[Mesa-dev] [PATCH] r600g: Include SH and SMX when invalidating read caches

[Martin Andersson]

On Sun, Jun 23, 2013 at 7:41 PM, Alex Deucher <alexdeucher at gmail.com> wrote:
> On Sat, Jun 22, 2013 at 11:53 AM, Martin Andersson <g02maran at gmail.com> wrote:
>> On Sat, Jun 22, 2013 at 12:22 PM, Marek Olšák <maraeo at gmail.com> wrote:
>>> Reviewed-by: Marek Olšák <maraeo at gmail.com>
>>>
>>> BTW, SMX is a write cache, to maybe it shouldn't be part of this patch.
>>
>> I made a little experiment where i ran
>> "ext_framebuffer_multisample-unaligned-blit 4 color downsample -auto"
>> 10000 times and found that without SMX the test failed 177 times and
>> with SMX it didn't fail at all. So I do think the SMX cache should be
>> invalidated somewhere.
>>
>> Before http://cgit.freedesktop.org/mesa/mesa/commit/?id=4539f8e20af286d1f521eb016c89c6d9af0b801c
>> it was under R600_CONTEXT_FLUSH_AND_INV, is that a better place?
>>
>> If that is the proper place for SMX should SH also be there, since it
>> was also there before the patch, or do you have any other suggestions?
>
> Does something like this help?  You might play with some variants of
> this patch.  This might break compute however as Tom had some problems
> with CB flushes on cayman class hw in the past.

I tested this patch as it is and I found that it helped alot. Here is
a comparison of three piglit runs. Take these numbers with a grain of
salt, since the failures are random and I only did one run on each
experiment.

master:       11055/11781
my patch:    11216/11781
your patch: 11319/11781

The random failures are all but gone. I did three runs with you patch
and only found three tests that still fail randomly:

spec/!OpenGL 3.0/gl-3.0-texture-integer
spec/glsl-1.40/uniform_buffer/fs-struct-pad
spec/glsl-1.40/uniform_buffer/vs-struct-pad

//Martin

> Alex
>
>
>>
>>> Marek
>>>
>>> On Sun, Jun 16, 2013 at 1:27 PM, Martin Andersson <g02maran at gmail.com> wrote:
>>>> Not including the SH and SMX caches when invalidating read caches causes
>>>> random failures on some piglit tests when VA is enabled.
>>>>
>>>> Since the failures are random, and there other problems also causing random
>>>> failures, it's hard to know exactly what tests were effected, but these
>>>> tests now consistently pass:
>>>>
>>>> fast_color_clear/all-colors
>>>> fast_color_clear/redundant-clear
>>>> spec/!OpenGL 1.1/draw-pixels samples={2,4,6,8}
>>>> spec/!OpenGL 1.1/drawbuffer-modes
>>>> ---
>>>>  src/gallium/drivers/r600/r600_hw_context.c | 2 ++
>>>>  1 file changed, 2 insertions(+)
>>>>
>>>> diff --git a/src/gallium/drivers/r600/r600_hw_context.c b/src/gallium/drivers/r600/r600_hw_context.c
>>>> index 944b666..df20e56 100644
>>>> --- a/src/gallium/drivers/r600/r600_hw_context.c
>>>> +++ b/src/gallium/drivers/r600/r600_hw_context.c
>>>> @@ -231,6 +231,8 @@ void r600_flush_emit(struct r600_context *rctx)
>>>>         if (rctx->flags & R600_CONTEXT_INVAL_READ_CACHES) {
>>>>                 cp_coher_cntl |= S_0085F0_VC_ACTION_ENA(1) |
>>>>                                 S_0085F0_TC_ACTION_ENA(1) |
>>>> +                               S_0085F0_SH_ACTION_ENA(1) |
>>>> +                               S_0085F0_SMX_ACTION_ENA(1) |
>>>>                                 S_0085F0_FULL_CACHE_ENA(1);
>>>>                 emit_flush = 1;
>>>>         }
>>>> --
>>>> 1.8.3
>>>>
>>>> _______________________________________________
>>>> mesa-dev mailing list
>>>> mesa-dev at lists.freedesktop.org
>>>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev