[Nouveau] Second copy engine on GF116

[Ilia Mirkin]

On Mon, Nov 24, 2014 at 8:33 PM, Andy Ritger <aritger at nvidia.com> wrote:
> On Fri, Nov 21, 2014 at 01:39:55AM -0500, Ilia Mirkin wrote:
>> On Fri, Nov 21, 2014 at 1:16 AM, Andy Ritger <aritger at nvidia.com> wrote:
>> > Hi Ilia,
>> >
>> > Actually 0x90b8 is different than copy engine.  I'm not very familiar
>> > with it, but 0x90b8 is an engine for performing LZO decompression as
>> > part of performing the copy.  It has a variety of limitations (e.g.,
>> > cannot handle blocklinear format), and was only in a few Fermi chips,
>> > as I understand it.
>>
>> According to our driver source, GF100, GF104, GF110, GF114, and GF116
>> all have it. [So GF106, GF108, GF117, GF119 don't have it.] We've only
>> had problems reported against GF116... and only for some people.
>
> Hmm, some of our internal documentation is inconsistent about whether it
> applies to GF100, but otherwise what I see matches your list.  I guess
> "few" was not entirely accurate.
>
>> > It is probably easiest to just ignore it.  You can distinguish this
>> > decompress engine from normal copy engine by looking at the CE capability
>> > register on falcon (0x00000650).  If bit 2 is '1', then the falcon is
>> > a decompress engine.
>>
>> I presume you mean a +0x650 register on the pcopy engines (0x104000
>> and 0x105000). I only have access to the GF108 right now, which
>> returns 3 for 0x104650 and 4 for 0x105650. We're using the engine at
>> 0x104000 for copy on the GF108...
>
> Yes, 0x104650 and 0x105650 are the right addresses, from what I can tell.
>
> FWIW, the other capability bits are:
> bit 0: "DMACOPY_SUPPORTED"
> bit 1: "PIXREMAP_SUPPORTED"
>
> (I think PIXREMAP_SUPPORTED is in reference to the component remapping
> controlled by methods 0x00000700, 0x00000704, and 0x00000708 in the
> copy engine class).

Neat. We went around and grabbed that 0x650 register on a bunch of
GPUs, see the CE* columns at:

http://envytools.readthedocs.org/en/latest/hw/gpu.html#fermi-kepler-maxwell-family

It looks like it's actually returning 0 on both "copy" engines for a
bunch of those cards -- GF100, GF104, GF114, probably GF110. But other
cards have them as either 3 or 4. I'm guessing that '0' should be
treated as if it were a '3' (or a '7')?

Curiously, a GF116 card that I thought was working fine on nouveau
actually has 3 for the first engine and 4 for the second. Perhaps it
just had enough VRAM that I never triggered the conditions required
for nouveau to use that second copy engine (we use it, when available,
for drm-initiated buffer moves).

>
>> From my admittedly limited understanding, both 0x104000 and 0x105000
>> appear to be falcon engines, where the fuc is presumably able to drive
>> some underlying hardware. The actual fifo methods are implemented in
>> the fuc, which in turn does iowr/etc commands.
>>
>> Are you saying that the "decompress" engine (at 0x105000 right?) has a
>> different piece of hardware behind it than the copy engine at
>> 0x104000, or does NVIDIA simply provide different fuc for it that
>> exposes somewhat different functionality via FIFO methods?
>
> There is definitely a falcon at the frontend, and there is different
> falcon ucode for "normal" copy engine versus the "decompress" engine.
> But, I don't know off hand what dedicated hardware, if any, is behind it.

Seems likely that the HW is different, since it'd be madness to try to
do decompression in the falcon code itself. (Not to say that the ISA
isn't suited to it, just they have relatively slow clocks.) mwk is in
the process of working it all out.

  -ilia