Systems and methods for systolic array design from a high-level program
Systems and methods for automated systolic array design from a high-level program are disclosed. One implementation of a systolic array design supporting a convolutional neural network includes a two-dimensional array of reconfigurable processing elements arranged in rows and columns. Each processing element has an associated SIMD vector and is connected through a local connection to at least one other processing element. An input feature map buffer having a double buffer is configured to store input feature maps, and an interconnect system is configured to pass data to neighboring processing elements in accordance with a processing element scheduler. A CNN computation is mapped onto the two-dimensional array of reconfigurable processing elements using an automated system configured to determine suitable reconfigurable processing element parameters.
1. A method for defining a computational system, comprising the steps of:
determining a feasible mapping of a computation onto a systolic array including rows and columns of processing elements, each processing element having an associated single instruction multiple data (SIMD) vector;
selecting a shape of a systolic array; and
determining a data tiling size using a loop tiling representation,
wherein the loop tiling representation includes a plurality of outer loops, a plurality of middle loops, and a plurality of inner loops;
wherein the outer loops are configured to computer a plurality of image blocks associated with the computational system; and
wherein the middle loops are configured to represent a sequential processing of feeding data from a plurality of input buffers associated with the computational system to the processing elements.
2. The method of claim 1 , wherein the inner loops are configured to represent parallelism in the systolic array.
3. The method of claim 1 , wherein bounds of the middle loops are defined according to a size of a data reuse buffer of the plurality of input buffers.
4. The method of claim 2 , wherein the inner loops are performed by a parallel digital signal processing (DSP) unit in the systolic array.
5. The method of claim 4 , wherein an architectural shape of the systolic array is defined according to bounds of the inner loops.