10

我有以下代码（正常、SSE 和 AVX）：

int testSSE(const aligned_vector & ghs, const aligned_vector & lhs) {
    int result[4] __attribute__((aligned(16))) = {0};
    __m128i vresult = _mm_set1_epi32(0);
    __m128i v1, v2, vmax;

    for (int k = 0; k < ghs.size(); k += 4) {
        v1 = _mm_load_si128((__m128i *) & lhs[k]);
        v2 = _mm_load_si128((__m128i *) & ghs[k]);
        vmax = _mm_add_epi32(v1, v2);
        vresult = _mm_max_epi32(vresult, vmax);
    }
    _mm_store_si128((__m128i *) result, vresult);
    int mymax = result[0];
    for (int k = 1; k < 4; k++) {
        if (result[k] > mymax) {
            mymax = result[k];
        }
    }
    return mymax;
}

 int testAVX(const aligned_vector & ghs, const aligned_vector & lhs) {
    int result[8] __attribute__((aligned(32))) = {0};
    __m256i vresult = _mm256_set1_epi32(0);
    __m256i v1, v2, vmax;

    for (int k = 0; k < ghs.size(); k += 8) {
        v1 = _mm256_load_si256((__m256i *) & ghs[ k]);
        v2 = _mm256_load_si256((__m256i *) & lhs[k]);
        vmax = _mm256_add_epi32(v1, v2);
        vresult = _mm256_max_epi32(vresult, vmax);
    }
    _mm256_store_si256((__m256i *) result, vresult);
    int mymax = result[0];
    for (int k = 1; k < 8; k++) {
        if (result[k] > mymax) {
            mymax = result[k];
        }
    }
    return mymax;
}

int testNormal(const aligned_vector & ghs, const aligned_vector & lhs) {
    int max = 0;
    int tempMax;
    for (int k = 0; k < ghs.size(); k++) {
        tempMax = lhs[k] + ghs[k];
        if (max < tempMax) {
            max = tempMax;
        }
    }
    return max;
}

所有这些功能都使用以下代码进行测试：

void alignTestSSE() {
    aligned_vector lhs;
    aligned_vector ghs;

    int mySize = 4096;
    int FinalResult;
    int nofTestCases = 1000;
    double time, time1, time2, time3;
    vector<int> lhs2;
    vector<int> ghs2;

    lhs.resize(mySize);
    ghs.resize(mySize);
    lhs2.resize(mySize);
    ghs2.resize(mySize);

    srand(1);
    for (int k = 0; k < mySize; k++) {
        lhs[k] = randomNodeID(1000000);
        lhs2[k] = lhs[k];
        ghs[k] = randomNodeID(1000000);
        ghs2[k] = ghs[k];
    }
    /* Warming UP */
    for (int k = 0; k < nofTestCases; k++) {
        FinalResult = testNormal(lhs, ghs);
    }

    for (int k = 0; k < nofTestCases; k++) {
        FinalResult = testSSE(lhs, ghs);
    }

    for (int k = 0; k < nofTestCases; k++) {
        FinalResult = testAVX(lhs, ghs);
    }

    cout << "===========================" << endl;
    time = timestamp();
    for (int k = 0; k < nofTestCases; k++) {
        FinalResult = testSSE(lhs, ghs);
    }
    time = timestamp() - time;
    time1 = time;
    cout << "SSE took " << time << " s" << endl;
    cout << "SSE Result: " << FinalResult << endl;

    time = timestamp();
    for (int k = 0; k < nofTestCases; k++) {
        FinalResult = testAVX(lhs, ghs);
    }
    time = timestamp() - time;
    time3 = time;
    cout << "AVX took " << time << " s" << endl;
    cout << "AVX Result: " << FinalResult << endl;



    time = timestamp();
    for (int k = 0; k < nofTestCases; k++) {
        FinalResult = testNormal(lhs, ghs);
    }
    time = timestamp() - time;
    cout << "Normal took " << time << " s" << endl;
    cout << "Normal Result: " << FinalResult << endl;
    cout << "SpeedUP SSE= " << time / time1 << " s" << endl;
    cout << "SpeedUP AVX= " << time / time3 << " s" << endl;
    cout << "===========================" << endl;
    ghs.clear();
    lhs.clear();
}

在哪里

inline double timestamp() {
    struct timeval tp;
    gettimeofday(&tp, NULL);
    return double(tp.tv_sec) + tp.tv_usec / 1000000.;
}

和

typedef vector<int, aligned_allocator<int, sizeof (int)> > aligned_vector;

是使用https://gist.github.com/donny-dont/1471329的 AlignedAllocator 的对齐向量

我有一个 intel-i7 haswell 4771，以及最新的 Ubuntu 14.04 64bit 和 gcc 4.8.2。一切都是最新的。我用 -march=native -mtune=native -O3 -m64 编译。

结果是：

SSE took 0.000375986 s
SSE Result: 1982689
AVX took 0.000459909 s
AVX Result: 1982689
Normal took 0.00315714 s
Normal Result: 1982689
SpeedUP SSE= 8.39696 s
SpeedUP AVX= 6.8647 s

这表明在 AVX2 上完全相同的代码比 SSE 慢 22%。我做错了什么还是这是正常行为？

4

3 回答 3

15

我将您的代码转换为更普通的 C++（纯数组、无向量等），对其进行清理并在禁用自动向量化的情况下对其进行测试，并获得了合理的结果：

#include <iostream>
using namespace std;

#include <sys/time.h>
#include <cstdlib>
#include <cstdint>

#include <immintrin.h>

inline double timestamp() {
    struct timeval tp;
    gettimeofday(&tp, NULL);
    return double(tp.tv_sec) + tp.tv_usec / 1000000.;
}

int testSSE(const int32_t * ghs, const int32_t * lhs, size_t n) {
    int result[4] __attribute__((aligned(16))) = {0};
    __m128i vresult = _mm_set1_epi32(0);
    __m128i v1, v2, vmax;

    for (int k = 0; k < n; k += 4) {
        v1 = _mm_load_si128((__m128i *) & lhs[k]);
        v2 = _mm_load_si128((__m128i *) & ghs[k]);
        vmax = _mm_add_epi32(v1, v2);
        vresult = _mm_max_epi32(vresult, vmax);
    }
    _mm_store_si128((__m128i *) result, vresult);
    int mymax = result[0];
    for (int k = 1; k < 4; k++) {
        if (result[k] > mymax) {
            mymax = result[k];
        }
    }
    return mymax;
}

int testAVX(const int32_t * ghs, const int32_t * lhs, size_t n) {
    int result[8] __attribute__((aligned(32))) = {0};
    __m256i vresult = _mm256_set1_epi32(0);
    __m256i v1, v2, vmax;

    for (int k = 0; k < n; k += 8) {
        v1 = _mm256_load_si256((__m256i *) & ghs[k]);
        v2 = _mm256_load_si256((__m256i *) & lhs[k]);
        vmax = _mm256_add_epi32(v1, v2);
        vresult = _mm256_max_epi32(vresult, vmax);
    }
    _mm256_store_si256((__m256i *) result, vresult);
    int mymax = result[0];
    for (int k = 1; k < 8; k++) {
        if (result[k] > mymax) {
            mymax = result[k];
        }
    }
    return mymax;
}

int testNormal(const int32_t * ghs, const int32_t * lhs, size_t n) {
    int max = 0;
    int tempMax;
    for (int k = 0; k < n; k++) {
        tempMax = lhs[k] + ghs[k];
        if (max < tempMax) {
            max = tempMax;
        }
    }
    return max;
}

void alignTestSSE() {

    int n = 4096;
    int normalResult, sseResult, avxResult;
    int nofTestCases = 1000;
    double time, normalTime, sseTime, avxTime;

    int lhs[n] __attribute__ ((aligned(32)));
    int ghs[n] __attribute__ ((aligned(32)));

    for (int k = 0; k < n; k++) {
        lhs[k] = arc4random();
        ghs[k] = arc4random();
    }

    /* Warming UP */
    for (int k = 0; k < nofTestCases; k++) {
        normalResult = testNormal(lhs, ghs, n);
    }

    for (int k = 0; k < nofTestCases; k++) {
        sseResult = testSSE(lhs, ghs, n);
    }

    for (int k = 0; k < nofTestCases; k++) {
        avxResult = testAVX(lhs, ghs, n);
    }

    time = timestamp();
    for (int k = 0; k < nofTestCases; k++) {
        normalResult = testNormal(lhs, ghs, n);
    }
    normalTime = timestamp() - time;

    time = timestamp();
    for (int k = 0; k < nofTestCases; k++) {
        sseResult = testSSE(lhs, ghs, n);
    }
    sseTime = timestamp() - time;

    time = timestamp();
    for (int k = 0; k < nofTestCases; k++) {
        avxResult = testAVX(lhs, ghs, n);
    }
    avxTime = timestamp() - time;

    cout << "===========================" << endl;
    cout << "Normal took " << normalTime << " s" << endl;
    cout << "Normal Result: " << normalResult << endl;
    cout << "SSE took " << sseTime << " s" << endl;
    cout << "SSE Result: " << sseResult << endl;
    cout << "AVX took " << avxTime << " s" << endl;
    cout << "AVX Result: " << avxResult << endl;
    cout << "SpeedUP SSE= " << normalTime / sseTime << endl;
    cout << "SpeedUP AVX= " << normalTime / avxTime << endl;
    cout << "===========================" << endl;

}

int main()
{
    alignTestSSE();
    return 0;
}

测试：

$ clang++ -Wall -mavx2 -O3 -fno-vectorize SO_avx.cpp && ./a.out
===========================
Normal took 0.00324106 s
Normal Result: 2143749391
SSE took 0.000527859 s
SSE Result: 2143749391
AVX took 0.000221968 s
AVX Result: 2143749391
SpeedUP SSE= 6.14002
SpeedUP AVX= 14.6015
===========================

我建议你试试上面的代码，使用-fno-vectorize（或者-fno-tree-vectorize如果使用 g++），看看你是否得到类似的结果。如果你这样做了，那么你可以向后工作到你的原始代码，看看不一致可能来自哪里。

于 2014-05-06T15:51:44.053 回答
6

在我的机器（核心 i7-4900M）上，基于来自Paul R的更新代码，使用 g++ 4.8.2 迭代 100,000 次而不是 1000 次，我得到以下结果：

g++ -Wall -mavx2 -O3 -std=c++11 test_avx.cpp && ./a.exe 
SSE took             508,029 us
AVX took           1,308,075 us
Normal took          297,017 us


g++ -Wall -mavx2 -O3 -std=c++11 -fno-tree-vectorize test_avx.cpp && ./a.exe 
SSE took             509,029 us
AVX took           1,307,075 us
Normal took        3,436,197 us

GCC 在优化“正常”代码方面做得非常出色。然而，“AVX”代码的缓慢性能可以通过以下几行来解释，这需要完整的 256 位存储（哎呀！），然后是 8 个整数的最大搜索。

_mm256_store_si256((__m256i *) result, vresult);
int mymax = result[0];
for (int k = 1; k < 8; k++) {
  if (result[k] > mymax) {
     mymax = result[k];
  }
}
return mymax;

最好继续使用 AVX 内在函数，最多 8 个。我可以提出以下更改

v1      = _mm256_permute2x128_si256(vresult,vresult,1);  // from ABCD-EFGH to ????-ABCD
vresult = _mm256_max_epi32(vresult, v1);
v1      = _mm256_permute4x64_epi64(vresult,1);  // from ????-ABCD to ????-??AB
vresult = _mm256_max_epi32(vresult, v1);
v1      = _mm256_shuffle_epi32(vresult,1); // from ????-???AB to ????-???A
vresult = _mm256_max_epi32(vresult, v1);

// no _mm256_extract_epi32 => need extra step
__m128i vres128 = _mm256_extracti128_si256(vresult,0);
return _mm_extract_epi32(vres128,0);

为了公平比较，我还更新了 SSE 代码，然后：

SSE took             483,028 us
AVX took             258,015 us
Normal took          307,017 us

AVX 时间减少了 5 倍！

于 2014-05-14T13:30:43.817 回答
0

手动进行循环展开可以加快 SSE/AVX 代码的速度。

我的 i5-5300U 上的原始版本：

Normal took 0.347 s
Normal Result: 2146591543
AVX took 0.409 s
AVX Result: 2146591543
SpeedUP AVX= 0.848411

手动循环展开后：

Normal took 0.375 s
Normal Result: 2146591543
AVX took 0.297 s
AVX Result: 2146591543
SpeedUP AVX= 1.26263
于 2017-05-03T11:12:59.080 回答