Figure 3

Update rule for matrix H. Matrix V is blockwise transferred, while W and H are fully loaded into the GPU memory at algorithm start. Nevertheless, both V and H are processed in portions of size n × b _m and k × b _m , respectively (b _m  ≤ m). Circled operations denote CUDA Kernels. Symbols “ .* ” and “ ./ ” denote pointwise matrix operations. Updated columns from H are marked with a big down arrow. Finally, the squared region at bottom left represents the reduction and accumulation of updated columns into a single k-length vector required by the next update-W rule.