soniumCompute backends · 34
BLAS, BEM kernels, and block-Krylov workspaces across CPU, CUDA, and Metal, with residency accounting and fault contracts. Method shown on hover; tolerance is the pass gate. back to QA
| Case | Tier | Reference | Gate |
|---|---|---|---|
compute.properties.v1 | smoke | independent checked-index arithmetic | exact |
compute.operation-matrix.v1 | smoke | independent scalar row-major GEMV/GEMM/reduction/local-contraction accumulation, precision cast round trips, gather index/weight accumulation, CSR SpMV/SpMM and stable transpose references, and view-index references | maximum relative element error <= 4e-5; at least 144 finite reduction, 60 reduction-failure, 12 successful gather, 12 gather-failure, 48 caller-owned multi-dot, 32 local-contract, 24 cast, 16 sparse, |
compute.strided-projection-contract.v1 | smoke | fixed-order Hermitian projection coefficients and unchanged destination on rejection | successful_coefficients >= 2; shape_rejected=1; non_finite_rejected=1; allocation_rejected=1; destination_preserved=1 |
compute.batched-dense-contract.v1 | smoke | independent scalar grouped GEMV/GEMM accumulation | relative error <= 5e-5; four scalar families, 12 GEMV cases, 12 GEMM cases, 8 typed rejections |
compute.block-workspace.v1 | smoke | independent scalar block-matrix multiplication | single and multi-RHS absolute error <= 1e-12; warm capacity/buffers stable; clear and layout rejection pass |
compute.block-krylov-workspace.v1 | smoke | independent dense 2x2 block-GMRES solution | solution error <= 1e-8; stable device buffers/capacity and host capacity; clear and dimension rejection pass |
compute.csr-spmv-contract.v1 | smoke | independent scalar CSR row accumulation | single/multi-RHS SpMV errors <= 1e-12; all rejection and no-mutation invariants pass |
compute.csr-spmm-contract.v1 | smoke | independent scalar CSR multi-RHS column-major accumulation | maximum relative error <= 2e-6; all rejection and padding invariants pass |
compute.fmm-contract.v1 | smoke | independent scalar complex row accumulation | maximum gather/scatter absolute error <= 2e-12 |
compute.memory-pool-contract.v1 | smoke | independent quota and allocation-accounting invariants | all five accounting invariants must pass |
compute.gmres-workspace.v1 | smoke | independent analytic 2x2 SPD solution | solution error <= 1e-11; stable capacity, clear, and dimension rejection pass |
compute.gmres-residual-contract.v1 | smoke | independent Complex32 CSR residual recomputation | reported-vs-independent residual error <= 2e-5; both strict cases remain non-converged |
compute.scheduler-contract.v1 | smoke | independent scheduler terminal-state invariants | all lifecycle flags >= 1; final bytes and active slots = 0 |
compute.transfer-stats.v1 | smoke | checked copy boundaries and independent round-trip data | all three transfer directions count one operation and 32 bytes; malformed copy does not count |
compute.perf-contract.v1 | smoke | independent integer accounting formulas | case_count >= 20; normal_cases >= 11; saturation_cases >= 9 |
compute.performance-regression-contract.v1 | smoke | sonium-bench empirical confidence-interval comparison | baseline_cells >= 3; accepted_cells >= 3; regression_cases >= 1; rejection_cases >= 2; protocol_checks >= 4 |
compute.cold-warm-contract.v1 | smoke | independent scalar row-major GEMM reference | warm_samples >= 5; max_abs_error <= 1e-12 |
compute.csr-prebake-contract.v1 | smoke | independent scalar CSR column application | all valid application errors <= 1e-12; malformed setup must fail |
compute.capability-contract.v1 | smoke | CPU capability table plus typed unsupported-operation contract | checked_capability_cells >= 64; unsupported_fft_probe >= 1 |
compute.capability-modes.v1 | smoke | CPU capability table joined with reduction-mode residency | mode_cells >= 12 |
compute.view-copy-contract.v1 | smoke | independent strided-view copy and rejection invariants | view_copy_cases >= 12; rejection_cases >= 4; native_d2d_operations >= 4 |
compute.nd-view-copy-contract.v1 | smoke | independent logical-index mapping and destination-preserving rejection | copy_cases >= 4; rejection_cases >= 4; CPU d2d_operations may be zero because the shared default is host-staged |
compute.compatibility-contract.v1 | smoke | independent exact Hermitian dot products | scalar_cases >= 12; valid_modes >= 8; rejection_cases >= 8 |
compute.fft-view-contract.v1 | smoke | shared rank/shape/stride/device/alias validation | valid_descriptor_cases >= 12; rejection_cases >= 16 |
compute.backend-parity.v1 | smoke | independent f64 GEMV/GEMM | 2e-5 absolute; 2e-6 relative |
compute.sparse-workload.v1 | smoke | analytical dense-system solution | 1e-8 |
compute.bem-kernel.v1 | smoke | sonium-wave 3-D Green function and regular_pair_scalar | point-source 2e-14; batch/scalar 2e-12 |
compute.blas-contracts.v1 | smoke | independent row-major GEMV/GEMM accumulation | 3e-12 |
compute.complex-gemv-parity.v1 | smoke | independent implementation parity between complex4m and complex_native | Complex32 <= 2e-5; Complex64 <= 2e-12 |
compute.operator-tape.v1 | smoke | independent scale-two forward and Hermitian-adjoint result | 1e-12; zero reported host/device transfer |
compute.reproducible-workload-bundle.v1 | smoke | analytical dense-system solution plus bit-stable CPU, FEM, and BEM replays | residual and solution error <= 1e-8; output/residual/footprint replay identical |
compute.fem-adjoint.v1 | smoke | independent CSR accumulation and Hermitian inner-product identity | inner-product 2e-12; block 2e-13; residuals 2e-12 |
compute.fem-adjoint-cuda.v1 | accelerator-gate | independent CPU CSR accumulation, f64/Complex64 residuals, block parity, and Hermitian inner-product identity | inner-product 2e-12; block 2e-13; f64/Complex64 residuals 2e-12; nonzero native CUDA work and artifact provenance |
compute.bem-adjoint.v1 | smoke | independent dense row accumulation and Hermitian inner-product identity | operator errors <= 2e-12; GMRES solution <= 1e-8 and residual <= 1e-10 |