Raspberry Pi 1 B vs. Raspberry Pi 2 B ASM speed difference

Question

I have the following code:

for (short l = j; l < j + input->w_small; l = l + 4){
  add_b = k * input->w_big + l;
  add_s = (k - i) * input->w_small + l - j;

  __asm__ __volatile__(
      "ldr %%r1, [%1];"
      "ldr %%r2, [%2];"
      "usada8 %0, %%r1, %%r2, %0;"
      :"+r" (sad)
      : "r" (input->pic_big + add_b), "r" (input->pic_small + add_s)
      : "r1", "r2"
      );
}

This is part of an image processing algorithm. The application runs 29.24 seconds on RPi 1 B and 7.65 seconds on RPi 2 B resulting in 3.82x speed-up. The question is, why? I understand, that there is an architectural change between the models, but I didn't find any reference regarding USADA8, that it should be significantly faster on ARMv7. Any ideas?

PS: Don't get me wrong, I am perfectly happy with the results, just being curious :)

Answer 1

There may be many reasons, but the main ones are probably (according to this ):

• the core frequency is not the same (900MHz for model 2B and 700MHz for model 1B)
• The L1 cache of 2B is twice the size og the L1 cache in model 1B (16kB vs 32 kB). I suspect that L2 caches and generally speaking the cache hierarchy is also different.
• You might have a different config for both (you can tweak the frequencies of various things)