Compile Linux version for PFGW

[nordi]

@paulunderwood: I worked around the -lgcc_s problem you had. Remove the " -Wl,-no_pie" option from the makefile. It seems "ld" only has "-pie", not "-no_pie".

[nordi]

I have a pretty well-automated version now: To build PFGW with Docker, get and extract the sources of PFGW. In the directory with the sources, create a Dockerfile with this content:

FROM gcc
RUN apt-get update && apt-get install -y libgmp-dev wget unzip

RUN mkdir /usr/src/mprime && \
    cd /usr/src/mprime && \ 
    # openpfgw insists on version 28.6
    wget -O mprime.zip --quiet https://www.mersenne.org/ftp_root/gimps/p95v286.source.zip && \
    # unzip produces some warnings and therefore has exit code 1.
    unzip mprime.zip ; \
    rm mprime.zip
RUN cd /usr/src/mprime/gwnum && \
    make -j -f make64

COPY . /usr/src/openpfgw
WORKDIR /usr/src/openpfgw
RUN cp /usr/lib/x86_64-linux-gnu/libgmp.a packages/gmp/64bit/
RUN cp /usr/src/mprime/gwnum/gwnum.a packages/gwnum/64bit/gwnum.a
RUN sed -i 's/-Wl,-no_pie//g' makefile

RUN make -j

Run "docker build -t pfgw ." (mind the dot at the end) to build a new image. Then "docker run -it pfgw" to start a container with that image. The executable is at /usr/src/openpfgw/pfgw64 and it actually works.

This assumes "#include <ctype.h>" fix to PFGW is already applied.

[paulunderwood]

Thanks for the tips, nordi

Incidentally, I added gwset_num_threads ( &gwdata, 4 ); to every line with gwinit2(&gwdata, sizeof(gwhandle), (char *) GWNUM_VERSION); and got a multi-threaded PFGW working at 300% on my Haswell 4770k. Good for speedy proofs, but not so good for overall number-crunching throughput

[henryzz]

Originally Posted by paulunderwood

Thanks for the tips, nordi

Incidentally, I added gwset_num_threads ( &gwdata, 4 ); to every line with gwinit2(&gwdata, sizeof(gwhandle), (char *) GWNUM_VERSION); and got a multi-threaded PFGW working at 300% on my Haswell 4770k. Good for speedy proofs, but not so good for overall number-crunching throughput

It is more efficient the larger the FFT. There would be people using it if this feature was added to the main source. Please do rouge.
A larger fft can fit in the L3 cache with 1x4threads than 4x1threads

[bur]

I tried compiling the sources from sourceforge, but get the following error:

g++: error: packages/gwnum/64bit/gwnum.a: File or directory not found
make: *** [makefile:8: pfgw64] Error 1

The gwnum/64bit/ directory is indeed empty. Any idea how to get it to compile? Or is it unnceccessary anyway and I should just use the precompiled binaries? I have an i9-10900k.

[rogue]

Originally Posted by bur

I tried compiling the sources from sourceforge, but get the following error:

g++: error: packages/gwnum/64bit/gwnum.a: File or directory not found
make: *** [makefile:8: pfgw64] Error 1

The gwnum/64bit/ directory is indeed empty. Any idea how to get it to compile? Or is it unnceccessary anyway and I should just use the precompiled binaries? I have an i9-10900k.

You have to d/l gwnum separately. This is part of Prime95, so d/l the Prime95 source. You will likely need to build the 64-bit Linux version of gwnum.a (if it isn't included) then copy to the directory specified above. There should be a linux makefile in the Prime95 sources.

[Prime95]

Originally Posted by rogue

You have to d/l gwnum separately. This is part of Prime95, so d/l the Prime95 source. You will likely need to build the 64-bit Linux version of gwnum.a (if it isn't included) then copy to the directory specified above. There should be a linux makefile in the Prime95 sources.

In the gwnum directory, type "make -f make64"

[bur]

Ok, thanks, it worked.

Are there any flags I could use to optimize e.g. for avx2? The makefile didn't look like it, but I'm far from being an expert there.

[rogue]

Originally Posted by bur

Ok, thanks, it worked.

Are there any flags I could use to optimize e.g. for avx2? The makefile didn't look like it, but I'm far from being an expert there.

No. gwnum will choose the best FFT based upon the available hardware. If the best FFT uses AVX2, then it will use that FFT. If AVX2 is not available, then any FFT that requires AVX2 will not be used.