Re: [PATCH v5 0/2] Fixes for MI_REPORT_PERF_COUNT

[Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>]

On Fri, Dec 20, 2024 at 11:54:04AM -0800, Souza, Jose wrote:
>On Fri, 2024-12-20 at 09:19 -0800, Umesh Nerlige Ramappa wrote:
>> OA programming sequence for query mode or MI_REPORT_PERF_COUNT requires
>> modifying some HW registers in the same hw context as the user exec
>> queue. User passes the exec_queue to the OA interface and OA
>> implementation submits an MI_LOAD_REGISTER_IMM to this queue to modify
>> the registers.
>>
>> The OA implementation submits a batch mapped in GGTT to the user exec
>> queue and hence, some plumbing is added into relevant code to enable
>> that (as per suggestions from Matthew Brost).
>>
>> v2: review rework
>> v3:
>> - review rework
>> - original patches squashed for porting to stable
>> - code cleanup
>>
>> v4:
>> - review rework/fixes
>>
>> v5:
>> - minor review rework/fixes
>
>v5 fixed the oops that I got in v4 and also fixed the GPU hang when closing OA stream.

Thanks a lot for catching that in v4 and ensuring v5 works.

Regards,
Umesh
>
>>
>> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
>>
>> Umesh Nerlige Ramappa (2):
>>   xe/oa: Fix query mode of operation for OAR/OAC
>>   xe/oa: Drop the unused logic to parse context image
>>
>>  drivers/gpu/drm/xe/xe_oa.c              | 224 +++++-------------------
>>  drivers/gpu/drm/xe/xe_oa_types.h        |   3 -
>>  drivers/gpu/drm/xe/xe_ring_ops.c        |   5 +-
>>  drivers/gpu/drm/xe/xe_sched_job_types.h |   2 +
>>  4 files changed, 51 insertions(+), 183 deletions(-)
>>
>