hwCrc32c

#ifdef __ARM_FEATURE_CRC32
#include <arm_acle.h>
#define _mm_crc32_u8 __crc32cb
#define _mm_crc32_u16 __crc32ch
#define _mm_crc32_u32 __crc32cw
#define _mm_crc32_u64 __crc32cd
#else
#include <nmmintrin.h>
#endif
#include <stdio.h>
#include <stdint.h>
#include <stddef.h>

/**
 * Calculates CRC-32C using hardware support
 *
 * @param initialCrc The initial CRC to use for the operation
 * @param buf The buffer that stores the data whose CRC is to be calculated
 * @param len The size of the buffer
 * @return The CRC-32C of the data in the buffer
 */
uint32_t hwCrc32c(uint32_t initialCrc, const char *buf, size_t len) {
    uint32_t crc = initialCrc;

    // XOR the initial CRC with INT_MAX
    crc ^= 0xFFFFFFFF;

    if (len >= 8) {
        // Align the input to the word boundary
        const auto offset = (intptr_t)buf & 0b111;
        switch (offset) {
            case 7:
                crc = _mm_crc32_u8(crc, *(uint8_t *)buf);
                break;
            case 6:
                crc = _mm_crc32_u16(crc, *(uint16_t *)buf);
                break;
            case 5:
                crc = _mm_crc32_u8(crc, *(uint8_t *)buf);
                crc = _mm_crc32_u16(crc, *(uint16_t *)(buf + 1));
                break;
            case 4:
                crc = _mm_crc32_u32(crc, *(uint32_t *)buf);
                break;
            case 3:
                crc = _mm_crc32_u8(crc, *(uint8_t *)buf);
                crc = _mm_crc32_u32(crc, *(uint32_t *)(buf + 1));
                break;
            case 2:
                crc = _mm_crc32_u16(crc, *(uint16_t *)buf);
                crc = _mm_crc32_u32(crc, *(uint32_t *)(buf + 2));
                break;
            case 1:
                crc = _mm_crc32_u8(crc, *(uint8_t *)buf);
                crc = _mm_crc32_u16(crc, *(uint16_t *)(buf + 2));
                crc = _mm_crc32_u32(crc, *(uint32_t *)(buf + 4));
                break;
        }

        buf += offset;
        len -= offset;

        for (; len >= 8; len -= 8) {
            crc = (uint32_t)_mm_crc32_u64(crc, *(uint64_t *)buf);
            buf += 8;
        }
    }

    switch (len) {
        case 1:
            crc = _mm_crc32_u8(crc, *(uint8_t *)buf);
            break;
        case 2:
            crc = _mm_crc32_u16(crc, *(uint16_t *)buf);
            break;
        case 3:
            crc = _mm_crc32_u16(crc, *(uint16_t *)buf);
            crc = _mm_crc32_u8(crc, *(uint8_t *)(buf + 2));
            break;
        case 4:
            crc = _mm_crc32_u32(crc, *(uint32_t *)buf);
            break;
        case 5:
            crc = _mm_crc32_u32(crc, *(uint32_t *)buf);
            crc = _mm_crc32_u8(crc, *(uint8_t *)(buf + 4));
            break;
        case 6:
            crc = _mm_crc32_u32(crc, *(uint32_t *)buf);
            crc = _mm_crc32_u16(crc, *(uint16_t *)(buf + 4));
            break;
        case 7:
            crc = _mm_crc32_u32(crc, *(uint32_t *)buf);
            crc = _mm_crc32_u16(crc, *(uint16_t *)(buf + 4));
            crc = _mm_crc32_u8(crc, *(uint8_t *)(buf + 6));
            break;
    }

    // XOR again with INT_MAX
    crc ^= 0xFFFFFFFF;
    return crc;
}

#include <string_view>

using namespace std;

uint32_t hwCrc32c(std::string_view sv) {
    return hwCrc32c(0, sv.data(), sv.length());
}

int main() {
    std::printf("%08X\n", hwCrc32c("input"sv));
    std::printf("%08X\n", hwCrc32c("123"sv));
    std::printf("%08X\n", hwCrc32c("12345678901234567890"));
}