gnfi-shift.cpp

#include <cstdint>
#include <cstdio>
#include <bitset>
#include <immintrin.h>

// Attempts at implementing _mm_srai_epi8, _mm_slli_epi8, and _mm_srli_epi8
// using affine galois field transformations(_mm_gf2p8affine_epi64_epi8, GNFI)
// Wed Nov  4 05:34:35 PM PST 2020 - wunkolo@gmail.com

inline __m128i _mm_srai_epi8(__m128i a, std::uint8_t imm8)
{
	const std::uint64_t sign_extend = ~(0xFFFFFFFFFFFFFFFF << (imm8 * 8)) & 0x8080808080808080;
	return _mm_gf2p8affine_epi64_epi8(
		a,
		// Perform a logical shift right, but shift in 0x80 bytes
		// to replicate the most significant bit
		_mm_set1_epi64x(
			0x0102040810204080 << (imm8 * 8)
			// Simulate shifting in 0x80 bits by doing a bit-wise OR
			// against masked 0x80 bytes
			| sign_extend
			//| _bzhi_u64(0x8080808080808080ull, imm8 * 8) // BMI2
		),
		0
	);
}

inline __m128i _mm_slli_epi8(__m128i a, std::uint8_t imm8)
{
	return _mm_gf2p8affine_epi64_epi8(
		a, _mm_set1_epi64x(0x0102040810204080 >> (imm8 * 8)), 0
	);
}

inline __m128i _mm_srli_epi8(__m128i a, std::uint8_t imm8)
{
	return _mm_gf2p8affine_epi64_epi8(
		a, _mm_set1_epi64x(0x0102040810204080 << (imm8 * 8)), 0
	);
}

void Print128(const __m128i& Vector)
{
	const std::int8_t* Bytes = reinterpret_cast<const std::int8_t*>(&Vector);
	std::putchar('\t');	for(std::size_t i=0;i<16;++i) std::printf("%s'",std::bitset<8>(Bytes[i]).to_string().c_str());	std::putchar('\n');
	std::putchar('\t');	for(std::size_t i=0;i<16;++i) std::printf("%d'",Bytes[i]);										std::putchar('\n');
}

int main()
{
	{
		const __m128i TestVector = _mm_set1_epi8(0b10010000);
		std::puts("Input");		Print128(TestVector);
		std::puts("SRA(3)");	Print128(_mm_srai_epi8(TestVector,3));
		std::puts("SLL(3)");	Print128(_mm_slli_epi8(TestVector,3));
		std::puts("SRL(3)");	Print128(_mm_srli_epi8(TestVector,3));
	}
}

// Input
// 	10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'10010000'
// 	-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'-112'
// SRA(3)
// 	11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'11110010'
// 	-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'-14'
// SLL(3)
// 	10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'10000000'
// 	-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'-128'
// SRL(3)
// 	00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'00010010'
// 	18'18'18'18'18'18'18'18'18'18'18'18'18'18'18'18'