[Feature]: Add comprehensive throughput benchmarks for kornia-tensor-ops

[Vinesh2929]

🚀 Feature Description

Add throughput benchmarks for the 11 unbenched operations in kornia-tensor-ops:
add, sub, mul, div, min, mul_scalar, abs, powf, powi, mean, and sum_elements.
Currently only dot_product and cosine_similarity are benchmarked.

📂 Feature Category

Performance Optimization

💡 Motivation

The GSoC 2026 GPU acceleration project requires CPU performance baselines before
GPU kernels can be written and speedups verified. Without benchmarks for these ops,
there is no way to measure the impact of GPU acceleration when it lands.

💭 Proposed Solution

Add ~220 lines to crates/kornia-tensor-ops/benches/bench_ops.rs covering all 11
unbenched ops. Use Throughput::Bytes (not Elements) so results are directly
comparable to hardware bandwidth specs and roofline analysis. Test sizes:
[64, 512, 4096, 65536] — spanning sub-cache through bandwidth-bound regimes.

📚 Library Reference

Criterion.rs benchmarking library (already a dependency). Pattern follows existing
bench_dot_product1 and bench_cosine_similarity in bench_ops.rs.

🔄 Alternatives Considered

Could use Throughput::Elements instead of Bytes, but Bytes allows direct comparison
to DRAM bandwidth in MB/s which is the standard language of roofline modeling.

🎯 Use Cases

When GPU backends are added to kornia-tensor, corresponding GPU benchmarks can reuse
this metric and speedup will be immediately visible as MB/s on the same roofline chart.

📝 Additional Context

I am a GSoC 2026 applicant targeting the GPU acceleration project. @edgarriba

🤝 Contribution Intent

• I plan to submit a PR to implement this feature
• I'm requesting this feature but not planning to implement it

[github-actions]

⚠️ Before Submitting a PR

Please wait for a maintainer to approve and assign this issue to you or someone else before submitting a PR. PRs submitted without prior approval and assignment may be closed. See our Contributing Guide for details.

[sidd-27]

the kornia-tensor-ops isnt really used anywhere so this doesn't make sense to have

[Vinesh2929]

the kornia-tensor-ops isnt really used anywhere so this doesn't make sense to have

Thanks for the feedback Sidd! That makes sense, would it be more useful to add throughput benchmarks for kornia-imgproc operations instead? Things like resize, gray_from_rgb, and the filter ops seem like natural candidates since those are the CPU baselines that would matter most for the GPU acceleration work. Happy to redirect the PR there if that's a better fit.

[namanguptagit]

the kornia-tensor-ops isnt really used anywhere so this doesn't make sense to have

Thanks for the feedback Sidd! That makes sense, would it be more useful to add throughput benchmarks for kornia-imgproc operations instead? Things like resize, gray_from_rgb, and the filter ops seem like natural candidates since those are the CPU baselines that would matter most for the GPU acceleration work. Happy to redirect the PR there if that's a better fit.

I think its a project for Gsoc and needs a lott of planning and discussion before doing so maybe they will plan these as tasks during gsoc. Just my opinion.