hum, might be some integer overflow. I don't think the original xptminer was intended to run that fast

sorry, i changed -d to -f to use -d for "device", but forgot the error.

About the code, it is a "pruned" SPH  code (removed all that was not used, removed all possible forks and loops, etc). I tried the reference implementation for shavite (the slower one) but could not make it compatible with the code from xptminer/SPH (for some reason hashes were always different, no matter what i did, even the "init" held different initial values). So i just gave up on other implementations and assumed they had changed something in the SPH for metiscoin. Metis "looks" vaguely like "fugue", but after the lack of success with shavite i didn't even try to use another implementation.

The code for the hashes is quite complex, so it takes a lot of time to compile on some cards (specially AMD). I had to wait up to 20 minutes on one of the test machines...

The quick fix that i applied today made me reduce the worksize from 0x80000 to 0x8000, which will introduce a HUUUUGE latency in faster cards. I'm working on a solution for this just about now.

Well, I was expecting people would report performance here. For now the only i know is amd 6970 runs at 4700 kh/s (could be 5300 if i figure out a way around those 0x8000 cycles, maybe tomorrow)

Probably yes. This is an OpenCL miner, so it works on virtually any GPU (And also with some CPUs). Just try it out and report the results.

hum.. gcc just ignores this. I'll fix it right now... (just commiited the fix, try pulling the code again)

binary is already up to date.

I had to do a quick fix to make it work. I'll look into the bug with more detail later today and see how to make it faster again.

sorry, i'll commit it right now... done...

redownload. This is fixed now.

looks like it's fixed. New version (build3) finds shares and they're all valid.

damn.. looks like i missed this little guy here: "block->nonce = n*0x10000+f;".. where did this 0x10000 come from?

This computer has no GPU on it, so i am unable to test it properly. Seems i'll only be able to fix it tomorrow

found the problem. I am repacking it and should be ready in 5 minutos or so.

I'm checking right now why no shares...

I Just released a GPU (openCL) miner for MTC. If anyone cares to test. Performance is not great though. I did it just so we have something to do with our GPUs while iddle.

Source code is available at:

https://github.com/girino/xptMiner

and Binaries (for windows only at present, will add linux later tonight)

https://www.dropbox.com/sh/2wv2awwgebpdo0m/STzOU5tlXa

Please let me know if it works for you and how fast (or slow) it is on your GPUs.

Edit: Linux, Windows and OSX binaries available. Same download location!

Edit: Build 5 fixes the reconnect bug and the speed reporting bug.

Edit: Build 7 incorporates gigawatt's speed improvements. developper fee is up to 3% by default so i can give gigawatt a share for his work. As always, binaries for linux, windows and osx are available on the links above.

Edit: The networking code from ypool's latest xptminer was merged back into Build 8. Several changes and tests were made to ensure stability (no more "0kh/s" after some hours, no more closing on network disconnect or server down). Faster than ever, and two new hidden options: -s to set the step size (in powers of 2, defaults to 2^20), and -a to set variations on algorithm. Currently supports "-a 1" for lookup tables in constant memory and "-a 2" (default) for tables in global memory. -a 2 is faster in most systems. The code for metis was reverted to a previous version since it is 2x slower on some NVidia GPUs. Binaries and code available, as always, on the links above.

Edit: Build 9 uses local memory more intensely and gets up to 35% speed improvements in GPUS with lots of local memory. For those with small local memory, the "-a 2" hidden option reverts to the old version of the opencl (Older cards or Amazon EC2 need this option). Also, ypool instability helped figure out more networking bugs, so networking is even more stable! More stable, UPGRADE TO AVOID LOST CONNECTIONS WHEN YPOOL DOWN!

Thanks in advance,
girino

I Just released a GPU (openCL) miner for MTC. If anyone cares to test. Performance is not great though. I did it just so we have something to do with our GPUs while iddle.

Sorce code is available at:

https://github.com/girino/xptMiner

and Binaries (for windows only at presente, will add linux later tonight)

https://www.dropbox.com/sh/2wv2awwgebpdo0m/STzOU5tlXa

Please let me know if it works for you and how fast (or slow) it is on your GPUs.

Thanks in advance,
girino.

Galera,

Eu sei que os protoshares já sairam de moda, mas como eu estava nessa onda desde quando "dava lucro", eu acabei entrando na corrida pra desenvolver um mineirador para PTS usando GPU.

Se alguém aqui estiver interessado: https://bitsharestalk.org/index.php?topic=2460.0

Ele cobra uma "developer fee" de 2,5% no pool ypool.net, que vem direto pra mim. Caso queiram minerar em outro pool sem "developer fee" e com taxas mais baixas, tem uma versão para ptsweb.beeeeer.org, mas como não ganho nada com isso, só forneço os fontes, sem binários por enquanto: https://bitsharestalk.org/index.php?topic=2374.0

Agradeço a todos que usarem, testarem e derem dicas de como melhorar o miner.

Abraços,
girino.

Romanos 12:21 - Não se deixem vencer pelo mal, mas vençam o mal com o bem.