堆栈内存溢出
  简体   繁体   English

如何在AVX2中实现车道交叉逻辑按位移位/旋转（左和右）

[英]How to implement lane crossing logical bit-wise shift/rotate (left and right) in AVX2

 原文  2018-03-19 16:26:29  7   2    c++/ c/ avx2

问题描述

As per this answer , I've created the following test program: 根据这个答案 ，我创建了以下测试程序： 

#include <iso646.h>
#include <immintrin.h>

#include <stdio.h>

#define SHIFT_LEFT( N ) \ 
\
    inline __m256i shift_left_##N ( __m256i A  ) { \
\
    if ( N == 0 ) return A; \
    else if ( N <  16 ) return _mm256_alignr_epi8 ( A, _mm256_permute2x128_si256 ( A, A, _MM_SHUFFLE ( 0, 0, 2, 0 ) ), ( uint8_t ) ( 16 - N ) ); \
    else if ( N == 16 ) return _mm256_permute2x128_si256 ( A, A, _MM_SHUFFLE ( 0, 0, 2, 0 ) ); \
    else return _mm256_slli_si256 ( _mm256_permute2x128_si256 ( A, A, _MM_SHUFFLE ( 0, 0, 2, 0 ) ), ( uint8_t ) ( N - 16 ) ); \
}

void print ( const size_t n ) {

    size_t i = 0x8000000000000000;

    while ( i ) {

        putchar ( ( int ) ( n & i ) + ( int ) ( 48 ) );
        i >>= 1;
        putchar ( ( int ) ( n & i ) + ( int ) ( 48 ) );
        i >>= 1;

        putchar ( ' ' );
    }
}

SHIFT_LEFT ( 2 );

int main ( ) {

    __m256i a = _mm256_set_epi64x ( 0x00, 0x00, 0x00, 0x03 );
    __m256i b = shift_left_2 ( a );

    size_t * c = ( size_t * ) &b;

    print ( c [ 3 ] ); print ( c [ 2 ] ); print ( c [ 1 ] ); print ( c [ 0 ] ); putchar ( '\n' );

    return 0;
}

The above program does not give the expected (by me) output, as far as I can see. 据我所知，上述程序没有给出预期的（由我）输出。 I'm stumped as to how these functions work together (read the descriptions). 我对这些功能如何协同工作感到难过（阅读说明）。 Am I doing something wrong, or is the implementation of shift_left() wrong? 我做错了什么，还是执行了shift_left（）错了？

EDIT1: I came to realize (and confirmed in the comments) that this code only intends to shift by max 32 (and are bytes), so it does not satisfy my goal. 编辑1：我开始意识到（并在评论中确认）此代码只打算移动最多32（并且是字节），因此它不能满足我的目标。 Which leaves the question, "How to implement lane crossing logical bit-wise shift (left and right) in AVX2". 这留下了一个问题，“如何在AVX2中实现车道交叉逻辑逐位移位（左和右）”。

EDIT2: Fast forward: In the meanwhile, I'm less stumped as to how it works and have coded what I needed. EDIT2：快进：与此同时，我对它的运作方式不太感兴趣，并编码了我需要的东西。 I've posted the code (shift and rotate) and accepted that as the answer. 我发布了代码（移位和旋转）并接受了答案。

2 个解决方案

解决方案1
 8  2018-03-20 23:21:17

Probably not the kind of answer that you're expecting. 可能不是你期望的那种答案。 But here's a reasonably efficient solution that actually works for a run-time shift amount. 但这是一个合理有效的解决方案，实际上适用于运行时移位量。

The costs are: 费用是：

  • Preprocess: ~12 - 14 instructions 预处理： ~12 - 14条指令
  • Rotation: 5 instructions 轮换： 5条说明
  • Shift: 6 instructions 班次： 6条指令

In order to shift or rotate anything, you must first preprocess the shift amount. 要移动或旋转任何东西，必须先预处理移位量。 Once you have that, you can efficiently perform shifts/rotations. 完成后，您可以有效地执行轮班/轮换。

Because the preprocessing step is so expensive, this solution utilizes an object to hold the preprocessed shift amount so that it can be reused many times when shifting by the same amount. 因为预处理步骤非常昂贵，所以该解决方案利用物体来保持预处理的移位量，使得当移位相同的量时可以多次重复使用。

For efficiency, the object should be on the stack in the same scope as the code that does the shifting. 为了提高效率，对象应该在与移位代码相同的范围内。 This allows the compiler to promote all the fields of the object into registers. 这允许编译器将对象的所有字段提升为寄存器。 Furthermore, it's recommended to force-inline all the methods of the class. 此外，建议强制内联类的所有方法。 

#include <stdint.h>
#include <immintrin.h>

class LeftShifter_AVX2{
public:
    LeftShifter_AVX2(uint32_t bits){
        //  Precompute all the necessary values.
        permL = _mm256_sub_epi32(
            _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7),
            _mm256_set1_epi32(bits / 32)
        );
        permR = _mm256_sub_epi32(permL, _mm256_set1_epi32(1));

        bits %= 32;
        shiftL = _mm_cvtsi32_si128(bits);
        shiftR = _mm_cvtsi32_si128(32 - bits);
        __m256i maskL = _mm256_cmpgt_epi32(_mm256_setzero_si256(), permL);
        __m256i maskR = _mm256_cmpgt_epi32(_mm256_setzero_si256(), permR);
        mask = _mm256_or_si256(maskL, _mm256_srl_epi32(maskR, shiftR));
    }

    __m256i rotate(__m256i x) const{
        __m256i L = _mm256_permutevar8x32_epi32(x, permL);
        __m256i R = _mm256_permutevar8x32_epi32(x, permR);
        L = _mm256_sll_epi32(L, shiftL);
        R = _mm256_srl_epi32(R, shiftR);
        return _mm256_or_si256(L, R);
    }
    __m256i shift(__m256i x) const{
        return _mm256_andnot_si256(mask, rotate(x));
    }

private:
    __m256i permL;
    __m256i permR;
    __m128i shiftL;
    __m128i shiftR;
    __m256i mask;
};

Test Program: 测试程序： 

#include <iostream>
using namespace std;

void print_u8(__m256i x){
    union{
        __m256i v;
        uint8_t s[32];
    };
    v = x;
    for (int c = 0; c < 32; c++){
        cout << (int)s[c] << " ";
    }
    cout << endl;
}

int main(){
    union{
        __m256i x;
        char buffer[32];
    };
    for (int c = 0; c < 32; c++){
        buffer[c] = (char)c;
    }
    print_u8(x);
    print_u8(LeftShifter_AVX2(0).shift(x));
    print_u8(LeftShifter_AVX2(8).shift(x));
    print_u8(LeftShifter_AVX2(32).shift(x));
    print_u8(LeftShifter_AVX2(40).shift(x));
}

Output: 输出： 

0 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 
0 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 
0 0 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 
0 0 0 0 0 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 
0 0 0 0 0 0 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

Right-shift is very similar. 右移非常相似。 I'll leave that as an exercise for the reader. 我将把它作为读者的练习。

解决方案2
 2 已采纳  2018-03-22 16:08:17

The following code implements lane-crossing logical bit-wise shift/rotate (left and right) in AVX2: 以下代码在AVX2中实现了车道交叉逻辑按位移位/旋转（左和右）： 

// Prototypes...

__m256i _mm256_sli_si256 ( __m256i, int );
__m256i _mm256_sri_si256 ( __m256i, int );
__m256i _mm256_rli_si256 ( __m256i, int );
__m256i _mm256_rri_si256 ( __m256i, int );


// Implementations...

__m256i left_shift_000_063 ( __m256i a, int n ) { // 6

    return _mm256_or_si256 ( _mm256_slli_epi64 ( a, n ), _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), _mm256_permute4x64_epi64 ( _mm256_srli_epi64 ( a, 64 - n ), _MM_SHUFFLE ( 2, 1, 0, 0 ) ), _MM_SHUFFLE ( 3, 3, 3, 0 ) ) );
}

__m256i left_shift_064_127 ( __m256i a, int n ) { // 7

    __m256i b = _mm256_slli_epi64 ( a, n );
    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 2, 1, 0, 0 ) );

    __m256i c = _mm256_srli_epi64 ( a, 64 - n );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 1, 0, 0, 0 ) );

    __m256i f = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), d, _MM_SHUFFLE ( 3, 3, 3, 0 ) );
    __m256i g = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), e, _MM_SHUFFLE ( 3, 3, 0, 0 ) ); // 6

    return _mm256_or_si256 ( f, g );
}

__m256i left_shift_128_191 ( __m256i a, int n ) { // 7

    __m256i b = _mm256_slli_epi64 ( a, n );
    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 1, 0, 0, 0 ) );

    __m256i c = _mm256_srli_epi64 ( a, 64 - n );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 1, 0, 0, 0 ) );

    __m256i f = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), d, _MM_SHUFFLE ( 3, 3, 0, 0 ) );
    __m256i g = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), e, _MM_SHUFFLE ( 3, 0, 0, 0 ) );

    return _mm256_or_si256 ( f, g );
}

__m256i left_shift_192_255 ( __m256i a, int n ) { // 5

    return _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), _mm256_slli_epi64 ( _mm256_permute4x64_epi64 ( a, _MM_SHUFFLE ( 0, 0, 0, 0 ) ), n ), _MM_SHUFFLE ( 3, 0, 0, 0 ) );
}

__m256i _mm256_sli_si256 ( __m256i a, int n ) {

    if ( n < 128 ) return n <  64 ? left_shift_000_063 ( a, n ) : left_shift_064_127 ( a, n % 64 );
    else           return n < 192 ? left_shift_128_191 ( a, n % 64 ) : left_shift_192_255 ( a, n % 64 );
}


__m256i right_shift_000_063 ( __m256i a, int n ) { // 6

    return _mm256_or_si256 ( _mm256_srli_epi64 ( a, n ), _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), _mm256_permute4x64_epi64 ( _mm256_slli_epi64 ( a, 64 - n ), _MM_SHUFFLE ( 0, 3, 2, 1 ) ), _MM_SHUFFLE ( 0, 3, 3, 3 ) ) );
}

__m256i right_shift_064_127 ( __m256i a, int n ) { // 7

    __m256i b = _mm256_srli_epi64 ( a, n );
    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 3, 3, 2, 1 ) );

    __m256i c = _mm256_slli_epi64 ( a, 64 - n );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 3, 3, 3, 2 ) );

    __m256i f = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), d, _MM_SHUFFLE ( 0, 3, 3, 3 ) );
    __m256i g = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), e, _MM_SHUFFLE ( 0, 0, 3, 3 ) );

    return _mm256_or_si256 ( f, g );
}

__m256i right_shift_128_191 ( __m256i a, int n ) { // 7

    __m256i b = _mm256_srli_epi64 ( a, n );
    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 3, 2, 3, 2 ) );

    __m256i c = _mm256_slli_epi64 ( a, 64 - n );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 3, 2, 1, 3 ) );

    __m256i f = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), d, _MM_SHUFFLE ( 0, 0, 3, 3 ) );
    __m256i g = _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), e, _MM_SHUFFLE ( 0, 0, 0, 3 ) );

    return _mm256_or_si256 ( f, g );
}

__m256i right_shift_192_255 ( __m256i a, int n ) { // 5

    return _mm256_blend_epi32 ( _mm256_setzero_si256 ( ), _mm256_srli_epi64 ( _mm256_permute4x64_epi64 ( a, _MM_SHUFFLE ( 0, 0, 0, 3 ) ), n ), _MM_SHUFFLE ( 0, 0, 0, 3 ) );
}

__m256i _mm256_sri_si256 ( __m256i a, int n ) {

    if ( n < 128 ) return n <  64 ? right_shift_000_063 ( a, n ) : right_shift_064_127 ( a, n % 64 );
    else           return n < 192 ? right_shift_128_191 ( a, n % 64 ) : right_shift_192_255 ( a, n % 64 );
}


__m256i left_rotate_000_063 ( __m256i a, int n ) { // 5

    return _mm256_or_si256 ( _mm256_slli_epi64 ( a, n ), _mm256_permute4x64_epi64 ( _mm256_srli_epi64 ( a, 64 - n ), _MM_SHUFFLE ( 2, 1, 0, 3 ) ) );
}

__m256i left_rotate_064_127 ( __m256i a, int n ) { // 6

    __m256i b = _mm256_slli_epi64 ( a, n );
    __m256i c = _mm256_srli_epi64 ( a, 64 - n );

    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 2, 1, 0, 3 ) );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 1, 0, 3, 2 ) );

    return _mm256_or_si256 ( d, e );
}

__m256i left_rotate_128_191 ( __m256i a, int n ) { // 6

    __m256i b = _mm256_slli_epi64 ( a, n );
    __m256i c = _mm256_srli_epi64 ( a, 64 - n );

    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 1, 0, 3, 2 ) );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 0, 3, 2, 1 ) );

    return _mm256_or_si256 ( d, e );
}

__m256i left_rotate_192_255 ( __m256i a, int n ) { // 5

    return _mm256_or_si256 ( _mm256_srli_epi64 ( a, 64 - n ), _mm256_permute4x64_epi64 ( _mm256_slli_epi64 ( a, n ), _MM_SHUFFLE ( 0, 3, 2, 1 ) ) );
}

__m256i _mm256_rli_si256 ( __m256i a, int n ) {

    if ( n < 128 ) return n <  64 ? left_rotate_000_063 ( a, n ) : left_rotate_064_127 ( a, n % 64 );
    else           return n < 192 ? left_rotate_128_191 ( a, n % 64 ) : left_rotate_192_255 ( a, n % 64 );
}


__m256i right_rotate_000_063 ( __m256i a, int n ) { // 5

    return _mm256_or_si256 ( _mm256_srli_epi64 ( a, n ), _mm256_permute4x64_epi64 ( _mm256_slli_epi64 ( a, 64 - n ), _MM_SHUFFLE ( 0, 3, 2, 1 ) ) );
}

__m256i right_rotate_064_127 ( __m256i a, int n ) { // 6

    __m256i b = _mm256_srli_epi64 ( a, n );
    __m256i c = _mm256_slli_epi64 ( a, 64 - n );

    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 0, 3, 2, 1 ) );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 1, 0, 3, 2 ) );

    return _mm256_or_si256 ( d, e );
}

__m256i right_rotate_128_191 ( __m256i a, int n ) { // 6

    __m256i b = _mm256_srli_epi64 ( a, n );
    __m256i c = _mm256_slli_epi64 ( a, 64 - n );

    __m256i d = _mm256_permute4x64_epi64 ( b, _MM_SHUFFLE ( 1, 0, 3, 2 ) );
    __m256i e = _mm256_permute4x64_epi64 ( c, _MM_SHUFFLE ( 2, 1, 0, 3 ) );

    return _mm256_or_si256 ( d, e );
}
__m256i right_rotate_192_255 ( __m256i a, int n ) { // 5

    return _mm256_or_si256 ( _mm256_slli_epi64 ( a, 64 - n ), _mm256_permute4x64_epi64 ( _mm256_srli_epi64 ( a, n ), _MM_SHUFFLE ( 2, 1, 0, 3 ) ) );
}

__m256i _mm256_rri_si256 ( __m256i a, int n ) {

    if ( n < 128 ) return n <  64 ? right_rotate_000_063 ( a, n      ) : right_rotate_064_127 ( a, n % 64 );
    else           return n < 192 ? right_rotate_128_191 ( a, n % 64 ) : right_rotate_192_255 ( a, n % 64 );
}

I have tried to make the _mm256_permute4x64_epi64 ops (when there in any case have to be two) to partially overlap, which should keep the overall latency to a minimum. 我试图使_mm256_permute4x64_epi64操作（在任何情况下必须是两个）部分重叠，这应该将整体延迟保持在最小。

Most of the suggestions and or clues given by commenters were helpful in putting together the code, thanks to those. 评论者提供的大多数建议和线索都有助于汇总代码，感谢这些。 Obviously, improvements and or any other comments are welcome. 显然，欢迎改进和/或任何其他意见。

I think that Mystical's answer is interesting, but too complicated to be used effectively for generalized shifting/rotating for use fe in a library. 我认为Mystical的答案很有趣，但是太复杂而无法有效地用于广义的移位/旋转以便在库中使用fe。

暂无
暂无

声明:本站的技术帖子网页，遵循CC BY-SA 4.0协议，如果您需要转载，请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.

相关问题 使用 AVX2 指令左移 128 位数字 - left shift of 128 bit number using AVX2 instruction 矩阵迭代的按位移位？ - Bit-wise shift for Matrix iteration? OpenGL着色器GLSL中的颜色按位移位 - Colour bit-wise shift in OpenGL shader GLSL AVX替代AVX2的矢量移位？ - AVX alternative of AVX2's vector shift? 如何在逐位操作中选择正确的左移？ - How to choose the correct left shift in bit wise operations? AVX2：AVX 寄存器中 8 位元素的 CountTrailingZeros - AVX2: CountTrailingZeros on 8 bit elements in AVX register 条件语句如何与按位运算符一起使用？ - How does condition statement work with bit-wise operators? 从移位运算符，逐位运算符和sizeof运算符产生的结果的值类别是什么？ - What is the value category of result yielded from shift operators, bit-wise operators, and sizeof operator? 有没有办法用AVX2写入_mm256_shldi_epi8（a，b，1）？ （在向量之间每8位元素移动一位） - Is there a way to write _mm256_shldi_epi8(a,b,1) with AVX2? (Shift one bit per 8-bit element between vectors) 如何在AVX2中将32位无符号整数转换为16位无符号整数？ - How to convert from 32-bit to 16-bit unsigned integers in AVX2?
相关标签
 
粤ICP备18138465号  © 2020-2024 STACKOOM.COM