Control: tag -1 confirmed Good day,
having a closer look at kmc, there is simde set up, and it looks
like enabling -march=x86-64-v2 leads through a buggy build path.
Some parts of the source code are designed to build against some
specific combinations of machine specific flags. In the present
case, if I inject -march=x86-64-v2, and some diagnostic output,
then a mismatch appears between code targetting the specific cpu
capability sse2, caused by availability of sse4.1 in build
options:
vvvv
In file included from kmer_counter/raduls_sse2.cpp:12:
kmer_counter/raduls_impl.h:734:2: warning: #warning
"RADULS_RADIX_SORT_FUNNAME=RadixSortMSD_SSE41" [-Wcpp]
^^^^^
thus, explaining the error reported by Matthias:
> ./kmer_counter/kmc.h:1132: undefined reference to `void
> RadulsSort::RadixSortMSD_SSE2<CKmer<2u> >(CKmer<2u>*, CKmer<2u>*, unsigned
> long long, unsigned int, unsigned int, CMemoryPool*)'
> collect2: error: ld returned 1 exit status
To a larger extent, this is bound to fail with additional
symptoms when using machine types x86-64-v3, as it appends
various avx support, without mentionning various combinations
thrown by -march=native. I see two options to mitigate this:
1. either disable build of raduls_sse2.cpp, which is unneeded
in x86-64-v2 context, since it is overred by the sse4.1
implementation raduls_sse41.cpp any ways;
2. or cheat with the C preprocessor macro definitions, to get
back the missing function in its sse2 only context.
Point 1 looks cleaner to me, but I only have been able to
implement point 2 successfully so far, without causing baseline
violations in normal builds, nor beating the purpose of
optimisations, and with support for further flags combinations
than just the baseline or -march=x86-64-v2. Will push changes
on Salsa this evening, most likely.
Have a nice day, :)
--
Étienne Mollier <[email protected]>
Fingerprint: 8f91 b227 c7d6 f2b1 948c 8236 793c f67e 8f0d 11da
Sent from /dev/pts/6, please excuse my verbosity.
signature.asc
Description: PGP signature
