Trial division with CUDA (mmff) -- used, but runs like new! - Page 26

flashjh · 2012-11-17, 19:33

Quote:

Originally Posted by MattcAnderson

Hi All,
I finally got mmff going on my GPU. The error I had was cudart64_42_9.dll needed. from post # 185 of this thread, I found mmff v 23 with cuda 5.0, which seems to work. This seemed to solve my problem, but the windows 64 download from doublemersenne.org did not. I think this is because it has cuda 4.2. I guess my system requires cuda 5.0 and won't work with cuda 4.2.

Regards,
Matt

CUDA 4.2 version requires the 42 dll files here. Just put them in the mmff directory.

ET_ · 2012-12-04, 17:54

Ubuntu 11.10 - Linux64, CUDA drivers 5.0, CUDA runtime version 4.1

Code:

uigi@luigi-ubuntu:~$ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/4.6.1/lto-wrapper
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu/Linaro 4.6.1-9ubuntu3' --with-bugurl=file:///usr/share/doc/gcc-4.6/README.Bugs --enable-languages=c,c++,fortran,objc,obj-c++,go --prefix=/usr --program-suffix=-4.6 --enable-shared --enable-linker-build-id --with-system-zlib --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --with-gxx-include-dir=/usr/include/c++/4.6 --libdir=/usr/lib --enable-nls --with-sysroot=/ --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --enable-plugin --enable-objc-gc --disable-werror --with-arch-32=i686 --with-tune=generic --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 4.6.1 (Ubuntu/Linaro 4.6.1-9ubuntu3)

As mmff 0.26 binary is compiled with CUDA runtime version 4.2, I couldn't run it out of the box, and tried recompiling it, but...

I had problems trying to compile mmff

It seems that some classical C math functions don't link correctly, even moving the -lm switch on the Makefile at the beginning of the line.

In any case, I am appending the compile log, hoping that someone of good will will throw an eye on it...

Thank you anyway.

Luigi

akruppa · 2012-12-04, 18:07

The GNU linker by default processes object files and libraries in the order as they are specified on the command line, remembers which unresolved references there are from each file it processes, and tries to resolve those references with symbols in object files and libraries that follow. This means putting -lm at the start of the command line has no effect: there are no unsatisfied references to math library functions at that point, and the unresolved references to libm from later files never get resolved because no -lm follows them.

Put -lm at the end, and generally a object/library file which contains functions that another file Y needs after the file Y. If you have circular dependencies, you can use some command line switch (forgot which) to make gnu ld re-scan the objects on the command line until all symbols are resolved, but you rarely need that.

ET_ · 2012-12-04, 18:26

Quote:

Originally Posted by akruppa

The GNU linker by default processes object files and libraries in the order as they are specified on the command line, remembers which unresolved references there are from each file it processes, and tries to resolve those references with symbols in object files and libraries that follow. This means putting -lm at the start of the command line has no effect: there are no unsatisfied references to math library functions at that point, and the unresolved references to libm from later files never get resolved because no -lm follows them.

Put -lm at the end, and generally a object/library file which contains functions that another file Y needs after the file Y. If you have circular dependencies, you can use some command line switch (forgot which) to make gnu ld re-scan the objects on the command line until all symbols are resolved, but you rarely need that.

Thank you Alex.

Update: I tried a make clean followed by a make all (I forgot the "all" thingie...

)

result:

Code:

nvcc fatal   : Unsupported gpu architecture 'compute_30'

So I just cancelled the portion that said

Code:

--generate-code arch=compute_30,code=sm_30

as I successfully did for mfaktc.

result:

Code:

#error -- unsupported GNU version! gcc 4.6 and up are not supported!

Question:

Does runtime version 5.0 correct this error, or I have to try some fancy link and download an older version of GCC?

Luigi

Batalov · 2012-12-04, 18:30

Quote:

Originally Posted by ET_

Question:

Does runtime version 5.0 correct this error, or I have to try some fancy link and download an older version of GCC?

Luigi

Good question. The answer is no.

Find this file and edit __GNUC_MINOR__ > 6 into __GNUC_MINOR__ > 9

Code:

/usr/local/cuda/include/host_config.h:
#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 9)

ET_ · 2012-12-04, 18:44

Quote:

Originally Posted by Batalov

Good question. The answer is no.

Find this file and edit __GNUC_MINOR__ > 6 into __GNUC_MINOR__ > 9

Code:

/usr/local/cuda/include/host_config.h:
#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 9)

It went farther, compiled the ptxtras (there is a __f variable set and unised), but still complains for the math functions.

Log appended.

There is a request to restart the system, maybe a pending update; I will tell you if after the reboot something changes.

Thank you Serge.

Luigi

akruppa · 2012-12-04, 18:50

Code:

gcc -fPIC -L/usr/local/cuda/lib64/ -lcudart -lm timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -o ../mmff.exe

ET_ · 2012-12-04, 19:22

Quote:

Originally Posted by akruppa

Code:

gcc -fPIC -L/usr/local/cuda/lib64/ -lcudart -lm timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -o ../mmff.exe

That's the line come from George's Makefile.

Should I modify it?

Luigi

frmky · 2012-12-04, 19:38

Newer versions of gcc no longer support putting the libraries at the beginning. Try

Code:

gcc -fPIC -L/usr/local/cuda/lib64/ -o ../mmff.exe timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -lcudart -lm

akruppa · 2012-12-04, 19:41

I just checked what I did on our work machine to make it compile:

Code:

kruppaal@quiche:~/mmff$ diff src/Makefile src.my/Makefile
14c14
< NVCCFLAGS += --generate-code arch=compute_20,code=sm_20 --generate-code arch=compute_30,code=sm_30
---
> NVCCFLAGS += --generate-code arch=compute_20,code=sm_20 # --generate-code arch=compute_30,code=sm_30
16c16
< NVCCFLAGS += --compiler-options=-Wall
---
> NVCCFLAGS += --compiler-options="-Wall -B/usr/lib/gcc/x86_64-linux-gnu/4.5.3/"
20c20,21
< LDFLAGS = -fPIC $(CUDA_LIB) -lcudart -lm
---
> LDFLAGS = -fPIC $(CUDA_LIB)
> MMFFLIB = -lcudart -lm
35c36
<       $(LD) $(LDFLAGS) $^ -o $@
---
>       $(LD) $(LDFLAGS) $^ $(MMFFLIB) -o $@

The -B/usr/lib/gcc/x86_64-linux-gnu/4.5.3/ is an override because 4.6 is the default gcc, but 4.5.3 is installed as well; with this option, nvcc uses the 4.5.3 version and stops complaining.

ET_ · 2012-12-04, 19:58

Thank you Greg, your line worked great!

Alex, now I found an old Makefile from the project GPU-ECM:

Code:

# CUDA does not support gcc >= 4.6
# this can be useful if your default gcc is >= 4.6
# if not just comment the following two lines.
#CC_BIN:=/tmp/gcc45
#USE_THIS_CC:=--compiler-bindir $(CC_BIN)

I totally forgot about it, but you were right: you (and I) had a lower version of GCC at that time. Thank you for yur last mssge, you gave me the clue to remember where I saw that message.

It's time to reserve some exponent... My mmff is already doing its selftest.

Thanks again!

Luigi

2012-12-04, 18:50	#282
akruppa "Nancy" Aug 2002 Alexandria 100110100011₂ Posts	Code: gcc -fPIC -L/usr/local/cuda/lib64/ -lcudart -lm timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -o ../mmff.exe

2012-12-04, 19:38	#284
frmky Jul 2003 So Cal 2·3⁴·13 Posts	Newer versions of gcc no longer support putting the libraries at the beginning. Try Code: gcc -fPIC -L/usr/local/cuda/lib64/ -o ../mmff.exe timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -lcudart -lm

2012-12-04, 19:58	#286
ET_ Banned "Luigi" Aug 2002 Team Italia 3²×5×107 Posts	Thank you Greg, your line worked great! Alex, now I found an old Makefile from the project GPU-ECM: Code: # CUDA does not support gcc >= 4.6 # this can be useful if your default gcc is >= 4.6 # if not just comment the following two lines. #CC_BIN:=/tmp/gcc45 #USE_THIS_CC:=--compiler-bindir $(CC_BIN) I totally forgot about it, but you were right: you (and I) had a lower version of GCC at that time. Thank you for yur last mssge, you gave me the clue to remember where I saw that message. It's time to reserve some exponent... My mmff is already doing its selftest. Thanks again! Luigi

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Mersenne trial division implementation	mathPuzzles	Math	8	2017-04-21 07:21
trial division over a factor base	Peter Hackman	Factoring	7	2009-10-26 18:27
P95 trial division strategy	SPWorley	Math	8	2009-08-24 23:26
Trial division software for Mersenne	SPWorley	Factoring	7	2009-08-16 00:23
Need GMP trial-division timings	ewmayer	Factoring	7	2008-12-11 22:12

2012-12-04, 18:07	#278
akruppa "Nancy" Aug 2002 Alexandria 2,467 Posts	The GNU linker by default processes object files and libraries in the order as they are specified on the command line, remembers which unresolved references there are from each file it processes, and tries to resolve those references with symbols in object files and libraries that follow. This means putting -lm at the start of the command line has no effect: there are no unsatisfied references to math library functions at that point, and the unresolved references to libm from later files never get resolved because no -lm follows them. Put -lm at the end, and generally a object/library file which contains functions that another file Y needs after the file Y. If you have circular dependencies, you can use some command line switch (forgot which) to make gnu ld re-scan the objects on the command line until all symbols are resolved, but you rarely need that.