[PATCH 1/3] riscv: optimized memcpy

[Matteo Croce]

On Fri, Jun 18, 2021 at 2:32 AM Matteo Croce <mcroce at linux.microsoft.com> wrote:
>
> On Thu, Jun 17, 2021 at 11:48 PM Matteo Croce
> <mcroce at linux.microsoft.com> wrote:
> >
> > On Thu, Jun 17, 2021 at 11:30 PM David Laight <David.Laight at aculab.com> wrote:
> > >
> > > From: Matteo Croce
> > > > Sent: 16 June 2021 19:52
> > > > To: Guo Ren <guoren at kernel.org>
> > > >
> > > > On Wed, Jun 16, 2021 at 1:46 PM Guo Ren <guoren at kernel.org> wrote:
> > > > >
> > > > > Hi Matteo,
> > > > >
> > > > > Have you tried Glibc generic implementation code?
> > > > > ref: https://lore.kernel.org/linux-arch/20190629053641.3iBfk9-
> > > > I_D29cDp9yJnIdIg7oMtHNZlDmhLQPTumhEc at z/#t
> > > > >
> > > > > If Glibc codes have the same performance in your hardware, then you
> > > > > could give a generic implementation first.
> > >
> > > Isn't that a byte copy loop - the performance of that ought to be terrible.
> > > ...
> > >
> > > > I had a look, it seems that it's a C unrolled version with the
> > > > 'register' keyword.
> > > > The same one was already merged in nios2:
> > > > https://elixir.bootlin.com/linux/latest/source/arch/nios2/lib/memcpy.c#L68
> > >
> > > I know a lot about the nios2 instruction timings.
> > > (I've looked at code execution in the fpga's intel 'logic analiser.)
> > > It is a very simple 4-clock pipeline cpu with a 2-clock delay
> > > before a value read from 'tightly coupled memory' (aka cache)
> > > can be used in another instruction.
> > > There is also a subtle pipeline stall if a read follows a write
> > > to the same memory block because the write is executed one
> > > clock later - and would collide with the read.
> > > Since it only ever executes one instruction per clock loop
> > > unrolling does help - since you never get the loop control 'for free'.
> > > OTOH you don't need to use that many registers.
> > > But an unrolled loop should approach 2 bytes/clock (32bit cpu).
> > >
> > > > I copied _wordcopy_fwd_aligned() from Glibc, and I have a very similar
> > > > result of the other versions:
> > > >
> > > > [  563.359126] Strings selftest: memcpy(src+7, dst+7): 257 Mb/s
> > >
> > > What clock speed is that running at?
> > > It seems very slow for a 64bit cpu (that isn't an fpga soft-cpu).
> > >
> > > While the small riscv cpu might be similar to the nios2 (and mips
> > > for that matter), there are also bigger/faster cpu.
> > > I'm sure these can execute multiple instructions/clock
> > > and possible even read and write at the same time.
> > > Unless they also support significant instruction re-ordering
> > > the trivial copy loops are going to be slow on such cpu.
> > >
> >
> > It's running at 1 GHz.
> >
> > I get 257 Mb/s with a memcpy, a bit more with a memset,
> > but I get 1200 Mb/s with a cyle which just reads memory with 64 bit addressing.
> >
>
> Err, I forget a mlock() before accessing the memory in userspace.
>
> The real speed here is:
>
> 8 bit read: 155.42 Mb/s
> 64 bit read: 277.29 Mb/s
> 8 bit write: 138.57 Mb/s
> 64 bit write: 239.21 Mb/s
>

Anyway, thanks for the info on nio2 timings.
If you think that an unrolled loop would help, we can achieve the same in C.
I think we could code something similar to a Duff device (or with jump
labels) to unroll the loop but at the same time doing efficient small copies.

Regards,

--
per aspera ad upstream