[Intel-gfx] DG1 VRAM question

[Matthew Auld]

On 01/07/2020 00:27, Dave Airlie wrote:
> On Sat, 27 Jun 2020 at 03:17, Daniele Ceraolo Spurio
> <daniele.ceraolospurio at intel.com> wrote:
>>
>>
>>
>> On 6/26/20 12:14 AM, Lucas De Marchi wrote:
>>> Cc Matt and Daniele
>>>
>>> On Thu, Jun 25, 2020 at 9:38 PM Dave Airlie <airlied at gmail.com> wrote:
>>>>
>>>> I can't figure this out easily so I'd thought I'd just ask, but does
>>>> DG1 have VRAM > PCIE aperture, I'm not sure I've see any mention of
>>>
>>> We'd need to go via lmem since there's no mappable aperture. There are
>>> a few patches in tree for that
>>> (see e.g. 54b512cd7a6d ("drm/i915: do not map aperture if it is not
>>> available.")) but more missing.
>>>
>>
>> To clarify, although the legacy aperture mapping that allowed the CPU to
>> access memory via the GGTT for swizzling is gone, VRAM/LMEM is still
>> cpu-mappable via pci bar.
>> Will leave the questions about possible trashing to Matt as he's more
>> familiar than me with how this works.
> 
> Matt?
> 
> Is DG1 assuming we can get 64-bit BARs always and the CPU will have
> access to the complete VRAM? or is there any ideas about what happens
> in those situations where 64-bit BARs aren't available and there is
> memory pressure on the PCI BAR space.
> 
> With other discrete GPUs we've got lots of things like visible VRAM
> limitations, writing page tables with GPU hw instead of from the CPU,
> having mapping bring things into the visible area, so you can stream
> something into VRAM, but then it'll migrated to non-visible area if
> it's unmapped and there is memory pressure.

Yes, we just assume that LMEM size == LMEMBAR size, where the whole 
thing is 1:1 mapped and CPU visible. We don't currently have the concept 
of CPU visible/non-visible LMEM.

> 
> Dave.
>