[PATCH 10/10] drm/amdgpu: add gang submit frontend

[Mohan Marimuthu, Yogesh]

[Public]

Hi Christian,

No other comments. With p->jobs[i] fixed, the test case worked. I have to clean up the code and send it for review.
I wanted to add comparing the time with and without gang submission and fail test case if former is slow. I will do this later. I will send the test case for review first.

Thank you,
Yogesh

-----Original Message-----
From: Koenig, Christian <Christian.Koenig at amd.com> 
Sent: Wednesday, June 1, 2022 5:42 PM
To: Mohan Marimuthu, Yogesh <Yogesh.Mohanmarimuthu at amd.com>; Christian König <ckoenig.leichtzumerken at gmail.com>; amd-gfx at lists.freedesktop.org; Olsak, Marek <Marek.Olsak at amd.com>
Subject: Re: [PATCH 10/10] drm/amdgpu: add gang submit frontend



Am 01.06.22 um 14:09 schrieb Mohan Marimuthu, Yogesh:
> [SNIP]
> -	/* Make sure all BOs are remembered as writers */
> -	amdgpu_bo_list_for_each_entry(e, p->bo_list)
> +	list_for_each_entry(e, &p->validated, tv.head) {
> +
> +		/* Everybody except for the gang leader uses BOOKKEEP */
> +		for (i = 0; i < (p->gang_size - 1); ++i) {
> +			dma_resv_add_fence(e->tv.bo->base.resv,
> +					   &p->jobs[i]->base.s_fence->finished,
> +					   DMA_RESV_USAGE_BOOKKEEP);
> +		}
> +
> +		/* The gang leader as remembered as writer */
>   		e->tv.num_shared = 0;
> +	}
>
>
> p->jobs[i] = NULL is done in previous loop. Now when running &p->jobs[i]->base.s_fence->finished there is NULL pointer crash.

Ah, yes good point. Going to fix that.

Any more comments on this? Did you finished the test cases?

Thanks,
Christian.