As shown in the screenshot above, there are always building erros when i use some NEON Instruction in float16_t type(such as vaddq_f16)~
I suspect it is ARMRAL_ARCH_COMPILE_OPTIONS issues which need i to modify "-march=armv8-a+crypto" CACHE INTERNAL ".
I tried many methods and the compilation error still exists。
Can anybody tell me how to resolve this erros?
Thanks a lot!!
Hi Zadoff,
Did you manage to solve the issue?
target specific option mismatch is indicating a mismatch between compile target and the SIMD instructions.
Since you have tried many methods, may I ask whether you have tried replacing "-march" flag with -mcpu=native yet ??
This would turn on everything your processor has and that should help the compiler pick the SIMD enabled architecture pieces.Please let me know if that hadn't worked already !?
Suyash
Sr. Applications Engineer - HPC
Hello again.On a second look at the error message, it appears that there is a possibility your target architecture may not be compatible with the ACLE vaddq_f16 accumulation function, which compiles to FADD in NEON, or even vmulq_f16 for that matter. This is because half precision floating point intrinsics were only enabled from ArmV8.2-A ISA onwards. Make sure that you have the required target architecture and if that is indeed the case, you can go ahead and use either of -mcpu/-march =native.Meanwhile I am running a diagnosis and trying to reproduce the error on my end.