gallium/u_threaded: don't map big VRAM buffers for the first upload directly

This improves Paraview "many spheres" performance 4x along with the radeonsi
commit.

Reviewed-by: Nicolai Hähnle <[email protected]>

2 files changed, 20 insertions, 0 deletions

diff --git a/src/gallium/auxiliary/util/u_threaded_context.c b/src/gallium/auxiliary/util/u_threaded_context.c
index 0f232580532..ccce12b00ce 100644
--- a/src/gallium/auxiliary/util/u_threaded_context.c
+++ b/src/gallium/auxiliary/util/u_threaded_context.c
@@ -1284,6 +1284,20 @@ tc_improve_map_buffer_flags(struct threaded_context *tc,
    if (usage & tc_flags)
       return usage;
 
+   /* Use the staging upload if it's preferred. */
+   if (usage & (PIPE_TRANSFER_DISCARD_RANGE |
+                PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) &&
+       !(usage & PIPE_TRANSFER_PERSISTENT) &&
+       /* Try not to decrement the counter if it's not positive. Still racy,
+        * but it makes it harder to wrap the counter from INT_MIN to INT_MAX. */
+       tres->max_forced_staging_uploads > 0 &&
+       p_atomic_dec_return(&tres->max_forced_staging_uploads) >= 0) {
+      usage &= ~(PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
+                 PIPE_TRANSFER_UNSYNCHRONIZED);
+
+      return usage | tc_flags | PIPE_TRANSFER_DISCARD_RANGE;
+   }
+
    /* Sparse buffers can't be mapped directly and can't be reallocated
     * (fully invalidated). That may just be a radeonsi limitation, but
     * the threaded context must obey it with radeonsi.
diff --git a/src/gallium/auxiliary/util/u_threaded_context.h b/src/gallium/auxiliary/util/u_threaded_context.h
index 8977b03cd20..ac7bc3dec73 100644
--- a/src/gallium/auxiliary/util/u_threaded_context.h
+++ b/src/gallium/auxiliary/util/u_threaded_context.h
@@ -241,6 +241,12 @@ struct threaded_resource {
     * pointers. */
    bool	is_shared;
    bool is_user_ptr;
+
+   /* If positive, prefer DISCARD_RANGE with a staging buffer over any other
+    * method of CPU access when map flags allow it. Useful for buffers that
+    * are too large for the visible VRAM window.
+    */
+   int max_forced_staging_uploads;
 };
 
 struct threaded_transfer {