[Mesa-dev] [PATCH] gallium/u_threaded: replace pipe_resource_reference with atomic_inc

[Marek Olšák]

On Tue, Jun 13, 2017 at 5:18 AM, Michel Dänzer <michel at daenzer.net> wrote:
> On 13/06/17 01:31 AM, Marek Olšák wrote:
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> ---
>>  src/gallium/auxiliary/util/u_threaded_context.c | 9 +++++++--
>>  1 file changed, 7 insertions(+), 2 deletions(-)
>>
>> diff --git a/src/gallium/auxiliary/util/u_threaded_context.c b/src/gallium/auxiliary/util/u_threaded_context.c
>> index 3038fc6..a97f016 100644
>> --- a/src/gallium/auxiliary/util/u_threaded_context.c
>> +++ b/src/gallium/auxiliary/util/u_threaded_context.c
>> @@ -195,22 +195,27 @@ _tc_sync(struct threaded_context *tc, const char *info, const char *func)
>>
>>     tc_debug_check(tc);
>>  }
>>
>>  #define tc_sync(tc) _tc_sync(tc, "", __func__)
>>  #define tc_sync_msg(tc, info) _tc_sync(tc, info, __func__)
>>
>>  static void
>>  tc_set_resource_reference(struct pipe_resource **dst, struct pipe_resource *src)
>>  {
>> -   *dst = NULL;
>> -   pipe_resource_reference(dst, src);
>> +   /* Unexpectedly, pipe_resource_reference is a bottleneck here, taking 3.7%
>> +    * of CPU time in a state-heavy microbenchmark. The problem here is that
>> +    * pipe_resource_reference isn't inlined here.
>> +    */
>> +   *dst = src;
>> +   if (src)
>> +      p_atomic_inc(&src->reference.count);
>>  }
>
> That's pretty ugly. Does the attached patch help as well? It allows
> pipe_resource_reference to be inlined again, and has a pretty dramatic
> effect on the size of the generated code for me:
>
>    text    data     bss     dec     hex filename
>   42586     512       0   43098    a85a u_threaded_context.o
>   35282     512       0   35794    8bd2 u_threaded_context.o.patched
>
>    text    data     bss     dec     hex filename
> 10434626         271176 2062272 12768074         c2d34a /tmp/radeonsi_dri.so
> 10001218         271176 2062272 12334666         bc364a /tmp/radeonsi_dri.so.patched
>
>
> It might also help for other pipe_resource_reference callers.

Yes, the patch helps.

Marek

>
>
> --
> Earthling Michel Dänzer               |               http://www.amd.com
> Libre software enthusiast             |             Mesa and X developer