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 <nicolai.haehnle@amd.com>

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.

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 {

src/gallium/drivers/radeon/r600_buffer_common.c

@@ -191,10 +191,15 @@ void si_init_resource_fields(struct r600_common_screen *rscreen,
 	res->vram_usage = 0;
 	res->gart_usage = 0;
 
-	if (res->domains & RADEON_DOMAIN_VRAM)
+	if (res->domains & RADEON_DOMAIN_VRAM) {
 		res->vram_usage = size;
-	else if (res->domains & RADEON_DOMAIN_GTT)
+
+		res->b.max_forced_staging_uploads =
+			rscreen->info.has_dedicated_vram &&
+			size >= rscreen->info.vram_vis_size / 4 ? 1 : 0;
+	} else if (res->domains & RADEON_DOMAIN_GTT) {
 		res->gart_usage = size;
+	}
 }
 
 bool si_alloc_resource(struct r600_common_screen *rscreen,
@@ -289,6 +294,7 @@ void si_replace_buffer_storage(struct pipe_context *ctx,
 	pb_reference(&rdst->buf, rsrc->buf);
 	rdst->gpu_address = rsrc->gpu_address;
 	rdst->b.b.bind = rsrc->b.b.bind;
+	rdst->b.max_forced_staging_uploads = rsrc->b.max_forced_staging_uploads;
 	rdst->flags = rsrc->flags;
 
 	assert(rdst->vram_usage == rsrc->vram_usage);