Deprecation of AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED

[Marek Olšák]

On Mon, Jul 3, 2017 at 12:08 PM, Michel Dänzer <michel at daenzer.net> wrote:
> On 30/06/17 08:43 PM, Marek Olšák wrote:
>>
>> I don't know what is being talked about here anymore, but I wouldn't
>> like to use CPU_ACCESS_REQUIRED or CPU_ACCESS_REALLY_REQUIRED in
>> userspace. The reason is that userspace doesn't and can't know whether
>> CPU access will be required, and the frequency at which it will be
>> required. 3 heaps {no CPU access, no flag, CPU access required} are
>> too many. Userspace mostly doesn't use the "no flag" heap for VRAM. It
>> uses "CPU access required" for almost everything except tiled
>> textures, which use "no CPU access".
>
> FWIW, the difference between setting CPU_ACCESS_REQUIRED and not setting
> it for a BO created in VRAM will be: If it's set, the BO is initially
> created in CPU visible VRAM, otherwise it's most likely created in CPU
> invisible VRAM.
>
> If userspace knows that a BO will likely be accessed by the CPU first,
> setting the flag could save a move from CPU invisible to CPU visible
> VRAM when the CPU access happens. Conversely, if a BO will likely never
> be accessed by the CPU, not setting the flag may reduce pressure on CPU
> visible VRAM.
>
> Not sure radeonsi can make this distinction though.

It can't.

Either all mappable BOs set CPU_ACCESS_REQUIRED, or all mappable BOs
don't set it. Either way, there is only one combination of flags for
mappable BOs in VRAM, and therefore only one kind of behavior the
kernel can follow.

>
>
>> I've been trying to trim down the number of heaps. So far, I have:
>> - VRAM_NO_CPU_ACCESS (implies WC)
>> - VRAM (implies WC)
>> - VRAM_GTT (combined, implies WC)
>
> Is this useful? It means:
>
> * The BO may be created in VRAM, or if there's no space, in GTT.
> * Once the BO is in GTT for any reason, it will never go back to VRAM.
>
> Such BOs will tend to end up in GTT after some time, at the latest after
> suspend/resume.
>
> I think it would be better for radeonsi to choose either VRAM or GTT as
> the preferred domain, and let the kernel handle it.

Currently, radeonsi on amdgpu doesn't use VRAM_GTT with the current kernel.

I'm aware of the limited usefulness.

Marek