Attachment 181440 Details for Bug 218203 – same benchmark but with 128k checksums

same benchmark but with 128k checksums

fletcher4.cpp (text/plain), 3.55 KB, created by Adam Stylinski on 2017-04-03 19:22:34 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Adam Stylinski

Created: 2017-04-03 19:22:34 UTC

Size: 3.55 KB

patch

obsolete

>#include <stdio.h>
>#include <stdlib.h>
>#include <stdint.h>
>#include <immintrin.h>
>#include <chrono>
>#include <iostream>
>
>#define NUMBYTES (1 << 17)
>
>void fletcher_4_nativeavx(const void *buf, uint64_t size, uint64_t *chksum)
>{   
>    const uint32_t *ip = (uint32_t*)buf;
>    const uint32_t *ipend = ip + (size / sizeof (uint32_t));
>
>    /* Initially the vector registers to zero for the checksums */
>    __m256i a = _mm256_setzero_si256();
>    __m256i b = _mm256_setzero_si256();
>    __m256i c = _mm256_setzero_si256();
>    __m256i d = _mm256_setzero_si256();
>
>    __m256i data;
>
>    uint64_t aSum, bSum, cSum, dSum;
>    uint64_t *aVec, *bVec, *cVec, *dVec;
>
>    /* Transform scalar loop into vector one. Don't have to worry
>     * about peeling off remainder, as the loop termination criteria
>     * for fletcher 4 ends abruptly when the size isn't a modulus
>     * of 16 bytes */
>    for (; ip < ipend; ip+=4) {
>        data = _mm256_cvtepu32_epi64(_mm_loadu_si128((__m128i*)ip));
>        a = _mm256_add_epi64(a, data);
>        b = _mm256_add_epi64(b, a);
>        c = _mm256_add_epi64(b, c);
>        d = _mm256_add_epi64(c, d);
>    }
>
>    /* From the following intel whitepaper: 
>     * https://software.intel.com/en-us/articles/fast-computation-of-fletcher-checksums
>     * They found the recurrence relation already, evidently
>     */
>
>    /* Since there's no hadd for epi64, we'll see what the compiler emits */
>    // This might be an overflow danger, as add_epi64 does signed integer math
>    aVec = (uint64_t*)&a;
>    bVec = (uint64_t*)&b;
>    cVec = (uint64_t*)&c;
>    dVec = (uint64_t*)&d;
>
>    aSum = aVec[0] + aVec[1] + aVec[2] + aVec[3];
>    bSum = -aVec[1] - 2 * aVec[2] - 3 * aVec[3] + 
>           4 * (bVec[0] + bVec[1] + bVec[2] + bVec[3]);
>    cSum = aVec[2] +  3 * aVec[3] 
>           -  6 * bVec[0] - 10 * bVec[1] - 14 * bVec[2] - 18 * bVec[3]
>           + 16 * (cVec[0] + cVec[1] + cVec[2] + cVec[3]);
>    dSum = -aVec[3]
>           +  4*bVec[0] + 10*bVec[1] + 20*bVec[2] + 34*bVec[3]
>           - 48*cVec[0] - 64*cVec[1] - 80*cVec[2] - 96*cVec[3]
>           + 64 * (dVec[0] + dVec[1] +dVec[2] + dVec[3]);
>
>    chksum[0] = aSum;
>    chksum[1] = bSum;
>    chksum[2] = cSum;
>    chksum[3] = dSum;
>}
>
>
>void fletcher_4_native(const void *buf, uint64_t size, uint64_t *chksum)
>{   
>    const uint32_t *ip = (uint32_t*)buf;
>    const uint32_t *ipend = ip + (size / sizeof (uint32_t));
>    uint64_t a, b, c, d;
>
>    for (a = b = c = d = 0; ip < ipend; ip++) {
>      a += ip[0];
>      b += a;
>      c += b;
>      d += c;
>    }
>
>    chksum[0] = a;
>    chksum[1] = b;
>    chksum[2] = c;
>    chksum[3] = d;
>}
>
>int main(void)
>{
>    char *buf = (char*)malloc(NUMBYTES);
>    FILE *randFile = fopen("/dev/urandom", "r");
>    fread(buf, sizeof(char), NUMBYTES, randFile);
>    uint64_t chksum[4];
>    uint64_t chksum2[4];
>    auto start = std::chrono::steady_clock::now();
>    fletcher_4_nativeavx(buf, NUMBYTES, chksum2); 
>    auto end = std::chrono::steady_clock::now();
>    auto diff = end - start;
>    std::cout << 
>        std::chrono::duration<double, std::milli>(diff).count() <<
>        " ms total (avx)" << std::endl;
>
>    start = std::chrono::steady_clock::now();
>    fletcher_4_native(buf, NUMBYTES, chksum); 
>    end = std::chrono::steady_clock::now();
>    diff = end - start;
>    std::cout << 
>        std::chrono::duration<double, std::milli>(diff).count() <<
>        " ms total (scalar)" << std::endl;
>
>    printf("a = %lu, b = %lu, c = %lu, d = %lu\n", chksum[0], chksum[1], chksum[2], chksum[3]);
>    printf("a = %lu, b = %lu, c = %lu, d = %lu\n", chksum2[0], chksum2[1], chksum2[2], chksum2[3]);
>
>    return 0;
>}

Actions: View

Attachments on bug 218203: 181322 | 181326 | 181440