asahi: pack blend key

reduce hashing for fs epilogs, hopefully

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29607>

2 files changed, 32 insertions, 10 deletions

diff --git a/src/asahi/lib/agx_linker.h b/src/asahi/lib/agx_linker.h
index fe2e40afbc0..c28036c25bb 100644
--- a/src/asahi/lib/agx_linker.h
+++ b/src/asahi/lib/agx_linker.h
@@ -85,13 +85,25 @@ struct agx_fs_prolog_key {
    unsigned cf_base;
 };
 
+struct agx_blend_rt_key {
+   enum pipe_blend_func rgb_func          : 3;
+   enum pipe_blendfactor rgb_src_factor   : 5;
+   enum pipe_blendfactor rgb_dst_factor   : 5;
+   enum pipe_blend_func alpha_func        : 3;
+   enum pipe_blendfactor alpha_src_factor : 5;
+   enum pipe_blendfactor alpha_dst_factor : 5;
+   unsigned colormask                     : 4;
+   unsigned pad                           : 2;
+};
+static_assert(sizeof(struct agx_blend_rt_key) == 4, "packed");
+
 struct agx_blend_key {
-   nir_lower_blend_rt rt[8];
-   unsigned logicop_func;
+   struct agx_blend_rt_key rt[8];
+   uint8_t logicop_func;
    bool alpha_to_coverage, alpha_to_one;
-   bool padding[2];
+   bool padding;
 };
-static_assert(sizeof(struct agx_blend_key) == 232, "packed");
+static_assert(sizeof(struct agx_blend_key) == 36, "packed");
 
 struct agx_fs_epilog_link_info {
    /* Base index of spilled render targets in the binding table */
diff --git a/src/asahi/lib/agx_nir_prolog_epilog.c b/src/asahi/lib/agx_nir_prolog_epilog.c
index a6f4bd2cd9c..c40b1043387 100644
--- a/src/asahi/lib/agx_nir_prolog_epilog.c
+++ b/src/asahi/lib/agx_nir_prolog_epilog.c
@@ -258,13 +258,13 @@ lower_tests_zs(nir_shader *s, bool value)
 }
 
 static inline bool
-blend_uses_2src(nir_lower_blend_rt rt)
+blend_uses_2src(struct agx_blend_rt_key rt)
 {
    enum pipe_blendfactor factors[] = {
-      rt.rgb.src_factor,
-      rt.rgb.dst_factor,
-      rt.alpha.src_factor,
-      rt.alpha.dst_factor,
+      rt.rgb_src_factor,
+      rt.rgb_dst_factor,
+      rt.alpha_src_factor,
+      rt.alpha_dst_factor,
    };
 
    for (unsigned i = 0; i < ARRAY_SIZE(factors); ++i) {
@@ -337,10 +337,20 @@ agx_nir_fs_epilog(nir_builder *b, const void *key_)
    };
 
    static_assert(ARRAY_SIZE(opts.format) == 8, "max RTs out of sync");
-   memcpy(opts.rt, key->blend.rt, sizeof(opts.rt));
 
    for (unsigned i = 0; i < 8; ++i) {
       opts.format[i] = key->rt_formats[i];
+      opts.rt[i] = (nir_lower_blend_rt){
+         .rgb.src_factor = key->blend.rt[i].rgb_src_factor,
+         .rgb.dst_factor = key->blend.rt[i].rgb_dst_factor,
+         .rgb.func = key->blend.rt[i].rgb_func,
+
+         .alpha.src_factor = key->blend.rt[i].alpha_src_factor,
+         .alpha.dst_factor = key->blend.rt[i].alpha_dst_factor,
+         .alpha.func = key->blend.rt[i].alpha_func,
+
+         .colormask = key->blend.rt[i].colormask,
+      };
    }
 
    /* It's more efficient to use masked stores (with