C (m x n) = A B (Realization.Nvidia.SM86.TiledProductSM86), each operand
read where it is stored: A as rows (m x k) or as columns (A^T, k x m), B
as rows (B^T, n x k) or as columns (k x n). The model's products are
three patterns: the forward's X W^T (rows, rows), a data gradient's dY W
(rows, columns) and a weight gradient's dY^T X (columns, columns). Each
product's tile and stages are a candidate measured on the card and
recorded (Coppelius.Build.TiledChoices, scripts/alpha/tiled-choices.py)
among the candidates below.
510def coppeliusTiledRows : (family TiledProductLayout) = (constructor TiledProductLayout TiledProductRows)The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.