asahi,nir: Move asahi dynamic clipz pass to common.

Acked-by: Alyssa Rosenzweig <alyssa@rosenz.ca>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41088>

src/compiler/nir/nir.h

@@ -5444,6 +5444,7 @@ bool nir_lower_vars_to_scratch_global(nir_shader *shader,
 bool nir_lower_scratch_to_var(nir_shader *nir);
 
 bool nir_lower_clip_halfz(nir_shader *shader);
+bool nir_lower_clip_halfz_dynamic(nir_shader *shader);
 
 void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
 void nir_gather_clip_cull_distance_sizes_from_vars(nir_shader *nir);

src/compiler/nir/nir_intrinsics.py

@@ -2335,7 +2335,7 @@ intrinsic("quad_ballot_agx", src_comp=[1], dest_comp=1, flags=[CAN_ELIMINATE])
 # 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)
+system_value("clip_z_coeff", 1)
 
 # True if drawing triangle fans with first vertex provoking, false otherwise.
 # This affects flatshading, which is defined weirdly for fans with first.

src/compiler/nir/nir_lower_clip_halfz.c

@@ -63,3 +63,42 @@ nir_lower_clip_halfz(nir_shader *shader)
                                      nir_metadata_control_flow,
                                      NULL);
 }
+
+/* Dynamic lowered I/O version of nir_lower_clip_halfz.
+ * nir_intrinsic_load_clip_z_coeff is used to load the dynamic clip coefficient.
+ */
+static bool
+lower_pos_write_dynamic(nir_builder *b, nir_intrinsic_instr *intr,
+                        UNUSED void *data)
+{
+   if (intr->intrinsic != nir_intrinsic_store_output)
+      return false;
+   if (nir_intrinsic_io_semantics(intr).location != VARYING_SLOT_POS)
+      return false;
+
+   assert(nir_intrinsic_component(intr) == 0 && "not yet scalarized");
+   b->cursor = nir_before_instr(&intr->instr);
+
+   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(b);
+
+   /* 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;
+}
+
+bool
+nir_lower_clip_halfz_dynamic(nir_shader *shader)
+{
+   if (shader->info.stage != MESA_SHADER_VERTEX &&
+       shader->info.stage != MESA_SHADER_GEOMETRY &&
+       shader->info.stage != MESA_SHADER_TESS_EVAL)
+      return false;
+
+   return nir_shader_intrinsics_pass(shader, lower_pos_write_dynamic,
+                                     nir_metadata_control_flow,
+                                     NULL);
+}

src/gallium/drivers/asahi/agx_nir_lower_sysvals.c

@@ -194,7 +194,7 @@ lower_intrinsic(nir_builder *b, nir_intrinsic_instr *intr,
       return load_sysval_root(b, 1, 16, &u->sprite_mask);
    case nir_intrinsic_load_shader_part_tests_zs_agx:
       return load_sysval_root(b, 1, 16, &u->no_epilog_discard);
-   case nir_intrinsic_load_clip_z_coeff_agx:
+   case nir_intrinsic_load_clip_z_coeff:
       return nir_f2f32(b, load_sysval_root(b, 1, 16, &u->clip_z_coeff));
    case nir_intrinsic_load_rasterization_stream:
       return nir_imm_int(b, 0);

src/gallium/drivers/asahi/agx_state.c

@@ -1389,30 +1389,6 @@ asahi_cs_shader_key_equal(const void *a, const void *b)
    return true;
 }
 
-/* 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)
-{
-   if (intr->intrinsic != nir_intrinsic_store_output)
-      return false;
-   if (nir_intrinsic_io_semantics(intr).location != VARYING_SLOT_POS)
-      return false;
-
-   assert(nir_intrinsic_component(intr) == 0 && "not yet scalarized");
-   b->cursor = nir_before_instr(&intr->instr);
-
-   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);
-
-   /* 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;
-}
-
 /*
  * To implement point sprites, we'll replace TEX0...7 with point coordinate
  * reads as required. However, the .zw needs to read back 0.0/1.0. This pass
@@ -1560,8 +1536,7 @@ agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
 
       if (key->hw) {
          NIR_PASS(_, nir, agx_nir_lower_point_size, true);
-         NIR_PASS(_, nir, nir_shader_intrinsics_pass, agx_nir_lower_clip_m1_1,
-                  nir_metadata_control_flow, NULL);
+         NIR_PASS(_, nir, nir_lower_clip_halfz_dynamic);
 
          NIR_PASS(_, nir, nir_lower_io_to_scalar, nir_var_shader_out, NULL,
                   NULL);
@@ -1661,8 +1636,7 @@ agx_compile_variant(struct agx_device *dev, struct pipe_context *pctx,
        */
       NIR_PASS(_, gs_copy, agx_nir_lower_point_size, false);
 
-      NIR_PASS(_, gs_copy, nir_shader_intrinsics_pass, agx_nir_lower_clip_m1_1,
-               nir_metadata_control_flow, NULL);
+      NIR_PASS(_, gs_copy, nir_lower_clip_halfz_dynamic);
 
       NIR_PASS(_, gs_copy, nir_lower_io_to_scalar, nir_var_shader_out, NULL,
                NULL);