Problems with installing git on windows 7

[mareichhoff]

Hello,

this is my error output (last lines of the output in cygwin):

/usr/bin/ar: strmm_kernel_LN.obj: No such file or directory
make[1]: *** [../Makefile.tail:40: libs] Error 1
make[1]: Leaving directory '/c/Users/meichho/Documents/openBlas/kernel'
make: *** [Makefile:139: libs] Error 1

meichho@SC-R90FWVVK-1 /cygdrive/c/Users/meichho/Documents/openBlas
$ make install
/usr/bin/make -f Makefile.install install
make[1]: Entering directory '/c/Users/meichho/Documents/openBlas'
Makefile.install:21: *** OpenBLAS: Please run "make" firstly.  Stop.
make[1]: Leaving directory '/c/Users/meichho/Documents/openBlas'
make: *** [Makefile:322: install] Error 2

Then I tried "make install", but it didn't work. I don't understand the error message above.

Thank you very much for your help!

Best regards,
Markus

[martin-frbg]

Could be that autodetection of your CPU failed for some reason, or your compiler version is too old to understand some of the assembly code used (e.g. AVX2 commands). Can you give more information about your system and compiler, and upload the full build log (or at least significantly more than you showed above) please ?

[mareichhoff]

I installed cygwin some days ago. Everything is new. I try to paste the output here in some minutes. Thank you for your help!

[martin-frbg]

I'll be offline now for an hour or two. In the meantime you could try building for a specific cpu - do a make clean first, and then do make TARGET=NEHALEM (should work on most older x86_64 targets) or make TARGET=HASWELL (should work on most any x86_64 "desktop category" cpu released in the last couple of years).

[brada4]

Can you capture full long output with "script" command:
make clean
script
make
make install
exit
Now upload long typesript file.

[mareichhoff]

Hello Martin, brada,

I looked in "system". My intel processor is "Haswell". So I tried "make clean" and then "make TARGET=HASWELL" without success. I can't upload the full output. These exceeds the space here. Or is it possible to mail it to you?

Best regards,
Markus

[mareichhoff]

Intel Core i5-4210M CPU

[brada4]

Should be detected fine as haswell correctly already few releases back.
You can attach output by renaming it to *.txt

[mareichhoff]

Ah I see..... drag & drop... I will do it now. After this I will try it with sript in between an put the output here again.
make TARGET=HASWELL.txt

[brada4]

I think you need to cross-compile if you want 32bit DLL (i.e i686-mingw-something-gcc etc)

gcc -O2 -DMS_ABI -mincoming-stack-boundary=2 -DMAX_STACK_ALLOC=2048 -Wall -m32 -DF_INTERFACE_GFORT -DNO_AVX -DSMP_SERVER -DNO_WARMUP -DMAX_CPU_NUMBER=4 -DASMNAME=_strmm_kernel_LN -DASMFNAME=_strmm_kernel_LN_ -DNAME=strmm_kernel_LN_ -DCNAME=strmm_kernel_LN -DCHAR_NAME=\"strmm_kernel_LN_\" -DCHAR_CNAME=\"strmm_kernel_LN\" -DNO_AFFINITY -I.. -UDOUBLE  -UCOMPLEX -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA ../kernel/x86/ -o strmm_kernel_LN.obj
gcc.exe: warning: ../kernel/x86/: linker input file unused because linking not done

[mareichhoff]

O.k., thank you very much! I will try it.

[martin-frbg]

I believe 32bit Haswell support is broken (or just not properly supported), make TARGET=SANDYBRIDGE would probably work

[brada4]

It is replaced by westmere/nehalem in make files (though nice to have somebody to confirm it is still the case)

[martin-frbg]

Trying to make a 32-bit "Haswell target" library on a 64-bit system automatically results in a Nehalem build, apparently by design. Trying to build a Haswell llibrary on a 32-bit system however leads to the error reported at the top of this thread.
Apparently the kernel/x86/KERNEL.HASWELL file needs xTRMMKERNEL definitions added, but choosing either of the various trmmkernel_ implementations available in "generic" gives the dreaded LESS THAN HALF ACCURATE errors in all xTRMM tests. Seems the USE_TRMM=1 for HASWELL in kernel/Makefile.L3 needs an additional "ifeq($ARCH), X86_64)" qualifier to get around this, but I am not yet through with my tests.

[martin-frbg]

Build issue fixed by #1333 now - note though that 32bit builds for Haswell and later will still be using only Penryn-level (SSE) optimizations internally.

[martin-frbg]

@mareichhoff did you get a chance to try ?