Re: [Intel-gfx] [PATCH v2] drm/i915/dg2: Add performance workaround 18019455067

[Lucas De Marchi <lucas.demarchi@intel.com>]

On Wed, Jun 29, 2022 at 03:16:09PM -0700, Matt Roper wrote:
>On Mon, Jun 27, 2022 at 03:59:28PM +0300, Lionel Landwerlin wrote:
>> The recommended number of stackIDs for Ray Tracing subsystem is 512
>> rather than 2048 (default HW programming).
>>
>> v2: Move the programming to dg2_ctx_gt_tuning_init() (Lucas)
>
>I'm not sure this is actually the correct move.  As far as I can see on
>bspec 46261, RT_CTRL isn't part of the engine's context, so we need to
>make sure it gets added to engine->wa_list instead of
>engine->ctx_wa_list, otherwise it won't be properly re-applied after
>engine resets and such.  Most of our other tuning values are part of the
>context image, so this one is a bit unusual.
>
>To get it onto the engine->wa_list, the workaround needs to either be
>defined via rcs_engine_wa_init() or general_render_compute_wa_init().
>The latter is the new, preferred location for registers that are part of
>the render/compute reset domain, but that don't live in the RCS engine's
>0x2xxx MMIO range (since all RCS and CCS engines get reset together, the
>items in general_render_compute_wa_init() will make sure it's dealt with
>as part of the handling for the first RCS/CCS engine, so that we won't
>miss out on applying it if the platform doesn't have an RCS).
>
>At the moment we don't have too many "tuning" values that we need to set
>that aren't part of an engine's context, so we don't yet have a
>dedicated "tuning" function for engine-style workarounds like we do with
>ctx-style workarounds.


what I meant on my review was not to move it to
dg2_ctx_gt_tuning_init(), but rather to follow the same logic: we need
an equivalent tuning version for engine wa.

Lucas De Marchi