AVX512F: _mm512_fmaddsub_pd is missing

[robinchrist]

It seem like the AVX512F _mm512_fmaddsub_pd intrinsic is missing.

Could it please be added? :)

[mr-c]

Hello @robinchrist ; I would accept a PR to add a simde_mm512_fmaddsub_pd in a new file named simde/x86/avx512/fmaddsub.h ; I can assist you in that if you would like. Take a look at the following:

• https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_fmaddsub_pd&ig_expand=3171
• simde_mm_fmaddsub_pd
• simde_mm256_fmaddsub_pd

[robinchrist]

@mr-c yeah this is way too complicated for me, honestly

The complexity lies in the addsub, e.g.

simde/simde/x86/avx.h

Lines 1734 to 1767 in fa6a869

	SIMDE_FUNCTION_ATTRIBUTES
	simde__m256d
	simde_mm256_addsub_pd (simde__m256d a, simde__m256d b) {
	#if defined(SIMDE_X86_AVX_NATIVE)
	return _mm256_addsub_pd(a, b);
	#else
	simde__m256d_private
	r_,
	a_ = simde__m256d_to_private(a),
	b_ = simde__m256d_to_private(b);
	#if defined(SIMDE_LOONGARCH_LASX_NATIVE)
	simde__m256d_private aev_, aod_, bev_, bod_;
	aev_.i256 = __lasx_xvpickev_d(a_.i256, a_.i256);
	aod_.i256 = __lasx_xvpickod_d(a_.i256, a_.i256);
	bev_.i256 = __lasx_xvpickev_d(b_.i256, b_.i256);
	bod_.i256 = __lasx_xvpickod_d(b_.i256, b_.i256);
	aev_.d256 = __lasx_xvfsub_d(aev_.d256, bev_.d256);
	aod_.d256 = __lasx_xvfadd_d(aod_.d256, bod_.d256);
	r_.i256 = __lasx_xvilvl_d(aod_.i256, aev_.i256);
	#elif SIMDE_NATURAL_VECTOR_SIZE_LE(128)
	r_.m128d[0] = simde_mm_addsub_pd(a_.m128d[0], b_.m128d[0]);
	r_.m128d[1] = simde_mm_addsub_pd(a_.m128d[1], b_.m128d[1]);
	#else
	SIMDE_VECTORIZE
	for (size_t i = 0 ; i < (sizeof(r_.f64) / sizeof(r_.f64[0])) ; i += 2) {
	r_.f64[ i ] = a_.f64[ i ] - b_.f64[ i ];
	r_.f64[i + 1] = a_.f64[i + 1] + b_.f64[i + 1];
	}
	#endif
	return simde__m256d_from_private(r_);
	#endif
	}

I can TRY, but don't expect a great solution. Especially not something that has LoongArch or NEON in it

[robinchrist]

Would you want the non-fma addsub in a separate file? Or in the same file as simde_mm512_fmaddsub_pd?

[robinchrist]

Hm, interesting. A _mm512_addsub_pd does not exist.
Should _mm512_addsub_pd be added nevertheless, or only with with the simde_ prefix?
Or should I simply embed the functionality of _mm512_addsub_pd in _mm512_fmaddsub_pd and not offer it separately?

[mr-c]

I can TRY, but don't expect a great solution. Especially not something that has LoongArch or NEON in it

We only need a plain implementation to merge; no SIMDE_*_NATIVE code needed. Others will come along with optimized implementations later. Often the compilers do a great job!

Hm, interesting. A _mm512_addsub_pd does not exist. Should _mm512_addsub_pd be added nevertheless, or only with with the simde_ prefix? Or should I simply embed the functionality of _mm512_addsub_pd in _mm512_fmaddsub_pd and not offer it separately?

Your choice. If you need it, then call it simde_x_mm512_addsub_pd, and it can go in the same file.

[robinchrist]

I've made a first PR #1246, no clue whether this works yet (tests should / will follow).
Can you please let me know in the PR whether this goes into the desired direction?

[mr-c]

Fixed in #1246 ; thank you!