we are using ATLAS successfully in our speech processing/recognition library. ATLAS helps us mainly for efficient implementation of classification by neural nets. We've done a benchmark comparing Intel's MKL, AMD's ACML and ATLAS, which proved to suit our needs the best. As MS Win is OS of our target audience, we've been using ATLAS compiled for Win as well. Everybody has been happy with a 32-bit version so far but a demand for 64-bit version of our library now emerges. So I began to investigate the possibility of installing 64-bit ATLAS on Win x64.
I tried to compile ATLAS using mingw-w64_x86-64_cygwin32. I spent a lot of time trying to pass through the whole build process but without success. I ended up with a configuration that used mingw-w64_x86-64 compilers for all the tasks but harness routines compilation (XCC). I used default (32-bit!) GCC built in Cygwin for that. I hope this is not a problem as it should serve only for building helper (not even the tune) tools for building process. The reason for using biult-in GCC was that mingw-w64_x86-64_cygwin32's standard functions behave like native windows and not cygwin/linux ones. Problems were with absolute paths used in fopen(), system(), etc.
I've passed through the configuration step. Then I've just slightly edited Make.inc to substitute 'gcc x86_64-w64-mingw32-gcc' by 'x86_64-w64-mingw32-gcc' in GOODGCC (have no idea how the original value with two compilers mixed into the single variable was assigned). Then I ran 'make build'. Making itself seemed to run smoothly but each (tuning) binary which was built using libatlas.a ended up with segmentation fault. Using GDB I was just able to find out that it failed during execution of some ASM code (e.g. in ATL_daxpy_xp1yp1aXbX.c).
I just wanted to ask if you have some experience with installing ATLAS on Win x64 or at least know that it should be somehow possible. Any help would be greatly appreciated. I am a bit afraid that problem might be in mingw-w64_x86-64 compiler which is probably not completely mature yet.
Thank you very much.