[Mesa-dev] Mesa (master): i965: Double the size of the state cache.

[Kenneth Graunke]

On 08/27/2012 03:00 PM, Stéphane Marchesin wrote:
> On Mon, Aug 27, 2012 at 2:24 PM, Kenneth Graunke
> <kwg at kemper.freedesktop.org> wrote:
>> Module: Mesa
>> Branch: master
>> Commit: 87cdefed405da1a57a006737297d20ba0c25fa1b
>> URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=87cdefed405da1a57a006737297d20ba0c25fa1b
>>
>> Author: Kenneth Graunke <kenneth at whitecape.org>
>> Date:   Sun Aug 26 00:07:56 2012 -0700
>>
>> i965: Double the size of the state cache.
>>
>> We probably want to do something more sophisticated here, but this at
>> least makes it through L4D2 without dumping the program cache.
>>
>> Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
>> Reviewed-by: Eric Anholt <eric at anholt.net>
>>
>> ---
>>
>>  src/mesa/drivers/dri/i965/brw_state_cache.c |    4 ++--
>>  1 files changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_state_cache.c b/src/mesa/drivers/dri/i965/brw_state_cache.c
>> index 57a5ee9..8823b22 100644
>> --- a/src/mesa/drivers/dri/i965/brw_state_cache.c
>> +++ b/src/mesa/drivers/dri/i965/brw_state_cache.c
>> @@ -372,10 +372,10 @@ brw_clear_cache(struct brw_context *brw, struct brw_cache *cache)
>>  void
>>  brw_state_cache_check_size(struct brw_context *brw)
>>  {
>> -   /* un-tuned guess.  Each object is generally a page, so 1000 of them is 4 MB of
>> +   /* un-tuned guess.  Each object is generally a page, so 2000 of them is 8 MB of
> 
> The problem is that this statement ("Each object is generally a page") is wrong.
> 
> A wm prog cache entry always contains prog_data as aux data.
> brw_wm_prog_data has those two fields:
> const float *param[MAX_UNIFORMS * 4]; /* should be: BRW_MAX_CURBE */
> const float *pull_param[MAX_UNIFORMS * 4];
> 
> MAX_UNIFORMS is 4096. So in 64 bit, brw_wm_prog_data is more than 256k.
> 
> Therefore by doubling the cache, you are going from 256 MB to 512 MB,
> not 4 MB to 8 MB. As far as I know, that cache is the biggest consumer
> of memory in i965.
> 
> Stéphane

Ouch!  Clearly, having 32768 pointers in the key is absurd...especially
since programs usually only use a few uniforms.

Thankfully, Eric has a three patch series which fixes the problem by
dynamically allocating the arrays:
http://lists.freedesktop.org/archives/mesa-dev/2012-August/026403.html

Hopefully that will land soon.  I can revert the patch in the meantime
if you would prefer.

I also think we probably want to do something smarter with the cache.
Evicting everything after 1000 is dumb.  Arguably, evicting everything
after 2000 is also dumb.  It was just meant as a stop-gap.  And
apparently not a very good one...