asahi: rewrite queries

1. always keep the query in gpu memory, so we can implement qbos properly.

2. use a lightweight data structure for tracking writers to reduce overhead

3. allow many writers per query to eliminate stalls

4. use context-wide occlusion heap, to satisfy #1 without introducing
   flushes or silly copies. this is what the pvr mesa driver does :-)

Signed-off-by: Alyssa Rosenzweig <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>

Author: alyssarosenzweig

src/gallium/drivers/asahi/agx_batch.c

@@ -24,12 +24,6 @@
          agx_msg("[Batch %u] " fmt "\n", agx_batch_idx(batch), ##__VA_ARGS__); \
    } while (0)
 
-static unsigned
-agx_batch_idx(struct agx_batch *batch)
-{
-   return batch - batch->ctx->batches.slots;
-}
-
 bool
 agx_batch_is_active(struct agx_batch *batch)
 {
@@ -125,9 +119,7 @@ agx_batch_init(struct agx_context *ctx,
 
    util_dynarray_init(&batch->scissor, ctx);
    util_dynarray_init(&batch->depth_bias, ctx);
-   util_dynarray_init(&batch->occlusion_queries, ctx);
-   util_dynarray_init(&batch->nonocclusion_queries, ctx);
-   util_dynarray_init(&batch->timestamp_queries, ctx);
+   util_dynarray_init(&batch->timestamps, ctx);
 
    batch->clear = 0;
    batch->draw = 0;
@@ -177,8 +169,6 @@ agx_batch_cleanup(struct agx_context *ctx, struct agx_batch *batch, bool reset)
    uint64_t begin_ts = ~0, end_ts = 0;
    /* TODO: UAPI pending */
    agx_finish_batch_queries(batch, begin_ts, end_ts);
-   batch->occlusion_buffer.cpu = NULL;
-   batch->occlusion_buffer.gpu = 0;
 
    if (reset) {
       int handle;
@@ -212,9 +202,7 @@ agx_batch_cleanup(struct agx_context *ctx, struct agx_batch *batch, bool reset)
 
    util_dynarray_fini(&batch->scissor);
    util_dynarray_fini(&batch->depth_bias);
-   util_dynarray_fini(&batch->occlusion_queries);
-   util_dynarray_fini(&batch->nonocclusion_queries);
-   util_dynarray_fini(&batch->timestamp_queries);
+   util_dynarray_fini(&batch->timestamps);
 
    if (!(dev->debug & (AGX_DBG_TRACE | AGX_DBG_SYNC))) {
       agx_batch_print_stats(dev, batch);
@@ -774,13 +762,6 @@ agx_batch_reset(struct agx_context *ctx, struct agx_batch *batch)
    agx_batch_cleanup(ctx, batch, true);
 }
 
-void
-agx_batch_add_timestamp_query(struct agx_batch *batch, struct agx_query *q)
-{
-   if (q)
-      util_dynarray_append(&batch->timestamp_queries, struct agx_query *, q);
-}
-
 /*
  * Timestamp queries record the time after all current work is finished,
  * which we handle as the time after all current batches finish (since we're a

src/gallium/drivers/asahi/agx_pipe.c

@@ -1333,19 +1333,6 @@ agx_flush_batch(struct agx_context *ctx, struct agx_batch *batch)
     */
    agx_batch_add_bo(batch, batch->vdm.bo);
 
-   /* Occlusion queries are allocated as a contiguous pool */
-   unsigned oq_count =
-      util_dynarray_num_elements(&batch->occlusion_queries, struct agx_query *);
-   size_t oq_size = oq_count * sizeof(uint64_t);
-
-   if (oq_size) {
-      batch->occlusion_buffer =
-         agx_pool_alloc_aligned(&batch->pool, oq_size, 64);
-      memset(batch->occlusion_buffer.cpu, 0, oq_size);
-   } else {
-      batch->occlusion_buffer.gpu = 0;
-   }
-
    if (batch->vs_scratch)
       agx_batch_add_bo(batch, ctx->scratch_vs.buf);
    if (batch->fs_scratch)