[Mesa-dev] [PATCH 8/9] drisw: Enable flush control for llvmpipe and softpipe

[Adam Jackson]

On Tue, 2015-05-12 at 23:34 +0200, Roland Scheidegger wrote:
> I like faster!
> 
> I don't really understand though where the wins come from.
> softpipe can't have much pending work in the first place (maybe tile
> cache?), llvmpipe potentially has the rasterizer threads I guess but 
> I'm not quite sure how this works...

Well the purpose of the series is eliminating the implicit glFlush when
switching contexts, so I think "calling flush way less" is probably the
entirety of the win.  There may not be much work to do, but the cost of
figuring that out is non-zero.  For softpipe, flushing less mostly
means reducing CPU cycle usage in an entirely cycle-bound execution
path.  For llvmpipe I think we're closer to being bound by data cache
throughput than by cycle count, so flushing less is reducing the amount
of redundant d$ work.

The testcase is admittedly pathological, the main loop looks like:

for each window
    minor setup
for each span in height
    for each window
        draw a bit
for each window
    swapbuffers

where you have three windows and a height of 100, and every window
switch is a MakeCurrent.  So eliminating the flush on ctxsw is
eliminating >300 flushes per frame.

- ajax