Degraded performance on Zen 4 CPUs

[MattCatz]

The handwritten kernel for the __AVX512F__ path uses the intrinsic _mm512_mask_compressstoreu_pd which maps to the instruction VPCOMPRESSPD. For Zen 4, that instruction is microcode and slower than the scaler path.

ALP/include/alp/encoder.hpp

Lines 351 to 361 in 31ca0ed

	#ifdef __AVX512F__
	if constexpr (std::is_same_v<PT, double>) {
	for (size_t i {0}; i < config::VECTOR_SIZE; i = i + 8) {
	__m512d l = _mm512_loadu_pd(ENCODED_VALUE_ARR + i);
	__m512d r = _mm512_loadu_pd(VALUE_ARR_WITHOUT_SPECIALS + i);
	__m512i index = _mm512_loadu_pd(DOUBLE_INDEX_ARR + i);
	auto is_exception = _mm512_cmpneq_pd_mask(l, r);
	_mm512_mask_compressstoreu_pd(TMP_INDEX_ARR + exceptions_idx, is_exception, index);
	exceptions_idx += LOOKUP_TABLE[is_exception];
	}
	} else {