void matrix2d_fill(uint8_t** ptr, const uint32_t x, const uint32_t y, const uint8_t val)
{
    const uint64_t val64 = (uint64_t)(((uint64_t)(val) << 56)   |
                                   ((uint64_t)(val)  << 48) |
                                   ((uint64_t)(val)  << 40) |
                                   ((uint64_t)(val)  << 32) |
                                   ((uint64_t)(val)  << 24) |
                                   ((uint64_t)(val)  << 16) |
                                   ((uint64_t)(val)  << 8)  |
                                   ((uint64_t)(val))        );
    uint64_t size       = x * y;
    uint64_t iterator64 = size / 8;
    uint64_t balance64  = size - (iterator64 * 8);
    uint64_t iterator32 = balance64 / 8;
    uint64_t balance32  = balance64 - 4;
    uint64_t iterator16 = balance32 / 8;
    uint64_t balance16  = balance32 - 2;
    uint64_t iterator8  = balance16 / 8;
    uint64_t balance8   = (balance16 == 1);
    uint64_t* begin64   = (uint64_t*)ptr;
    uint64_t* end64     = (uint64_t*)(ptr + balance64);
    uint32_t* begin32   = (uint32_t*)end64;
    uint32_t* end32     = (uint32_t*)(begin32 + balance32);
    uint16_t* begin16   = (uint16_t*)end32;
    uint16_t* end16     = (uint16_t*)(begin16 + balance16);
    uint8_t* begin8     = (uint8_t*)end16;
    uint8_t* end8       = (uint8_t*)(begin8 + balance8);
    for(;begin64 < end64; begin64++)
        *begin64 = val64;
    for(;begin32 < end32; begin32++)
        *begin32 = (uint32_t)val64;
    for(;begin16 < end16; begin16++)
        *begin16 = (uint16_t)val64;
    for(;begin8 < end8; begin8++)
        *begin8 = (uint8_t)val64;
}