Re: [PATCH v2 5/7] HID: ft260: improve i2c large reads performance

[Michael Zaidman <michael.zaidman@gmail.com>]

On Tue, Oct 04, 2022 at 08:15:56PM +0200, Enrik Berkhan wrote:
> Hi Michael,
> 
> On Wed, 2022-09-28 at 17:48 +0300, Michael Zaidman wrote:
> > After:
> > 
> > $ sudo ./i2cperf -d 2 -o 2 -s 128 -r 0-0xff 13 0x51 -S
> > 
> >   Read block via i2ctransfer by chunks
> >   -------------------------------------------------------------------
> >   data rate(bps)  efficiency(%)  data size(B)  total IOs   IO size(B)
> >   -------------------------------------------------------------------
> >   49316           85             256           2           128
> > 
> > Kernel log:
> > 
> > [  +1.447360] ft260_i2c_write_read: off 0x0 rlen 128 wlen 2
> > [  +0.000002] ft260_i2c_write: rep 0xd0 addr 0x51 off 0 len 2 wlen 2 flag 0x2 d[0] 0x0
> > [  +0.001633] ft260_xfer_status: bus_status 0x41, clock 100
> > [  +0.000190] ft260_xfer_status: bus_status 0x40, clock 100
> > [  +0.000001] ft260_i2c_read: rep 0xc2 addr 0x51 len 128 rlen 128 flag 0x7
> > [  +0.008617] ft260_raw_event: i2c resp: rep 0xde len 60
> > [  +0.008033] ft260_raw_event: i2c resp: rep 0xde len 60
> > [  +0.000954] ft260_raw_event: i2c resp: rep 0xd1 len 8
> 
> As the ft260 can pack up to 60 bytes into one report, would it make
> sense to use a multiple-of-60 size (120 or 180)? Might reduce overhead
> by another tiny bit ...
> 
> Cheers,
> Enrik
> 
The size of the Read IO to perform is given to the driver by the upper
layer. So it's up to him how to align the IO request size.

When we read from the EEPROM, we want to issue the read requests with
EEPROM page size granularity. The I2C EEPROMs page sizes are usually a
power of 2 aligned.

Please see the examples of reading 4K bytes from the 24C512 EEPROM
first by the read requests of EEPROM page size granularity of 128 bytes
and the second time of the 120 bytes (a multiple of 60 bytes granularity).
In the power of 2 aligned cases, we issued lesser Read IOs (I2C combined
transactions - write address read data) than when we did it with the 60
bytes alignment. Hence the performance gain.

$ sudo ./i2cperf -d 2 -o 2 -s 128 -r 0-0xfff 13 0x51 -S
  Read block via i2ctransfer by chunks
  -------------------------------------------------------------------
  data rate(bps)  efficiency(%)  data size(B)  total IOs   IO size(B)
  -------------------------------------------------------------------
  49581           85             4096          32          128

$ sudo ./i2cperf -d 2 -o 2 -s 120 -r 0-0xfff 13 0x51 -S
  Read block via i2ctransfer by chunks
  -------------------------------------------------------------------
  data rate(bps)  efficiency(%)  data size(B)  total IOs   IO size(B)
  -------------------------------------------------------------------
  48816           85             4096          35          120

Thanks,
Michael

>