Abstract: Systems, processors and methods are disclosed for organizing processing datapaths to perform operations in parallel while executing a single program. Each datapath executes the same sequence of instructions, using a novel instruction sequencing method. Each datapath is implemented through a processor having a data memory partitioned into identical regions. A master processor fetches instructions and conveys them to the datapath processors. All processors are connected serially by an instruction pipeline, such that instructions are executed in parallel datapaths, with execution in each datapath offset in time by one clock cycle from execution in adjacent datapaths. The system includes an interconnection network that enables full sharing of data in both horizontal and vertical dimensions, with the effect of coupling any datapath to the memory of any other datapath without adding processing cycles in common usage.

Abstract: Methods and systems are disclosed for increased speed of processing operations on data in two-dimensional arrays, and for detecting a feature in an image. A method for detecting a feature in an image comprises storing, in a set of data memories within a parallel processing system, first image data representing pixels of a first image. The method further comprises storing, in a memory of a host processor coupled to the parallel processing system, feature kernel data representing a set of weight matrices. A method for increased speed of processing operations on data in two-dimensional arrays comprises storing, in a set of data memories within a parallel processing system, first array data representing elements of a first array. The method further comprises, for each of multiple selected elements within the first array, performing a processing operation on the selected element to produce an output element corresponding to the selected element.

Abstract: Systems, processors and methods are disclosed for organizing processing datapaths to perform operations in parallel while executing a single program. Each datapath executes the same sequence of instructions, using a novel instruction sequencing method. Each datapath is implemented through a processor having a data memory partitioned into identical regions. A master processor fetches instructions and conveys them to the datapath processors. All processors are connected serially by an instruction pipeline, such that instructions are executed in parallel datapaths, with execution in each datapath offset in time by one clock cycle from execution in adjacent datapaths. The system includes an interconnection network that enables full sharing of data in both horizontal and vertical dimensions, with the effect of coupling any datapath to the memory of any other datapath without adding processing cycles in common usage.

Abstract: Systems, processors and methods are disclosed for organizing processing datapaths to perform operations in parallel while executing a single program. Each datapath executes the same sequence of instructions, using a novel instruction sequencing method. Each datapath is implemented through a processor having a data memory partitioned into identical regions. A master processor fetches instructions and conveys them to the datapath processors. All processors are connected serially by an instruction pipeline, such that instructions are executed in parallel datapaths, with execution in each datapath offset in time by one clock cycle from execution in adjacent datapaths. The system includes an interconnection network that enables full sharing of data in both horizontal and vertical dimensions, with the effect of coupling any datapath to the memory of any other datapath without adding processing cycles in common usage.

Abstract: Methods, devices and systems are disclosed that interface a host computer to a specialized processor. In an embodiment, an instruction generation unit comprises attribute, decode, and instruction buffer stages. The attribute stage is configured to receive a host-program operation code and a virtual host-program operand from the host computer and to expand the virtual host-program operand into an operand descriptor. The decode stage is configured to receive the first operand descriptor and the host-program operation code, convert the host-program operation code to one or more decoded instructions for execution by the specialized processor, and allocate storage locations for use by the specialized processor. The instruction buffer stage is configured to receive the decoded instruction, place the one or more decoded instructions into one or more instruction queues, and issue decoded instructions from at least one of the one or more instruction queues for execution by the specialized processor.

Abstract: Systems, processors and methods are disclosed for organizing processing datapaths to perform operations in parallel while executing a single program. Each datapath executes the same sequence of instructions, using a novel instruction sequencing method. Each datapath is implemented through a processor having a data memory partitioned into identical regions. A master processor fetches instructions and conveys them to the datapath processors. All processors are connected serially by an instruction pipeline, such that instructions are executed in parallel datapaths, with execution in each datapath offset in time by one clock cycle from execution in adjacent datapaths. The system includes an interconnection network that enables full sharing of data in both horizontal and vertical dimensions, with the effect of coupling any datapath to the memory of any other datapath without adding processing cycles in common usage.