ggml.c: error: 'SYS_getcpu' undeclared did you mean 'SYS_get_cpu'

[themanyone]

On both projects, whisper.cpp, and llama.cpp, we get this compilation error with latest git when compiling with CUBLAS.

WHISPER_CUBLAS=1 make -j

ggml.c: In function 'ggml_numa_init':
ggml.c:2088:26: error: 'SYS_getcpu' undeclared (first use in this function); did you mean 'SYS_get_cpu'?
 2088 |     getcpu_ret = syscall(SYS_getcpu,&current_cpu,&g_state.numa.current_node);
      |                          ^~~~~~~~~~
      |                          SYS_get_cpu
ggml.c:2088:26: note: each undeclared identifier is reported only once for each function it appears in
make: *** [Makefile:317: ggml.o] Error 1

The fix is very straightforward. We simply drop in SYS_get_cpu.

[petterreinholdtsen]

I created bull request #1897 based on your description, and hope it will be included in the master branch soon.

[ggerganov]

Hm, that's strange. We should be able to check which of the 2 is available (SYS_getcpu or SYS_get_cpu) and use the correct one

cc @bmtwl Any ideas?

Edit: there is already a discussion in llama.cpp: ggml-org/llama.cpp#5694 so let's figure it out there

[petterreinholdtsen]

[Georgi Gerganov]

Hm, that's strange. We should be able to check which of the 2 is available (`SYS_getcpu` or `SYS_get_cpu`) and use the correct one

I tried tracking down if and when the define changed in glibc, without any luck, so I can not tell which version of glibc to use as a separator. On the other hand, the code in /usr/include/x86_64-linux-gnu/bits/syscall.h look like this: #ifdef __NR_getcpu # define SYS_getcpu __NR_getcpu #endif It is thus very simple to use a construct like this: #if defined(SYS_getcpu) ... use SYS_getcpu #elif defined(SYS_get_cpu) ... use SYS_get_cpu #endif

…

-- Happy hacking Petter Reinholdtsen

[themanyone]

That should do it. I'm stuck using gcc-12 for acceleration, because Nvidia cuBLAS requires it for their old glibc workarounds (Nvidia's workarounds conflict with gcc-13's hot patches which [re]address the same issue, leading to conflicts).

[ggerganov]

Ok, would like to hear if @cebtenzzre agrees before merging

[bmtwl]

I tested the proposed strategy on a private branch and it worked through the permutations I tried.

+#if defined SYS_get_cpu
+ getcpu_ret = syscall(SYS_get_cpu,&current_cpu,&g_state.numa.current_node);
+#elif defined SYS_getcpu
getcpu_ret = syscall(SYS_getcpu,&current_cpu,&g_state.numa.current_node);
+#endif

Has a PR been submitted for this issue, or is there still a need for one?

[petterreinholdtsen]

[bmwl]

Has a PR been submitted for this issue, or is there still a need for one?

I created #1897 with my proposed solution. Feel free to adjust it if you want a different approach.

…

-- Happy hacking Petter Reinholdtsen