Re: [PATCH 00/18] Add Hantro regmap and VC8000 h264 decode support

[Ezequiel Garcia]

Hi Robin,

On Thu, 2020-10-29 at 14:15 +0000, Robin Murphy wrote:
> On 2020-10-29 13:07, Ezequiel Garcia wrote:
> > Hello Adrian,
> > 
> > On Mon, 2020-10-12 at 23:59 +0300, Adrian Ratiu wrote:
> > > Dear all,
> > > 
> > > This series introduces a regmap infrastructure for the Hantro driver
> > > which is used to compensate for different HW-revision register layouts.
> > > To justify it h264 decoding capability is added for newer VC8000 chips.
> > > 
> > > This is a gradual conversion to the new infra - a complete conversion
> > > would have been very big and I do not have all the HW yet to test (I'm
> > > expecting a RK3399 shipment next week though ;). I think converting the
> > > h264 decoder provides a nice blueprint for how the other codecs can be
> > > converted and enabled for different HW revisions.
> > > 
> > > The end goal of this is to make the driver more generic and eliminate
> > > entirely custom boilerplate like `struct hantro_reg` or headers with
> > > core-specific bit manipulations like `hantro_g1_regs.h` and instead rely
> > > on the well-tested albeit more verbose regmap subsytem.
> > > 
> > > To give just two examples of bugs which are easily discovered by using
> > > more verbose regmap fields (very easy to compare with the datasheets)
> > > instead of relying on bit-magic tricks: G1_REG_DEC_CTRL3_INIT_QP(x) was
> > > off-by-1 and the wrong .clk_gate bit was set in hantro_postproc.c.
> > > 
> > > Anyway, this series also extends the MMIO regmap API to allow relaxed
> > > writes for the theoretical reason that avoiding unnecessary membarriers
> > > leads to less CPU usage and small improvements to battery life. However,
> > > in practice I could not measure differences between relaxed/non-relaxed
> > > IO, so I'm on the fence whether to keep or remove the relaxed calls.
> > > 
> > > What I could masure is the performance impact of adding more sub-reg
> > > field acesses: a constant ~ 20 microsecond bump per G1 h264 frame. This
> > > is acceptable considering the total time to decode a frame takes three
> > > orders of magnitude longer, i.e. miliseconds ranges, depending on the
> > > frame size and bitstream params, so it is an acceptable trade-off to
> > > have a more generic driver.
> > > 
> > 
> > Before going forward with using regmap, I would like to have a sense
> > of the footprint it adds, and see if we can avoid that 20 us penalty.
> > 
> > I'd also like to try another approach, something that has less
> > memory footprint and less runtime penalty.
> > 
> > How about something like this:
> > 
> > #define G1_PIC_WIDTH 4, 0xff8, 23
> > #define ...
> >                                   
> > struct hantro_swreg {
> >          u32 value[399 /*whatever size goes here*/];
> > };
> >                                                                                   
> > void hantro_reg_write(struct hantro_swreg *r,
> >                        unsigned int swreg, u32 mask, u32 offset, u32 new_val)
> > {
> >          r->value[swreg] = (r->value[swreg] & ~(mask)) |
> >                            ((new_val << offset) & mask);
> > }
> > 
> > Which you can then use in a very similar way as the current proposal:
> > 
> > hantro_reg_write(&swreg, G1_PIC_WIDTH, width);
> > 
> > The first advantage here is that we no longer have any
> > footprint for the fields.
> > 
> > The ugly macros for "4, 0xff8, 23" can be auto-generated from
> > existing vendor headers, when possible, so that shouldn't
> > bother us.
> > 
> > The register set is "flushed" using _relaxed, but it
> > could be still costly.
> > 
> > If that is indeed costly, perhaps we can avoid writing
> > the entire set by having a dirty bit somewhere.
> > 
> > In any case, it's worth exploring our options first, I think.
> 
> Or maybe the regmap API itself deserves extending with a "deferred" 
> operating mode where updates to the cached state can be separated from 
> committing that state to the underlying hardware.
> 
> ...which, after a brief code search out of curiosity, apparently already 
> exists in the form of regcache_cache_only()/regcache_sync(), so there's 
> probably no need to reinvent it :)
> 

To be fair, and despite it could seem an anti-pattern, this particular
wheel is so tiny and trivial, that I'm starting to seriously consider
reinventing it.

I've been thinking long about this but just can't seem to see exactly
what benefit we're getting from using MMIO regmaps here,
as opposed to just a simple macro with an index, a mask, and an offset.

Ezequiel