Scalable and Parameterized VLSI Architecture for Compressive Sensing Sparse Approximation

An integrated circuit architecture for compressive sensing (CS) reconstruction comprising:

• a vector computation core comprising a plurality of processing elements inter-connected in parallel that can be configured to perform vector operations;
• a scalar computation core configured to perform scalar operations;
• a shift register logic configured to provide optional latency in feedback paths associated with the plurality of processing elements and in a feedback path between the vector and scalar computation cores;
• a plurality of data-path memories each comprising a memory controller and configured to connect the vector and scalar computation cores, where the plurality of data-path memories are configured to provide data-flow control using the plurality of memory controllers and the plurality of data-path memories comprises:
• at least one data-path memory configured to receive and store CS sampled input data;
• at least one data-path memory configured to store an active set of values;
• at least one data-path memory configured to store a residual error estimate;
• and
• at least one data-path memory configured to store an estimated output data signal;
• a global control unit configured to synchronize and coordinate the plurality of memory controllers to jointly perform a selective set of vector and scalar operations using:
• a first atom searching configuration of the vector and scalar computation cores, and the plurality of data-path memories in which a single atom is selected by maximizing its correlation coefficients with the residual error estimate and added to the active set of values associated with the estimated output data signal;
• a second least squares solving configuration in which the shift register logic, the vector and scalar computation cores, and the plurality of data-path memories are configured to solve a series of linear equations relating the output data signal to the CS sampled input data to determine updates for the estimated output data signal that minimizes an error measure determined using the updated estimated output data signal and the CS sampled input data;
• and
• a third estimation update configuration of the vector and scalar computation cores, and the plurality of data-path memories, in which an output data signal estimate is updated based upon the solution to the series of linear equations and the residual error estimate;
• wherein the global control unit is configured to:
• set the residual error estimate to an initial value equal to the CS sampled input data;
• set the local controllers of the plurality of data-path memories and the vector and scalar computation cores to their initial states;
• and
• iterate between the first, second and third configurations until the updated residual estimate falls below a predetermined threshold.