asahi: pack tilebuffer usc word ahead-of-time

reduce draw time overhead.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29179>

src/asahi/lib/agx_tilebuffer.c

@@ -46,6 +46,19 @@ agx_select_tile_size(unsigned bytes_per_pixel)
    unreachable("No supported tile size meets the bytes per pixel requirement");
 }
 
+static unsigned
+agx_shared_layout_from_tile_size(struct agx_tile_size t)
+{
+   if (t.width == 32 && t.height == 32)
+      return AGX_SHARED_LAYOUT_32X32;
+   else if (t.width == 32 && t.height == 16)
+      return AGX_SHARED_LAYOUT_32X16;
+   else if (t.width == 16 && t.height == 16)
+      return AGX_SHARED_LAYOUT_16X16;
+   else
+      unreachable("Invalid tile size");
+}
+
 struct agx_tilebuffer_layout
 agx_build_tilebuffer_layout(const enum pipe_format *formats, uint8_t nr_cbufs,
                             uint8_t nr_samples, bool layered)
@@ -120,6 +133,8 @@ agx_build_tilebuffer_layout(const enum pipe_format *formats, uint8_t nr_cbufs,
    tib.sample_size_B = ALIGN_POT(offset_B, 8);
 
    tib.tile_size = agx_select_tile_size(tib.sample_size_B * nr_samples);
+
+   agx_tilebuffer_pack_usc(&tib);
    return tib;
 }
 
@@ -142,19 +157,6 @@ agx_tilebuffer_supports_mask(struct agx_tilebuffer_layout *tib, unsigned rt)
    return agx_internal_format_supports_mask((enum agx_internal_formats)fmt);
 }
 
-static unsigned
-agx_shared_layout_from_tile_size(struct agx_tile_size t)
-{
-   if (t.width == 32 && t.height == 32)
-      return AGX_SHARED_LAYOUT_32X32;
-   else if (t.width == 32 && t.height == 16)
-      return AGX_SHARED_LAYOUT_32X16;
-   else if (t.width == 16 && t.height == 16)
-      return AGX_SHARED_LAYOUT_16X16;
-   else
-      unreachable("Invalid tile size");
-}
-
 uint32_t
 agx_tilebuffer_total_size(struct agx_tilebuffer_layout *tib)
 {
@@ -163,13 +165,18 @@ agx_tilebuffer_total_size(struct agx_tilebuffer_layout *tib)
 }
 
 void
-agx_usc_tilebuffer(struct agx_usc_builder *b, struct agx_tilebuffer_layout *tib)
+agx_tilebuffer_pack_usc(struct agx_tilebuffer_layout *tib)
 {
-   agx_usc_pack(b, SHARED, cfg) {
-      cfg.uses_shared_memory = true;
-      cfg.layout = agx_shared_layout_from_tile_size(tib->tile_size);
-      cfg.sample_stride_in_8_bytes = tib->sample_size_B / 8;
-      cfg.sample_count = tib->nr_samples;
-      cfg.bytes_per_threadgroup = agx_tilebuffer_total_size(tib);
+   agx_pack(&tib->usc, USC_SHARED, cfg) {
+      if (tib->nr_samples > 0) {
+         cfg.uses_shared_memory = true;
+         cfg.layout = agx_shared_layout_from_tile_size(tib->tile_size);
+         cfg.sample_stride_in_8_bytes = tib->sample_size_B / 8;
+         cfg.sample_count = tib->nr_samples;
+         cfg.bytes_per_threadgroup = agx_tilebuffer_total_size(tib);
+      } else {
+         cfg.layout = AGX_SHARED_LAYOUT_VERTEX_COMPUTE;
+         cfg.bytes_per_threadgroup = 65536;
+      }
    }
 }

src/asahi/lib/agx_tilebuffer.h

@@ -9,6 +9,7 @@
 #include <stdbool.h>
 #include <stdint.h>
 #include "util/format/u_formats.h"
+#include "agx_pack.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -23,7 +24,6 @@ extern "C" {
 struct nir_shader;
 struct nir_def;
 struct nir_builder;
-struct agx_usc_builder;
 
 struct agx_tile_size {
    uint8_t width;
@@ -57,6 +57,9 @@ struct agx_tilebuffer_layout {
 
    /* Selected tile size */
    struct agx_tile_size tile_size;
+
+   /* USC word corresponding to this configuration of the tilebuffer */
+   struct agx_usc_shared_packed usc;
 };
 
 /*
@@ -104,9 +107,6 @@ bool agx_nir_lower_alpha_to_coverage(struct nir_shader *shader,
 
 bool agx_nir_lower_alpha_to_one(struct nir_shader *shader);
 
-void agx_usc_tilebuffer(struct agx_usc_builder *b,
-                        struct agx_tilebuffer_layout *tib);
-
 uint32_t agx_tilebuffer_total_size(struct agx_tilebuffer_layout *tib);
 
 enum pipe_format
@@ -115,6 +115,8 @@ agx_tilebuffer_physical_format(struct agx_tilebuffer_layout *tib, unsigned rt);
 bool agx_tilebuffer_supports_mask(struct agx_tilebuffer_layout *tib,
                                   unsigned rt);
 
+void agx_tilebuffer_pack_usc(struct agx_tilebuffer_layout *tib);
+
 #ifdef __cplusplus
 } /* extern C */
 #endif

src/gallium/drivers/asahi/agx_state.c

@@ -2977,7 +2977,7 @@ agx_build_pipeline(struct agx_batch *batch, struct agx_compiled_shader *cs,
    }
 
    if (stage == PIPE_SHADER_FRAGMENT) {
-      agx_usc_tilebuffer(&b, &batch->tilebuffer_layout);
+      agx_usc_push_packed(&b, SHARED, &batch->tilebuffer_layout.usc);
    } else if (stage == PIPE_SHADER_COMPUTE || stage == PIPE_SHADER_TESS_CTRL) {
       unsigned size = cs->b.info.local_size + variable_shared_mem;
 
@@ -3167,7 +3167,7 @@ agx_build_meta(struct agx_batch *batch, bool store, bool partial_render)
       }
    }
 
-   agx_usc_tilebuffer(&b, &batch->tilebuffer_layout);
+   agx_usc_push_packed(&b, SHARED, &batch->tilebuffer_layout.usc);
 
    /* Get the shader */
    key.reserved_preamble = uniforms;