https://github.com/eriksjolund/compile-time-simd-blend-mask

Compile-time blend masks that unifies _mm256_blend_epi8, _mm256_blend_epi16, _mm256_blend_epi32
https://github.com/eriksjolund/compile-time-simd-blend-mask

avx2 blend boost-hana compile-time

Last synced: 30 days ago
JSON representation

Compile-time blend masks that unifies _mm256_blend_epi8, _mm256_blend_epi16, _mm256_blend_epi32

• Host: GitHub
• URL: https://github.com/eriksjolund/compile-time-simd-blend-mask
• Owner: eriksjolund
• License: bsl-1.0
• Created: 2018-12-17T19:04:33.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2018-12-19T21:12:03.000Z (about 6 years ago)
• Last Synced: 2024-10-31T17:45:19.908Z (3 months ago)
• Topics: avx2, blend, boost-hana, compile-time
• Language: C++
• Homepage:
• Size: 25.4 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# compile-time-simd-blend-mask

Compile-time blend masks that unifies _mm256_blend_epi8, _mm256_blend_epi16, _mm256_blend_epi32

by using the C++ library [boost::hana](http://boostorg.github.io/hana)

## Introduction

The intrinsics functions

* [__m256i _mm256_blendv_epi8(__m256i v1, __m256i v2, __m256i mask)](https://software.intel.com/en-us/node/523908)

* [__m256i _mm256_blend_epi16(__m256i a, __m256i b, const int imm8)](https://software.intel.com/en-us/cpp-compiler-developer-guide-and-reference-mm-blend-epi32-mm256-blend-epi16-32)

* [__m256i _mm256_blend_epi32(__m256i a, __m256i b, const int imm8)](https://software.intel.com/en-us/cpp-compiler-developer-guide-and-reference-mm-blend-epi32-mm256-blend-epi16-32)

serve similar purpose, but they take different arguments. The first function encodes the blend mask in the SIMD vector `__256i`, 

but the two latter encode the blend mask in an `int`. The value of such an `int` needs to be known at compile-time (unlike the `__m256i` mask for the first function).

## Example

This project implements a common API that can be used like this

```C++

    auto mask = hana::make_tuple(

      hana::false_c,  hana::false_c,

      hana::false_c, hana::false_c, 

      hana::true_c, hana::true_c, 

      hana::true_c, hana::true_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::false_c, hana::false_c, 

      hana::true_c, hana::true_c, 

      hana::true_c, hana::true_c, 

      hana::true_c, hana::true_c, 

      hana::true_c, hana::true_c

    );

    __m256i a = _mm256_set1_epi8(0);

    __m256i b = _mm256_setr_epi8( 1,  2,  3,  4,  5,  6,  7,  8,

                                  9, 10, 11, 12, 13, 14, 15, 16,

                                 17, 18, 19, 20, 21, 22, 23, 24,

                                 25, 26, 27, 28, 29, 30, 31, 32);

  auto blend_result = blend256(a, b, mask);

```

At compile-time the mask will be analyzed and the fastest possible intrinsics function will be chosen (in this case 

`             _mm256_blend_epi32()`).

 If the mask allows it  `_mm256_blend_epi32()` will be used.

 Otherwise  if the  mask allows it  `_mm256_blend_epi16()` will be used.

 Otherwise `_mm256_blend_epi8()` will be used.

 

# Implementation details

The file

[src/blend256.h](src/blend256.h)

contains the implementation of the `blend256()` function.

The implementation makes use of the C++ library [boost::hana](http://boostorg.github.io/hana).