Count uint8

unrolled4.cpp

#include <iostream>
#include <unistd.h>
#include <sys/mman.h>
#include <cstdint>
#include <immintrin.h>

using namespace std;

int main() {
    off_t fsize = lseek(0, 0, SEEK_END);
    char* buffer = (char*)mmap(0, fsize, PROT_READ, MAP_PRIVATE | MAP_POPULATE, 0, 0);

    __m256i v1 = _mm256_set1_epi8(127); // set a 256-bit value with all bytes being 127

    uint32_t count = 0;

    int i;
    for (i = 0; i + 128 <= fsize; i += 128) {
        __m256i v2_1 = _mm256_load_si256((__m256i*)&buffer[i]); // load 32 bytes from buffer
        __m256i v2_2 = _mm256_load_si256((__m256i*)&buffer[i+32]); // load the next 32 bytes from buffer
        __m256i v2_3 = _mm256_load_si256((__m256i*)&buffer[i+64]); // load the next 32 bytes from buffer
        __m256i v2_4 = _mm256_load_si256((__m256i*)&buffer[i+96]); // load the next 32 bytes from buffer

        __m256i v3_1 = _mm256_cmpeq_epi8(v1, v2_1);
        __m256i v3_2 = _mm256_cmpeq_epi8(v1, v2_2);
        __m256i v3_3 = _mm256_cmpeq_epi8(v1, v2_3);
        __m256i v3_4 = _mm256_cmpeq_epi8(v1, v2_4);

        uint32_t mask1 = _mm256_movemask_epi8(v3_1);
        uint32_t mask2 = _mm256_movemask_epi8(v3_2);
        uint32_t mask3 = _mm256_movemask_epi8(v3_3);
        uint32_t mask4 = _mm256_movemask_epi8(v3_4);

        count += _mm_popcnt_u32(mask1) + _mm_popcnt_u32(mask2) + _mm_popcnt_u32(mask3) + _mm_popcnt_u32(mask4);
    }

    // Process the remaining bytes
    for (; i < fsize; i++) {
        if (buffer[i] == 127) count++;
    }

    cout << count << endl;
    return 0;
}

unrolled8.cpp

#include <iostream>
#include <unistd.h>
#include <sys/mman.h>
#include <cstdint>
#include <immintrin.h>

using namespace std;

int main() {
    off_t fsize = lseek(0, 0, SEEK_END);
    char* buffer = (char*)mmap(0, fsize, PROT_READ, MAP_PRIVATE | MAP_POPULATE, 0, 0);

    __m256i v1 = _mm256_set1_epi8(127); // set a 256-bit value with all bytes being 127

    uint32_t count = 0;

    int i;
    for (i = 0; i + 256 <= fsize; i += 256) {
        __m256i v2_1 = _mm256_load_si256((__m256i*)&buffer[i]); // load 32 bytes from buffer
        __m256i v2_2 = _mm256_load_si256((__m256i*)&buffer[i+32]); // load the next 32 bytes from buffer
        __m256i v2_3 = _mm256_load_si256((__m256i*)&buffer[i+64]); // load the next 32 bytes from buffer
        __m256i v2_4 = _mm256_load_si256((__m256i*)&buffer[i+96]); // load the next 32 bytes from buffer
        __m256i v2_5 = _mm256_load_si256((__m256i*)&buffer[i+128]); // load the next 32 bytes from buffer
        __m256i v2_6 = _mm256_load_si256((__m256i*)&buffer[i+160]); // load the next 32 bytes from buffer
        __m256i v2_7 = _mm256_load_si256((__m256i*)&buffer[i+192]); // load the next 32 bytes from buffer
        __m256i v2_8 = _mm256_load_si256((__m256i*)&buffer[i+224]); // load the next 32 bytes from buffer

        __m256i v3_1 = _mm256_cmpeq_epi8(v1, v2_1);
        __m256i v3_2 = _mm256_cmpeq_epi8(v1, v2_2);
        __m256i v3_3 = _mm256_cmpeq_epi8(v1, v2_3);
        __m256i v3_4 = _mm256_cmpeq_epi8(v1, v2_4);

        uint32_t mask1 = _mm256_movemask_epi8(v3_1);
        uint32_t mask2 = _mm256_movemask_epi8(v3_2);
        uint32_t mask3 = _mm256_movemask_epi8(v3_3);
        uint32_t mask4 = _mm256_movemask_epi8(v3_4);

        count += _mm_popcnt_u32(mask1) + _mm_popcnt_u32(mask2) + _mm_popcnt_u32(mask3) + _mm_popcnt_u32(mask4);

        __m256i v3_5 = _mm256_cmpeq_epi8(v1, v2_5);
        __m256i v3_6 = _mm256_cmpeq_epi8(v1, v2_6);
        __m256i v3_7 = _mm256_cmpeq_epi8(v1, v2_7);
        __m256i v3_8 = _mm256_cmpeq_epi8(v1, v2_8);

        uint32_t mask5 = _mm256_movemask_epi8(v3_5);
        uint32_t mask6 = _mm256_movemask_epi8(v3_6);
        uint32_t mask7 = _mm256_movemask_epi8(v3_7);
        uint32_t mask8 = _mm256_movemask_epi8(v3_8);

        count += _mm_popcnt_u32(mask5) + _mm_popcnt_u32(mask6) + _mm_popcnt_u32(mask7) + _mm_popcnt_u32(mask8);
    }

    // Process the remaining bytes
    for (; i < fsize; i++) {
        if (buffer[i] == 127) count++;
    }

    cout << count << endl;
    return 0;
}