Message-ID: <20150502023427.GA2485@openwall.com>
Date: Sat, 2 May 2015 05:34:27 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: yescrypt on GPU

Hi,

The yescrypt cryptocoin stuff is starting to pay off.  djm34 has just
implemented support for BSTY mining on GPU, in both OpenCL and CUDA
(with tiny bits of inline PTX assembly, even - for things such as the
pwxform MULs):

https://bitcointalk.org/index.php?topic=775289.msg11252741#msg11252741

The code is still very dirty.  I expect it won't build or work for most
people as-is, yet.  However, it looks reasonably well optimized, and
specialized to the yescrypt settings that BSTY uses (e.g., loop counts
are precomputed and hard-coded, etc.)

My own experience with it so far is this: the modified sgminer built
fine on Linux.  When trying to build the yescrypt OpenCL kernel with an
older AMD Catalyst I had installed on that machine, the OpenCL compiler
segfaulted.  Telling it to use another OpenCL platform number, to target
an NVIDIA card in the same machine, it sort of worked - the kernel built
and started mining, but instead of accepted shares only the HW error
counter increased slowly - roughly at the pace I would have expected
valid shares to be generated at the reported hashes per second rate.
Speaking of which, it was 198 h/s on a GTX TITAN.  (Indeed, this doesn't
mean much until the code starts to actually work correctly on that
system.)  For comparison, a (much cheaper) quad-core CPU does ~3400 h/s.

The modified sgminer includes a copy of yescrypt-opt.c to validate the
shares found on GPU.  I guess this is how the "HW errors" were detected.

The CUDA code is present only in the "windows" branch of djm34's fork of
ccminer-tpsp.  With a few easy changes here and there, it mostly built
for me on Linux - however, cuda_yescrypt.cu has been compiling for 1.5
hours already as I am typing this message.  (No idea if that build will
finish or not.  The compiler's memory usage is slowly growing, which
might be a good sign.)

bsty     27511 99.7  0.1 242304 216548 pts/1   R+   04:47 100:23 cicc -arch compute_35 -m64 -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 -maxreg 128 -nvvmir-library /usr/local/cuda/bin/../nvvm/libdevice/libdevice.compute_35.10.bc --orig_src_file_name yescrypt/cuda_yescrypt.cu /tmp/tmpxft_00006af6_00000000-9_cuda_yescrypt.cpp3.i -o /tmp/tmpxft_00006af6_00000000-5_cuda_yescrypt.ptx

So I have no performance numbers from actually running it yet.

Unfortunately, neither djm34 nor anyone else has posted any specific
performance numbers yet.  djm34 only made this comment: "It is difficult
to go higher in intensity due to the high mem requirement. (also current
speed is rather low... for the same reasons)".  Obviously.

To remind, BSTY uses yescrypt v0 (with an unimportant high-level bug
introduced when merging into BSTY wallet, and now preserved as part of
BSTY-specific yescrypt) at 2 MB, with 1 KB blocks (r=8).  yescrypt v1
should be almost same speed (no changes to the performance critical
parts were made between v0 and v1).

Alexander

Powered by blists - more mailing lists