Skip to main content

Module outer_product

Module outer_product 

Source
Expand description

ADR-020 iter-11h-c2 — vector outer product forward + backward.

Forward: y[i, j] = lhs[i] · rhs[j] Backward: dlhs[i] = Σ_j dy[i, j] · rhs[j] drhs[j] = Σ_i dy[i, j] · lhs[i]

Distinct from matmul: matmul kernel has a 32-element floor on each dim (M, N, K ≥ 32 for dW backward); outer products have inner-dim = 1, falling below that floor.

Statics§

OUTER_PRODUCT_SHADER_SOURCE

Functions§

dispatch_outer_product_backward_lhs_f32
dispatch_outer_product_backward_rhs_f32
dispatch_outer_product_f32
register