For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo trait
AOperandLoader
One BK-strip A contribution for the shared Apple GEMM body.
accumulate_strip is the single seam between the plain-GEMM and fused-conv
kernels: for strip index k_strip, it loads this strip's A side (slab tile
vs online im2col gather) and runs the MMA into accum. Both impls consume
the same pre-tiled b_sub and write the same accum; only the A side
differs. Any per-simdgroup slab is built once by the caller before the
K-loop and held in the impl (see DenseALoader), so accumulate_strip only
adds this strip's K offset (k_strip * BK).
The loader is parametrized on the GEMM in_type, so the seam method names
the concrete _BodyMma[Self.in_type] / its AccumType. The shared body
infers the loader's CONCRETE type (infer-only L) and keys its own
mma_op/accum/B on L.in_type so both sides of the erased dispatch agree
(KB patterns/trait-type-erasure-and-stride-layout-workaround); at the
run/run_conv call sites L.in_type == Self.in_type is proven, so there
is no runtime cost.
Implemented traits
comptime members
b_type
comptime b_type
in_type
comptime in_type
num_m_mmas
comptime num_m_mmas
num_n_mmas
comptime num_n_mmas
Required methods
accumulate_strip
def accumulate_strip[bounded: Bool, b_layout: TensorLayout](mut self, mut mma_op: MmaOpApple[DType.float32, Self.in_type, _Self.num_m_mmas, _Self.num_n_mmas, b_type=Self.b_type], mut accum: Array[SIMD[DType.float32, SIMDLength(8)], MmaOpApple[DType.float32, Self.in_type, _Self.num_m_mmas, _Self.num_n_mmas, b_type=Self.b_type].num_accum], b_sub: TileTensor[Self.b_type, b_layout, ImmutAnyOrigin], conv: ConvIm2colParams, k_strip: Int32, *, a_valid_rows: Int32, b_valid_cols: Int, k_valid: Int)