I agree that 17985.4 MFLOPS needs very specific coding. But the 1733.4 MFLOPS solution is a really naive one. Data locality and cache friendliness are two of t

I agree that 17985.4 MFLOPS needs very specific coding. But the 1733.4 MFLOPS solution is a really naive one. Data locality and cache friendliness are two of the first things taught in high performance classes, with an obvious trick : iterating using row-major order like in the second case. This trivial optimization gives a 9419.8 MFLOPS throughput. I don't think "everyday code" like that is so inefficient given the maximum result which can be obtained (17985.4 MFLOPS).