Execute benchmarks - Githubissues

ComputationalRadiationPhysics / imresh

Shrink-Wrap Phase Reconstruction Algorithm

MIT License

3 stars 2 forks source link

Execute benchmarks #11

Closed mxmlnkn closed 8 years ago

mxmlnkn commented 8 years ago

This may be necessary while already developing:

[x] benchmark all parts on each own. If there are multiple versions of one part, then choose the fastest.
[x] benchmark all fastest working together in shrink wrap
[x] do profiling?

mxmlnkn commented 8 years ago

the vectorReduce benchmarks are quite dreamy ... array length 4686239 float elements:

cudaVectorMaxSharedMemoryReduce: 1.18ms
cudaVectorMaxWarpReduce: 0.36ms
cudaVectorMaxWarpAndSharedMemoryReduce: 0.38ms
vectorMaxOpenMP: 10269.7ms Measurement doesn't include cudaMemcpy (because the shrink wrap algorithm calls this function roughly 10 times with data already in graphic memory).

mxmlnkn commented 8 years ago

Well was/is done pretty extensively in the unit tests mainTestAll-GTX760.txt I'm sorry and I know this makes it pretty useless, but please note, that some headers might not be in the correct order, please consult the source code and correct it there.