Skip to content

core-performance.md

Latest commit

 

History

History
70 lines (55 loc) · 3.53 KB

core-performance.md

File metadata and controls

70 lines (55 loc) · 3.53 KB
Raw

Common-kernel performance experiment

This records the shared-kernel baseline. See the subsequent C view experiment for scalar dispatch and bounded-memory widening improvements.

Apple M1 Pro, CPython 3.13.15, NumPy 2.5.3, one thread. Baseline: b793d1f. Times below are microseconds, medians of seven calibrated samples (at least 3 ms per sample, except the loop-count cap). Inputs and results are checked against NumPy before timing. Python and NumPy controls use the same harness in both runs.

Five-bit arrays, 65,536 values

API Operation Before µs After µs Speedup
array-api get-scalar 12.638 0.782 16.2×
array-api set-restore 55299.041 3.780 14631.3×
array-api assign-32 27577.125 4.179 6598.3×
array-api small-view-to-numpy 11.067 0.947 11.7×
array-api sum 26.205 10.003 2.6×
array-api strided-sum 25.014 20.965 1.2×
native-packed sum 340.958 8.839 38.6×
native-aligned sum 289.953 3.209 90.3×

set-restore includes an integer read and two assignments. assign-32 writes 32 strided elements. No timed assignment increases storage width. The small conversion copies a preexisting 32-element view. Native sum changes from Python sum(a) to the new a.sum() kernel; Array API retains the same xp.sum(a) call.

Continuous sum across widths, 65,536 values

Bits NumPy uint8 µs Native packed µs Native aligned µs Array API µs
1 12.529 0.218 0.218 1.402
2 12.618 1.538 1.539 2.709
3 12.548 6.297 2.339 7.467
4 12.517 2.413 2.414 3.598
5 12.543 8.839 3.209 10.003
6 12.553 10.657 3.845 12.598
7 12.517 8.908 4.286 10.415
8 12.525 2.001 2.003 3.187

Scalar complexity and traced allocation

Values Read µs Set/restore µs Read peak B Set peak B Sum peak B
128 0.770 3.736 305 402 556
65,536 0.782 3.780 337 434 556
1,048,576 0.781 3.790 337 434 556

These are tracemalloc peaks per operation, not process RSS or total native memory. C stack storage is excluded: the view descriptor uses approximately 1 KiB at its maximum supported rank. Reduction creates no decoded array. Retained payload layouts are unchanged (5-bit packed uses approximately 5/8 of uint8 payload).

Limits

  • Array API scalar operations still create Python view/dtype objects and remain slower than NumPy scalars.
  • The 6-bit Array API sum is approximately tied with NumPy in this run; small differences are not a robust win.
  • Strided sum remains slower than NumPy, although the first experimental regression was removed.
  • Widening storage is still O(n) and temporarily decodes the root; the constant-time mutation claim excludes widening.
  • Other axis/dtype reductions and advanced indexing still use NumPy fallbacks.
  • This warm-cache deterministic periodic workload is not a replacement for random/skewed competitor benchmarks.

Reproduction

OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 python benchmarks/core_experiment.py --label current --output current.json
python benchmarks/report_core_experiment.py

Run the same benchmark script with baseline and current wheels in the same environment. Raw data: before, after. Architecture and rejected alternatives: shared kernels.