Not all aarch64 support 8.2 instructions

[Ramon Fried]

Build assumes that aarch64 is armv8.2, but ARM a53, and a72 are armv8.0
IF(XNNPACK_TARGET_PROCESSOR MATCHES "^(aarch64|arm64)$" OR IOS_ARCH MATCHES "^arm64.*") SET_PROPERTY(SOURCE ${XNNPACK_AARCH64_NEONFP16ARITH_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -march=armv8.2-a+fp16 ") SET_PROPERTY(SOURCE ${XNNPACK_NEONDOT_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -march=armv8.2-a+dotprod ") SET_PROPERTY(SOURCE ${XNNPACK_AARCH64_ASM_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -march=armv8.2-a+fp16+dotprod ")
This cause the following error when compiled using a53 cross-compiler:
| cc1: warning: switch '-mcpu=cortex-a72.cortex-a53' conflicts with '-march=armv8.2-a+fp16+dotprod' switch | xnnpack/src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S: Assembler messages: | xnnpack/src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S:125: Error: selected processor does not support fmla v20.8h,v16.8h,v0.h[0]'
| xnnpack/src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S:126: Error: selected processor does not support fmla v22.8h,v16.8h,v1.h[0]'

[Marat Dukhan]

Don't build with custom -mcpu flags. XNNPACK builds microkernels for all variants of AArch64, and choose which one to use in runtime.

[Ramon Fried]

Thanks. That's Interesting, I'm working on integrating Tensorflow lite to Yocto build, and it passes this.