Quote from: dga on January 13, 2014, 05:54:03 pm

Quote from: archit on January 13, 2014, 05:52:25 pm

Why not AVX2 for SHA256 as well?

Why bother?

Won't improve anything?

There are 4-10 hash collisions per group of 2^23 SHA512 hashes that have to be pushed through SHA256.  Making SHA256 faster would make 1/1,000,000th of the computation faster. :-)

Not too much of a big deal, still functional. Although well done for pointing it out I'm sure he will appreciate it

Yup, It's just a comment typo.    It just makes it a little harder to understand the PoW function when reading the source.

Hi, FreeTrade -

Line 94 of momentum.cpp has an incorrect comment:

//use last 4 bits of first cache as next location

4 should be 14 (or abstracted in terms of the other constants).

Looked like it was faster than yam on my two junk servers, but slower on the rest

it was also dumping cores everywhere with mmap failing

Thanks for the report.  The mmap failure is just a warning - it falls back to malloc.

To silence - and run a little faster with both yam and my code - run:

echo "2048" > /proc/sys/vm/nr_hugepages

But the dumping cores is bad.  Could you send me a little more detail, or a stack trace?  (And on what kind of machine?)

The slower on the rest isn't too surprising.  There are a lot of optimizations to be done yet, particularly for huge servers with respect to thread affinity and other things.  And the SHA512 code is virtually untouched from the Intel release.  The goal isn't to beat yam with this release, it's just to start the ball rolling a little bit.

There are some constants to play with to tune for different platforms, but it's not worth going there yet (unless you're interested in poking in the code).

  -Dave

Following up on yvg1900's release of yam, I figured I'd improve the state of the art of the open source versions a bit:

https://github.com/dave-andersen/ptsminer

I haven't made it build yet on windows (it just needs to compile the avx2 assembly code - should be straightforward if someone wants to clue me in on how to appropriately invoke gcc there), but it should work on other platforms.  As a warning, I've only really tried it on avx2, since I'm a fan of Haswell.  THIS SOFTWARE SHOULD BE CONSIDERED A BETA QUALITY RELEASE.  At best. 

As with my GPU release, this one is based very directly on ptsminer, so it's tied to beeeeer for the moment.  I plan to fix that and let it be used with other pools in the near future, but that's going to take some dev work.  sigh.

There's a lot of optimization to be done, but this gets the basics as far as memory subsystem optimization, and bridges a lot of the gap between the old OSS version and yam M7i.  I haven't tried out M7j, mind you -- it's probably a bit faster still, but this release should bridge the gap considerably.

It incorporates the same optional, extendible 1% dev fee that the gpu miner does.  Prior ptsminer devs, if you feel like you should be in the list, please PM me and I'll get you added!

With gratitude to FreeTrade for the donation that kept me interested in hacking on and releasing this stuff, and to yvg1900 for some very engaging unofficial competition. *grin*

  -Dave

Quote from: dga on January 12, 2014, 10:32:57 pm

Quote from: phrozenspite on January 12, 2014, 08:30:49 pm

1gh v1.2 - windows 8 64 bit.  sapphire 7970 ghz edition 3gb, 1040 core 1500 ram clock speed. ~1120 CPM

Just for fun:

cudapts (linux) v2014-01-12 GTX690, stock clock settings, 1780 c/m.

Do you mind to try 1gh miner on this one? That would give a general idea how much OpenCL is inferior to CUDA.

Unfortunately, that machine also has one of my coin wallets on it, and I don't run untrusted binaries on it.

(Mind you, I'm not saying there's anything wrong with the 1gh miner - but there's been a lot of malware floating around.)

Quote from: girino on January 12, 2014, 07:41:40 pm

Quote from: girino on January 12, 2014, 04:39:52 pm

(...)

I am adapting my version of ptsminer to work with opencl. I can't get the same performance as cudapts yet (on an amazon g2.2xlarge his version gets 550 cpm and mine gets 290 cpm, so lots of room to improve). But you are welcome to try it (on AMD GPUs, for example, where there's no CUDA), and of course, suggest improvements.

https://github.com/girino/ptsminer

Got 480 CPM on amazon g2.2xlarge. Still not as good as cudapts, but getting close!

What is cudapts cpm there?

It's currently about 750-780.  480 is pretty good from OpenCL on Nvidia.

  -Dave