On 10/13/2017 04:11 AM, Lukas Wunner wrote:
[cc += William Breathitt Gray, Geert Uytterhoeven, Phil Reid,
David Daney, Iban Rodriguez]
The drivers
gpio-104-dio-48e.c, gpio-74x164.c, gpio-gpio-mm.c,
gpio-pca953x.c, gpio-thunderx.c, gpio-ws16c48.c,
gpio-xilinx.c
currently use an inefficient algorithm for their ->set_multiple
callback: They iterate over every chip (or bank or gpio pin),
check if any bits are set in the mask for this particular chip,
and if so, update the affected GPIOs. If the mask is sparsely
populated, you'll waste CPU time checking chips even though
they're not affected by the operation at all.
Iterating over the chips is backwards, it is more efficient to
iterate over the bits set in the mask, identify the corresponding
chip and update its affected GPIOs. The gpio-max3191x.c driver
I posted yesterday contains an example for such an algorithm and
you may want to improve your ->set_mutiple implementation accordingly:
Do you have any profiling results that demonstrate significant
system-wide performance improvements by making such a change?
For the gpio-thunderx.c driver, the words in the bits/mask exactly match
the banks, so the number of iterations doesn't change with the approach
you suggest.
In fact, an argument could be made in the other direction. The
increased ICache pressure from code bloat and branch prediction misses
resulting from extra testing of the mask bits could easily make the
system slower using your suggestion.
Thanks,
David Daney
--
To unsubscribe from this list: send the line "unsubscribe linux-gpio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
