International Business Machines Corporation

Paul C Stabler

Clarence Rosser Ogilvie

Bruce G Hazelzet

Gordon Taylor Davis

Timothy Jay Dell

Mark William Kellogg

Llewellyn Bradley Marshall, Iv

Christopher P Miller

Memory card with signal processing element

A serial bus and connection to a device on a computer system through a system memory controller is provided on a memory card having a DSP and a memory bus controller to allow the DSP on the memory card to gain access to the system device without using the system memory bus. The serial bus is a two wire serial bus connecting the device to the DSP through the system memory controller. If more than one memory card is present with DSPs or more than one device is contending for access, the system memory controller or arbitrate the access of each memory card or contending device. In such case the serial bus will signal the system memory controller when it wants access to the particular device, and the system memory controller will act as arbitrator to grant or not grant access to the particular memory card or device requesting access. If access is granted the bus memory controller outputs the required control or command word on the serial bus followed by the address and the required data. This serial information is received by the system memory controller which packets it, and, upon completion, outputs the information rapidly on a parallel bus, e.g. a PCI bus to the device which needs the information.

Memory card utilizing two wire bus

A cache based processing system is provided with a loop detection circuit for detecting the entry into and termination of program loops and for enabling peripheral device access to the main memory after completion of the first pass through the loop and terminating access when the program leaves the loop.

System and method for controlling peripheral device memory access in a data processing system

A method and system for prefetching data from storage and storing the data in a cache memory for use by an executing program includes means for detecting when a program has entered a processing loop and has completed at least one pass through the processing loop. At the completion of one pass through the processing loop, determining the requirement for additional data and prefetching the required data. Monitoring the operation of the program to detect termination of loop processing and terminating the prefetch of data from storage until the detection of a subsequent program loop.

System and method for prefetching data

A computer system having interrupts synchronized to data storage by having an interrupt data signal (interrupt packet) follow the path of the data to an interrupt receiver, which interrupts the processor to execute an interrupt service routine. Rather than having a dedicated interrupt line from a peripheral device to a processor, the peripheral device sends the interrupt across a bus from the peripheral to the processing unit via an interrupt receiver. The system can include a processor; a memory circuit in circuit communication with the processor; a peripheral device in circuit communication with the memory circuit via a bus; and an interrupt circuit in circuit communication with the peripheral circuit via the bus and in circuit communication with the processor via an interrupt bus; and wherein the peripheral device transmits data to the memory circuit across the bus and then transmits a predetermined interrupt data signal to the interrupt circuit across the bus; and wherein the interrupt circuit asserts an interrupt signal onto the interrupt bus to interrupt the processor responsive to receiving the predetermined interrupt data signal from the bus, thereby assuring that the processor is interrupted after the data is transmitted to the memory circuit.

Computer system generating a processor interrupt in response to

A method and apparatus for parallel 'normalize-round-normalize' floating point arithmetic. The rounding-normalizer of the invention receives as an input an infinitely precise mantissa which is the result of a floating point operation. This infinitely precise result mantissa is broken into two fields, the Close Enough Bits, and the Picky Bits. These bits are selectively passed to four parallel data paths for taking close enough bits and picky bits and producing a correctly rounded mantissa. The paths are, respectively; (a) a 1X.XX . . . X data path for mantissas greater than or equal to 2 but less than 4, the data path right shifting the upper bits and adjusting the exponent by +1 or +2; (b) a 01.XX . . . X data path for mantissas greater than or equal to 1 but less than 2, the data path right shifting the upper bits and adjusting the exponents by +0 or +1; (c) a 00.1X . . . X data path for mantissas greater than or equal to 1/2 but less than 1, the data path left shifting bits and adjusting the exponent by -1 or +0; and (d) a 00.01X . . . X data path for mantissas less than 1/2, the data path left shifting the bits and adjusting the exponent by the shift amount. Each of the data paths produces a normalized and rounded mantissa for its subset of mantissas.

Growing community of inventors

South Burlington, VT, United States of America

Paul C Stabler