CPU Permutation Benchmark Results

• Target profile: mac-cpu
• Suite: cpu/permutation
• Suite file: benchmarks/cpu/permutation.yaml
• Timestamp: 2026-09-23T12:05:30.508140+00:00
• tenferro-rs commit: 8a1839febeb3c868a502e26397ca10761bbc568d

CPU Information

• Model: Apple M5 Max
• Vendor: Apple
• Logical CPUs: 18
• Physical CPUs: 18
• Sockets: 1
• Cores per socket: 18
• Threads per core: 1
• NUMA nodes: 1
• Performance levels: Super: 6 physical / 6 logical (L1i 192 KiB, L1d 128 KiB, L2 16 MiB, 6 CPUs/L2); Performance: 12 physical / 12 logical (L1i 128 KiB, L1d 64 KiB, L2 8 MiB, 6 CPUs/L2)
• Python platform: macOS-26.5.1-arm64-arm-64bit-Mach-O

CPU pinning is unavailable on macOS; thread counts are controlled only via RAYON_NUM_THREADS / OMP_NUM_THREADS / JULIA_NUM_THREADS.

tenferro-rs measures TypedTensorView::transpose_view followed by BackendSession::to_contiguous_read; one shared session is entered before timing, and input view descriptors and metadata-only transpose_view are built outside the timed region. Materialization accepts arbitrary source strides. Every backend allocates a fresh destination inside each timed call, so the table compares allocation-inclusive end-to-end materialization rather than destination-reuse copy kernels. hptt only participates in patterns with a contiguous source and destination. Correctness is verified against an internal, untimed odometer reference before any timing; a FAILED cell means that backend's output did not match the reference for that pattern.

Threads: 1

Median (p25 / p75) in ms. Missing backends are shown as -; the fastest backend per pattern is bolded.

Threads: 4

Median (p25 / p75) in ms. Missing backends are shown as -; the fastest backend per pattern is bolded.