Haswell上的AVX2比SSE慢

Question

我有以下代码(普通,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();
}

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

这表明完全相同的代码在AVX2上比SSE慢22%.我做错了什么还是这种正常行为？

Answer 1

我将您的代码转换为更多的vanilla C++(普通数组,无向量等),清理它并在禁用自动向量化的情况下对其进行测试并获得合理的结果:

#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
===========================

我建议你尝试使用上面的代码-fno-vectorize(或者-fno-tree-vectorize如果使用g ++),看看你是否得到了类似的结果.如果您这样做,那么您可以向后查找原始代码,以查看可能出现的不一致之处.

Answer 2

在我的机器(核心i7-4900M)上,基于Paul R的更新代码,g ++ 4.8.2,100,000次迭代而不是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用下面的行来解释,这需要一个完整的256位存储(哎哟!),然后是一个超过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,然后我:

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

AVX时间减少了5倍!