ENH: benchmark for iterators doing a static cast of pixel

[blowekamp]

No description provided.

[dzenanz]

This looks good on a glance, I have not tried running it (yet).

[blowekamp]

Here is the performance I get on my platform.

Processor:           Apple M1 Max
    Cache:           131072
    Clock:           0
    Physical CPUs:   10
    Logical CPUs:    10
    Virtual Memory:  Total: 1024            Available: 745
    Physical Memory: Total: 65536           Available: 29087
OSName:              macOS
    Release:         14.8
    Version:         23J21
    Platform:        arm64
    Operating System is 64 bit
ITK Version: 6.0.0
Name Of Probe (Time)          Iterations     Total (s)      Min (s)        Mean (s)       Max (s)        StdDev (s)     
IFf3->IFd3-ImageAlgorithm     25             0.0420382      0.0012722      0.00168153     0.00651908     0.00116535     
IFf3->IFd3-Range              25             0.0455065      0.00154901     0.00182026     0.0054009      0.000873138    
IFf3->IFd3-RangeForLoop       25             0.0472529      0.00160909     0.00189012     0.00570703     0.000914247    
IFf3->IFd3-RegionIterator     25             0.146089       0.0054369      0.00584356     0.00982189     0.000977947    
IFf3->IFd3-ScanlineIterator   25             0.134336       0.00511503     0.00537344     0.008497       0.000764301    
IVf->IVd-ImageAlgorithm       25             0.0428834      0.00131106     0.00171534     0.00652003     0.00114749     
IVf->IVd-Range                25             1.31563        0.051743       0.0526253      0.056469       0.00111167     
IVf->IVd-RangeForLoop         25             1.32139        0.051738       0.0528556      0.0596611      0.0019433      
IVf->IVd-RegionIterator       25             1.35381        0.0531781      0.0541523      0.0612571      0.0017682      
IVf->IVd-ScanlineIterator     25             1.33279        0.051826       0.0533116      0.0563779      0.00140679     
Iu2->Ii2-ImageAlgorithm       25             0.00393724     0.000125885    0.00015749     0.000872135    0.00014889     
Iu2->Ii2-Range                25             0.0285077      0.00107479     0.00114031     0.00161409     0.000106848    
Iu2->Ii2-RangeForLoop         25             0.210796       0.00799298     0.00843185     0.00861001     0.000214828    
Iu2->Ii2-RegionIterator       25             0.127022       0.00462103     0.00508087     0.00763893     0.000759563    
Iu2->Ii2-ScanlineIterator     25             0.00756073     0.000286818    0.000302429    0.000630856    6.85298e-05  

updated

I would be curious how MSVC compares.

[dzenanz]

Running b8d6f5d, compiled with VS2026:

System:              Ryzenator
Processor:           AMD Ryzen 9 5900 12-Core Processor
    Cache:           64
    Clock:           3001
    Physical CPUs:   12
    Logical CPUs:    24
    Virtual Memory:  Total: 147357          Available: 111060
    Physical Memory: Total: 130973          Available: 97578
OSName:              Windows
    Release:         6.2.9200
    Version:         Windows 8
    Platform:        AMD64
    Operating System is 64 bit
ITK Version: 6.0.0
Name Of Probe (Time)          Iterations     Total (s)      Min (s)        Mean (s)       Max (s)        StdDev (s)
IFf3->IFd3-Range              3              0.015655       0.00506568     0.00521835     0.00548363     0.000230611
IFf3->IFd3-RangeForLoop       3              0.0137331      0.00445437     0.00457772     0.00474715     0.000151734
IFf3->IFd3-RegionIterator     3              0.0164073      0.00500798     0.00546908     0.00620317     0.000642661
IFf3->IFd3-ScanlineIterator   3              0.0153575      0.0046649      0.00511916     0.00597906     0.000745089
IVf->IVd-Range                3              0.111175       0.0358229      0.0370584      0.0387268      0.0014996
IVf->IVd-RangeForLoop         3              0.0901725      0.0296252      0.0300575      0.0303664      0.000385727
IVf->IVd-RegionIterator       3              0.0869195      0.0286071      0.0289732      0.0296695      0.000603314
IVf->IVd-ScanlineIterator     3              0.0908246      0.0290203      0.0302749      0.0314319      0.00120876
Iu2->Ii2-Range                3              0.00413585     0.00125837     0.00137862     0.00148511     0.000113992
Iu2->Ii2-RangeForLoop         3              0.00282955     0.000879765    0.000943184    0.000981092    5.52715e-05
Iu2->Ii2-RegionIterator       3              0.00654435     0.001266       0.00218145     0.00364113     0.00127767
Iu2->Ii2-ScanlineIterator     3              0.00458884     0.000971317    0.00152961     0.0018959      0.000491289

[blowekamp]

I updated the PR, and my posed performance benchmark to include comparison to the ImageAlgorithm::Copy ( which is an std::transform ).

There are a couple observations here:

• The ImageAlgorithm::Copy is significantly faster.
• There seems to be difference between how efficiently the range based for loop is optimized between compilers. Particular on my bench mark the conversion of scalars is notably slower.
• The conversion of VariableLengthVectors is note able slower with iterators ~10x+ due to the static_cast to the output creating a dynamic memory allocation.