crypto: twofish-avx - tune assembler code for more performance
authorJussi Kivilinna <jussi.kivilinna@mbnet.fi>
Tue, 28 Aug 2012 11:24:43 +0000 (14:24 +0300)
committerHerbert Xu <herbert@gondor.apana.org.au>
Thu, 6 Sep 2012 20:17:04 +0000 (04:17 +0800)
commitf94a73f8dd5644f45f9d2e3139608ca83b932d93
tree7b4c39aa93c9d699aaed30e80f82da0bf730e4d1
parent49d30d3d5fd133a6243acf76c98242a68408a31a
crypto: twofish-avx - tune assembler code for more performance

Patch replaces 'movb' instructions with 'movzbl' to break false register
dependencies and interleaves instructions better for out-of-order scheduling.

Tested on Intel Core i5-2450M and AMD FX-8100.

tcrypt ECB results:

Intel Core i5-2450M:

size    old-vs-new      new-vs-3way     old-vs-3way
        enc     dec     enc     dec     enc     dec
256     1.12x   1.13x   1.36x   1.37x   1.21x   1.22x
1k      1.14x   1.14x   1.48x   1.49x   1.29x   1.31x
8k      1.14x   1.14x   1.50x   1.52x   1.32x   1.33x

AMD FX-8100:

size    old-vs-new      new-vs-3way     old-vs-3way
        enc     dec     enc     dec     enc     dec
256     1.10x   1.11x   1.01x   1.01x   0.92x   0.91x
1k      1.11x   1.12x   1.08x   1.07x   0.97x   0.96x
8k      1.11x   1.13x   1.10x   1.08x   0.99x   0.97x

[v2]
 - Do instruction interleaving another way to avoid adding new FPU<=>CPU
   register moves as these cause performance drop on Bulldozer.
 - Further interleaving improvements for better out-of-order scheduling.

Tested-by: Borislav Petkov <bp@alien8.de>
Cc: Johannes Goetzfried <Johannes.Goetzfried@informatik.stud.uni-erlangen.de>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
arch/x86/crypto/twofish-avx-x86_64-asm_64.S