testfloat_expr_cli tutorial

This tutorial shows how to execute a Berkeley TestFloat vector file with the binary runner. The runner is reached as floating-conformance --backend testfloat; below, testfloat abbreviates that command.

Quick start

Generate vectors with TestFloat’s testfloat_gen (installed by just conformance fetch binary) and run them:

testfloat_gen -level 1 -rnear_even -tininessafter f64_mul > f64_mul.tv
testfloat --function f64_mul --rounding rnear_even --tininess after f64_mul.tv

TestFloat execution summary
function: f64_mul
rounding: rnear_even
tininess: after
cases: …
selected cases: …
passed cases: …
failed cases: 0

Everyday tasks

Match the generator options

Pass the same function, rounding and tininess as the generator, and --exact when the vectors were generated with -exact:

testfloat_gen -rminMag -exact f32_roundToInt > r.tv
testfloat --function f32_roundToInt --rounding rminMag --exact r.tv

Shard a large file

testfloat --function f128_mulAdd --shard-count 8 --shard-index 3 --json big.tv

Shard i runs the vectors whose index is i modulo the shard count.

JSON for scripts

--json prints function, rounding, tininess, exact, totalCases, selectedCases, passedCases, failedCases and failedIds (FUNCTION:LINE).

Going further

just conformance run binary --level 1 --tininess after --tininess before plans the full matrix, streams testfloat_gen output in chunks into temporary files, runs this runner on each chunk, and remaps the line-based ids to global vector numbers.
testfloat_expr tutorial explains the pass rule.

Common pitfalls

Exactly one file. A second path is an error.
Rounding defaults to rnear_even. Always pass the generator’s mode.
Ids are line numbers of the file you pass. When you split a stream into chunks, keep track of the offset yourself.