fix(bf16): rewrite encoder/decoder as standard IEEE 754 bit ops#51
Closed
gHashTag wants to merge 5 commits into
Closed
fix(bf16): rewrite encoder/decoder as standard IEEE 754 bit ops#51gHashTag wants to merge 5 commits into
gHashTag wants to merge 5 commits into
Conversation
The test file does #include <goldenfloat/gf16.hpp> but the cpp/include directory was not in the include path, causing a build failure. Closes #36
- Fix root Cargo.toml malformed cfg: cfg(unix(all(not(...)))) -> cfg(unix) - Add #include "gf16.h" to Go cgo preamble (was missing, all C names unresolved) - Add zig-out/bin to Python library search paths for CI compatibility
Closes #28 — replace goto-bus/setup-zig with mlugg/setup-zig in both test-bindings.yml and release.yml. Closes #35 — add #![allow(non_camel_case_types)] to goldenfloat-sys lib.rs for FFI-style type names. Closes #38 — GF16.fromF32 now maps NaN to exp=0x3F,mant=1 (not inf). GF16.toF32 returns std.math.nan when exp=0x3F and mant!=0. Closes #36 — add cpp/include to CMakeLists include_directories so <goldenfloat/gf16.hpp> resolves. Closes #37 — remove import "C" from gf16_test.go (Go forbids cgo in test files). Rewrite tests as exported Test* functions using the Go wrapper API from gf16.go.
Previous implementation had multiple bugs: - Exponent clamped to ±7 instead of IEEE-754 ±127 (7-bit range vs 8-bit) - Wrong mantissa width (8 bits instead of 7) - Wrong bit layout for decode (7-bit exp + 8-bit mantissa instead of 8+7) - Wrong bias handling in frexp path Standard BF16 is simply the top 16 bits of an IEEE 754 f32: [S:1][E:8][M:7] = bits 31..16 of f32 This replaces 54 lines of broken frexp-based code with 2 lines of correct bit manipulation, matching every major BF16 implementation (PyTorch, TensorFlow, MLX, etc.). Added 5 new BF16 tests covering: 1.0, 100.0, 1e10, small values, and special values (inf, NaN, ±0). Closes #22
gHashTag
force-pushed
the
fix/issue-22-bf16-encoder
branch
from
accc088 to
733470f
Compare
Owner
Author
|
Replaced by #52 (clean rebase on main) |
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
1 participant
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.
Summary
Replaces the broken
f32ToBf16/bf16ToF32implementation with the standard 2-line bit manipulation approach.Bugs Fixed
The previous implementation had 4 bugs identified in #22:
Fix
Standard BF16 is simply the top 16 bits of an IEEE 754 f32:
f32ToBf16(a)=@bitCast(a) >> 16bf16ToF32(x)=@bitCast(x << 16)This matches every major BF16 implementation (PyTorch, TensorFlow, MLX).
Impact
This will change bf16 benchmark results — the broken codec artificially made bf16 look like GF16 on narrow ranges. With the fix, bf16 will correctly show its full IEEE-754 dynamic range (±127 exponent), making the GF16 vs bf16 comparison honest.
Tests Added
BF16: roundtrip 1.0— verifies exact encoding (0x3F80)BF16: roundtrip 100.0— verifies dynamic rangeBF16: roundtrip 1e10— verifies large values work (was broken before)BF16: roundtrip small values— 8 values from ±0.5 to ±3.14 to ±1e-10BF16: special values— inf, NaN, ±0 preservationCloses #22