Abstract: A neural processing device, a processing element included therein and a method for operating various formats of the neural processing device are provided. The neural processing device includes at least one neural processor, a shared memory shared by the at least one neural processor, and a global interconnection configured to transmit data between the at least one neural processor and the shared memory, wherein each of the at least one neural processor comprises at least one processing element, each of the at least one processing element receives an input in a first format and thereby performs an operation, and receives an input in a second format that is different from the first format and thereby performs an operation if a format conversion signal is received, and the first format and the second format have a same number of bits.

Abstract: An adder and a method for calculating 2n+x are provided, where x is a variable input expressed in a floating point format and n is an integer. The adder comprises: a first path configured to calculate 2n+x for x<0 and 2n?1?|x|<2n+1; a second path configured to calculate 2n+x for |x|<2n; a third path configured to calculate 2n+x for |x|?2n; and selection logic configured to cause the adder to output a result from one of the first, second, and third paths in dependence on the values of x and n.

Abstract: An information handling system may include a processor, one or more accelerators communicatively coupled to the processor, and a management controller communicatively coupled to the processor and the one or more accelerators and configured for out-of-band management of the information handling system, the management controller further configured to receive information regarding the one or more accelerators, determine a criticality factor for each of the one or more accelerators based on the information, determine an accelerator health status for each of the one or more accelerators, and determine an overall system health of the information handling system based on the criticality factors and the accelerator health statuses.

Abstract: A hardware module comprising at least one of: one or more field programmable gate arrays and one or more application specific integrated circuits configured to: receive a number in floating-point representation at a first precision level, the number comprising an exponent and a first mantissa; apply a first random number to the first mantissa to generate a first carry; truncate the first mantissa to a level specified by a second precision level; add the first carry to the least significant bit of the mantissa truncated to the level specified by the second precision level to form a mantissa for the number in floating-point representation at the second precision level.

Abstract: Systems, apparatuses, and methods utilizing an elastic floating-point encoding format are described. In particular, at least one operand of an instruction is to store, or stores, data in the elastic floating-point encoding format. In some implementations, the floating-point encoding format includes a sign bit, a self-identifying field, a mantissa, and a non-overlapping exponent range.

Abstract: A data processing apparatus includes difference circuitry that calculates a difference between exponents of a first floating-point operand and a second floating-point operand. Shift circuitry generates a fractional string by shifting fractional bits of a selected operand of the first floating-point operand and the second floating-point operand based on the difference. Logic circuitry generates an integer-bit string representing an integer-bit of the selected operand having been shifted based on the difference. Combining circuitry combines the fractional string and the integer-bit string to produce a significand string representing the selected operand having been shifted based on the difference. The logic circuitry generates the integer-bit string using operations other than shifting.

Abstract: A technique is provided for performing a mixed precision estimate. A processing circuit receives an input of a first precision having a wide precision value. The processing circuit computes an output in an output exponent range corresponding to a narrow precision value based on the input having the wide precision value.

Abstract: A computation processor outputs whether a carry-out is generated, by incrementing a result of computation by 1, during rounding of the result of the computation. The computation processor includes a computing unit that performs the computation; a shift amount calculating unit that calculates a shift amount of the result of the computation; a normalizing unit that performs normalization of the result of the computation, by using the shift amount; a predicting unit that, when the result of the computation is shifted by an amount equal to or more than a predetermined shift amount by using the shift amount, predicts whether each of bits in a predetermined region of a shift result is 1, in parallel with the normalization; and a detecting unit that detects a generation of the carry-out, by receiving a normalized result from the normalizing unit and a predicted result from the predicting unit.

Abstract: A round-far-reround mode (preferably in a BID encoded Decimal format) of a floating point instruction prepares a result for later rounding to a variable number of digits by detecting that the least significant digit may be a 0, and if so changing it to 1 when the trailing digits are not all 0. A subsequent reround instruction is then able to round the result to any number of digits at least 2 fewer than the number of digits of the result. An optional embodiment saves a tag indicating the fact that the low order digit of the result is 0 or 5 if the trailing bits are non-zero in a tag field rather than modify the result. Another optional embodiment also saves a half-way-and-above indicator when the trailing digits represent a decimal with a most significant digit having a value of 5. An optional subsequent rewound instruction is able to round the result to any number of digits fewer or equal, to the number of digits of the result using the saved tags.

Abstract: A floating point execution unit is capable of selectively repurposing a subset of the significand bits in a floating point value for use as additional exponent bits to dynamically provide an extended range for floating point calculations. A significand field of a floating point operand may be considered to include first and second portions, with the first portion capable of being concatenated with the second portion to represent the significand for a floating point value, or, to provide an extended range, being concatenated with the exponent field of the floating point operand to represent the exponent for a floating point value.

Abstract: A method, computer program product and a system for identifying decimal floating point addition operations that guarantee operand alignment and do not require alignment, normalization or rounding are provided. The method includes: receiving an instruction to perform an addition of a first operand and a second operand; extracting a first exponent (EXP) and a first most significant digit (MSD) from the first operand; extracting a second EXP and a second MSD from the second operand; and determining whether alignment between the first operand and the second operand is guaranteed, based on the first EXP, the first MSD, the second EXP and the second MSD.

Abstract: A method for controlling rounding modes in a single instruction multiple data (SIMD) floating-point unit is disclosed. The SIMD floating-point unit includes a floating-point status-and-control register (FPSCR) having a first rounding mode bit field and a second rounding mode bit field. The SIMD floating-point unit also includes means for generating a first slice and a second slice. During a floating-point operation, the SIMD floating-point unit concurrently performs a first rounding operation on the first slice and a second rounding operation on the second slice according to a bit in the first rounding mode bit field and a bit in the second rounding mode bit field within the FPSCR, respectively.

Abstract: Methods and arrangements to correct for double rounding errors when rounding floating point numbers to nearest away are described. Embodiments include transformations, code, state machines or other logic to perform a floating point operation on one or more floating point numbers of precision P1 in base b, producing positive result res0 of precision greater than precision P1; rounding positive result res0 to precision P1 to the nearest away, producing positive result res1; and rounding the result res1 to precision P2 to the nearest away, where P2 is narrower than P1, producing result res2. The embodiments may also include correcting res2 for double rounding errors. The correcting may include determining that res1 is midway between two consecutive floating point numbers of precision P2, the larger being res2, determining that rounding res0 to produce res1 involved rounding up, and decrementing the significand of res2 to obtain the corrected result.

Abstract: Methods and arrangements to correct for double rounding errors when rounding floating point numbers to nearest even are described. Embodiments include transformations, code, state machines or other logic to perform a floating point operation on one or more floating point numbers of precision P1 in base b, producing positive result res0 of precision greater than precision P1; rounding positive result res0 to precision P1 to the nearest even, producing positive result res1; and rounding the result res1 to precision P2 to the nearest even, where P2 is narrower than P1, producing result res2. The embodiments may also include correcting res2 for double rounding errors. The correcting may include determining that res1 is midway between two consecutive floating point numbers of precision P1, the larger (smaller) being res2, determining that rounding res0 to produce res1 involved rounding up (down), and decrementing (incrementing) the significand of res2 to obtain the corrected result res2?.

Abstract: A method and system to perform shifting and rounding operations within a microprocessor, such as, for example, a digital signal processor, during execution of a single instruction are described. An instruction to shift and round data within a source register unit of a register file structure is received within a processing unit. The instruction includes a shifting bit value indicating the bit amount for a right shift operation and is subsequently executed to shift data within the source register unit to the right by an encoded bit value, calculated by subtracting a single bit from the shifting bit value contained within the instruction. A predetermined bit extension is further inserted within the vacated bit positions adjacent to the shifted data. Subsequently, an addition operation is performed on the shifted data and a unitary integer value is added to the shifted data to obtain resulting data.

Abstract: An arithmetic program conversion apparatus, an arithmetic program conversion program and an arithmetic program conversion method that can convert the floating-point arithmetic of an arithmetic program into a fixed-point arithmetic without degrading the accuracy.

Abstract: An apparatus for controlling rounding modes in a single instruction multiple data (SIMD) floating-point unit is disclosed. The SIMD floating-point unit includes a floating-point status-and-control register (FPSCR) having a first rounding mode bit field and a second rounding mode bit field. The SIMD floating-point unit also includes means for generating a first slice and a second slice. During a floating-point operation, the SIMD floating-point unit concurrently performs a first rounding operation on the first slice and a second rounding operation on the second slice according to a bit in the first rounding mode bit field and a bit in the second rounding mode bit field within the FPSCR, respectively.

Abstract: One embodiment of the present invention provides a system that facilitates performing exception-free arithmetic operations within a computer system. During execution of a computer program, the system receives an instruction to perform an arithmetic operation that involves manipulating floating-point values. If the arithmetic operation manipulates a floating-point value representing {+0}, the arithmetic operation is performed in a manner consistent with {+0} representing a set containing a single value “?0”, wherein “?0” is the limit of a sequence of values that approaches zero only from above. Similarly, if the arithmetic operation manipulates a floating-point value representing {?0}, the arithmetic operation is performed in a manner consistent with {?0} representing a set containing a single value “+0”, wherein “+0” is the limit of a sequence of values that approaches zero only from below.

Abstract: One embodiment of the present invention provides a system that solves a global optimization problem specified by a function ƒ and a set of equality constraints q1(x)=0 (i=1, . . . , r), wherein ƒ is a scalar function of a vector x=(x1, x2, x3, . . . xn). During operation, the system receives a representation of the function ƒ and the set of equality constraints and stores the representation in a memory within a computer system. Next, the system and performs an interval global optimization process to compute guaranteed bounds on a globally minimum value of the function ƒ(x) subject to the set of equality constraints. Performing this interval global optimization process involves, applying term consistency to the set of equality constraints over a subbox X, and excluding portions of the subbox X that violate the set of equality constraints.

Abstract: The present invention provides for saving power in a floating point unit. Bypass logic is coupled to the input of the aligner and the multiplier. An aligner bypass is coupled to the output of the aligner and an output of the bypass logic. A multiplier bypass is coupled to the output of the multiplier and an output of the bypass logic. The aligner bypass and the multiplier bypass transmit the output of the aligner and multiplier, or the bypass logic, as a function of an aligner bypass signal and a multiplier bypass signal, respectively. An adder is coupled to the output of the aligner bypass and the multiplier bypass. Clock disable logic is used to selectively enable and disable at least portions of the aligner, multiplier and bypass logic. This is done based on the operation and on the value of the operands.

Abstract: The processing elements if a single instruction multiple data (SIMD) massively parallel processor (MPP) are provided with two register blocks. One register block includes logic for performing limited left shifting, while the other register block includes logic for performing limited right shifting. A method is disclosed for using the registers blocks with their associated logic to perform floating point significand alignment and normalization. The limited shifting logic occupies less die space than a full feature barrel shifter, thereby permitting a greater number of processing elements.

Abstract: A data calculating device preferably used to improve the calculation precision when fixed-point calculation is performed by block-floating-point system. Each piece of data of a data group is calculated, the minimum scale factor representative of the calculated data is detected as a group scale factor (GSF), and calculated data is subjected to scaling based on the detected GSFs. These processing are applied to each data group of a data block. The minimum GSF out of the detected GSFs is detected as a block scale factor (BSF). When calculation of the calculated data is performed again, the calculated data of the data group is subjected to scaling according to the GSFs and BSF before the calculation performed again.

Abstract: A floating-point unit of a computer includes a floating-point computation unit, floating-point registers and a floating-point status register. The floating-point status register may include a main status field and one or more alternate status fields. Each of the status fields contains flag and control information. Different floating-point operations may be associated with different status fields. Subfields of the floating-point status register may be updated dynamically during operation. The control bits of the alternate status fields may include a trap disable bit for deferring interruptions during speculative execution. A widest range exponent control bit in the status fields may be used to prevent interruptions when the exponent of an intermediate result is within the range of the register format but exceeds the range of the memory format. The floating-point data may be stored in big endian or little endian format.

Abstract: A method and apparatus that performs anticipatory rounding of intermediate results in a floating point arithmetic system while the intermediate results are being normalized is disclosed. One embodiment of the present invention includes four logic levels, implemented in N-NARY logic. In the first three logic levels, propagation information is gathered for preselected bit groups from the coarse and medium shift output of the normalizer as those results become available. In the fourth level, an incremented, normalized intermediate single-precision or double-precision mantissa result is produced by combining fine shift output bit values with propagation information for the appropriate top bit group, middle bit group, and bottom bit group. The appropriate bit groups are determined by examining the value of the fine shift select signal.

Abstract: A reconfigurable register file integrated in an instruction set architecture capable of extended precision operations, and also capable of parallel operation on lower precision data is described. A register file is composed of two separate files with each half containing half as many registers as the original. The halves are designated even or odd by virtue of the register addresses which they contain. Single width and double width operands are optimally supported without increasing the register file size and without increasing the number of register file ports. Separate extended registers are also employed to provide extended precision for operations such as multiply-accumulate operations.

Abstract: The present invention is a method and apparatus for rounding a result operand of a floating-point (FP) operation which causes an underflow. The FP operation is recomputed using a truncate rounding mode to generate an underflowed operand. The underflowed operand is denormalized and providing characteristic bits. A rounding bit is generated based on the characteristic bits. The rounding bit is merged with the denormalized operand to generate the rounded result operand.

Abstract: A floating point arithmetic unit provides consistent propagation of NaNs le performing high precision calculations on hardware designed to perform lower precision calculations. In one embodiment, the floating point arithmetic unit is provided with a microcode memory that stores more than one set of NaN propagation rules. In operation, the floating point arithmetic unit accesses one of the sets of NaN propagation rules according to the precision of the calculation being performed. A method of performing calculations in a floating point arithmetic unit includes dynamically determining if a calculation to be performed is to be a quad precision calculation or a double precision calculation. If it is determined that a quad precision calculation is to be performed, quad precision NaN propagation rules are selected and a quad precision calculation is performed using the selected quad precision NaN propagation rules.

Abstract: A floating point unit (60) capable of executing projection instructions provides performance improvement in multiple precision floating point arithmetic. The projection instructions provide for obtaining partial sequences of numbers, products, and sums which have definite alignments and widths which a programmer can set. This allows very fast computation of both individual intermediate computations and final results. A range projection instruction (210, 410) builds a mask with an exponent from one source (230, 430) and a mantissa from another (240, 440). A project instruction (610) builds a result by masking (660) mantissa bits in a source operand after alignment (630) with a mask. Projection multiply (810), add (1000), and subtract instructions build results by masking (850, 1070) mantissa bits of unrounded partial results after alignment (830, 1020, 1040) with a mask.

Abstract: A format for representing floating point numbers reduces the overhead typically associated with parsing floating point numbers and thereby provides for significantly improved processing speeds, particularly for bit-serial processors. According to an exemplary single-precision embodiment, numbers are represented using a 36-bit data format. Extra bits in the representation according to the invention allow certain conditions, such as overflow/underflow and the zero-ness of a number, to be detected and asserted quickly. Other conditions, such as denormalization are subsumed into normal processing through the extension of an exponent range in the representation.

Abstract: A reconfigurable register file integrated in an instruction set architecture capable of extended precision operations, and also capable of parallel operation on lower precision data is described. A register file is composed of two separate files with each half containing half as many registers as the original. The halves are designated even or odd by virtue of the register addresses which they contain. Single width and double width operands are optimally supported without increasing the register file size and without increasing the number of register file ports. Separate extended registers are also employed to provide extended precision for operations such as multiply-accumulate operations.