Date: Sat, 10 Aug 2013 17:10:31 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: Parallella: Litecoin mining Rafael, On Sat, Aug 10, 2013 at 03:14:26AM +0100, Rafael Waldo Delgado Doblas wrote: > Indeed, but finally I could decrease the TMTO up to 6 and now the > performance is ~1.150Kh/s. This is good news, although the target performance is much higher - I'd guesstimate 3 kh/s with pure C and 4-5 kh/s with asm code. (We had a smaller difference between pure C and asm for Katja's bcrypt code - with best asm speed being about 20% higher than best C speed - but for scrypt I'd expect a bigger difference due to potential use of IMADD for bit rotates, and due to Salsa20 containing relatively more ADDs than XORs, as compared to Blowfish.) > TMTO 5 fails, I need to check why. OK. At a later time, you could even try 4.5, if you make your code really compact. (Of course, supporting a TMTO factor that is not a whole number requires slightly different code.) I think at this time your priority should be to find out why you're not getting more than 1.15 kh/s at TMTO factor 6. You could get something like 2+ kh/s with pure C code with it, with more optimal code (this is just my gut feeling, though). Is the code based off scrypt's -ref or -nosse? Anyhow, even -nosse performed a lot of unnecessary data copying. This is something I optimized in escrypt, but only in -sse code. If the problem is indeed still present in the code you're using, you could compare escrypt's crypto_scrypt-sse.c against the original in scrypt-1.1.6.tgz and implement similar optimizations (I mean removing blkxor and blkcpy) for whatever code revision you have. > size for TMTO 5 SCRATCHBUF_SIZE 26828 is > text data bss dec hex filename > 7264 2584 8 9856 2680 epiphany-scrypt.elf > > size for TMTO 6 SCRATCHBUF_SIZE 22464 is > text data bss dec hex filename > 7264 2584 8 9856 2680 epiphany-scrypt.elf > > size for TMTO 100 SCRATCHBUF_SIZE 1952 is > text data bss dec hex filename > 7192 2584 8 9784 2638 epiphany-scrypt.elf > > Is this a normal output? It looks reasonable to me, except that I did not expect you'd have this much pre-initialized data (2.5 KB). What is it? > Why the only information is about text, data and bss sections?. What else did you expect to see there? Your scratch space that you use for scrypt's V? How do you allocate it? Alexander
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.