Skip to content

PORTING_STATUS.md

Latest commit

 

History

History
58 lines (48 loc) · 3.77 KB

PORTING_STATUS.md

File metadata and controls

58 lines (48 loc) · 3.77 KB
Raw

Cross-platform porting status

The library builds its complete API on both Windows (x64) and Linux/SysV. On Windows every function behaves as it did upstream. On Linux, correctness depends on whether a module has been migrated to the cross-platform ABI layer (src/asm/abi.inc) and validated against a scalar reference.

Windows x64: every module below works (the ABI shims are no-ops there).

Linux / System V: only the ✅ modules are verified correct. The ⏳ modules still use the Windows-only calling convention and will return incorrect results if called on Linux — do not use them on Linux yet.

✅ Ported and test-verified on Linux

Module(s) Functions
fp_core_reductions* (10 widths) fp_reduce_{add,min,max,mul,product}_*, fp_reduce_add_f64_where
fp_core_fused_folds* (10 widths) fp_fold_{dotp,sumsq,sad}_*
fp_core_fused_maps* (all widths) fp_map_{scale,offset,axpy}_*, fp_zip_add_* (i8/i16/i32/i64, u8/u16/u32/u64, f32/f64)
fp_core_predicates fp_pred_all_eq_const_i64, fp_pred_any_gt_const_i64, fp_pred_all_gt_zip_i64
fp_core_tier3 fp_count_i64, fp_range_i64, fp_iterate_{add,mul}_i64, fp_group_i64, fp_run_length_encode_i64, fp_zip_with_index_i64, fp_reduce_{and,or}_bool, fp_replicate_f64
3d_math_kernels fp_map_transform_vec3_f32, fp_zipWith_vec3_add_f32, fp_map_quat_rotate_vec3_f32, fp_reduce_vec3_add_f32, fp_fold_vec3_dot_f32, fp_quat_normalize_asm, fp_quat_to_mat4
fp_core_matrix fp_mat4_identity, fp_mat4_mul, fp_mat4_mul_vec3, fp_mat4_transpose, fp_mat4_mul_vec3_batch
fp_core_essentials fp_contains_i64, fp_find_index_i64, fp_take_n_i64, fp_drop_n_i64, fp_slice_i64, fp_reverse_i64, fp_concat_i64, fp_reduce_product_{i64,f64}, fp_replicate_i64
fp_core_simple_maps fp_map_abs_{i64,f64}, fp_map_sqrt_f64, fp_map_clamp_{i64,f64}
fp_core_scans fp_scan_add_{i64,f64} (prefix sum)
fp_core_descriptive_stats fp_moments_f64, fp_descriptive_stats_f64
fp_core_percentiles fp_percentile_sorted_f64, fp_percentiles_sorted_f64, fp_quartiles_sorted_f64
fp_core_compaction fp_filter_gt_i64_{simple,simd}, fp_take_while_gt_i64, fp_drop_while_gt_i64, fp_partition_gt_i64 (7-arg)
fp_core_tier2 fp_unique_i64, fp_union_i64, fp_intersect_i64 (set ops)

Covered by tests/test_reductions.c, test_int_families.c, test_maps.c, test_game_math.c, test_essentials.c (run make test).

✅ Migration complete

Every in-scope assembly module is now ported to the cross-platform ABI and test-verified on Linux (make test runs 12 suites). BLAKE3 hashing (fp_blake3_avx2.asm) was removed — it is cryptographic code from the upstream vector database, out of scope for an FP + game-math library.

Bugs fixed so far

  • AVX-512 instruction in an AVX2 library — vpbroadcastq ymm, r64 (4 sites) raised SIGILL on AVX2-only CPUs. Replaced with vmovq+vpbroadcastq.
  • fp_reduce_vec3_add_f32 returned (0,0,0) for arrays shorter than 8 (tail accumulator was zeroed before use).
  • fp_mat4_transpose produced a wrong permutation (not a valid transpose); replaced with a correct 4×4 transpose.
  • fp_reverse_i64 used an in-place pair-swap loop while writing to a separate output buffer, interleaving the two ends instead of reversing. Rewritten as out[i] = in[n-1-i].
  • Narrow-width fp_core_fused_maps_* clobbered callee-saved r12/r13/r14 (+ r15 in some axpy, rbx in axpy_u64, and xmm6/7/15 on Windows) without preserving them — corrupting the caller on every platform (segfaults on Linux under -O2). Every narrow map function now uses a uniform frame that preserves rbx/r12–r15 (and, on Win64 only, xmm6/xmm7/xmm15).