asahi: make clip_halfz dynamic

we could move this to the linker but meh, this is good enough for now

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

@@ -1741,6 +1741,12 @@ intrinsic("load_fb_layers_v3d", dest_comp=1, flags=[CAN_ELIMINATE, CAN_REORDER])
 # Equivalent to popcount(ballot(true) & ((1 << subgroup_invocation) - 1))
 system_value("active_subgroup_invocation_agx", 1)
 
+# With [0, 1] clipping, no transform is needed on the output z' = z. But with [-1,
+# 1] clipping, we need to transform z' = (z + w) / 2. We express both cases as a
+# lerp between z and w, where this is the lerp coefficient: 0 for [0, 1] and 0.5
+# for [-1, 1].
+system_value("clip_z_coeff_agx", 1)
+
 # mesa_prim for the input topology (in a geometry shader)
 system_value("input_topology_agx", 1)
 

src/gallium/drivers/asahi/agx_nir_lower_sysvals.c

@@ -179,6 +179,8 @@ 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_clip_z_coeff_agx:
+      return nir_f2f32(b, load_sysval_root(b, 1, 16, &u->clip_z_coeff));
    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;

src/gallium/drivers/asahi/agx_state.c

@@ -1675,7 +1675,7 @@ agx_link_varyings_vs_fs(struct agx_pool *pool, struct agx_varyings_vs *vs,
    return ptr.gpu;
 }
 
-/* nir_lower_clip_halfz analogue for lowered I/O */
+/* Dynamic lowered I/O version of nir_lower_clip_halfz */
 static bool
 agx_nir_lower_clip_m1_1(nir_builder *b, nir_intrinsic_instr *intr,
                         UNUSED void *data)
@@ -1691,8 +1691,10 @@ agx_nir_lower_clip_m1_1(nir_builder *b, nir_intrinsic_instr *intr,
    nir_def *pos = intr->src[0].ssa;
    nir_def *z = nir_channel(b, pos, 2);
    nir_def *w = nir_channel(b, pos, 3);
+   nir_def *c = nir_load_clip_z_coeff_agx(b);
 
-   nir_def *new_z = nir_fmul_imm(b, nir_fadd(b, z, w), 0.5f);
+   /* Lerp. If c = 0, reduces to z. If c = 1/2, reduces to (z + w)/2 */
+   nir_def *new_z = nir_ffma(b, nir_fneg(b, z), c, nir_ffma(b, w, c, z));
    nir_src_rewrite(&intr->src[0], nir_vector_insert_imm(b, pos, new_z, 2));
    return true;
 }
@@ -1893,12 +1895,8 @@ agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
       if (key->next_stage == ASAHI_VS_FS) {
          NIR_PASS(_, nir, agx_nir_lower_point_size,
                   key->next.fs.fixed_point_size);
-
-         if (!key->next.fs.clip_halfz) {
-            NIR_PASS(_, nir, nir_shader_intrinsics_pass,
-                     agx_nir_lower_clip_m1_1,
-                     nir_metadata_block_index | nir_metadata_dominance, NULL);
-         }
+         NIR_PASS(_, nir, nir_shader_intrinsics_pass, agx_nir_lower_clip_m1_1,
+                  nir_metadata_block_index | nir_metadata_dominance, NULL);
       } else if (key->next_stage == ASAHI_VS_GS) {
          NIR_PASS(_, nir, agx_nir_lower_sysvals, PIPE_SHADER_VERTEX, false);
          NIR_PASS(_, nir, agx_nir_lower_vs_before_gs, dev->libagx,
@@ -2080,12 +2078,8 @@ agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
 
       /* TODO: deduplicate */
       NIR_PASS(_, gs_copy, agx_nir_lower_point_size, key->fixed_point_size);
-
-      if (!key->clip_halfz) {
-         NIR_PASS(_, gs_copy, nir_shader_intrinsics_pass,
-                  agx_nir_lower_clip_m1_1,
-                  nir_metadata_block_index | nir_metadata_dominance, NULL);
-      }
+      NIR_PASS(_, gs_copy, nir_shader_intrinsics_pass, agx_nir_lower_clip_m1_1,
+               nir_metadata_block_index | nir_metadata_dominance, NULL);
 
       base_key.vs.outputs_flat_shaded = key->outputs_flat_shaded;
       base_key.vs.outputs_linear_shaded = key->outputs_linear_shaded;
@@ -2434,7 +2428,7 @@ agx_update_vs(struct agx_context *ctx, unsigned index_size_B)
    /* Only proceed if the shader or anything the key depends on changes
     *
     * vb_mask, attributes, vertex_buffers: VERTEX
-    * clip_halfz: RS
+    * point_size_per_vertex: RS
     * outputs_{flat,linear}_shaded: FS_PROG
     */
    if (!((ctx->dirty & (AGX_DIRTY_VS_PROG | AGX_DIRTY_VERTEX | AGX_DIRTY_XFB |
@@ -2456,8 +2450,6 @@ agx_update_vs(struct agx_context *ctx, unsigned index_size_B)
    };
 
    if (key.next_stage == ASAHI_VS_FS) {
-      key.next.fs.clip_halfz = ctx->rast->base.clip_halfz;
-
       /* If we are not rasterizing points, don't set fixed_point_size to
        * eliminate the useless point size write.
        */
@@ -2533,7 +2525,6 @@ agx_update_gs(struct agx_context *ctx, const struct pipe_draw_info *info,
       .rasterizer_discard = ctx->rast->base.rasterizer_discard,
 
       /* TODO: Deduplicate */
-      .clip_halfz = ctx->rast->base.clip_halfz,
       .fixed_point_size = !ctx->rast->base.point_size_per_vertex &&
                           rasterized_prim == MESA_PRIM_POINTS,
       .outputs_flat_shaded =

src/gallium/drivers/asahi/agx_state.h

@@ -150,6 +150,9 @@ struct PACKED agx_draw_uniforms {
 
    /* Nonzero for indexed draws, zero otherwise */
    uint16_t is_indexed_draw;
+
+   /* Zero for [0, 1] clipping, 0.5 for [-1, 1] clipping. */
+   uint16_t clip_z_coeff;
 };
 
 struct PACKED agx_stage_uniforms {
@@ -457,7 +460,6 @@ struct asahi_vs_shader_key {
       } gs;
 
       struct {
-         bool clip_halfz;
          bool fixed_point_size;
          uint64_t outputs_flat_shaded;
          uint64_t outputs_linear_shaded;
@@ -507,12 +509,11 @@ struct asahi_gs_shader_key {
    /* Rasterizer shader key */
    uint64_t outputs_flat_shaded;
    uint64_t outputs_linear_shaded;
-   bool clip_halfz;
    bool fixed_point_size;
 
    /* If true, this GS is run only for its side effects (including XFB) */
    bool rasterizer_discard;
-   bool padding[5];
+   bool padding[6];
 };
 static_assert(sizeof(struct asahi_gs_shader_key) == 24, "no holes");
 

src/gallium/drivers/asahi/agx_uniforms.c

@@ -6,6 +6,7 @@
 #include "asahi/lib/agx_pack.h"
 #include "pipe/p_state.h"
 #include "util/format/u_format.h"
+#include "util/half_float.h"
 #include "util/macros.h"
 #include "agx_state.h"
 #include "pool.h"
@@ -102,9 +103,14 @@ agx_upload_uniforms(struct agx_batch *batch)
    batch->uniforms.tables[AGX_SYSVAL_TABLE_ROOT] = root_ptr.gpu;
    batch->uniforms.sample_mask = ctx->sample_mask;
 
-   batch->uniforms.sprite_mask = (batch->reduced_prim == MESA_PRIM_POINTS)
-                                    ? ctx->rast->base.sprite_coord_enable
-                                    : 0;
+   assert(_mesa_float_to_half(0.5) == 0x3800);
+   batch->uniforms.clip_z_coeff =
+      (ctx->rast && !ctx->rast->base.clip_halfz) ? 0x3800 : 0x0;
+
+   batch->uniforms.sprite_mask =
+      (batch->reduced_prim == MESA_PRIM_POINTS && ctx->rast)
+         ? ctx->rast->base.sprite_coord_enable
+         : 0;
 
    memcpy(root_ptr.cpu, &batch->uniforms, sizeof(batch->uniforms));
 }