asahi: Add flat/linear shaded varyings mask to the VS shader key

We need this information in order to arrange varyings properly, which
means we need shader variants. Add this to the shader key, taking the
value from the FS input info.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23998>

src/asahi/compiler/agx_compile.h

@@ -146,6 +146,17 @@ enum agx_format {
    AGX_NUM_FORMATS,
 };
 
+struct agx_vs_shader_key {
+   /* The GPU ABI requires all smooth shaded varyings to come first, then all
+    * flat shaded varyings, then all linear shaded varyings, as written by the
+    * VS. In order to correctly remap the varyings into the right order in the
+    * VS, we need to propagate the mask of flat/linear shaded varyings into the
+    * compiler.
+    */
+   uint64_t outputs_flat_shaded;
+   uint64_t outputs_linear_shaded;
+};
+
 struct agx_fs_shader_key {
    /* Normally, access to the tilebuffer must be guarded by appropriate fencing
     * instructions to ensure correct results in the presence of out-of-order
@@ -172,6 +183,7 @@ struct agx_shader_key {
    unsigned reserved_preamble;
 
    union {
+      struct agx_vs_shader_key vs;
       struct agx_fs_shader_key fs;
    };
 };

src/gallium/drivers/asahi/agx_state.c

@@ -1455,7 +1455,9 @@ agx_compile_variant(struct agx_device *dev, struct agx_uncompiled_shader *so,
       }
 
       /* Clip plane lowering creates discard instructions, so run that before
-       * lowering discards.
+       * 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_V(nir, nir_lower_clip_fs, key->clip_plane_enable, false);
@@ -1494,6 +1496,11 @@ agx_compile_variant(struct agx_device *dev, struct agx_uncompiled_shader *so,
    if (nir->info.stage == MESA_SHADER_FRAGMENT)
       base_key.fs.nr_samples = key_->fs.nr_samples;
 
+   if (nir->info.stage == MESA_SHADER_VERTEX) {
+      base_key.vs.outputs_flat_shaded = key_->vs.outputs_flat_shaded;
+      base_key.vs.outputs_linear_shaded = key_->vs.outputs_linear_shaded;
+   }
+
    NIR_PASS_V(nir, agx_nir_lower_sysvals, compiled,
               &base_key.reserved_preamble);
 
@@ -1713,13 +1720,19 @@ agx_update_vs(struct agx_context *ctx)
     *
     * vb_mask, attributes, vertex_buffers: VERTEX
     * streamout.active: XFB
+    * outputs_{flat,linear}_shaded: FS_PROG
     */
-   if (!(ctx->dirty & (AGX_DIRTY_VS_PROG | AGX_DIRTY_VERTEX | AGX_DIRTY_XFB)))
+   if (!(ctx->dirty & (AGX_DIRTY_VS_PROG | AGX_DIRTY_VERTEX | AGX_DIRTY_XFB |
+                       AGX_DIRTY_FS_PROG)))
       return false;
 
    struct asahi_vs_shader_key key = {
       .vbuf.count = util_last_bit(ctx->vb_mask),
       .xfb = ctx->streamout.key,
+      .outputs_flat_shaded =
+         ctx->stage[PIPE_SHADER_FRAGMENT].shader->info.inputs_flat_shaded,
+      .outputs_linear_shaded =
+         ctx->stage[PIPE_SHADER_FRAGMENT].shader->info.inputs_linear_shaded,
    };
 
    memcpy(key.vbuf.attributes, ctx->attributes,

src/gallium/drivers/asahi/agx_state.h

@@ -286,6 +286,8 @@ struct agx_blend {
 struct asahi_vs_shader_key {
    struct agx_vbufs vbuf;
    struct agx_xfb_key xfb;
+   uint64_t outputs_flat_shaded;
+   uint64_t outputs_linear_shaded;
 };
 
 struct asahi_fs_shader_key {