[PATCH v4 1/9] drm/xe/xelpg: Move Wa_14016712196 to the invalidate path
Tvrtko Ursulin
tvrtko.ursulin at igalia.com
Tue Apr 1 08:46:34 UTC 2025
On 31/03/2025 19:54, Rodrigo Vivi wrote:
> On Fri, Mar 28, 2025 at 04:35:28PM +0000, Tvrtko Ursulin wrote:
>> According to i915 Wa_14016712196 needs to be emmited before a
>> pipe control which contains a post sync operation.
>>
>> Therefore move it from flush (no post sync) to invalidate (post sync).
>>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at igalia.com>
>> Fixes: 8c5fe7d88bc1 ("drm/xe: Add Wa_16021333562 and Wa_14016712196")
>> Cc: Tejas Upadhyay <tejas.upadhyay at intel.com>
>> Cc: Aradhya Bhatia <aradhya.bhatia at intel.com>
>> Cc: Matt Roper <matthew.d.roper at intel.com>
>> Cc: Lucas De Marchi <lucas.demarchi at intel.com>
>> ---
>> Please double check.
>
> It looks like both options are possible.
>
> We either insert a PIPE_CONTROL with "**Depth Flush** " post any state that
> will send an implicit depth flush and prior to any PIPE_CONTROL. This
> PIPE_CONTROL is not required if 3DPRIMITIVE, 3DMESH or PIPE_CONTROL that
> doesn't require end of the pipe drain is programmed prior to a PIPE_CONTROL
> that requires end of the pipe drain and hits this issue.
>
> or
>
> We insert a this pipe_control prior to PIPE_CONTROL which will hit the issue. For
> timestamp, this could be a replicated pipe_control as it will write the
> correct value the 2nd time. For post sync with write immediate, SW would have
> to allocate a dummy address.
>
> So, is changing the order really helping your case or needed for your case?
No, I don't have access to Meteorlake so I was just comparing all the
differences between i915 and xe in this area. Given Juha-Pekka reported
the previous version of this series did not work there. I suspect that
problem was I simply miscounted MAX_JOB_SIZE_DW. That should be fixed in
this v4. So yeah, if you say this patch is wrong or not needed I will
drop it.
Regards,
Tvrtko
>> ---
>> drivers/gpu/drm/xe/xe_ring_ops.c | 16 +++++++++-------
>> 1 file changed, 9 insertions(+), 7 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_ring_ops.c b/drivers/gpu/drm/xe/xe_ring_ops.c
>> index 917fc16de866..88591b7a7715 100644
>> --- a/drivers/gpu/drm/xe/xe_ring_ops.c
>> +++ b/drivers/gpu/drm/xe/xe_ring_ops.c
>> @@ -134,8 +134,9 @@ emit_pipe_control(u32 *dw, int i, u32 bit_group_0, u32 bit_group_1, u32 offset,
>> return i;
>> }
>>
>> -static int emit_pipe_invalidate(u32 mask_flags, bool invalidate_tlb, u32 *dw,
>> - int i)
>> +static int
>> +emit_pipe_invalidate(struct xe_gt *gt, u32 mask_flags, bool invalidate_tlb,
>> + u32 *dw, int i)
>> {
>> u32 flags = PIPE_CONTROL_CS_STALL |
>> PIPE_CONTROL_COMMAND_CACHE_INVALIDATE |
>> @@ -152,6 +153,10 @@ static int emit_pipe_invalidate(u32 mask_flags, bool invalidate_tlb, u32 *dw,
>>
>> flags &= ~mask_flags;
>>
>> + if (XE_WA(gt, 14016712196))
>> + i = emit_pipe_control(dw, i, 0, PIPE_CONTROL_DEPTH_CACHE_FLUSH,
>> + LRC_PPHWSP_FLUSH_INVAL_SCRATCH_ADDR, 0);
>> +
>> return emit_pipe_control(dw, i, 0, flags, LRC_PPHWSP_FLUSH_INVAL_SCRATCH_ADDR, 0);
>> }
>>
>> @@ -173,10 +178,6 @@ static int emit_render_cache_flush(struct xe_sched_job *job, u32 *dw, int i)
>> bool lacks_render = !(gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK);
>> u32 flags;
>>
>> - if (XE_WA(gt, 14016712196))
>> - i = emit_pipe_control(dw, i, 0, PIPE_CONTROL_DEPTH_CACHE_FLUSH,
>> - LRC_PPHWSP_FLUSH_INVAL_SCRATCH_ADDR, 0);
>> -
>> flags = (PIPE_CONTROL_CS_STALL |
>> PIPE_CONTROL_TILE_CACHE_FLUSH |
>> PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
>> @@ -361,7 +362,8 @@ static void __emit_job_gen12_render_compute(struct xe_sched_job *job,
>> mask_flags = PIPE_CONTROL_3D_ENGINE_FLAGS;
>>
>> /* See __xe_pt_bind_vma() for a discussion on TLB invalidations. */
>> - i = emit_pipe_invalidate(mask_flags, job->ring_ops_flush_tlb, dw, i);
>> + i = emit_pipe_invalidate(gt, mask_flags, job->ring_ops_flush_tlb, dw,
>> + i);
>>
>> /* hsdes: 1809175790 */
>> if (has_aux_ccs(xe))
>> --
>> 2.48.0
>>
More information about the Intel-xe
mailing list