Repository navigation
defk-vcx.5: exact float32 as a documented recipe, not a builtin - #83
Merged
Merged
Conversation
The owner chose the recipe over a builtin: no catalog workload needs one. docs/language.md gets an 'Exact float32 arithmetic' section and a rewrite row, the cheatsheet a rewrite row, performance.md 6.44 the decision, and element_types_test a mul_add32 kernel with ties to even, negative, product and subnormal ties against an Elixir float32 reference.
DanielLuning
pushed a commit
that referenced
this pull request
Oct 26, 2026
…rktree ../defk-vcx.5). mix ci passed. Close after merge; then defk-vcx can close."
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changed
The owner chose option 1 of defk-vcx.5: a recipe, no builtin. No catalog workload needs a builtin today.
docs/language.md: a new section "Exact float32 arithmetic" under Float arithmetic. It states the one-element[f32]scratch pattern, why it is exact for+ - * /andsqrt(an f64 holds more than twice an f32's bits), thatexp/log/poware rounded twice, and that the write raises where f32 hardware gives inf. Two run-checked kernels (add32,dot32) show a tie and an f32 dot product. A row in "Unsupported, write this instead" forx :: f32/<<y::float-32>> = <<x::float-32>>.docs/guides/cheatsheet.cheatmd: one row in Common rewrites.docs/performance.md6.44: records the decision.test/element_types_test.exs: kernelmul_add32(with an agreement property) and a test of ties to even:2^24 + 1,2^24 + 3, the negative mirrors, two product ties, and subnormal ties, each checked in both modes and against an Elixir float32 reference.Results
No measurement. The numbers in 6.44 are unchanged.
How it was checked
mix cion an M-series Mac in the worktree: 2312 tests passed. (The first run failed the two knownReportTest.Fullcache tests from the copied_build; after removing those cache files, everything passed.)CHANGELOG.mdentry.test/stability/? No.