Bigger and better GPU sieving drive: Discussion

[mdettweiler]

Originally Posted by Ken_g6

Either (1) get gcc 4.3, or (2) remove the -mtune directive. I sorta thought gcc 4.3 or later would be required to install the CUDA SDK.

What Linux version are you using?

This computer is running Ubuntu 8.04 (64-bit). Here's what gcc reports as its version:

gary@buttford:~$ gcc --version
gcc (GCC) 4.2.4 (Ubuntu 4.2.4-1ubuntu4)
Copyright (C) 2007 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

I couldn't even find gcc 4.3 in the package manager...

I'll give it a try without -mtune.

[mdettweiler]

Okay, that resolved the earlier error; now I have a new one:

gary@buttford:~/Desktop/gpu-stuff/ppsieve31$ make
cd pps; make
make[1]: Entering directory `/home/gary/Desktop/gpu-stuff/ppsieve31/pps'
Makefile:77: Extraneous text after `else' directive
gcc -c -m32 -march=i586 -o ../main.o -fomit-frame-pointer -s -O3 -DNDEBUG -D_REENTRANT -I. -I.. ../main.c
In file included from /usr/include/features.h:354,
                 from /usr/include/assert.h:37,
                 from ../main.c:17:
/usr/include/gnu/stubs.h:7:27: error: gnu/stubs-32.h: No such file or directory
make[1]: *** [../main.o] Error 1
make[1]: Leaving directory `/home/gary/Desktop/gpu-stuff/ppsieve31/pps'
make: *** [all] Error 2

Possibly some sort of library I need to install first?

[Ken_g6]

Try just making the specific binary for your platform. (Linux64? "make ppsieve-cuda-x86_64-linux". For Linux32, "make ppsieve-cuda-x86-linux".)

Edit: By the way, did you notice that the latest CUDA SDK is made for Ubuntu 10.04? (Yes, I need to upgrade too.)

Meanwhile, I've built an ATI OpenCL client that might work, but nobody's come by to test it yet. (Except for one guy who couldn't get it to run. )

[frmky]

I'm sure it'll be slower, but should the OpenCL code work on nVidia as well? Would it be useful to test it?

[Ken_g6]

It might work, and it's open-source so I'm not stopping anyone from trying; but I rather doubt it would be useful. Unless it makes a GTX 460 faster. But I was just thinking, if you run two processes on the same GPU, that might work too.

[frmky]

Is the OpenCL source available? The link at http://sites.google.com/site/kenscode/prime-programs points to the CUDA source.

[Ken_g6]

Check the "redcl" branch. The codebases should merge in the not-too-distant future.

[mdettweiler]

Originally Posted by Ken_g6

Try just making the specific binary for your platform. (Linux64? "make ppsieve-cuda-x86_64-linux". For Linux32, "make ppsieve-cuda-x86-linux".)

Edit: By the way, did you notice that the latest CUDA SDK is made for Ubuntu 10.04? (Yes, I need to upgrade too.)

Meanwhile, I've built an ATI OpenCL client that might work, but nobody's come by to test it yet. (Except for one guy who couldn't get it to run. )

Cool, that worked. I'll give it a try vs. the CUDA 2.3 binary shortly.

Regarding CUDA 3.2 being made for Ubuntu 10.04: I wonder if it really does need 10.04, or if it can work OK with the older 8.04? The CUDA 3.1 Ubuntu package is supposedly made for 7.10, but it works fine on 8.04.

[frmky]

Originally Posted by Ken_g6

It might work

No luck. First I needed to add #include <string> and change CL/cl.hpp to CL/cl.h in appcl.cpp. I then added the correct include directories and disabled the OpenCL BOINC build in the Makefile. With these changes, I successfully created a binary. But when I tried to use it, I got

Error: Building Program (clBuildProgram): Program build failure
:110: error: no matching overload found for arguments of type 'int, size_t, uint __attribute__((address_space(1)))*'
VSTORE(0, i, factor_found_arr);
^~~~~~

Edit: Running two CUDA processes on the same GTX 480 just halves the speed of each of them. There's no net gain.

[mdettweiler]

I just now got the chance to try the ppsieve-cuda-x86_64-linux binary I built earlier today with CUDA 3.1, and lo and behold, it's MUCH faster! On the 11T-12T range I just started on Gary's GTX 460, it gets ~2.8M p/sec. as opposed to ~1.7M p/sec. with the CUDA 2.3 binary.

I've attached the binary here for anyone else who wants it. If anyone needs a 32-bit binary, let me know.

[Ken_g6]

Whoa!

Can someone with a GTX 470 or 480 (frmky?) confirm whether or not this applies to the higher-end Fermis? I'm guessing that it does, but to a lesser degree.

Edit: Also, have you tested that it's outputting correct results? Testing thread here, and/or test it against the version I compiled.