Thread 1 "main" received signal SIGILL, Illegal instruction.

[biemster]

Seemed to come from here:

0x00005555555586fb in _mm256_fmadd_ps (__C=..., __B=..., __A=...) at /usr/lib/gcc/x86_64-linux-gnu/7/include/fmaintrin.h:65
65	  return (__m256)__builtin_ia32_vfmaddps256 ((__v8sf)__A, (__v8sf)__B,

with backtrace

(gdb) bt
#0  0x00005555555586fb in _mm256_fmadd_ps (__C=..., __B=..., __A=...) at /usr/lib/gcc/x86_64-linux-gnu/7/include/fmaintrin.h:65
#1  ggml_vec_dot_f16 (n=96, s=0x7ffffffe4e54, x=0x7fff646b6ee0, y=0x7fff64746ee0) at ggml.c:375
#2  0x0000555555564766 in ggml_compute_forward_conv_1d_1s_f16_f32 (params=0x7ffffffe51c0, src0=0x7fff9025f0f0, src1=0x7fff65482030, dst=0x7fff6556c6f0) at ggml.c:4668
#3  0x0000555555564f40 in ggml_compute_forward_conv_1d_1s (params=0x7ffffffe51c0, src0=0x7fff9025f0f0, src1=0x7fff65482030, dst=0x7fff6556c6f0) at ggml.c:4806
#4  0x0000555555568707 in ggml_compute_forward (params=0x7ffffffe51c0, tensor=0x7fff6556c6f0) at ggml.c:5809
#5  0x000055555556a6ec in ggml_graph_compute (ctx=0x5555557f3b48 <g_state+104>, cgraph=0x7ffffffe5340) at ggml.c:6611
#6  0x0000555555580cb2 in whisper_encode (model=..., n_threads=4, mel_offset=0, mel_inp=..., features=std::vector of length 0, capacity 0) at main.cpp:1353
#7  0x0000555555584664 in main (argc=5, argv=0x7fffffffdb78) at main.cpp:2225

On Ubuntu 18.04, gcc 7.5.0, on an Intel(R) Core(TM) i5-3470 CPU @ 3.20GHz

[biemster]

It seems my CPU does not support FMA, whatever that is :(
There might be a workaround described here: https://stackoverflow.com/a/42854597/11545161
I'm going to try that. Or throw this machine out the window

[biemster]

I see now you did some changes in the rpi4 branch in this area, since that also does not seem to have FMA, I'm going to have a look there too. I seem to be a bit too trigger happy on opening new issues today, sorry for the noise.

[ggerganov]

Yes, currently on x86 architectures your CPU must support the FMA and F16C instruction sets. You can check if your CPU supports it with this command:

tail /proc/cpuinfo | grep flags

The flags avx2, fma and f16c should be listed there.

The rpi4 branch is for Arm architecture, so it won't work on your Intel CPU.

The workaround you can try is to remove the -mavx -mavx2 -mfma -mf16c flags from the Makefile and rebuild. ggml should fallback to non-SIMD implementation which is much slower, but it should at least work on every machine.

[WilliamTambellini]

Using/linking with
https://github.com/oneapi-src/oneDNN
should resolve these kind of issues.