US 5,675,823 AGranted

Grain structured processing architecture device and a method for processing three dimensional volume element data

AssigneeGeneral Electric Company

InventorMeng-Ling Hsiao

Life of this patent

Estimating the term…

Loading patent…

Abstract

According to the present invention, a 3D connectivity-conserved grain-structured processing architecture uses connectable massively parallel processors. A 3D grain-structured processing architecture is provided. The 3D links of the grain-structured processing architecture provide direct local communication as well as global communication for voxel processing and analysis tasks. A novel transport naming scheme which is scalable in any three-dimension direction and the local/global communication protocol are disclosed. The 3D volumetric data set is first divided into a set of voxel sub-cubes. Each voxel subcube is processed by a processor element of the grain-structured processing architecture. Data discontinuity is produced after performing local operations in a massively parallel processors environment and requires replacing the resulted voxel data set in the overlap region for each voxel sub-cube. An inter-processor communication procedure for overcoming this problem in the 3D connected grain-structured processors architecture is also disclosed.

Claims4

Dependent claims

A method of processing a data set of voxels arranged in a three-dimensional cubic grid-like structure using a parallel computer comprising a plurality of processor elements constructed in a cube-like structure having a plurality of rows, columns, and planes, each processor element comprising a memory, a CPU, a float point unit, and six communication ports, each processor element connected via said communication ports to its nearest neighbor processor element, wherein said processor elements comprise internal, vertex, edge, and face processor elements and wherein said vertex, edge and face processor elements are in a cross-row, cross-column, and cross-plane communicable relationship with other vertex, edge and face processor elements, in said cross-row communicable relationship, a first processor element in a row is in direct communication with a last processor element in a previous row except for a first row, in said cross-column communicable relationship, a first processor element in a column is in direct communication with a last processor element in a previous column except for a first column, in said cross-plane communicable relationship, each of the processor elements in a first plane except processor elements in a first column of the first plane is in direct communication with a processor element from a previous column of the same row in a last plane, the method comprising the steps of:

decomposing said data set into a plurality of sub-cubes;
mapping said voxels within said sub-cubes into individual processor elements in said parallel computer;
performing local kernel operations on said voxels within said sub-cubes with said parallel computer;
mapping said sub-cubes into individual processor elements in said parallel computer;
padding each sub-cube with voxel data from an adjacent sub-cube into an overlap region in each of the x, y, and z directions in relation to said sub-cube;
performing local operations on said sub-cubes;
and
swapping data within said overlap regions for each of the sub-cubes in the x, y and z directions in a sequential manner.

5 more elements

A method of processing a data set of voxels arranged in a three dimensional cubic grid-like structure using a grain structured processing architecture device comprising a plurality of processor elements constructed in a grain structured architecture having a plurality of rows, columns, and planes, each processor element communicating in a cross-row, cross-column, and cross-plane communicable relationship with other processor elements, in said cross-row relationship, a first processor element in a row is in direct communication with a last processor element in a previous row except for a processor element in a first row, in said cross-column relationship, a first processor element in a column is in direct communication with a last processor element in a previous column except for a processor element in a first column, in said cross-plane communicable relationship, each of the processor elements in a first plane except processor elements in a first column of the first plane is in direct communication with a processor element from a previous column of the same row in a last plane, the method comprising the steps:

decomposing said data set into a plurality of sub-cubes;
mapping said voxels within said sub-cubes into individual processor elements in said grain structured processing architecture device;
performing local kernel operations on said voxels within said sub-cubes with said grain structured processing architecture device;
mapping said sub-cubes into individual processor elements in said grain structured processing architecture device;
padding each sub-cube with voxel data from an adjacent sub-cube into an overlap region in each of the x, y, and z directions in relation to said sub-cube;
performing local operations on said sub-cubes;
and
swapping data within said overlap regions for each of the sub-cubes in the x, y and z directions in a sequential manner.

5 more elements

Dependencies are inferred from the claim text.