-m 0(cache_conflict) | 1(no_cache_conflict)
-t thread_count
-s matrix_size
-d (dumping matrixs)

Multiple threads run on different cpus(set by cpu affinity) execute matrix mul-
tiplication, when they are designed to write the adjacent memory address in ma-
trix C, there will be a waste of time for each cpu to keep the cache line cons-
istence. 

So it's better to partition the threads to write different cache line avoiding 
the consistence waste.

I run the program on an AMD 4-physical core with -t 4 (4 threads each bonding 
to a different cpu), -s 1024 (1024 square matrix), the -m 1 (no_cache_conflict)
had a 50% performance promoting than -m 0 (cache conflict).

gcc matrix.c -pthread