Disable the NEON optimisations on gcc <= 8

[JAicewizard]

GCC7 doesnt support all the necessary NEON intrinsics, which is really a shame. However this means that for aarch64 GCC cannot compile faiss with neon intrinsics, so we should avoid using them.

This is similar to #3860, build issues on GCC7, which I need. This one is a bit uglier, since GCC7 does support NEON just not all of the intrinsics.

[alexanderguzhva]

@JAicewizard this should not affect clang, correct? because clang may pretend in certain situations

[JAicewizard]

I just checked on CC explorer here and even with the oldest aarch64 clang there were no issues. So there might be an old clang compiler that doesn't work, but I am not in a position to test all of them. I suspect they were just always supported.

[facebook-github-bot]

@junjieqi has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[junjieqi]

@JAicewizard we received bunch of errors internally. In case, you already know the issue before I look into it.

fbcode/faiss/impl/ScalarQuantizer.cpp:712:25: error: unknown type name 'float32x4x2_t'
[CONTEXT]   712 |     FAISS_ALWAYS_INLINE float32x4x2_t
[CONTEXT]       |                         ^
fbcode/faiss/impl/ScalarQuantizer.cpp:714:9: error: unknown type name 'uint8x8_t'; did you mean 'uint8_t'?
[CONTEXT]   714 |         uint8x8_t x8 = vld1_u8((const uint8_t*)(code + i));
[CONTEXT]       |         ^~~~~~~~~
fbcode/faiss/impl/ScalarQuantizer.cpp:714:24: error: use of undeclared identifier 'vld1_u8'
[CONTEXT]   714 |         uint8x8_t x8 = vld1_u8((const uint8_t*)(code + i));
[CONTEXT]       |                        ^
fbcode/faiss/impl/ScalarQuantizer.cpp:715:9: error: unknown type name 'uint16x8_t'; did you mean 'uint16_t'?
[CONTEXT]   715 |         uint16x8_t y8 = vmovl_u8(x8); // convert uint8 -> uint16
[CONTEXT]       |         ^~~~~~~~~~
fbcode/faiss/impl/ScalarQuantizer.cpp:715:25: error: use of undeclared identifier 'vmovl_u8'
[CONTEXT]   715 |         uint16x8_t y8 = vmovl_u8(x8); // convert uint8 -> uint16
[CONTEXT]       |                         ^
fbcode/faiss/impl/ScalarQuantizer.cpp:716:9: error: unknown type name 'uint16x4_t'
[CONTEXT]   716 |         uint16x4_t y8_0 = vget_low_u16(y8);
[CONTEXT]       |         ^
fbcode/faiss/impl/ScalarQuantizer.cpp:716:27: error: use of undeclared identifier 'vget_low_u16'
[CONTEXT]   716 |         uint16x4_t y8_0 = vget_low_u16(y8);
[CONTEXT]       |                           ^
fbcode/faiss/impl/ScalarQuantizer.cpp:717:9: error: unknown type name 'uint16x4_t'
[CONTEXT]   717 |         uint16x4_t y8_1 = vget_high_u16(y8);
[CONTEXT]       |         ^
fbcode/faiss/impl/ScalarQuantizer.cpp:717:27: error: use of undeclared identifier 'vget_high_u16'
[CONTEXT]   717 |         uint16x4_t y8_1 = vget_high_u16(y8);
[CONTEXT]       |                           ^
fbcode/faiss/impl/ScalarQuantizer.cpp:719:9: error: unknown type name 'float32x4_t'
[CONTEXT]   719 |         float32x4_t z8_0 = vcvtq_f32_u32(
[CONTEXT]       |         ^
fbcode/faiss/impl/ScalarQuantizer.cpp:721:9: error: unknown type name 'float32x4_t'
[CONTEXT]   721 |         float32x4_t z8_1 = vcvtq_f32_u32(vmovl_u16(y8_1));
[CONTEXT]       |         ^
fbcode/faiss/impl/ScalarQuantizer.cpp:724:33: error: use of undeclared identifier 'vmovq_n_f32'
[CONTEXT]   724 |         return {vsubq_f32(z8_0, vmovq_n_f32(128.0)),
[CONTEXT]       |                                 ^
fbcode/faiss/impl/ScalarQuantizer.cpp:725:33: error: use of undeclared identifier 'vmovq_n_f32'
[CONTEXT]   725 |                 vsubq_f32(z8_1, vmovq_n_f32(128.0))};

[JAicewizard]

This should be fixed now

[junjieqi]

github action failed. @JAicewizard could you help fix them before I import again? Thanks!