README

This is for conv/gemm optimization test.
Run:
g++ test_gemm.cpp -o demo -Ofast -std=c++11 -lpthread
./demo

gemm thread test:
Only gemm_O5_MThread has improvement. Because the problemsize for one thread should be big enough to hide the time for create thread, destory thread and data loading.
What's the problemsize like? Need to do more test.
peak performance for 1024 1024 1024 0.024s