virgl: use virgl_transfer_inline_write even less
authorGurchetan Singh <gurchetansingh@chromium.org>
Thu, 7 Feb 2019 00:26:18 +0000 (16:26 -0800)
committerGert Wollny <gert.wollny@collabora.com>
Fri, 15 Feb 2019 10:19:05 +0000 (11:19 +0100)
We've noticed the Team Fortress 2 engine seems to do many small
calls to glSubData(..). Let's pick our heuristic based on the
resource base width, not the size of a particular upload.
This will cause transfers to be batched together in the transfer
queue.

Revelant glbench microbenchmark --

Before: buffer_upload_dynamic_element_array_131072 = 131.17 mbytes_sec
After: buffer_upload_dynamic_element_array_131072 = 6828.24 mbytes_sec
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
src/gallium/drivers/virgl/virgl_resource.c

index a21e911..2a1e0c1 100644 (file)
@@ -147,7 +147,7 @@ static void virgl_buffer_subdata(struct pipe_context *pipe,
 
    u_box_1d(offset, size, &box);
 
-   if (size >= (VIRGL_MAX_CMDBUF_DWORDS * 4))
+   if (resource->width0 >= getpagesize())
       u_default_buffer_subdata(pipe, resource, usage, offset, size, data);
    else
       virgl_transfer_inline_write(pipe, resource, 0, usage, &box, data, 0, 0);