MatrixVectorProduct

Published configurations

Runtime-shape spatial matrix-vector products used by gradient mapping.

Measured operation

Measures the canonical fixed-capacity product with runtime extents.

Architecture

All rectangular shapes with zero through three rows/columns and all real/complex operand pairs are measured. Input padding is poisoned; the logical contraction must not consume it and inactive result storage must be zero. Bounded dyadic operands give an exact independent reference sum, including genuinely complex values and mixed scalar promotion.

Timed calls include result construction and its storage initialization. Escaping operand/result addresses and memory barriers prevent constant folding and discarded padding stores. Guards and operand setup are untimed. The owning Variational physical-gradient benchmark measures the consumer; these primitive timings are not assembly or gradient-sweep timings.

Recorded results

Each case retains its original inputs, iteration policy and correctness checks. Raw reports preserve counters, repetitions and the Google Benchmark context. CPU, wall and manually supplied times are labeled separately; unlike operations and configurations are not averaged together.

Timing history

RodinSpatialMatrixVectorBenchmarks · auto · regular · ranks: 1

Benchmark machine and measurement conditions

Published CI benchmarks run on standard GitHub-hosted ubuntu-24.04 virtual machines for public repositories.

System	CPU	Memory	SSD
Ubuntu 24.04, x64	4 vCPUs	16 GB	14 GB

The benchmark builds use GCC 14 and Release mode. Each job runs one benchmark executable at a time. OpenMP, OpenBLAS and MKL use one thread; MPI ranks share the same runner. The selected configuration identifies the build and rank count.

The CPU model and host load can vary between virtual machines, so small timing differences should be checked across repeated runs. GitHub runner specifications describe the available hardware. The source workflow is BenchmarkGroup.yml.