Arm Limited

International Business Machines Corporation

Advanced Micro Devices Corporation

David Terrence Matheny

David Vivian Jaggar

Christopher Neal Hinds

David James Seal

Michael P Taborn

Gerard Richard Williams

David Walter Flynn

Paul K Miller

Wade A Walker

Matthew Paul Elwood

Steven Michael Burchfiel

Tan Ba Tran

A data processor corecomprising: a memory access interface portionoperable to perform data transfer operations between an external data source and at least one memoryassociated with said data processor core; a data processing portionoperable to perform data processing operations; a read/write portoperable to transfer data from said processor core to at least two busesA,B said at least two buses being operable to provide data communication between said processor coreand said at least one memory, said at least one memorycomprising at least two portionsA,B, each of said at least two busesA,B being operable to provide data access to respective ones of said at least two portionsA,B; arbitration logicassociated with said read/write port; wherein said arbitration logic is operable to route a data access request requesting access of data in one portion of said at least one memory received from said memory access interface to one of said at least two buses providing access to said one portion of said at least one memory and to route a further data access request requesting access of data in a further portion of said at least one memory received from said data processing portion to a further one of said at least two buses providing access to said further portion of said at least one memory, said routing of said data access requests being performed during the same clock cycle.

Arbitration of data transfer requests

Data processing apparatus and method for performing multiply-accumulate operations

A floating point unit is provided with a register bank comprising 32 registers that may be used as either vector registers of scalar registers. A data processing instruction includes at least one register specifying field pointing to a register containing a data value to be used in that operation. An increase in the instruction bit space available to encode more opcodes or to allow for more registers is provided by encoding whether a register is to be treated as a vector or a scalar within the register field itself. Further, the register field for one register of the instruction may encode whether another register is a vector or a scalar. The registers can be initially accessed using the values within the register fields of the instruction independently of the opcode allowing for easier decode.

Apparatus and method for processing data having a mixed vector/scalar register file

A data processing apparatus and method is provided, the apparatus comprising an execution unit having a plurality of pipelined stages for executing instructions, such that a maximum of 'n' instructions can be being executed simultaneously within the execution unit. Further, a set of at least 'n' logical exception registers are provided, each exception register being capable of storing a number of exception attributes associated with an instruction for which an exception has been detected during execution by the execution unit. In the event of an exception being detected during execution of a first instruction, the execution unit is arranged to: (i) store in a first of said exception registers said exception attributes associated with said first instruction; and (ii) to continue executing any remaining instructions already in the pipelined stages at the time the exception was detected. The execution unit is further arranged to store in said exception registers the exception attributes associated with any of the remaining instructions for which an exception is detected during execution, whereby the exception attributes stored in the exception registers can be provided to an exception processing tool for use in recovering from any exceptions occurring during processing of said first instruction and said remaining instructions. By this approach, when the exception processing tool is invoked, then it can deal with any exceptions arising from the instructions executed by the pipeline, and the data processing apparatus can then continue with the next instruction, without the need to re-execute any of the instructions that were in the pipeline at the time the first exception was detected.

Handling exceptions in a pipelined data processing apparatus

A floating point unit having a register bank containing a plurality of registers supports vector operations that execute a specified operation a plurality of times upon a sequence of data values form different registers. The register bank is divided into subsets and with the sequence of registers used in a vector operation wrapping within a subset. The subsets comprise disjoint, contiguous ranges of register numbers. The wrapping within ranges allows compact code and efficient to be provided for performing DSP operations, such as FIR filtering and matrix transformations.

Recirculating register file

A floating point unit is described that performs addition operations. An adder 16 within the floating point unit receives a first input and a second input to generate a sum. This sum is subject to subsequent normalization by a normalizer 60 and rounding by an incrementer 64. If an operation is performed that is immediately followed by an addition operation using the result of the preceding operation, then the normalized but unrounded sum is fed back to the adder 16 together with an indication of its rounding requirement. This rounding requirement can be performed by the adder 16 in parallel with the execution of the following addition by using the carry-in bit of the adder 16 to apply any increment required to rounding of the preceding result.

Round increment in an adder circuit

A floating point unit 10 provides a multiply-accumulate operation to determine a result B+(A*C). The multiplier 20 takes several processing cycles to determine the product (A*C). Whilst the multiplier 20 and its subsequent carry-save-adder 26 operate, an aligned value B' of the addend B is generated by an alignment-shifter 34. The aligned-addend B' may only partially overlap with the product (A*C) to which it is to be added using an adder 44. Any high-order-portion HOP of the aligned-addend B' that does not overlap with the product (A*C) must be subsequently concatenated with the output of the adder 44 that sums the product (A*C) with the overlapping portion of the aligned-addend B'. If the sum performed by the adder 44 generates a carry then it is an incremented version IHOP of the high-order-portion that should be concatenated with the output of the adder 44. This incremented-high-order-portion is generated by the adder 44 during otherwise idle processing cycles present due to the multiplier 20 operating over multiple cycles.

Floating point multiply-accumulate unit

An iterative division and/or iterative square root circuit 20 uses quotient digits q.sub.j+1 within the calculation that are dependent upon the input divisor D or radicand A and current partial remainder or partial radicand P.sub.j for the cycle reached. As the input divisor D or radicand A is fixed throughout the calculation, the critical path through the iterative circuit may be speeded up by preselecting and storing a subset QC of quotient digit values using a primary quotient digit selecting circuit 18, 22 operating in dependence upon the divisor D or radicand A and independently of the partial remainder or partial radicand P.sub.j. Within the iterative circuit 20, the quotient digits q.sub.j+1 to be used for each cycle can then be selected from this subset QC by a secondary quotient digit selecting circuit 24 in dependence upon the partial remainder or partial radicand P.sub.j and independent of the divisor D or radicand A.

Division and/or square root calculating circuit

A superscalar microprocessor includes a reorder buffer to correctly handle dependency checking and multiple updates to the same destination. The reorder buffer stores instructions in program order, and retires instructions that have executed and the results obtained. When a instruction is retired, the results of the instruction are stored and the memory space in the reorder buffer is deallocated. The results of the retired instructions are stored to a register file via a retire bus. If the results of two or more retired instructions output to the same register in the register file, then only the newest instruction, the later instruction in the original program sequence, is stored to the program register. The register file has a plurality of write ports for the transfer of data via the retire bus. If two retired instructions output to the same register, then a write port is not utilized. The retire window is the number of instructions monitored for retirement. The present invention advantageously increases the size of the retire window. Accordingly, if two or more retired instructions output to the same register, an additional instruction from the retire window can be retired. The additional instruction utilizes the write port not used by the older of the instructions that output to the same register. Thereby, the reorder buffer is emptied at a faster rate and causes less instruction dispatch stalls.

Apparatus and method for retiring instructions in excess of the number

A system and method for denormalizing a floating point result is disclosed. Denormalized operands are capable of representing much smaller values than can be represented by a number normalized under the ANSI/IEEE standard 754-1985 that governs the representation of numbers in floating point notation to ensure uniformity among floating point notation users. The majority of results will be normalized operands and therefore the floating point unit pipeline is optimized to produce normalized results but contains wider exponent fields in order to represent values received as denormalized numbers. In order to return the result as a denormalized number with the smaller ANSI/IEEE exponent field, denormalization is accomplished by using the same pipeline resources by means of the floating point unit feedback path and uses one of the exponent equalizing alignment shifters and an incrementor in order to round the denormalized result. In this way, denormalized results can be provided without stopping the dispatching of instructions, without providing status bits in the register files and rename registers and without the hold signals often present in other floating point units to accomplish denormalization.

Denormalization system and method of operation

An apparatus for rounding an answer produced during the execution of an operation by a multiple stage execution pipeline includes first circuitry for detecting when the operation is iterative and accuracy bits associated with the answer to determine if a rounding calculation is required. When the first circuitry detects a rounding calculation is required, it sets a correction factor for the answer in accordance with a rounding mode and the detected accuracy. In a method practiced by the apparatus, the rounding an answer produced during the execution of an operation occurs through the steps of detecting, when the operation is iterative, accuracy bits associated with the answer to determine if a rounding calculation is required, and, when the rounding calculation is required, setting a correction factor for the answer in accordance with a rounding mode and the detected accuracy bits.

Growing community of inventors

Austin, TX, United States of America

David Terrence Matheny