Analog Devices Global

Analog Devices,inc.

Analog Devices International Unlimited Company

Analog Devices Global Hamilton

Michael G Perkins

John L Redford

Kaushal Sanghai

Andrew James Higham

Boris Lerner

Michael S Allen

System and techniques for creating a single-instruction multiple-data (SIMD) processor predicate based on stride length are described herein. When an instruction for a SIMD processor is received, and the instruction has a specified stride length, a predicate memory can be read to obtain a current predicate. A new predicate can be determined based on the stride length and the current predicate. The new predicate is written to the predicate memory. When an instance of the instruction is executed by the SIMD processor, the execution is performed on a subset of data loaded into the SIMD processor based on the new predicate read from the predicate memory.

Stride length predicate creation

In one particular example, this disclosure provides an efficient mechanism to determine the degree of parallelization possible for a loop in the presence of possible memory aliases that cannot be resolved at compile-time. Hardware instructions are provided that test memory addresses at run-time and set a mode or register that enables a single instance of a loop to run the maximum number of SIMD (Single Instruction, Multiple Data) lanes to run in parallel that obey the semantics of the original scalar loop. Other hardware features that extend applicability or performance of such instructions are enumerated.

Paralleizing loops in the presence of possible memory aliases

The present disclosure provides a processor, and associated method, for performing parallel processing within a register. An exemplary processor may include a processing element having a compute unit and a register file. The register file includes a register that is divisible into lanes for parallel processing. The processor may further include a mask register and a predicate register. The mask register and the predicate register respective include a number of mask bits and predicate bits equal to a maximum number of divisible lanes of the register. A state of the mask bits and predicate bits is set to respectively achieve enabling/disabling of the lanes from executing an instruction and conditional performance of an operation defined by the instruction. Further, the processor is operable to perform a reduction operation across the lanes of the processing element and/or generate an address for each of the lanes of the processing element.

Processor architecture and method for simplifying programming single instruction, multiple data within a register

In an example, a system and method are provided for predicting in which way a requested memory address is most likely to be held in a multi-way cache, based on the last way accessed by the specified address register if available. If not available, then the system may determine that no best prediction is available. In that case, each way is read, and the superfluous values are disregarded, or a cache fill is performed as necessary. In certain embodiments, only a portion of the least significant bits of an add operation are used for way prediction in base-plus-offset addressing modes. This enables the decision to be made before the full-width add is complete, so that the clock cycle length is not unnecessarily lengthened by the prediction operation.

Cache way prediction

According to an example embodiment, a processor such as a digital signal processor (DSP), is provided with a register acting as a predicate counter. The predicate counter may include more than two useful values, and in addition to acting as a condition for executing an instruction, may also keep track of nesting levels within a loop or conditional branch. In some cases, the predicate counter may be configured to operate in single-instruction, multiple data (SIMD) mode, or SIMD-within-a-register (SWAR) mode.

Predicate counter

The present disclosure provides a memory interconnection architecture for a processor, such as a vector processor, that performs parallel operations. An example processor may include a compute array that includes processing elements; a memory that includes memory banks; and a memory interconnect network architecture that interconnects the compute array to the memory. In an example, the memory interconnect network architecture includes a switch-based interconnect network and a non-switch based interconnect network. The processor is configured to synchronously load a first data operand to each of the processing elements via the switch-based interconnect network and a second data operand to each of the processing elements via the non-switch-based interconnect network.

Memory interconnect network architecture for vector processor

Loop instructions are analyzed and assigned stage numbers based on dependencies between them and machine resources available. The loop instructions are selectively executed based on their stage numbers, thereby eliminating the need for explicit loop set-up and tear-down instructions. On a Single Instruction, Multiple Data machine, the final instance of each instruction may be executed on a subset of the processing elements or vector elements, dependent on the number of iterations of the original loop.

Growing community of inventors

Edinburgh, United Kingdom

Michael G Perkins