Investigating speed decrease between Jan 2019 and Jan 2020

[fangq]

reported by @ShijieYan:

A speed drop was noted compared to the earlier version of MCX(Jan 2019). For instance:

For the mcx_gpu_contest.m on a GTX 1080Ti (Ubuntu 18.04, nvcc 9.0, Nvidia Driver 390.129):

1. The latest MCXLAB (commit: 485b1ea) report benchmark score of 70523 (-26.6%). The additional speed drop is likely due to the branching added for the support of the hybrid optical property;
2. The MCXLAB on Jan 25 2019 (prior to the support of hybrid optical property) reports 83281 (-13.4%);
3. The MCXLAB on Dec 8 2019 reports 83281 (-13.4%);
4. The MCXLAB on Sep 6 2019 reports 82946 (-13.7%);
5. The MCXLAB during our workshop (Aug 8 2019, commit: 0bdbef6) reports 89599 (-6.8%);
6. The MCXLAB v2019.4 (commit: 711bf81) reports 91757 (-4.5%);
7. The MCXLAB v2019.3 (commit: ac76ff7) reports 97156 (no decrease);
8. The version back on Jan 18 2019 (commit: 335c9a0) reports 96115 (baseline);

For the demo_photon_sharing.m on a GTX 1080Ti (Ubuntu 18.04, nvcc 9.0, Nvidia Driver 390.129):

1. The latest MCXLAB (commit: 485b1ea) reports 14522 photon/ms (-45.2%)
2. The MCXLAB on Jan 25 2019 (prior to the support of hybrid optical property) reports 14094.43 photon/ms (-46.8%);
3. The MCXLAB during our workshop (Aug 8 2019, commit: 0bdbef6) reports 18796.99 photon/ms (-29.0%);
4. The MCXLAB v2019.4 (commit: 711bf81) reports 18338.53 photon/ms (-30.7%);
5. The MCXLAB v2019.3 (commit: ac76ff7) reports 23769.91 photon/ms (-10.2%);
6. The version back on Jan 18 2019 (commit: 335c9a0) reports 26476.04 photon/ms (baseline);

[fangq]

Qianqian Fang, Project Admin, Feb 8

Shijie, I found two commits are responsible for this speed decrease:

In Aug. 15, I removed the source type template, and replaced with 3 flag templates

1482b0a

this has caused the biggest speed drop based on my test

the 2nd one is less as important, but also contributes to the slow down, i.e. the change to support continuous media

ac10b73

I am going to revert the first one.

---

Qianqian Fang, Project Admin, Feb 8

#89 is fixed

I added the source type template back, and also added a new template for label-based medium to disable updatepropery().

this brings b1 back to 73kp/ms, b2 to 36kp/ms, b3 to 39kp/ms.
before this change, b1 was 59kp/ms, b2 was 29kp/ms, b3 was 30kp/ms.

the downside of having a lot of template is the long compile time - about 20 second -30 second on i7. the file size is also big - 13MB.

at some point, I got even higher speed - for B1, I got 97kp/ms, but I was not able to reproduce it.