Mesa (master): r300g: implement flush+sync (AKA glFinish)

[Marek Olšák]

Module: Mesa
Branch: master
Commit: 0b94c05c2827336e2abf46629590edf05a6b3f64
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=0b94c05c2827336e2abf46629590edf05a6b3f64

Author: Marek Olšák <maraeo at gmail.com>
Date:   Sun May 16 18:43:44 2010 +0200

r300g: implement flush+sync (AKA glFinish)

See comments in the code.

---

 src/gallium/drivers/r300/r300_flush.c |   36 +++++++++++++++++++++++++++++++++
 1 files changed, 36 insertions(+), 0 deletions(-)

diff --git a/src/gallium/drivers/r300/r300_flush.c b/src/gallium/drivers/r300/r300_flush.c
index 3d22027..d6876c1 100644
--- a/src/gallium/drivers/r300/r300_flush.c
+++ b/src/gallium/drivers/r300/r300_flush.c
@@ -1,5 +1,6 @@
 /*
  * Copyright 2008 Corbin Simpson <MostAwesomeDude at gmail.com>
+ * Copyright 2010 Marek Olšák <maraeo at gmail.com>
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -37,6 +38,8 @@ static void r300_flush(struct pipe_context* pipe,
     struct r300_context *r300 = r300_context(pipe);
     struct r300_query *query;
     struct r300_atom *atom;
+    struct pipe_framebuffer_state *fb;
+    unsigned i;
 
     CS_LOCALS(r300);
     (void) cs_count;
@@ -72,6 +75,39 @@ static void r300_flush(struct pipe_context* pipe,
     foreach(query, &r300->query_list) {
         query->flushed = TRUE;
     }
+
+    /* XXX
+     *
+     * This is a preliminary implementation of glFinish. Note that st/mesa
+     * uses a non-null fence when glFinish is called and then waits for
+     * the fence. Instead of returning the actual fence, we do the sync
+     * directly.
+     *
+     * The ideal implementation should use something like EmitIrqLocked and
+     * WaitIrq, or better, real fences.
+     *
+     * This feature degrades performance to the level of r300c for games that
+     * use glFinish a lot, even openarena does. Ideally we wouldn't need
+     * glFinish at all if we had proper throttling in swapbuffers so that
+     * the CPU wouldn't outrun the GPU by several frames, so this is basically
+     * a temporary fix for the input lag. Once swap&sync works with DRI2,
+     * I'll be happy to remove this code.
+     *
+     * - M. */
+    if (fence && r300->fb_state.state) {
+        fb = r300->fb_state.state;
+
+        for (i = 0; i < fb->nr_cbufs; i++) {
+            if (fb->cbufs[i]->texture) {
+                r300->rws->buffer_wait(r300->rws,
+                    r300_texture(fb->cbufs[i]->texture)->buffer);
+            }
+            if (fb->zsbuf) {
+                r300->rws->buffer_wait(r300->rws,
+                    r300_texture(fb->zsbuf->texture)->buffer);
+            }
+        }
+    }
 }
 
 void r300_init_flush_functions(struct r300_context* r300)