[PATCH 1/2] md/raid6: use faster multiplication for ARM NEON delta syndrome

[Ard Biesheuvel]

On 13 July 2017 at 18:51, Markus Stockhausen <stockhausen at collogia.de> wrote:
>> Von: Ard Biesheuvel [ard.biesheuvel at linaro.org]
>> Gesendet: Donnerstag, 13. Juli 2017 19:16
>> An: linux-arm-kernel at lists.infradead.org; linux-raid at vger.kernel.org
>> Cc: shli at kernel.org; Markus Stockhausen; linux at armlinux.org.uk; will.deacon at arm.com; catalin.marinas at arm.com; Ard Biesheuvel
>> Betreff: [PATCH 1/2] md/raid6: use faster multiplication for ARM NEON delta syndrome
>>
>> The P/Q left side optimization in the delta syndrome simply involves
>> repeatedly multiplying a value by polynomial 'x' in GF(2^8). Given
>> that 'x * x * x * x' equals 'x^4' even in the polynomial world, we
>> can accelerate this substantially by performing up to 4 such operations
>> at once, using the NEON instructions for polynomial multiplication.
>>
>> Results on a Cortex-A57 running in 64-bit mode:
>>
>>   Before:
>>   -------
>>   raid6: neonx1   xor()  1680 MB/s
>>   raid6: neonx2   xor()  2286 MB/s
>>   raid6: neonx4   xor()  3162 MB/s
>>   raid6: neonx8   xor()  3389 MB/s
>>
>>   After:
>>   ------
>>   raid6: neonx1   xor()  2281 MB/s
>>   raid6: neonx2   xor()  3362 MB/s
>>   raid6: neonx4   xor()  3787 MB/s
>>   raid6: neonx8   xor()  4239 MB/s
>
> Nice optimiziation. Nevertheless the test algorithm favours this implementation. See:
>
> int start = (disks>>1)-1, stop = disks-3; /* work on the second half of the disks */
>
> What gives the before/after test if you work on the middle data disks and not on
> the right ones? In the 4K page size this should be  start = 3, stop = 11 instead of
> start = 7, stop = 13. Given the large gain you see the impact should be lower but
> at least in the >10% range.
>

Relative before and after (using raid6test rather than the kernel
module this time, so they should not be compared with the numbers
above)

before
raid6: neonx1   xor()  1773 MB/s
raid6: neonx2   xor()  2362 MB/s
raid6: neonx4   xor()  3223 MB/s
raid6: neonx8   xor()  3375 MB/s

after
raid6: neonx1   xor()  2259 MB/s
raid6: neonx2   xor()  2975 MB/s
raid6: neonx4   xor()  3404 MB/s
raid6: neonx8   xor()  3788 MB/s

So your estimate is correct: 12% speedup for neonx8 in the 'start = 7,
stop = 13' case

-- 
Ard.