mesa/st: short-circuit glFinish calls on WIN32 only

Windows opengl32.dll calls glFinish prior to every swapbuffers, which
makes it pretty hard to get decent performance...

Work around by mapping finish to flush on PIPE_OS_WINDOWS.  This is
conformant, though it might confuse poorly-written benchmarks which
attempt to measure a single event rather than figuring out the rate of
continuous processing.

1 files changed, 12 insertions, 0 deletions

diff --git a/src/mesa/state_tracker/st_cb_flush.c b/src/mesa/state_tracker/st_cb_flush.c
index fbaffd154f..f7de4e7a4d 100644
--- a/src/mesa/state_tracker/st_cb_flush.c
+++ b/src/mesa/state_tracker/st_cb_flush.c
@@ -152,4 +152,16 @@ void st_init_flush_functions(struct dd_function_table *functions)
 {
    functions->Flush = st_glFlush;
    functions->Finish = st_glFinish;
+
+   /* Windows opengl32.dll calls glFinish prior to every swapbuffers.
+    * This is unnecessary and degrades performance.  Luckily we have some
+    * scope to work around this, as the externally-visible behaviour of
+    * Finish() is identical to Flush() in all cases - no differences in
+    * rendering or ReadPixels are visible if we opt not to wait here.
+    *
+    * Only set this up on windows to avoid suprise elsewhere.
+    */
+#ifdef PIPE_OS_WINDOWS
+   functions->Finish = st_glFlush;
+#endif
 }