Richardson-Lucy deconvolution: allow single-precision computation

[grlee77]

Description

This PR improves the performance of deconvolution by keeping the computations in single-precision when the input image is single-precision.

It also avoids unnecessary copies of the input arrays by adding copy=False to existing .astype() calls.

This can give up to a factor 2 performance improvement when calling Richardson Lucy either single-threaded or in various multi-threaded scenarios. See the timings for double-precision in #4083 (comment) vs. single-precision in #4083 (comment).

If this seems reasonable to others, I can also adapt the example from the comments into an ASV benchmark.

Checklist

• Docstrings for all functions
• Gallery example in ./doc/examples (new features only)
• Benchmark in ./benchmarks, if your changes aren't covered by an
  existing benchmark
• Unit tests
• Clean style in the spirit of PEP8

For reviewers

• Check that the PR title is short, concise, and will make sense 1 year
  later.
• Check that new functions are imported in corresponding __init__.py.
• Check that new features, API changes, and deprecations are mentioned in
  doc/release/release_dev.rst.

[emmanuelle]

Thanks for the PR! An ASV benchmark would be great if you find the time :-).

[grlee77]

I added some simple ASV benchmarks for richardson_lucy. There is also a separate commit that fixed the existing restoration benchmarks, which were currently failing to run due to a misnamed variable!

As expected, float32 operation is now faster. This PR also reduced peak memory use a bit in the float64 case by avoiding internal copies.

Result of benchmarks for release 0.17.2

[  0.00%] ·· Benchmarking existing-py_home_lee8rx_miniconda3_envs_pyir_bin_python3.7
[ 40.00%] ··· Running (benchmark_restoration.DeconvolutionSuite.time_denoise_nl_means_f32--)..
[ 60.00%] ··· benchmark_restoration.DeconvolutionSuite.peakmem_denoise_nl_means_f32                       293M
[ 70.00%] ··· benchmark_restoration.DeconvolutionSuite.peakmem_richardson_lucy_f64                        293M
[ 80.00%] ··· benchmark_restoration.DeconvolutionSuite.peakmem_setup                                      124M
[ 90.00%] ··· benchmark_restoration.DeconvolutionSuite.time_denoise_nl_means_f32                    2.32±0.01s
[100.00%] ··· benchmark_restoration.DeconvolutionSuite.time_richardson_lucy_f64                     2.33±0.02s

Result of benchmarks for this PR

[  0.00%] ·· Benchmarking existing-py_home_lee8rx_miniconda3_envs_pyir_bin_python3.7
[ 40.00%] ··· Running (benchmark_restoration.DeconvolutionSuite.time_denoise_nl_means_f32--)..
[ 60.00%] ··· benchmark_restoration.DeconvolutionSuite.peakmem_richardson_lucy_f32                        200M
[ 70.00%] ··· benchmark_restoration.DeconvolutionSuite.peakmem_richardson_lucy_f64                        276M
[ 80.00%] ··· benchmark_restoration.DeconvolutionSuite.peakmem_setup                                      124M
[ 90.00%] ··· benchmark_restoration.DeconvolutionSuite.time_richardson_lucy_f32                     1.64±0.01s
[100.00%] ··· benchmark_restoration.DeconvolutionSuite.time_richardson_lucy_f64                     2.30±0.01s

[grlee77]

For the benchmarks, I ended up making a new DeconvolutionSuite class since the setup method is different to the one in the existing RestorationSuite class (need to store a PDF and convolve it with the input image).

[grlee77]

I think maybe the doc failure might be because I haven't rebased on master since that PR was merged?

[emmanuelle]

Thanks @grlee77 !

[jni]

@grlee77 why is time RL f32 not shown in your benchmarks result?

[grlee77]

@grlee77 why is time RL f32 not shown in your benchmarks result?

It is there, but was misnamed as "time_denoise_nl_means_f32" due to a copy paste error that was later fixed in bf38f4c.

I will edit the comment above to reflect the correct naming.

[jni]

@grlee77 this is then ready after that fix and a rebase/merge on master!

[alexdesiqueira]

Thank you @grlee77!