GPU Permutation Benchmark Results

• Target profile: nvidia-gpu
• Suite: gpu/permutation
• Suite file: benchmarks/gpu/permutation.yaml
• Timestamp: 2026-10-06T19:49:20.873768+00:00
• tenferro-rs commit: 5cf78c7ec0ad9516dd78bab546d5e7bd42fa3102

GPU Information

• Device: cuda:0
• Name: NVIDIA A100 80GB PCIe
• UUID: GPU-530977e1-4968-9283-4129-9fbec3e66542
• Memory: 80 GiB
• Driver version: 580.126.09
• CUDA version: 13.0
• CUDA runtime: 12.9
• cuDNN version: 92700

tenferro-cuda-transpose is the eager TensorStructural::transpose op on CudaBackend (compact col-major input only); tenferro-cuda-to-contiguous is TypedTensor::backend_region_view (source layout) + TypedTensorView::transpose_view(perm) + TensorViewCanonicalization::to_contiguous (accepts arbitrary source strides). For framework comparisons, tenferro-cuda-to-contiguous is the primary like-for-like column for PyTorch's view/permute-then-materialize semantics; allocation and reuse timings are shown separately. tenferro-cuda-transpose is the direct structural-permutation comparison for primitive/kernel-oriented backends such as cuTENSOR. Both public tenferro columns allocate a fresh device tensor on every call. tenferro-cuda-destination-reuse calls the public CudaBackend::copy_read_into override from a compact source into an inverse-permuted caller-owned destination view allocated once outside timing. The production path uses the native CUDA copy kernel, so it is distinct from the raw cuTENSOR control. Its per-call destination-view metadata construction is included in the timed public API dispatch. cutensor, pytorch-cuda, and memcpy-d2d also reuse a destination buffer allocated once per pattern. memcpy-d2d only participates in the contiguous identity-permutation baseline pattern. Correctness is verified against a host-computed naive reference, downloaded outside any timed region, before any timing; a FAILED cell means that backend's output did not match the reference for that pattern. A skipped cell means the backend does not participate in that pattern's semantics (or, for cutensor, that the installed cuTENSOR library rejected the pattern's rank at runtime) -- both are reported as skipped rather than a failure.

The GPU-only pattern set is intentionally larger than the CPU permutation set: most rows contain 2^29 f64 elements (4 GiB per tensor) so A100-class measurements exercise steady-state device throughput at roughly the 10 ms scale instead of launch/synchronization overhead.

Device-copy baseline: memcpy-d2d median 5.132 ms, 1673.74 GB/s for memcpy_24d_64x2. It is a bandwidth reference, not a permutation participant, so it is not shown as a table column.

Allocating output

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

Reusing output

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