[sdebionne]

Description

This is a revival of PR #525. It adds a number of benchmark for reference, mostly on geometry transformation .

This MR mainly introduces a specialization of algorithms for transposed image view. The implementation uses a blocking algorithm to improve cache efficiency.

Performance of copy_pixels and transform_pixels on transposed view vary from x2 faster for smaller image to x3 faster on larger one.

Run on (64 X 3200 MHz CPU s)
CPU Caches:
L1 Data 32 KiB (x32)
L1 Instruction 32 KiB (x32)
L2 Unified 1024 KiB (x32)
L3 Unified 22528 KiB (x2)

current implementation

---------------------------------------------------------
Benchmark               Time             CPU   Iterations
---------------------------------------------------------
transpose/256       28564 ns        28564 ns        24407
transpose/512      132876 ns       132873 ns         5287
transpose/1024     548770 ns       548773 ns         1294
transpose/2048    2255201 ns      2255131 ns          302
transpose/4096   10757579 ns     10757075 ns           64
transpose/8192   63558779 ns     63556336 ns           11

blocked implementation

---------------------------------------------------------
Benchmark               Time             CPU   Iterations
---------------------------------------------------------
transpose/256       14612 ns        14612 ns        48072
transpose/512       57761 ns        57758 ns        11879
transpose/1024     446445 ns       446419 ns         1597
transpose/2048    3551008 ns      3550669 ns          195
transpose/4096   25210960 ns     25206203 ns           25
transpose/8192  175161964 ns    175150467 ns            4

References

See the docs in benchmark/README.md to build and run the benchmarks with CMake and pixi.

See #525 for the previous effort on the topic (benchmarks only).

Tasklist

• Add test case(s)
• Ensure all CI builds pass
• Review and approve