipe_poly_stipple *state)
 {
+   struct agx_context *ctx = agx_context(pctx);
+
+   memcpy(ctx->poly_stipple, state->stipple, sizeof(ctx->poly_stipple));
+   ctx->dirty |= AGX_DIRTY_POLY_STIPPLE;
 }
 
 static void
@@ -1758,6 +1762,46 @@ agx_compile_nir(struct agx_device *dev, nir_shader *nir,
    return compiled;
 }
 
+/*
+ * Insert code into a fragment shader to lower polygon stipple. The stipple is
+ * passed in a sideband, rather than requiring a texture binding. This is
+ * simpler for drivers to integrate and might be more efficient.
+ */
+static bool
+agx_nir_lower_poly_stipple(nir_shader *s)
+{
+   assert(s->info.stage == MESA_SHADER_FRAGMENT);
+
+   /* Insert at the beginning for performance. */
+   nir_builder b_ =
+      nir_builder_at(nir_before_impl(nir_shader_get_entrypoint(s)));
+   nir_builder *b = &b_;
+
+   /* The stipple coordinate is defined at the window coordinate mod 32. It's
+    * reversed along the X-axis to simplify the driver, hence the NOT.
+    */
+   nir_def *raw = nir_u2u32(b, nir_load_pixel_coord(b));
+   nir_def *coord = nir_umod_imm(
+      b,
+      nir_vec2(b, nir_inot(b, nir_channel(b, raw, 0)), nir_channel(b, raw, 1)),
+      32);
+
+   /* Load the stipple pattern for the row */
+   nir_def *pattern = nir_load_polygon_stipple_agx(b, nir_channel(b, coord, 1));
+
+   /* Extract the column from the packed bitfield */
+   nir_def *bit = nir_ubitfield_extract(b, pattern, nir_channel(b, coord, 0),
+                                        nir_imm_int(b, 1));
+
+   /* Discard fragments where the pattern is 0 */
+   nir_discard_if(b, nir_ieq_imm(b, bit, 0));
+   s->info.fs.uses_discard = true;
+
+   nir_metadata_preserve(b->impl,
+                         nir_metadata_dominance | nir_metadata_block_index);
+   return true;
+}
+
 /* Does not take ownership of key. Clones if necessary. */
 static struct agx_compiled_shader *
 agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
@@ -1891,6 +1935,11 @@ agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
                   key->cull_distance_size);
       }
 
+      /* Similarly for polygon stipple */
+      if (key->polygon_stipple) {
+         NIR_PASS_V(nir, agx_nir_lower_poly_stipple);
+      }
+
       /* Discards must be lowering before lowering MSAA to handle discards */
       NIR_PASS(_, nir, agx_nir_lower_discard_zs_emit);
 
@@ -2377,6 +2426,12 @@ 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 &&
+         rast_prim(batch->reduced_prim, ctx->rast->base.fill_front) ==
+            MESA_PRIM_TRIANGLES,
+
       .nr_samples = nr_samples,
 
       /* Only lower sample mask if at least one sample is masked out */
@@ -4328,6 +4383,13 @@ agx_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info,
       batch->uniforms.fixed_point_size = ctx->rast->base.point_size;
    }
 
+   if (IS_DIRTY(POLY_STIPPLE)) {
+      STATIC_ASSERT(sizeof(ctx->poly_stipple) == 32 * 4);
+
+      batch->uniforms.polygon_stipple = agx_pool_upload_aligned(
+         &batch->pool, ctx->poly_stipple, sizeof(ctx->poly_stipple), 4);
+   }
+
    if (IS_DIRTY(VS) || IS_DIRTY(FS) || ctx->gs || IS_DIRTY(VERTEX) ||
        IS_DIRTY(BLEND_COLOR) || IS_DIRTY(RS) || IS_DIRTY(PRIM)) {
 

src/gallium/drivers/asahi/agx_state.h

@@ -109,6 +109,9 @@ struct PACKED agx_draw_uniforms {
    /* Address of geometry param buffer if geometry shaders are used, else 0 */
    uint64_t geometry_params;
 
+   /* Address of polygon stipple mask if used */
+   uint64_t polygon_stipple;
+
    /* Blend constant if any */
    float blend_constant[4];
 
@@ -396,6 +399,7 @@ struct asahi_fs_shader_key {
     * don't want to emit lowering code for it, since it would disable early-Z.
     */
    bool api_sample_mask;
+   bool polygon_stipple;
 
    uint8_t cull_distance_size;
    uint8_t clip_plane_enable;
@@ -449,6 +453,7 @@ enum agx_dirty {
    AGX_DIRTY_XFB = BITFIELD_BIT(14),
    AGX_DIRTY_SAMPLE_MASK = BITFIELD_BIT(15),
    AGX_DIRTY_BLEND_COLOR = BITFIELD_BIT(16),
+   AGX_DIRTY_POLY_STIPPLE = BITFIELD_BIT(17),
 };
 
 /* Maximum number of in-progress + under-construction GPU batches.
@@ -535,6 +540,8 @@ struct agx_context {
    uint16_t sample_mask;
    struct pipe_framebuffer_state framebuffer;
 
+   uint32_t poly_stipple[32];
+
    struct pipe_query *cond_query;
    bool cond_cond;
    enum pipe_render_cond_flag cond_mode;