Re: [PATCH] drm/amdgpu: optimize the padding with hw optimization

["Christian König" <christian.koenig@amd.com>]

[-- Attachment #1: Type: text/plain, Size: 1375 bytes --]

Am 01.08.24 um 08:53 schrieb Marek Olšák:
> On Thu, Aug 1, 2024, 00:28 Khatri, Sunil <sukhatri@amd.com> wrote:
>
>
>     On 8/1/2024 8:49 AM, Marek Olšák wrote:
>     >> +       /* Header is at index 0, followed by num_nops - 1 NOP
>     packet's */
>     >> +       for (i = 1; i < num_nop; i++)
>     >> +               amdgpu_ring_write(ring, ring->funcs->nop);
>     > This loop should be removed. It's unnecessary CPU overhead and we
>     > should never get more than 0x3fff NOPs (maybe use BUG_ON).
>     Leaving the
>     > whole packet body uninitialized is the fastest option.
>     That was the original intent to just move the WPTR for the no of nops
>     and tried too. Based on Christian inputs we should not let the
>     nops packet
>
>     as garbage or whatever was there originally as a threat/safety
>     measure.
>
>
> It doesn't help safety. It can only be read by the GPU with 
> kernel-level permissions.
>
> Initializing the packet body is useless and adds CPU overhead, 
> especially with the 256 NOPs or so that we use for no reason.

Not filling the remaining ring buffers with NOPs is a pretty clear NAK 
from my side. Leaving garbage in the ring buffer is not even remotely 
defensive.

What we can do is to optimize filling N DWs into the ring buffer without 
updating the WPTR each time.

Regards,
Christian.

>
> Marek
>

[-- Attachment #2: Type: text/html, Size: 2886 bytes --]