Raspberry Pi Zero linear memory block diff

raspberry_pi_zero_linear_diff.cpp

#include <stdio.h>
#include <stdlib.h>
#include <memory.h>
#include <syslog.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sys/mman.h>
#include <errno.h>
#include <stdint.h>
#include <bcm_host.h>

volatile uint64_t *systemTimerRegister = 0;
#define tick() (*systemTimerRegister)

uint8_t a[320*480*2];
uint8_t b[320*480*2];

int diff(uint8_t *a, uint8_t *b)
{
  uint8_t *endPtr;
  asm volatile(
    "mov r1, %[a]\n" // pointer to buffer 1 to diff
    "add r0, r1, #307200\n" // end pointer of buffer 1, loop finish condition
    "mov r2, %[b]\n" // pointer to buffer 2 to diff

  "start2:\n"
    "pld [r1, #128]\n" // preload data caches for both buffers 128 bytes ahead of time
    "pld [r2, #128]\n"

    "ldmia r1!, {r3,r4,r5,r6}\n" // load 4x32-bit elements of buffer 1
    "ldmia r2!, {r7,r8,r9,r10}\n" // corresponding elements from buffer 2
    "cmp r3, r7\n" // compare all elements for diff
    "bne end2\n"
    "cmp r4, r8\n"
    "bne end2\n"
    "cmp r5, r9\n"
    "bne end2\n"
    "cmp r6, r10\n"
    "bne end2\n"

    // unroll once to process 8x32bits = 32 bytes = L1 data cache line size of Raspberry Pi Zero in one iteration
    "ldmia r1!, {r3,r4,r5,r6}\n"
    "ldmia r2!, {r7,r8,r9,r10}\n"
    "cmp r3, r7\n"
    "bne end2\n"
    "cmp r4, r8\n"
    "bne end2\n"
    "cmp r5, r9\n"
    "bne end2\n"
    "cmp r6, r10\n"
    "bne end2\n"

    "cmp r0, r1\n" // test loop end condition
    "bne start2\n"

  "end2:\n"
    "mov %[result], r1\n\t"
    : [result]"=r"(endPtr)
    : [a]"r"(a), [b]"r"(b)
    : "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "r8", "r9", "r10"
  );
  return endPtr-a;
}

#define BCM2835_TIMER_BASE                   0x3000

int main()
{
  int mem_fd = open("/dev/mem", O_RDWR|O_SYNC);
  if (mem_fd < 0) return -1;
  void *bcm2835 = mmap(NULL, bcm_host_get_peripheral_size(), (PROT_READ | PROT_WRITE), MAP_SHARED, mem_fd, bcm_host_get_peripheral_address());
  if (bcm2835 == MAP_FAILED) return -1;
  systemTimerRegister = (volatile uint64_t*)((uintptr_t)bcm2835 + BCM2835_TIMER_BASE + 0x04); // Generates an unaligned 64-bit pointer, but seems to be fine.

  for(int i = 0; i < 320*480*2; ++i)
    if (tick() == 0) a[i] = 0; else a[i] = 0xEE;
  for(int i = 0; i < 320*480*2; ++i)
    if (tick() == 0) b[i] = 0; else b[i] = 0xEE;

  uint64_t t0 = tick();
  int d = diff(a, b);
  // diff is coarse, find the exact position outside hot loop
  if (d != 307200)
  {
    d -= 32;
    while(a[d] == b[d]) ++d;
  }
  uint64_t t1 = tick();
  double clocks = (double)(t1-t0)*1000.0;
  int bytes = sizeof(a);
  double clocksPerByte = clocks / bytes;
  printf("bytes diffed: %d, diff at pos: %d, usecs taken: %llu, %f clocks/byte\n", bytes, d, t1-t0, clocksPerByte);
}

// g++ -I/opt/vc/include -L/opt/vc/lib -lbcm_host -O3 -o test test.cpp
// sudo ./test

bytes diffed: 307200, diff at pos: 307200, usecs taken: 1466, 4.772135 clocks/byte