Support for ternary logic in SSE, XOP, AVX2 and x86 programs
C++ Python Makefile
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
py
.gitignore
Makefile
README.rst
ternary_avx2.cpp
ternary_avx512.cpp
ternary_sse.cpp
ternary_x86_32.cpp
ternary_x86_64.cpp
ternary_xop.cpp
validate_avx2.cpp
validate_sse.cpp
validate_x86.cpp
validate_xop.cpp

README.rst

Ternary functions for SSE and AVX2

Introduction

AVX512F introduced instruction vpternlog (intrinsic _mm512_ternarylogic_epi{32,64}), which evaluates a three-argument boolean function for each bit of operands. The function is given as an integer number from 0 to 255.

This repository contains a kind of library which allows a programmer to use ternary functions in SSE, XOP, AVX2 and x86 in similar manner, by choosing the function by its number.

In native AVX512F code you would write:

// 0x96 = A xor B xor C
const __m512i res = _mm512_ternarylogic_epi32(A, B, C, 0x96);

With help of the library you can write an AVX2 program:

const __m256i res = ternarylogic::avx2::ternary<0x96>(A, B, C);

or a SSE program:

const __m128i res = ternarylogic::sse::ternary<0x96>(A, B, C);

Details

The function number comes from the result column of the function's truth table. The number is formed from bits a, b, ... h. You can find more details and examples in my article.

A B C function
0 0 0 a
0 0 1 b
0 1 0 c
0 1 1 d
1 0 0 e
1 0 1 f
1 1 0 g
1 1 1 h

Intel has published list of all functions in form of logical expression in their monumental document "Intel® 64 and IA-32 Architectures Software Developer’s Manual Volume 2 (2A, 2B, 2C & 2D): Instruction Set Reference, A-Z".

I extracted these functions and created application which generates SSE and AVX2 code. List of function is in the file py/data/intel.txt, I also put there more details. Later I manually optimized some of these function. And in the end I developed a program which found many better representations of functions for both Intel and AMD flavours of SIMD ISA.

Usage

You need python and make.

Type make, then following files will be generated:

  • ternary_sse.cpp,
  • ternary_avx2.cpp,
  • ternary_avx512.cpp (use only two-argument logic instructions),
  • ternary_xop.cpp,
  • ternary_x86_32.cpp,
  • ternary_x86_64.cpp.

You can include them directly into your application.

Programs validate_sse, validate_avx2 and validate_xop test if all generated functions are correct.

See also