AMD's VLIW weirdness

[atom]

Guess what, I've installed Catalyst 13.4 and the speed dropped!

It was a boring day so I took a look at the new kernel .ISA files produced by Catalyst it turned out AMD changed the VLIW handling. I assume this change is related to how the optimizer in the JiT-compiler is changed over time.

So I readjusted the VLIW settings. That means that I had to change some kernels from VLIW1 to VLIW2, some from VLIW2 to VLIW4, some from VLIW4 to VLIW5 but also some from VLIW2 to VLIW1. Even more strange, I had to change the GCN cards (hd7xxx) to VLIW2. Yes, that means they require to run on vector-datatypes now!

So here are the results of my hand-optimized VLIW settings:

hd5xxx / hd6xxx:

MD5: 10845 -> 10961 = 1.07%
MySQL: 1320 -> 1694 = 28.33%
MD4: 21238 -> 21705 = 2.20%
NTLM: 20424 -> 20884 = 2.25%
DCC: 5746 -> 5779 = 0.57%
SHA256: 1396 -> 1455 = 4.23%
DESCRYPT: 74108 -> 90702 = 22.39%
SHA512: 185 -> 220 = 18.92%
Cisco PIX: 7303 -> 7357 = 0.74%
LM: 981 -> 1056 = 7.65%
Oracle old: 221 -> 234 = 5.88%
SHA3 KECCAK: 144 -> 156 = 8.33%
Half MD5: 6165 -> 6196 = 0.50%

hd7xxx:

SHA1: 2153 -> 2543 = 18.11%
MySQL: 1047 -> 1198 = 14.42%
DESCRYPT: 60508 -> 91901 = 51.88%
SHA512: 75793 -> 79662 = 5.10%
LM: 1310 -> 1332 = 1.68%
SHA3 KECCAK: 70 -> 138 = 97.14%
NetNTLMv2: 396 -> 548 = 38.38%

You should see those improvements in next release on your AMD card, too.