Module tests/routines

Per-routine test structure for wgblas BLAS Level 1 routines.

Every routine follows the same three-test pattern. Each test (test.<routine>.js) imports from the shared helpers in tests/helpers/ and, for FMA routines, from its own helpers.js.

1 — Validation

Calls runValidation with the routine's validationSpecs object. Each spec is a loaded JSON file from tests/validation/params/ describing what inputs must throw and what boundary inputs must not throw.

This layer catches type errors, range errors, and size errors — the guard checks at the JS boundary before any GPU work begins.

2 — Fixtures

Calls runFixtures with 100 fast-check random inputs. For each input the GPU result is compared against the @stdlib CPU reference using a routine-specific error metric:

• exact routines (isamax, scopy, sswap, sscal) — threshold 0, result must match bit-for-bit.
• reduction routines (sasum, snrm2) — threshold 50 ULP, allowing for the different summation order between GPU tree-reduction and sequential CPU accumulation.
• FMA routines (saxpy, sdot, srot, srotm) — threshold 1 forward error factor, tolerating the difference between a fused and an unfused multiply-add. See tests/routines/helpers.mjs for the formula.

3 — Edge Cases

Deterministic assert.strictEqual / assert.deepEqual checks that target behaviour the random test cannot reliably hit:

• n=1 — single-element operation.
• alpha/c/s=0 — no-op or identity; catches missing short-circuit guards.
• incx≠incy — independent strides; catches index-mixing bugs.
• gap elements — 99s placed at non-strided positions must survive unchanged (scopy, sscal, sswap, srot, srotm, saxpy).
• ties — smallest-index tie-breaking in isamax.
• Pythagorean triple ([3,4]→5) — exact integer result from sqrt in snrm2.
• flag per value — each of -2/-1/0/1 tested separately for srotm.