BendMlTensorArray.gemm_any3: def of bend-ml-tensor-array@0.1.19.0
Import: import bend-ml-tensor-array@0.1.19.0/main.bend as BendMlTensorArray
Lines 1152-1162 of bend-ml-tensor-array@0.1.19.0/main.bend: https://jend.leeeet.dev/read.txt?path=bend-ml-tensor-array@0.1.19.0/main.bend&start_line=1152&end_line=1162

import Base

# one row (matrix · vector): the scalar loop at every depth; two or more rows: cells of 8. Up to 0.1.11.0 one row in
# parallel ran gemm_par_cols, which clones both operands at every node of the fork tree: a 50,257 x 768 weight matrix at
# depth 4 ran out of the default heap (5.1 GB) and 8192 x 2048 took 124 ms against 10 ms sequentially. Packing the
# matrix into cells of 8 instead costs more than the product, which reads each weight once (64 ms against 24 at
# 50,257 x 768). Matrix · vector in parallel is Bands.matvec. Measured 2026-10-16 (NOTES.md).
def gemm_any3(+ac: Bool, one: Bool, d: Nat, +n: Nat, +m: Nat, +k: Nat, +arow: U32, +sa: U32, +sb: U32, +bcol: U32, a: Array<F32>, b: Array<F32>, c: Array<F32>) -> G:
  match one:
    case True{}:
      gemm_ac(ac, n, m, k, arow, sa, sb, bcol, a, b, c)
    case False{}:
      gemm_t8_ac(ac, d, n, m, k, arow, sa, sb, bcol, a, b, c)

Uses (8):
  https://jend.leeeet.dev/definition.txt?path=Base/base.bend&name=Array
  https://jend.leeeet.dev/definition.txt?path=Base/base.bend&name=Bool
  https://jend.leeeet.dev/definition.txt?path=Base/base.bend&name=F32
  https://jend.leeeet.dev/definition.txt?path=Base/base.bend&name=Nat
  https://jend.leeeet.dev/definition.txt?path=Base/base.bend&name=U32
  https://jend.leeeet.dev/definition.txt?path=bend-ml-tensor-array@0.1.19.0/main.bend&name=G
  https://jend.leeeet.dev/definition.txt?path=bend-ml-tensor-array@0.1.19.0/main.bend&name=gemm_ac
  https://jend.leeeet.dev/definition.txt?path=bend-ml-tensor-array@0.1.19.0/main.bend&name=gemm_t8_ac