Author

Topic: cudaminer on Linux (Read 752 times)

newbie
Activity: 5
Merit: 0
August 06, 2013, 06:39:39 PM
#1
Sorry this is in the newbies section, but I registered to this forum specifically to post this, and I'm not yet allowed to post on the cudaminer thread. Maybe an admin can move this post to the proper place if it is deemed worthy.

I just managed to get cudaminer to build on Linux, and figured I should document what I had to change to make things work because the Linux instructions in the cudaminer thread don't work for me. This worked for cudaminer-2013-07-13.zip. You will (obviously) need the NVidia CUDA SDK installed and "nvcc" in your path.

First, after unzipping the archive and un7zing the src archive. I had to remove all of the ^Ms from the line endings or autotools gets confused and generates a configure that doesn't work:

dos2unix *

Then the main file is called "cpu-miner.c" but autotools seems to be set up for a file called "cuda-miner.c" We'll just copy it over to make autotools happy:

cp cpu-miner.c cuda-miner.c

Now you should be able to generate the configure script using autogen.sh:

chmod 755 ./autogen.sh
./autogen.sh

The configure script picks the wrong C compiler by default. It tries to use gcc, but we really need g++ for this because there are a couple of C++-isms in the source that gcc doesn't like. I also had a problem where it couldn't find the CUDA libraries, but that might just be my cuda-sdk install or something wonky with my distro, but it can't hurt to specify. So, we'll set CC and LDFLAGS to compensate when we run ./configure

CC=g++ LDFLAGS=/opt/cuda/lib64 ./configure

Hopefully, things will go well and you will have a Makefile.

Now, there are still a couple of things that I needed to tweak in the code. First off, I was continually getting "'sleep' was not declared in this scope" errors, and similar errors for other functions in unistd.h. I still have no idea why. The #include should have worked, but I had to specify the full path to make it work for some reason. Only do this if you are getting the sleep errors. It might just be something wrong with my environment, but oddly it only affected unistd.h. Anyway, in the cuda-miner.c file, I changed:

#include

to

#include "/usr/include/unistd.h"

Then, there's a call to _strdup() on line 1251 in that same file. I don't really know why there's an underscore there, but I don't have a library with a function by that name and the non-underscore version seems to work. So change:

device_config[cnt++] = _strdup(pch);

to

device_config[cnt++] = strdup(pch);

We should be done with cuda-miner.c now.

The last probelm I had was a call to alloca() in util.c that doesn't cast the void* to a char* to match the type of the pointer. We just need to add a cast to make the compiler happy, so on line 84 in util.c change:

buf = alloca(len);

to

buf = (char *) alloca(len);

And everything should be good to go. Type "make" and if you are very lucky (and I didn't change anything else I've forgotten about) it should compile for a while then you should have an executable file called "cudaminer" in the current directory. I've been running it from there, and did not run a "make install." It's probably not needed.

I've been running it for a bit, and it seems to be working. At least the mining pool I use has been accepting the work; that's as far as I know how to test it. It's probably a 2x speed improvement over cgminer on my GTX 670, which is awesome!

Oh, and thanks for your work on this cbuchner1. I will probably donate some LTC to you, once I've mined enough to cash out some LTCs of my own.
Jump to: