purcaro / straylight

I optimized some code for the European Space Agency, achieving significant speedups - using OpenMP, SSE, Eigen and CUDA. This is the resulting code, with complete documentation on the optimizations performed.

Executive summary

This is the software I developed for Task 2.2 of the "TASTE Maintenance and Evolutions" project of the European Space Agency. It is a port of the IDL implementation of the StrayLight algorithm to C++, and it utilizes OpenMP, Eigen and CUDA to achieve much better speeds than the original IDL code (execution time in gdl: 18 seconds - execution time of my C++ code: 169 ms)

Results for single precision speed - time per frame:

$ ./configure && make && ./src/strayLight -b
Execution time per image : 169 ms
Average and std deviation: 171.23 +/- 1.6 ms

Results for double precision speed - time per frame:

$ ./configure --enable-double && make && ./src/strayLight -b
Execution time per image : 222 ms
Average and std deviation: 224.46 +/- 1.9 ms

Documentation

I wrote extensive documentation of the optimizations I performed to achieve these speedups inside the deliverable in the doc/ folder.

Usage:

$ ./src/strayLight -h
strayLight 1.1c (Thu May 30 14:57:08 2013)

Usage: strayLight [OPTIONS]

  -h            this help
  -v            increase verbosity
  -V            show version
  -b            run benchmark (50 images)
  -i filename   instead of the ESA test image, process this file
  -c channel    process this channel
  -d N          dump log files from computation stages >= N
                (default: 13, i.e. the final result is saved,
                 and never used during benchmarking (no output).

Validation of results

To create a log of the output image generated by the StrayLight algorithm:

$ mkdir -p output
$ ./src/strayLight

This will use the default value of option -d, and generate the stage13 output file. If you wish to generate outputs from all stages, starting with stage 10:

$ ./src/strayLight -d 10

This will generate log files for stages 10, 11, 12 and 13.

To compare the final output (stage13) with the outputs generated by IDL, run this:

$ ./contrib/elementsDiff.py \
    ./output/stage13_1 \
    ./output.from.IDL/stage13_1 | ./contrib/stats.py | grep Overall
Overall: 6.61607071392e-06 +/- 375.7%

...which will verify that the difference between the values in the two outputs (from IDL and from C++) is indeed minimal.

Contact point

Thanassis Tsiodras, Dr.-Ing. [e-mail] / [e-mail] / [e-mail]