Altera Corporation

Intel Corporation

Tomasz Czajkowski

Davor Capalija

Andrew Ling

John Stuart Freeman

Andrei Mihai Hagiescu Miriste

Martin Langhammer

Eriko Nurvitadhi

Deshanand Pratap Singh

Deborah T Marr

David Neto

Dmitry Nikolai Denisenko

Peter Yiannacouras

Tian Yi David Han

Jack Z Yinger

Matrix multiplication systolic array feed methods and related processing element (PE) microarchitectures for efficiently implementing systolic array generic matrix multiplier (SGEMM) in integrated circuits is provided. A systolic array architecture may include a processing element array, a column feeder array, and a row feeder array. A bandwidth of external memory may be reduced by a factor of reduction based on interleaving of the matrix data via a feeding pattern of the column feeder array and the row feeder array.

Memory-size- and bandwidth-efficient method for feeding systolic array matrix multipliers

Power dissipation in integrated circuits may be reduced by efficient implementation of high level programming on the integrated circuits. As the high level programming logic is implemented on the integrated circuits, data inputs are disabled based upon branches and/or data that is not used by the high level programming.

Toggle rate reduction in high level programming implementations

A method of compiling program code includes determining if the program code controls a programmable logic device to execute other program code. The program code is a parallel program having a barrier function call for a group of threads. If it is determined that program code is to control the programmable logic device, then the program code is transformed by replacing the barrier function call with control logic inserted into the program code such that the transformed program code remains a parallel program and maintains synchronization among the group of threads. A compiler system that compiles program code with a barrier function call for a group of threads is also described.

M/A for compiling parallel program having barrier synchronization for programmable hardware

Systems and methods for calculating a dot product using digital signal processing units that are organized into a dot product processing unit for dot product processing using multipliers and adders of the digital signal processing units.

Dot product based processing elements

Systems and methods for using single-precision floating-point operation digital signal processing (DSP) blocks in conjunction to perform double-precision floating-point operations.

Double-precision floating-point operation

An integrated circuit may be provided with a specialized processing block that performs floating-point addition and subtraction operations. For this purpose, the specialized processing block includes a fused adder and subtractor stage with an adder circuit and a subtractor circuit. The adder and subtractor circuits share an alignment stage for aligning the mantissas of incoming floating-point numbers and provide a simplified normalization stage with one right shifter and one left shifter. The specialized processing blocks may be arranged in rows or columns such that an input of a first specialized processing block is directly coupled to an output of a second specialized processing block and an input of the second specialized processing block is directly coupled to an output of the first specialized processing block.

Fused floating-point arithmetic circuitry

Systems and methods for enhancing fixed-point operations, floating-point operations, or a combination thereof for programs implemented on an integrated circuit (IC) are provided. Portions of these operations may be shared among the operations. Accordingly, the embodiments described herein enhance these fixed-point operations, floating-point operations, or a combination thereof based upon these portions of the operations that may be shared.

Fixed-point and floating-point optimization

An integrated circuit is provided that performs floating-point addition or subtraction operations involving at least three floating-point numbers. The floating-point numbers are pre-processed by dynamically extending the number of mantissa bits, determining the floating-point number with the biggest exponent, and shifting the mantissa of the other floating-point numbers to the right. Each extended mantissa has at least twice the number of bits of the mantissa entering the floating-point operation. The exact bit extension is dependent on the number of floating-point numbers to be added. The mantissas of all floating-point numbers with an exponent smaller than the biggest exponent are shifted to the right. The number of right shift bits is dependent on the difference between the biggest exponent and the respective floating-point exponent.

Floating-point adder circuitry

Circuitry for dynamically ordering the execution of multiple threads in parallel is presented. The circuitry may include a control circuit that controls the execution of multiple subsets of threads using multiple processing units in parallel. Each of the plurality of processing units may be associated with an adjustable order thread issuer that may receive a subset of threads and an order in which to execute the subset of threads from the control circuit. The adjustable order thread issuer may manage the processing unit by providing each thread from the subset of threads for execution to the processing unit in the specified order. The adjustable order thread issuer may adjust the order in which threads are issued in an effort to optimize shared resource usage and thus improve the performance of a multithreaded application.

Repartitioning and reordering of multiple threads into subsets based on possible access conflict, for sequential access to groups of memory banks in a shared memory

Systems and methods for resource sharing of pipelined circuitry of an integrated circuit (IC) are provided. For example, in one embodiment, a method for sharing a functional unit of an integrated circuit (IC) includes receiving two or more threads configured to access the functional unit through two or more data entry points associated with corresponding data exit points configured to receive processed thread data. The method further includes arbitrating the processing of the two or more threads by the functional unit to obtain the processed thread data. To arbitrate, the exit points that cannot receive additional data are determined. Threads are only received from data entry points with corresponding data exit points that can receive additional data. The processed output data is provided to a corresponding exit point.

Multi-cycle resource sharing

Systems and methods for enhancing performance of programs implemented on an integrated circuit (IC) are provided. A forward-flow selector may determine a common branch for adding a data set to and removing a data set from. By selecting a common branch for adding and removing a data set, there will be a pipeline stage for data flowing into the branch. Accordingly, the embodiments described herein enhance throughput by increasing the number of datasets that may enter a branched pipeline without stalling.

Forward-flow selection

An integrated circuit is provided that performs floating-point operations involving at least two successive computational steps. Two floating-point numbers entering any additional computational step after the first computational step are aligned dynamically by shifting the mantissa of the floating-point number with the greater exponent to the left and the mantissa of the floating-point number with the smaller exponent to the right. The number of left shift bits is dependent on the magnitude of the difference between the two floating-point exponents and the number of leading zeroes in the mantissa with the greater exponent. The number of right shift bits is dependent on the magnitude of the difference between the two floating-point exponents and the number of left shift bits.

Methods and apparatus for performing dynamic data alignment for floating-point operations

Systems and methods for limiting resource usage of a kernel of an integrated circuit are provided. For example, in one embodiment a method for limiting a number of workgroups that may simultaneously access a kernel of an integrated circuit (IC) includes determining a threshold number of workgroups that may access the kernel simultaneously. A thread of execution is received. The thread of execution is allowed to access the kernel when the threshold number of workgroups would not be exceeded by the thread of execution accessing the kernel.

Growing community of inventors

Toronto, Canada

Tomasz Czajkowski