asahi: Track batch masks on ZS/blend CSO

Adapted from panfrost, with the work happening at CSO create time instead of
draw time allowing us to do more sophisticated analysis. We'll use these for
accurate masks in a moment.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20123>

diff --git a/src/gallium/drivers/asahi/agx_pipe.c b/src/gallium/drivers/asahi/agx_pipe.c
index 1334e501ba0c9a2006f51d14080172e23fc1f794..5758f062cad7ceee883ad62065770f296b067b99 100644 (file)
--- a/src/gallium/drivers/asahi/agx_pipe.c
+++ b/src/gallium/drivers/asahi/agx_pipe.c
@@ -954,6 +954,7 @@ agx_clear(struct pipe_context *pctx, unsigned buffers, const struct pipe_scissor
    }
 
    batch->clear |= fastclear;
+   batch->resolve |= buffers;
    assert((batch->draw & slowclear) == slowclear);
 }
 

diff --git a/src/gallium/drivers/asahi/agx_state.c b/src/gallium/drivers/asahi/agx_state.c
index c7933731cd332d799f3cbb5d1058a2a4502117f1..33cc5011ce79b1a18a7d6ee9e7de74622eb00e7c 100644 (file)
--- a/src/gallium/drivers/asahi/agx_state.c
+++ b/src/gallium/drivers/asahi/agx_state.c
@@ -163,6 +163,9 @@ agx_create_blend_state(struct pipe_context *ctx,
       }
 
       so->rt[i].colormask = rt.colormask;
+
+      if (rt.colormask)
+         so->store |= (PIPE_CLEAR_COLOR0 << i);
    }
 
    return so;
@@ -249,6 +252,22 @@ agx_create_zsa_state(struct pipe_context *ctx,
       so->back_stencil = so->front_stencil;
    }
 
+   if (state->depth_enabled) {
+      if (state->depth_func != PIPE_FUNC_NEVER &&
+          state->depth_func != PIPE_FUNC_ALWAYS) {
+
+         so->load |= PIPE_CLEAR_DEPTH;
+      }
+
+      if (state->depth_writemask)
+         so->store |= PIPE_CLEAR_DEPTH;
+   }
+
+   if (state->stencil[0].enabled) {
+      so->load |= PIPE_CLEAR_STENCIL; /* TODO: Optimize */
+      so->store |= PIPE_CLEAR_STENCIL;
+   }
+
    return so;
 }
 
@@ -1974,9 +1993,18 @@ agx_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info,
    if (reduced_prim != batch->reduced_prim) ctx->dirty |= AGX_DIRTY_PRIM;
    batch->reduced_prim = reduced_prim;
 
-   /* TODO: masks */
-   batch->draw |= ~0;
-   batch->load |= ~0;
+   /* Update batch masks based on current state */
+   if (ctx->dirty & AGX_DIRTY_BLEND) {
+      /* TODO: Any point to tracking load? */
+      batch->draw |= ctx->blend->store;
+      batch->resolve |= ctx->blend->store;
+   }
+
+   if (ctx->dirty & AGX_DIRTY_ZS) {
+      batch->load |= ctx->zs->load;
+      batch->draw |= ctx->zs->store;
+      batch->resolve |= ctx->zs->store;
+   }
 
    /* TODO: These are expensive calls, consider finer dirty tracking */
    if (agx_update_vs(ctx))

diff --git a/src/gallium/drivers/asahi/agx_state.h b/src/gallium/drivers/asahi/agx_state.h
index 888ab3fc6c7ae11ddd1d60861334dc6c4d790dfb..48da69d7aa3cca63d1d5f708b355ccf0b4c085df 100644 (file)
--- a/src/gallium/drivers/asahi/agx_state.h
+++ b/src/gallium/drivers/asahi/agx_state.h
@@ -97,7 +97,7 @@ struct agx_batch {
    struct agx_tilebuffer_layout tilebuffer_layout;
 
    /* PIPE_CLEAR_* bitmask */
-   uint32_t clear, draw, load;
+   uint32_t clear, draw, load, resolve;
 
    /* Base of uploaded texture descriptors */
    uint64_t textures;
@@ -133,12 +133,18 @@ struct agx_zsa {
    struct pipe_depth_stencil_alpha_state base;
    struct agx_fragment_face_packed depth;
    struct agx_fragment_stencil_packed front_stencil, back_stencil;
+
+   /* PIPE_CLEAR_* bitmask corresponding to this depth/stencil state */
+   uint32_t load, store;
 };
 
 struct agx_blend {
    bool logicop_enable, blend_enable;
    nir_lower_blend_rt rt[8];
    unsigned logicop_func;
+
+   /* PIPE_CLEAR_* bitmask corresponding to this blend state */
+   uint32_t store;
 };
 
 struct asahi_shader_key {