[PATCH 3/3] drm/amdgpu: enable gfx wave limiting for high priority compute jobs

Nirmoy nirmodas at amd.com
Mon Feb 1 13:56:07 UTC 2021


On 2/1/21 1:19 PM, Christian König wrote:
> Am 01.02.21 um 13:07 schrieb Nirmoy Das:
>> Enable gfx wave limiting for gfx jobs before pushing high priority
>> compute jobs so that high priority compute jobs gets more resources
>> to finish early.
>>
>> Signed-off-by: Nirmoy Das <nirmoy.das at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 9 +++++++++
>>   1 file changed, 9 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>> index 024d0a563a65..ee48989dfb4c 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>> @@ -195,6 +195,10 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, 
>> unsigned num_ibs,
>>       if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && 
>> ring->funcs->emit_mem_sync)
>>           ring->funcs->emit_mem_sync(ring);
>>   +    if (ring->funcs->emit_wave_limit && job &&
>> +        job->base.s_priority >= DRM_SCHED_PRIORITY_HIGH)
>> +        ring->funcs->emit_wave_limit(ring, true);
>
> Since we can only do this for one ring anyway we should probably check 
> the ring priority instead of the job priority.



>
> Alternatively you could put this into begin_use and end_use callbacks 
> of the ring instead of adding an extra callback for this.


I didn't know about begin_use callbacks. I will resend with above 
suggested changes.


Thanks,

Nirmoy


>
> Christian.
>
>> +
>>       if (ring->funcs->insert_start)
>>           ring->funcs->insert_start(ring);
>>   @@ -295,6 +299,11 @@ int amdgpu_ib_schedule(struct amdgpu_ring 
>> *ring, unsigned num_ibs,
>>       ring->current_ctx = fence_ctx;
>>       if (vm && ring->funcs->emit_switch_buffer)
>>           amdgpu_ring_emit_switch_buffer(ring);
>> +
>> +    if (ring->funcs->emit_wave_limit && job &&
>> +        job->base.s_priority >= DRM_SCHED_PRIORITY_HIGH)
>> +        ring->funcs->emit_wave_limit(ring, false);
>> +
>>       amdgpu_ring_commit(ring);
>>       return 0;
>>   }
>


More information about the amd-gfx mailing list