Faster GPU-ECM with CGBN

[Denial140]

My understanding was that the process for building it for GPU is the same, expect for additional parameters to the configure step. If this is wrong, then that may well explain it.

I managed to get fully through INSTALL-ecm when not trying to add GPU capability.

When I set the compiler to gcc (have tried gcc, gcc-10 and gcc-9 commands), the configure fails with "cuda.h present but cannot be compiled". I tried also to get gcc-8 as suggested in readme.gpu, but installing gcc-8 seems to uninstall the Nvidia toolkit (whether I use version 12.0 or 11.2 of the toolkit), and vice versa. If instead I set it to nvcc, I get a similar error for gmp.h. The only thing that gets past this stage is setting it to g++, which causes issues further down the line.

For this, it fails in the mentioned ways when running

./configure --enable-gpu --with-cgbn-include=../CGBN/include/cgbn CC="g++ -fpermissive" CFLAGS="-02 -pedantic -fomit-rame-pointer -m64 -mtune=skylake -march=broadwell"

, or when removing the --with-cgbn-include parameter and/or the CFLAGS.

The configure works fine with g++, but make gives several implicit type conversion errors (presumably as a result of treating main.c as C++? This was the impression I got from google.) Adding -fpermissive bypasses these, but gives a couple of errors about definitions crossing label jumps in cudawrapper.c. After adding some variable declarations to account for this, I get stuck with the following errors:

This, and the code block below it, happened when running make after the above configure.

[paulunderwood]

What arguments do you give ./configure when compiling ECM and what arguments do you give it when compiling for GPU capability?

(ps. surely CC="g++...." is plain wrong!)

[Denial140]

I have successfully compiled with GPU capability and CGBN - the important thing I was missing was the line "egrep -w '(CC|CFLAGS)' config.log" in INSTALL-ecm to get the compiler options used to compile GMP. My initial attempts at adding these failed, as I had unwittingly recompiled GMP with g++, which seems to leave certain things missing given the errors I was getting. The successful line in the end was

./configure --enable-gpu --with-cgbn-include=../../gmp-ecm/CGBN/include/cgbn --with-cuda=/usr/local/cuda-12.0 CFLAGS="-O2 -pedantic -fomit-frame-pointer -m64 -mtune=skylake -march=broadwell"

You may well be right that CC="g++ -fpermissive" is wrong - I must admit I have no idea what I'm doing, but it did reduce the type conversion errors to warnings.