#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <immintrin.h>

FILE *fp;
unsigned int output_file_offset = 0;
unsigned int input_file_offset = 0;
char *buffer;
uint64_t *out_buffer;
unsigned int bytes_read = 0;

void copyAndFixBuffer() {
    char* rawbuff = (char*)out_buffer;
    int s,e;
    for(s=0,e=0; s<output_file_offset*9;) {
        if(s % 9) {
            buffer[s]=rawbuff[e];
            s++,
            e++;
        }
        else {
            buffer[s] = '\n';
            s++;
        }
    }
    buffer[s]='\n';
}

void flushdata() {
    copyAndFixBuffer();
    fwrite(buffer, output_file_offset*9+1,1,stdout );
}

int main ()
{
    fp = fopen("/mnt/c/programming/randomdata.bin", "r");
    buffer = (char*) malloc(1010000000);
    out_buffer = (uint64_t*)malloc(1010000000);

    while(bytes_read = fread(buffer, 1, 100000000, fp)) {
        output_file_offset = 0;
        input_file_offset = 0;
        while(1) {
            if(input_file_offset == bytes_read) break;
            char c = buffer[input_file_offset];
            input_file_offset++;
            uint64_t out = __builtin_bswap64(_pdep_u64(c,0x0101010101010101)|0x3030303030303030);
            out_buffer[output_file_offset] = out;
            output_file_offset++;
        }
        flushdata();
    }
    return 0;
}

//>inb4 muh RAM
//it's faster this way