asahi: Use util_blitter_clear

To avoid flush, use util_blitter_clear to clear with rectangles when
there is already content in the scene. This ~doubles the performance of
Inochi2D, which does frequent stencil clears in the middle of the screen
to implement masking.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18380>

src/gallium/drivers/asahi/agx_pipe.c

@@ -362,20 +362,33 @@ agx_clear(struct pipe_context *pctx, unsigned buffers, const struct pipe_scissor
 {
    struct agx_context *ctx = agx_context(pctx);
 
-   /* TODO: support partial clears */
-   if (ctx->batch->clear | ctx->batch->draw)
-      pctx->flush(pctx, NULL, 0);
+   unsigned fastclear = buffers & ~(ctx->batch->draw | ctx->batch->load);
+   unsigned slowclear = buffers & ~fastclear;
 
-   ctx->batch->clear |= buffers;
+   assert(scissor_state == NULL && "we don't support PIPE_CAP_CLEAR_SCISSORED");
 
-   if (buffers & PIPE_CLEAR_COLOR0)
+   /* Fast clears configure the batch */
+   if (fastclear & PIPE_CLEAR_COLOR0)
       memcpy(ctx->batch->clear_color, color->f, sizeof(color->f));
 
-   if (buffers & PIPE_CLEAR_DEPTH)
+   if (fastclear & PIPE_CLEAR_DEPTH)
       ctx->batch->clear_depth = depth;
 
-   if (buffers & PIPE_CLEAR_STENCIL)
+   if (fastclear & PIPE_CLEAR_STENCIL)
       ctx->batch->clear_stencil = stencil;
+
+   /* Slow clears draw a fullscreen rectangle */
+   if (slowclear) {
+      agx_blitter_save(ctx, ctx->blitter, false /* render cond */);
+      util_blitter_clear(ctx->blitter, ctx->framebuffer.width,
+            ctx->framebuffer.height,
+            util_framebuffer_get_num_layers(&ctx->framebuffer),
+            slowclear, color, depth, stencil,
+            util_framebuffer_get_num_samples(&ctx->framebuffer) > 1);
+   }
+
+   ctx->batch->clear |= fastclear;
+   assert((ctx->batch->draw & slowclear) == slowclear);
 }
 
 
@@ -554,6 +567,7 @@ agx_flush(struct pipe_context *pctx,
    agx_pool_init(&ctx->batch->pipeline_pool, dev, AGX_MEMORY_TYPE_CMDBUF_32, true);
    ctx->batch->clear = 0;
    ctx->batch->draw = 0;
+   ctx->batch->load = 0;
    ctx->batch->encoder_current = ctx->batch->encoder->ptr.cpu;
    ctx->batch->scissor.count = 0;
    ctx->dirty = ~0;

src/gallium/drivers/asahi/agx_state.c

@@ -1706,6 +1706,7 @@ agx_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info,
 
    /* TODO: masks */
    ctx->batch->draw |= ~0;
+   ctx->batch->load |= ~0;
 
    /* TODO: Dirty track */
    agx_update_vs(ctx);

src/gallium/drivers/asahi/agx_state.h

@@ -98,7 +98,7 @@ struct agx_batch {
    struct pipe_surface *zsbuf;
 
    /* PIPE_CLEAR_* bitmask */
-   uint32_t clear, draw;
+   uint32_t clear, draw, load;
 
    float clear_color[4];
    double clear_depth;