feat(quantized): F16 type + ScalarOperand impl for burn NdArrayElement (sprint A2 partial)#125
Conversation
…perand impls (sprint A2 partial) Items (1)+(2) partial — F16 type + round-to-nearest-even conversion shipped. BF16x16/F16x16 SIMD vector types deferred (simd_half module WIP, not yet complete). What ships: - F16 struct with from_f32 (round-to-nearest-even), from_f32_truncate, to_f32 - F16 constants: ZERO, ONE, NEG_ONE, INFINITY, NEG_INFINITY, NAN - Slice converters: f32_to_f16_slice, f16_to_f32_slice, f32_vec_to_f16, f16_vec_to_f32 - Add/Sub/Mul/Div/Neg operator impls (via to_f32→op→from_f32) - ScalarOperand impl (unblocks burn's NdArrayElement for F16) - 11 new tests (round-trip, boundary values, NaN preservation, arithmetic) What's deferred (Wave 3): - BF16x16 / F16x16 SIMD vector types (simd_half module) - Slice-level add_bf16_inplace / cast_f16_to_f32_batch https://claude.ai/code/session_01NYGrxVopyszZYgLBxe4hgj
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 53c950e7ef
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| #[allow(clippy::all, missing_docs, dead_code, unused_variables, unused_imports)] | ||
| // pub mod simd_half; // TODO: BF16x16/F16x16 SIMD vectors (A2 WIP) |
There was a problem hiding this comment.
allow on backend module
Because pub mod simd_half is commented out, the outer attribute #[allow(clippy::all, missing_docs, ...)] now applies to the next real item (backend) when feature = "std" is enabled. This unintentionally suppresses linting and documentation checks for the entire backend module, which can hide real issues and makes maintenance harder.
Useful? React with 👍 / 👎.
2 participants
Summary
Sprint A2 (partial) — F16 type + conversions + ScalarOperand impls. Salvaged from the agent's stash after the agent timed out before pushing.
Closes item (1) of the parity list (ScalarOperand/LinalgScalar on half types). Item (2) (BF16x16/F16x16 SIMD vectors) deferred to Wave 3 — the
simd_halfmodule was started but not completed before timeout.What ships (+471/-2 LOC in
src/hpc/quantized.rs)F16(pub u16)— IEEE 754 binary16 withfrom_f32(round-to-nearest-even),from_f32_truncate,to_f32ZERO,ONE,NEG_ONE,INFINITY,NEG_INFINITY,NANf32_to_f16_slice,f16_to_f32_slice,f32_vec_to_f16,f16_vec_to_f32Add/Sub/Mul/Div/Negoperator impls (viato_f32→op→from_f32)ScalarOperandimpl — unblocks burn'sNdArrayElementfor F16What's deferred (Wave 3)
BF16x16/F16x16SIMD vector types (simd_halfmodule — WIP, not yet complete)add_bf16_inplace/cast_f16_to_f32_batchVerification
cargo build: cleancargo test --lib: 1781 passed, 0 failed (sprint total: +81 tests over master baseline)https://claude.ai/code/session_01NYGrxVopyszZYgLBxe4hgj
Generated by Claude Code