Optimize voice loading about mimic1

[zeehio]

I will write what I would do, but feel free to take the approach you prefer:

In order to optimize only what needs to be optimized, I would write a simple unittest-like program that loads the Mycroft voice from a file and prints the time it takes to load it. Then I would run it several times to get an estimate of the performance on my computer.
Then I would try those optimizations from easier to harder.

• inline functions seems easy but if the compiler is smart enough it should do it automatically so I don't expect a huge gain there. I don't know how smart compilers are, so I may be wrong here.
• I have never used setvbuf but if it is easy to setup it should reduce many internal read calls. It may be good.
• Reducing the number of cst_fread calls is possible up to a limit due to how the file format is set up. Strings are of variable length so I don't know how much can that be optimized.

I believe the slowness is due to either:

• too many small reads from file (would be solved by setvbuf)
• too many small cst_alloc calls when creating the cart trees. The solution for this would be to have a big malloc and put several tree nodes all together in there somehow...
• too many function calls in general. If that is the case we will need to think of a better format for dumping/loading cart trees...

Hope I can be of help!

[forslund]

Thanks for the suggestions. I am also sceptical to inlining the function calls as you say the compiler should be smart enough. Maybe just make them static so the compiler can optimize correctly even when making shared library. (Maybe it already can without static, I'm not up to date with how compiler works nowadays)

I've tried two things so far:

setting a 64k buffer with cstvbuf() reduced read operations from ~17000 down to ~1500 which may be significant for platforms like the raspberry pi. But I could only see a ~5% improvement in loading times on my computer.

Replace the four initial cst_fread() with a single in cst_read_tree_nodes(). No effect on the number of read operations but according to my time measurements the speed improvement is ~10%

I'll look into some sort of block memory allocation.

[forslund]

Small update, I've tried a bunch of minor changes. On my machine I can see ~25% performance improvement.

What I've done:

• Some more minor function optimizations except for the ones stated above gave a couple of percent each
• Replace cst_alloc() (which uses calloc()) with malloc() further improves things.

Next steps will be

• a simple block allocator
• try to mmap the file

[nodered]

I have around 300MB of free memory on my pi to play with...would a ramdisk help the loading time of the libraries?

I ran a trace with this
strace -o strace ./mimic test

Then did...
cat strace | grep /home/pi/mycroft-core/mimic/lib
cat strace | grep /home/pi/mycroft-core/mimic/.lib

I don't know much about the program but if some of the libraries can be put onto a ramdisk it might help with the performance.