← GPU latency results

gpu/reuse_contraction

Legacy report: ingestion provenance is unrecorded; no raw artifact is inferred.

Original report

gpu/reuse_contraction: CUDA public API benchmark

Measured cells: 1068. Unsupported cells: 572, including 0 numerically verified but timing-excluded cells. Unsupported cells do not count as measured coverage.

Raw run: data/results/nvidia-gpu/gpu/reuse_contraction/20261013_202222/.

Values are median (IQR) µs per operation. Samples time batches of declared operations plus device synchronization; retained output memory is limited to 256 MiB. Batch counts and raw durations are preserved in JSONL.

Setup, input transfers, library/plan initialization, JIT compilation, validation downloads and output destruction are outside timing. Intrinsic output allocation is inside timing; output-reuse rows instead use preallocated destinations across samples. Accumulating destinations are restored from prepared GPU seeds and synchronized before every batch clock, including initialization and calibration. Direct and eager each use one entered session across samples; direct FFT uses cached FftExecutor. JAX eager and compiled JIT are separate arms.

All measured output elements are checked against NumPy/SciPy double-precision reference results before and after sampling. Unsupported rows include unavailable APIs and APIs whose internal setup cannot be excluded; these are not performance measurements. Short batches limited by output memory remain identifiable in the sampling table.

Transposed borrowed-read cases use prepared native CUDA views in tenferro and PyTorch. JAX compares the same logical values in its native representation; its columns do not claim an equivalent public strided-view contract. View creation and input conversion are outside timing.

Unit-coefficient AXPBY uses native JAX Ref.addupdate with caller-owned fixed buffers. Its synchronization includes one GPU scalar completion token per destination and block_until_ready; tokens are retained through clock stop and included in the logical retention bound. Helper compilation and seed restoration are untimed. Full Ref snapshots are used only for untimed inspection. This scalar-token synchronization cost is included in the reported latency; it performs no full output copy or D2H transfer inside timing.

Generalized batch dot preserves native output order: tenferro [m,n,b], PyTorch/JAX [b,m,n]. Input permutation views and dimension-number configurations are prepared before timing. Output axes are normalized only on downloaded CPU arrays for verification; no timed output transpose or copy is added.

Prepared einsum uses the public ConcreteEinsumPlan for ik,kj->ij, built once outside execution and sampling. Native matrix multiplication/reuse/accumulation references match the selected mathematical and output-ownership contracts; they do not assert an equivalent prepared-plan interface. Transposed input views/values, read/write handles and accumulation seed restoration are untimed.

Sampling