porting @Microsoft SPTAG (native aliases enabled) (continued): ): _mm512_madd_epi16 wrong args, __mmask64 typo, wrong definition, compile warnings

[pabs3]

@mr-c Thanks for the upload of 0.7.4 to Debian experimental and fixing #961 but unfortunately that wasn't quite enough for me to build SPTAG.

Firstly commit ae545ce broke the definition of _mm512_madd_epi16, it mistakenly added two arguments to the start of the arguments, it should have two arguments not four, like the simde_mm512_madd_epi16 function does.

Secondly there was a typo in commit d850b83, __mask64 should be __mmask64 instead.

Thirdly the typedef uint64_t __mmask64; definition of __mmask64 conflicts with the one in GCC from Debian bookworm, which defines it with typedef unsigned long long __mmask64; in avx512bwintrin.h.

Fourthly the code successfully compiled but with a couple of GCC warnings:

In file included from /usr/include/simde/x86/avx512/mov.h:32,
                 from /usr/include/simde/x86/avx512/dpwssd.h:5,
                 from /usr/include/simde/x86/avx512/4dpwssd.h:5,
                 from /usr/include/simde/x86/avx512.h:33,
                 from /home/pabs/sptag-test-simde/SPTAG.git/AnnService/inc/Core/Common/InstructionUtils.h:16,
                 from /home/pabs/sptag-test-simde/SPTAG.git/AnnService/inc/Core/Common/DistanceUtils.h:11,
                 from /home/pabs/sptag-test-simde/SPTAG.git/AnnService/src/Core/Common/DistanceUtils.cpp:4:
/usr/include/simde/x86/avx512/cast.h: In function 'simde__m512 simde_mm512_castsi512_ps(simde__m512i)':
/usr/include/simde/x86/avx512/cast.h:105:1: note: the ABI for passing parameters with 64-byte alignment has changed in GCC 4.6
  105 | simde_mm512_castsi512_ps (simde__m512i a) {
      | ^~~~~~~~~~~~~~~~~~~~~~~~
/home/pabs/sptag-test-simde/SPTAG.git/AnnService/src/Core/Common/DistanceUtils.cpp: In function 'simde__m512i simde_mm512_add_epi32(simde__m512i, simde__m512i)':
/home/pabs/sptag-test-simde/SPTAG.git/AnnService/src/Core/Common/DistanceUtils.cpp:1051:1: warning: AVX512F vector return without AVX512F enabled changes the ABI [-Wpsabi]
 1051 | }
      | ^

[mr-c]

You are welcome!

Fixes for items 1 and 2 are inbound.

Thirdly the typedef uint64_t _mmask64; definition of _mmask64 conflicts with the one in GCC from Debian bookworm, which defines it with typedef unsigned long long __mmask64; in avx512bwintrin.h.

Looks like C99 & C++11 guarantee that "unsigned long long" is at least 64 bits in any data model, but could be longer. I think defining __mask64 as uint64_t is fine, yes? Is that giving you an error message, or were you (understably) just being extra cautious?

More digging: 64 bits minimum is also in C++03 TR1

https://en.cppreference.com/w/cpp/language/types

Fourthly the code successfully compiled but with a couple of GCC warnings:

simde/simde/simde-diagnostic.h

Lines 82 to 89 in 314452b

	/* GCC emits a lot of "notes" about the ABI being different for things
	* in newer versions of GCC. We don't really care because all our
	* functions are inlined and don't generate ABI. */
	#if HEDLEY_GCC_VERSION_CHECK(7,0,0)
	#define SIMDE_DIAGNOSTIC_DISABLE_PSABI_ _Pragma("GCC diagnostic ignored \"-Wpsabi\"")
	#else
	#define SIMDE_DIAGNOSTIC_DISABLE_PSABI_
	#endif

Feel free to compile with -Wno-psabi to avoid this, or prefix your usage with SIMDE_DIAGNOSTIC_DISABLE_PSABI_

[pabs3]

Thanks. Before I was getting an error with `uint64_t` `__mmask64` but just now I retried everything from scratch and didn't get any error. `SIMDE_DIAGNOSTIC_DISABLE_PSABI_` is automatically enabled for GCC version 7 and above, since I'm using GCC 12, so it should be enabled already but I am still seeing the two warnings.

…

-- bye, pabs https://bonedaddy.net/pabs3/

[mr-c]

FYI: In e8390a3 I did add a unsigned long long variant for gcc.

SIMDE_DIAGNOSTIC_DISABLE_PSABI_ is automatically enabled for GCC version 7 and above, since I'm using GCC 12, so it should be enabled already but I am still seeing the two warnings.

Hmm.. I don't know what to say, sorry. Can you try adding SIMDE_DIAGNOSTIC_DISABLE_PSABI_ to the SPTAG source files that generate warnings??

[pabs3]

I think the `unsigned long long __mmask64` is not actually needed, but I guess it probably causes no harm anyway. I forgot that the diagnostic disabling was only per source file. Adding `SIMDE_DISABLE_UNWANTED_DIAGNOSTICS` after I include the simde headers got rid of the warnings originating from those files, but there is still the warning from `avx.h`, which already uses the macro and the warning doesn't go away even when I add the macro just before the function that causes the warning. For some reason none of the other functions in the `avx.h` file generate the warning. Anyway, if the warning issue isn't of concern to you, then you can close this bug report now.

…

-- bye, pabs https://bonedaddy.net/pabs3/

[mr-c]

Yeah, I think some of the fixes from SIMDE_DISABLE_UNWANTED_DIAGNOSTICS don't stick for #defined functions.

Thanks again for the feedback!