Figure 7
Architecture of GPU-accelerated chunk computation for large-scale multi-parameter problems. In this figure, we assume M = 2 and denote the two scattering vectors by qx and qy. Chunking is performed along these two dimensions. All the model-parameter dimensions are conceptually represented by the p dimension. Left: given a SAS model and its parameter space, we compute G in chunks on a GPU and store it on disk if needed. Right: to compute any term in , Js or Hs that requires successive inner products with G, we chunk it along the q dimensions and load the corresponding chunk of G on a GPU to perform the inner products; the assembled results are then fed to the trust-region method to update the variables. |