Re: [Linux-fbdev-devel] [RFC 2.6.28 1/1] gpiolib: add set/get batch v3

[Dave H. <dhy...@gm...>]

Hi Jaya,

...snip...
> There was a question about why have both bitmask AND width when width
> could be worked out from bitmask. The reason for having both bitmask and
> length is for performance. Let's compare the following two scenarios:
>
> /* flash to black */
> for (i=0; i < 800*600; i++)
>        gpio_set_batch(DB0, 0x00000000, 0x0000FFFF)
>
> Internally, the above implementation has to work out the length of bits
> that are being set in bitmask in order to figure out which register
> boundaries are being crossed. That means looping through the bitmask
> and counting bits. That would need to be repeated for every call of
> gpio_set_batch in the above loop and would be a relatively high expense.
>
> /* flash to black */
> for (i=0; i < 800*600; i++)
>        gpio_set_batch(DB0, 0x00000000, 0x0000FFFF, 16)
>
> In the 2nd case, the implementation is able to skip all the bit counting.
> This usage method is also the intuitive way for drivers that are trying
> to push data across a gpio bus since they explicitly know the bus length.
...snip...

I haven't been following too closely, but what about

        gpio_set_batch(DB0, 0x00000000, 0x00FF00FF, 16)

That also has 16 bits set, but in different positions.

Shouldn't you just not bother with the length and optimize particular
bitmasks (i.e. 0xFF, 0xFFFF, 0xFFFFFF, 0xFFFFFFFF)?

-- 
Dave Hylands
Shuswap, BC, Canada
http://www.DaveHylands.com/