On Monday, April 3, 2023 9:48:40 AM PDT Ville Syrjälä wrote:
> On Mon, Apr 03, 2023 at 09:35:32AM -0700, Matt Roper wrote:
> > On Mon, Apr 03, 2023 at 07:02:08PM +0300, Ville Syrjälä wrote:
> > > On Fri, Mar 31, 2023 at 11:38:30PM -0700, fei.yang@xxxxxxxxx wrote:
> > > > From: Fei Yang <fei.yang@xxxxxxxxx>
> > > >
> > > > To comply with the design that buffer objects shall have immutable
> > > > cache setting through out its life cycle, {set, get}_caching ioctl's
> > > > are no longer supported from MTL onward. With that change caching
> > > > policy can only be set at object creation time. The current code
> > > > applies a default (platform dependent) cache setting for all objects.
> > > > However this is not optimal for performance tuning. The patch extends
> > > > the existing gem_create uAPI to let user set PAT index for the object
> > > > at creation time.
> > >
> > > This is missing the whole justification for the new uapi.
> > > Why is MOCS not sufficient?
> >
> > PAT and MOCS are somewhat related, but they're not the same thing. The
> > general direction of the hardware architecture recently has been to
> > slowly dumb down MOCS and move more of the important memory/cache
> > control over to the PAT instead. On current platforms there is some
> > overlap (and MOCS has an "ignore PAT" setting that makes the MOCS "win"
> > for the specific fields that both can control), but MOCS doesn't have a
> > way to express things like snoop/coherency mode (on MTL), or class of
> > service (on PVC). And if you check some of the future platforms, the
> > hardware design starts packing even more stuff into the PAT (not just
> > cache behavior) which will never be handled by MOCS.
>
> Sigh. So the hardware designers screwed up MOCS yet again and
> instead of getting that fixed we are adding a new uapi to work
> around it?
>
> The IMO sane approach (which IIRC was the situation for a few
> platform generations at least) is that you just shove the PAT
> index into MOCS (or tell it to go look it up from the PTE).
> Why the heck did they not just stick with that?
There are actually some use cases in newer APIs where MOCS doesn't
work well. For example, VK_KHR_buffer_device_address in Vulkan 1.2:
https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VK_KHR_buffer_device_address.html
It essentially adds "pointers to buffer memory in shaders", where apps
can just get a 64-bit pointer, and use it with an address. On our EUs,
that turns into A64 data port messages which refer directly to memory.
Notably, there's no descriptor (i.e. SURFACE_STATE) where you could
stuff a MOCS value. So, you get one single MOCS entry for all such
buffers...which is specified in STATE_BASE_ADDRESS. Hope you wanted
all of them to have the same cache & coherency settings!
With PAT/PTE, we can at least specify settings for each buffer, rather
than one global setting.
Compression has also been moving towards virtual address-based solutions
and handling in the caches and memory controller, rather than in e.g.
the sampler reading SURFACE_STATE. (It started evolving that way with
Tigerlake, really, but continues.)
> > Also keep in mind that MOCS generally applies at the GPU instruction
> > level; although a lot of instructions have a field to provide a MOCS
> > index, or can use a MOCS already associated with a surface state, there
> > are still some that don't. PAT is the source of memory access
> > characteristics for anything that can't provide a MOCS directly.
>
> So what are the things that don't have MOCS and where we need
> some custom cache behaviour, and we already know all that at
> buffer creation time?
For Meteorlake...we have MOCS for cache settings. We only need to use
PAT for coherency settings; I believe we can get away with deciding that
up-front at buffer creation time. If we were doing full cacheability,
I'd be very nervous about deciding performance tuning at creation time.
--Ken
Attachment:
signature.asc
Description: This is a digitally signed message part.
