gcc-8: _mm_loadu_si64 missing

[jploski]

The library does not compile using gcc<9 (e.g. on Debian 10) because of missing _mm_loadu_si64 intrinsic (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78782):

[  6%] Built target ggml_utils
[  6%] Building C object src/CMakeFiles/ggml.dir/ggml.c.o
/tmp/ggml/src/ggml.c: In function ‘bytes_from_nibbles_16’:
/tmp/ggml/src/ggml.c:476:19: warning: implicit declaration of function ‘_mm_loadu_si64’; did you mean ‘_mm_loadl_epi64’? [-Wimplicit-function-declaration]
     __m128i tmp = _mm_loadu_si64( ( const __m128i* )rsi );
                   ^~~~~~~~~~~~~~
                   _mm_loadl_epi64
/tmp/ggml/src/ggml.c:476:19: error: incompatible types when initializing type ‘__m128i’ {aka ‘__vector(2) long long int’} using type ‘int’
make[2]: *** [src/CMakeFiles/ggml.dir/build.make:63: src/CMakeFiles/ggml.dir/ggml.c.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:91: src/CMakeFiles/ggml.dir/all] Error 2
make: *** [Makefile:141: all] Error 2

Not sure if this is a bug since Debian 10 is pretty old, but if there is an easy fix and no other reasons that require gcc-9, it may be worth staying backward-compatible. Otherwise, maybe fail more gracefully (and/or mention gcc-9 in the docs).

[jploski]

Note: this now also affects llama.cpp (as of commit d40fded) - which used to compile using gcc-8.3.0 before...

[prusnak]

llama.cpp issue: ggml-org/llama.cpp#1120

[jploski]

I noticed that simply replacing the invocation of _mm_loadu_si64 by _mm_loadl_epi64 in ggml.c:439 allows the code to compile using gcc-8 again - and the quantized models seem to run just as well as before. This appears to be the only reference in the entire code base - perhaps you could add some #ifdef magic to make us folks with old compilers happy?...

[ggerganov]

PR welcome - not sure what #ifdef to use

[ggerganov]

The fix was suggested by @sw in the llama.cpp repo: ggml-org/llama.cpp#1154

[jploski]

Great, if no #ifdef necessary, so much the better.