type Eight

bendtorch-core@0.1.0.0 · src/tensor/matmul.bend:485

import Base
import bendtorch-core@0.1.0.0/src/tensor/parallel.bend as Parallel

485# Products fill their cells as `Parallel.build` does, with one change: a
486# subtree of eight cells that lie below the count and in one output row runs
487# as one loop, `row_run`, with eight accumulators. Term k reads the left
488# word of the row once and the eight right words of the cells, and adds each
489# product to its own cell's accumulator, so every cell still adds its terms
490# in increasing k from +0.0 and gets the word `product_dot` computes. The
491# eight chains of additions are independent, so a core runs them side by
492# side, where one cell's chain waits for each addition before the next.
493type Eight is Data:
494  Eight{w0: F32, w1: F32, w2: F32, w3: F32, w4: F32, w5: F32, w6: F32, w7: F32}