[Mesa-dev] [PATCH] intel: Fix broxton 2x6 way size computation

Anuj Phogat anuj.phogat at gmail.com
Wed Jun 7 18:44:49 UTC 2017


On Tue, Jun 6, 2017 at 4:51 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> On Tuesday, June 6, 2017 4:34:36 PM PDT Anuj Phogat wrote:
>> This patch is undoing the changes to way size computation
>> in broxton 2x6, made by below commit:
>>
>> Commit: 0d576fbfbe912cf3fb9ab594bb31eb58bccf2138
>> Author:     Anuj Phogat <anuj.phogat at gmail.com>
>> i965: Simplify l3 way size computations
>>
>> By making use of l3_banks field in gen_device_info struct
>> l3_way_size for gen7+ = 2 * l3_banks.
>>
>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101306
>> Signed-off-by: Anuj Phogat <anuj.phogat at gmail.com>
>> Cc: Jason Ekstrand <jason at jlekstrand.net>
>> Cc: Mark Janes <mark.a.janes at intel.com >
>> Cc: Francisco Jerez <currojerez at riseup.net>
>> ---
>> Note: Above bugzilla exposed a bug in our l3 allocation for
>> broxton 2x6. We need more changes to fix l3 config. I'll send
>> them later to the list. For now this patch brings things back
>> to where they were for bxt and unblocks the CI system to be
>> utilized for the performance work going on at present.
>> ---
>>  src/intel/common/gen_l3_config.c | 4 ++++
>>  1 file changed, 4 insertions(+)
>>
>> diff --git a/src/intel/common/gen_l3_config.c b/src/intel/common/gen_l3_config.c
>> index e0825e9..2520838 100644
>> --- a/src/intel/common/gen_l3_config.c
>> +++ b/src/intel/common/gen_l3_config.c
>> @@ -255,6 +255,10 @@ static unsigned
>>  get_l3_way_size(const struct gen_device_info *devinfo)
>>  {
>>     assert(devinfo->l3_banks);
>> +
>> +   if (devinfo->is_broxton)
>> +      return 4;
>> +
>>     return 2 * devinfo->l3_banks;
>>  }
>>
>>
>
> Yeah...it's strange, the docs indicate that there's only 1 bank of L3
> on Broxton 2x6, but it seems to have been working with 2...
>
> Acked-by: Kenneth Graunke <kenneth at whitecape.org>
>
> Your patches also changed the number of L3 banks in Kabylake GT 1.5.
> It now has more of them, matching GT2 instead of GT1.  I think that's
> correct by the documentation.
Right. It got doubled. I think Kabylake GT 1.5 performance will be
benefited by increased URB size. Adding mark in Cc.


More information about the mesa-dev mailing list