Altera Corporation

Intel Corporation

Meta System S.p.a.

Carl Ebeling

Dana How

Herman Henry Schmit

David Lewis

Mahesh A Iyer

David Wolk Mendel

Thiam Khean Hah

Vamsi Nalluri

Audrey Catherine Kertesz

Ramanand Venkata

Scott J Weber

Frederic Reblewski

Olivier Lepape

Jeffrey Christopher Chromczak

Jean Barbier

Saurabh Adya

Richard Grenier

Simon Peter Finn

Vadim Gutnik

Pohrong Rita Chu

Yeong Tat Liew

Jason Gee Hock Ong

An integrated circuit that includes common factor mass multiplier (CFMM) circuitry is provided that multiplies a common factor operand by a large number of multiplier operands. The CFMM circuitry may be implemented as a instance specific version (where at least some portion of the hardware has to be redesigned if the multipliers change) or a non-instance specific version (where the CFMM circuitry can work with arbitrary multipliers without having to redesign the hardware). Either version can be formed on a programmable integrated circuit or an application-specific integrated circuit. The CFMM circuitry may include a multiplier circuit that effectively multiplies the common factor by predetermined fixed constants to generate partial products and may further include shifting and add/subtract circuits for processing and combining the partial products to generate corresponding final output products. CFMM circuitry configured in this way can be used to support convolution neural networks or any operation that requires a straight common factor multiply.

Common factor mass multiplication circuitry

An integrated circuit that includes common factor mass multiplier (CFMM) circuitry is provided that multiplies a common factor operand by a large number of multiplier operands. The CFMM circuitry may be implemented as an instance specific version or a non-instance specific version. The instance specific version might also be fully enumerated so that the hardware doesn't have to be redesigned assuming all possible unique multiplier values are implemented. Either version can be formed on a programmable integrated circuit or an application-specific integrated circuit. CFMM circuitry configured in this way can be used to support convolution neural networks or any operation that requires a straight common factor multiply. Any adder component with the CFMM circuitry may be implemented using bit-serial adders. The bit-serial adders may be further connected in a tree in CNN applications to sum together many input streams.

An integrated circuit includes a signal network and a phase detector circuit. The signal network includes an adjustable delay circuit. The adjustable delay circuit is coupled at an intersection in the signal network between branches of the signal network. The signal network generates a first signal at a first leaf node of the signal network in response to a second signal. The signal network generates a third signal at a second leaf node of the signal network in response to the second signal. The phase detector circuit compares phases of the first and third signals to generate a phase detection signal. The adjustable delay circuit adjusts a delay provided to the first signal relative to the second signal to reduce a skew between the first and third signals based on the phase detection signal indicating that the first and third signals have the skew.

Techniques for signal skew compensation

A method for designing a system on a target device includes modifying a circuit to enable the circuit to support a plurality of threads at an instant of time. An interface is generated that enables one or more of the plurality of threads to be swapped out of an execution queue for accessing the circuit and that enables one or more other threads to be swapped into the execution queue for accessing the circuit, wherein at least one of the modifying and the generating is performed by a processor.

Method and apparatus for supporting temporal virtualization on a target device

An integrated circuit may have pipelined interconnects that are configurable to operate in registered single data rate mode, registered double data rate mode, or in combinational mode. The pipelined interconnect may include routing multiplexers for selecting incoming signals, circuitry for serialization and de-serialization, and memory elements that are configurable to store one or two signals per clock period. Operating the pipeline interconnects in double data rate mode may provide a trade-off between reducing the number of physical wires that are required to implement a design at a constant bandwidth or increasing the bandwidth while keeping the number of physical wires constant.

Pipelined interconnect circuitry with double data rate interconnections

An integrated circuit may have configurable storage blocks. A configurable storage block may include a memory array, and arithmetic and control circuitry. The arithmetic and control circuitry may be used to determine whether to operate the configurable storage block in a first mode which may provide random access to the memory array or in a second mode which may provide access to the memory array in a predefined order. Thus, the configurable storage block may implement simple first-in first-out modules and shift registers in addition to implementing memory modules with random access. Arithmetic and control circuitry may include a multiplexer that determines whether the configurable storage block is implementing simple first-in first-out modules or shift registers. When the configurable storage block implements first-in first-out modules, an up-down counter may be activated to generate a count value received at the multiplexer.

Configurable storage blocks having simple first-in first-out enabling circuitry

Systems and methods for phase detection are disclosed. Phase alignment between first and second clock signals is detected using a comparison of outputs from a collapsible pipeline and a non-collapsible pipeline.

Systems and methods for clock alignment using pipeline stages

A clock-tree construction method for a configurable clock grid structure having a plurality of sectors and a plurality of wire segments includes defining a clock region within the clock grid structure and constructing an H-tree that has a smallest size to cover the clock region. The method further includes aligning the clock region within the H-tree, pruning the H-tree and removing an unused segment from the H-tree. The method further includes performing a tree height reduction procedure to the pruned H-tree, and generating a clock tree with a reduced size or a reduced height from the tree height reduction procedure.

Sector-based clock routing methods and apparatus

An integrated circuit may have configurable storage blocks. A configurable storage block may include a memory array, an arithmetic circuit, and a control circuit. The control circuit may be used to determine whether to operate the configurable storage block in a first mode which may provide random access to the memory array or in a second mode which may provide access to the memory array in a predefined order. Thus, the configurable storage block may implement first-in first-out modules, shift registers, or delay-line modules in addition to implementing memory modules with random access.

Configurable storage blocks with embedded first-in first-out and delay line circuitry

Circuitry accepts an input signal and distributes the input signal to a plurality of locations within the circuitry. The circuitry includes a first circuit element and a second circuit element. The circuitry further includes a first plurality of wire segments that are substantially aligned to form a first bundle, and include a first wire segment. The circuitry further includes a second plurality of wire segments that are substantially aligned to form a second bundle, and have a second wire segment. An intersection element of the first bundle and the second bundle includes a first interconnecting wire segment that connects the first wire segment and the second wire segment, and the input signal is routed from the first wire segment to the second wire segment via the first interconnecting wire segment. The input signal is further transmitted to the second element from the second wire segment.

Configurable clock grid structures

Enhanced memory circuits are described that maintain coherency between concurrent memory reads and writes in a pipelined memory architecture. The described memory circuits can maintain data coherency regardless of the amount of pipelining applied to the memory inputs and/or outputs. Moreover, these memory circuits may be implemented as dedicated hard circuits in a field programmable gate array (FPGA) or other programmable logic device (PLD), and can be supplemented with user-configurable logic to achieve coherency in a variety of applications.

Systems and methods for maintaining memory access coherency in embedded memory blocks

Systems and methods for phase detection are disclosed. A collapsible three-stage pipeline includes a first register in a first stage having a first clock signal having first clock edges, a second register in a second stage that receives a first signal from the first stage, and having a second clock signal having second clock edges, and a third register in a third stage that receives a second signal from the second stage, and having a third clock signal having third clock edges, wherein each second clock edge has a corresponding first clock edge and a corresponding third clock edge. The circuitry may further include a two-stage pipeline including fourth and fifth stages, a counter that provides an input signal into the collapsible three-stage pipeline and the two-stage pipeline, and a comparator that compares a first output of the collapsible three-stage pipeline and a second output of the two-stage pipeline.

An integrated circuit may have circuitry that includes a storage circuit, a processing circuit, and at least one state register to implement a finite-state machine. The storage circuit may store base addresses and output data for each state of the finite-state machine. The storage circuit may further store offset values that are based on the input data to the finite-state machine and the state transition from a current state to a next state caused by the input data. The processing circuit may compute the address of the storage circuit location where the output data of the next state is stored. The computation of this address may depend on the offset value and base address of the current state. The state register may receive the address from the processing circuit, store the address, and perform the corresponding memory access operation on the storage circuit.

Growing community of inventors

Redwood City, CA, United States of America

Carl Ebeling