hk: fix tessellation + clipper queries

fixes upcoming cts

Backport-to: 25.1
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34585>
(cherry picked from commit d959557669f9deb5647e2481afbe1f3d55aa79f9)

.pick_status.json

@@ -1434,7 +1434,7 @@
         "description": "hk: fix tessellation + clipper queries",
         "nominated": true,
         "nomination_type": 4,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null,
         "notes": null

src/asahi/libagx/geometry.cl

@@ -758,7 +758,8 @@ libagx_prefix_sum_geom(constant struct agx_geometry_params *p)
 }
 
 KERNEL(1024)
-libagx_prefix_sum_tess(global struct libagx_tess_args *p)
+libagx_prefix_sum_tess(global struct libagx_tess_args *p, global uint *c_prims,
+                       global uint *c_invs, uint increment_stats__2)
 {
    _libagx_prefix_sum(p->counts, p->nr_patches, 1 /* words */, 0 /* word */);
 
@@ -786,6 +787,19 @@ libagx_prefix_sum_tess(global struct libagx_tess_args *p)
    desc[2] = alloc_B / elsize_B; /* start */
    desc[3] = 0;                  /* index_bias */
    desc[4] = 0;                  /* start_instance */
+
+   /* If necessary, increment clipper statistics too. This is only used when
+    * there's no geometry shader following us. See agx_nir_lower_gs.c for more
+    * info on the emulation. We just need to calculate the # of primitives
+    * tessellated.
+    */
+   if (increment_stats__2) {
+      uint prims = p->points_mode ? total
+                   : p->isolines  ? (total / 2)
+                                  : (total / 3);
+
+      increment_counters(c_prims, c_invs, NULL, prims);
+   }
 }
 
 uintptr_t

src/asahi/libagx/tessellator.h

@@ -97,6 +97,7 @@ struct libagx_tess_args {
     */
    enum libagx_tess_partitioning partitioning;
    uint32_t points_mode;
+   uint32_t isolines;
 
    /* When fed into a geometry shader, triangles should be counter-clockwise.
     * The tessellator always produces clockwise triangles, but we can swap
@@ -104,4 +105,4 @@ struct libagx_tess_args {
     */
    uint32_t ccw;
 } PACKED;
-static_assert(sizeof(struct libagx_tess_args) == 35 * 4);
+static_assert(sizeof(struct libagx_tess_args) == 36 * 4);

src/asahi/vulkan/hk_cmd_draw.c

@@ -1230,6 +1230,7 @@ hk_upload_tess_params(struct hk_cmd_buffer *cmd, struct libagx_tess_args *out,
       .tcs_per_vertex_outputs = tcs->info.tess.tcs_per_vertex_outputs,
       .partitioning = partitioning,
       .points_mode = gfx->tess.info.points,
+      .isolines = gfx->tess.info.mode == TESS_PRIMITIVE_ISOLINES,
    };
 
    if (!args.points_mode && gfx->tess.info.mode != TESS_PRIMITIVE_ISOLINES) {
@@ -1538,7 +1539,7 @@ hk_launch_gs_prerast(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
 
 static struct agx_draw
 hk_launch_tess(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
-               struct agx_draw draw)
+               struct agx_draw draw, uint64_t c_prims, uint64_t c_inv)
 {
    struct hk_device *dev = hk_cmd_buffer_device(cmd);
    struct hk_graphics_state *gfx = &cmd->state.gfx;
@@ -1621,7 +1622,7 @@ hk_launch_tess(struct hk_cmd_buffer *cmd, struct hk_cs *cs,
                      LIBAGX_TESS_MODE_COUNT, state);
 
    libagx_prefix_sum_tess(cmd, agx_1d(1024), AGX_BARRIER_ALL | AGX_PREGFX,
-                          state);
+                          state, c_prims, c_inv, c_prims || c_inv);
 
    libagx_tessellate(cmd, grid_tess, AGX_BARRIER_ALL | AGX_PREGFX, info.mode,
                      LIBAGX_TESS_MODE_WITH_COUNTS, state);
@@ -3373,8 +3374,6 @@ hk_ia_update(struct hk_cmd_buffer *cmd, struct agx_draw draw,
    /* Clipper counters depend on geom/tess outputs and must be written with the
     * geom/tess output. They are updated as IA counters only when geom/tess is
     * not used.
-    *
-    * TODO: Tessellation clipper counters not actually wired up, pending CTS.
     */
    if (geom || tess) {
       c_prims = 0;
@@ -3471,7 +3470,8 @@ hk_draw(struct hk_cmd_buffer *cmd, uint16_t draw_id, struct agx_draw draw_)
       }
 
       if (tess) {
-         draw = hk_launch_tess(cmd, ccs, draw);
+         draw = hk_launch_tess(cmd, ccs, draw, geom ? 0 : stat_c_prims,
+                               geom ? 0 : stat_c_inv);
       }
 
       if (geom) {

src/gallium/drivers/asahi/agx_state.c

@@ -3907,13 +3907,14 @@ agx_ia_update(struct agx_batch *batch, const struct pipe_draw_info *info,
    uint64_t c_invs = agx_get_query_address(
       batch, ctx->pipeline_statistics[PIPE_STAT_QUERY_C_INVOCATIONS]);
 
-   /* With a geometry shader, clipper counters are written by the pre-GS kernel
+   /* With a geometry/tessellation shader, clipper counters are written by the
-    * since they depend on the output on the geometry shader. Without a geometry
+    * pre-GS/tess prefix sum kernel since they depend on the output on the
-    * shader, they are written along with IA.
+    * geometry/tessellation shader.  Without a geometry/tessellation shader,
-    *
+    * they are written along with IA.
-    * TODO: Broken tessellation interaction, but nobody cares.
     */
-   if (ctx->stage[PIPE_SHADER_GEOMETRY].shader) {
+   if (ctx->stage[PIPE_SHADER_GEOMETRY].shader ||
+       ctx->stage[PIPE_SHADER_TESS_EVAL].shader) {
+
       c_prims = 0;
       c_invs = 0;
    }
@@ -4620,6 +4621,7 @@ agx_draw_patches(struct agx_context *ctx, const struct pipe_draw_info *info,
       .patch_coord_buffer = agx_resource(ctx->heap)->bo->va->addr,
       .partitioning = partitioning,
       .points_mode = point_mode,
+      .isolines = mode == TESS_PRIMITIVE_ISOLINES,
    };
 
    if (!point_mode && tes->tess.primitive != TESS_PRIMITIVE_ISOLINES) {
@@ -4736,10 +4738,24 @@ agx_draw_patches(struct agx_context *ctx, const struct pipe_draw_info *info,
 
    batch->uniforms.vertex_output_buffer_ptr = 0;
 
+   uint64_t c_prims = agx_get_query_address(
+      batch, ctx->pipeline_statistics[PIPE_STAT_QUERY_C_PRIMITIVES]);
+   uint64_t c_invs = agx_get_query_address(
+      batch, ctx->pipeline_statistics[PIPE_STAT_QUERY_C_INVOCATIONS]);
+
+   /* If there's a geometry shader, it will increment the clipper stats.
+    * Otherwise, we do when tessellating.
+    */
+   if (ctx->stage[PIPE_SHADER_GEOMETRY].shader) {
+      c_prims = 0;
+      c_invs = 0;
+   }
+
    /* Generate counts, then prefix sum them, then finally tessellate. */
    libagx_tessellate(batch, tess_grid, AGX_BARRIER_ALL, mode,
                      LIBAGX_TESS_MODE_COUNT, state);
-   libagx_prefix_sum_tess(batch, agx_1d(1024), AGX_BARRIER_ALL, state);
+   libagx_prefix_sum_tess(batch, agx_1d(1024), AGX_BARRIER_ALL, state, c_prims,
+                          c_invs, c_prims || c_invs);
    libagx_tessellate(batch, tess_grid, AGX_BARRIER_ALL, mode,
                      LIBAGX_TESS_MODE_WITH_COUNTS, state);
 
@@ -4949,6 +4965,7 @@ agx_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info,
         ctx->pipeline_statistics[PIPE_STAT_QUERY_VS_INVOCATIONS] ||
         ((ctx->pipeline_statistics[PIPE_STAT_QUERY_C_PRIMITIVES] ||
           ctx->pipeline_statistics[PIPE_STAT_QUERY_C_INVOCATIONS]) &&
+         !ctx->stage[PIPE_SHADER_TESS_EVAL].shader &&
          !ctx->stage[PIPE_SHADER_GEOMETRY].shader))) {
 
       uint64_t ptr;