ONNX op wishlist: typed Where, pair compares, copy-free matmul_f32, gather_rows_f32, float accumulation order by AdaWorldAPI · Pull Request #400 · AdaWorldAPI/ndarray

simd: blend_f32, blend_f32_scalar, blend_i64, blend_i64_scalar (ONNX Where); one generic body for the blend family

claude

committed

f9fa67c

simd: two-array compares to a mask (f32/i32/i64: eq, lt, le, gt, ge) and i64 scalar compares

claude

committed

1717946

hpc::amx_matmul: matmul_f32 reads inputs in place via strides (no packing, no result buffer); matmul_f32_batched

claude

committed

92eb4c1

hpc::gather: gather_rows_f32 (ONNX Gather axis 0) with the [-rows, rows) index rule, refusing instead of panicking

claude

committed

8113282

simd_stride: module doc names lut9's second input as the DR512 presence/polarity plane, not a sign plane

claude

committed

3403345

simd_masking_ops tests: type aliases for the compare tables (clippy type_complexity)

claude

committed

fa2c290

examples: float_accum_probe, bit hashes of gemm_f32 and mean_axis for comparing CPU tiers

claude

committed

ec5bb38

examples/float_accum_probe: rustfmt

claude

committed

f89fa33

docs+tests: float accumulation order across CPU tiers (wishlist item 4)

claude

committed

e057579

numeric_util test: as_chunks instead of chunks_exact (clippy)

claude

committed

8ec8e46

fix: fixture arithmetic in u64 (no usize overflow on 32-bit); matmul_f32_batched validates inner shapes even for an empty batch

claude

committed

6187f02

Navigation Menu

Repository navigation

ONNX op wishlist: typed Where, pair compares, copy-free matmul_f32, gather_rows_f32, float accumulation order - #400

ONNX op wishlist: typed Where, pair compares, copy-free matmul_f32, gather_rows_f32, float accumulation order#400
AdaWorldAPI merged 11 commits into
masterfrom
ccr-45a41fb4-0j3i0l

Commits on Oct 30, 2026

simd: blend_f32, blend_f32_scalar, blend_i64, blend_i64_scalar (ONNX Where); one generic body for the blend family

simd: two-array compares to a mask (f32/i32/i64: eq, lt, le, gt, ge) and i64 scalar compares

hpc::amx_matmul: matmul_f32 reads inputs in place via strides (no packing, no result buffer); matmul_f32_batched

hpc::gather: gather_rows_f32 (ONNX Gather axis 0) with the [-rows, rows) index rule, refusing instead of panicking

simd_stride: module doc names lut9's second input as the DR512 presence/polarity plane, not a sign plane

simd_masking_ops tests: type aliases for the compare tables (clippy type_complexity)

examples: float_accum_probe, bit hashes of gemm_f32 and mean_axis for comparing CPU tiers

examples/float_accum_probe: rustfmt

docs+tests: float accumulation order across CPU tiers (wishlist item 4)

numeric_util test: as_chunks instead of chunks_exact (clippy)

fix: fixture arithmetic in u64 (no usize overflow on 32-bit); matmul_f32_batched validates inner shapes even for an empty batch