pitfalls avoided by matmul.py

[singam-sanjay]

Helli @yaroslavvb,

In https://stackoverflow.com/questions/41804380/testing-gpu-with-tensorflow-matrix-multiplication, you claimed that your matmul_benchmark.py avoided common pitfalls. Can you please list them out ?

I can see that you have,

1. Avoided counting time take to transfer data to the device by making the first call outside the timing region. But, is it guaranteed that Tensorflow will not move data back to the host unless it is accessed ?
2. Averaging the time take over many iterations.

Thanks,
Sanjay

[yaroslavvb]

1. Use pre-warming since first call is slower due to initializations (ie PTX compilation)
2. Use sess.run(a.op) instead of sess.run(a) to avoid transferring any data from TF runtime
3. Make sure op inputs are on the same device as the op.
4. Use variables instead of constants and turn off optimizations (tf.OptimizerOptions.L0)