In digital processing systems, numerical data is typically expressed using integer or floating-point representation. Floating-point representation is preferred in many applications because of its ability to express a wide range of values and its ease of manipulation for some specified operations. A floating-point representation typically includes three components: a sign bit (sign), a mantissa (mant) that is sometimes referred to as a significand, and an exponent (exp). The represented floating-point number can be expressed as (-l) 'mant*2'\ Floating-point representations are also defined by"IEEE Standard for Binary Floating-Point Arithmetic,"which is referred to herein as the IEEE- 754 standard (or simply the IEEE standard) and incorporated herein by reference in its entirety for all purposes.

Many operations can be performed on floating-point numbers, including arithmetic operations such as addition, subtraction, and multiplication. For arithmetic operations, the IEEE standard provides guidelines to be followed to generate a unique answer for each floating-point operation. In particular, the IEEE standard describes the processing to be performed on the result from a particular operation (e. g., multiply, add), the precision of the resultant output, and the data format to be used. For example, the IEEE standard defines several rounding modes available for the results from add and multiply operations, and the bit position at which the rounding is to be performed. The requirements ensure identical results from different implementations of IEEE-compliant floating-point processors.

Many applications perform multiplication on two operands and addition (or subtraction) of the resultant product with a third operand. This multiply-add (or Madd) operation is common, for example, in digital signal processing where it is often used for computing filter functions, convolution, correlation, matrix transformations, and other functions. The Madd operation is also commonly used in geometric computation for (3-D) graphics applications.

Conventionally, a Madd operation can be achieved by sequentially performing a multiply (MUL) operation followed by an add (ADD) operation.

Performing the operations sequentially results in long processing delay. Improved performance can often be obtained by performing the Madd operation using a specially designed unit that also supports conventional floating-point multiplication and addition.

For Madd operations, post-processing is typically performed on the intermediate result from the multiply portion. To obtain a final Madd output that is fulfills IEEE rounding requirement, the post-processing includes possible denormalization and rounding of the intermediate result in accordance with one of the rounding modes defined by the IEEE standard. Denormalization is performed on a denormalized number (i. e., a non-zero number between the smallest positive representable normalized number, +amjn, and the smallest negative representable normalized number,-amin) to place the denormalized number in a proper format such that rounding can be performed at the bit location specified by the IEEE standard. The post- processing (or more specifically, the denormalization and rounding) to generate an IEEE- compliant Madd result typically lead to reduced accuracy (since some bits may be discarded during the denormalization and rounding), increased hardware complexity, and increased processing time. To reduce hardware complexity and improve processing time, some Madd architectures provide an additional operating mode in which numbers (e. g., intermediate results) smaller than the smallest representable normalized number are set or flushed to zero, or some other values such as amin. However, the flush-to-zero mode suffers from a higher loss in accuracy since the mantissa is replace with zero or some other predefined minimum value.

Accordingly, for Madd operations, techniques that increase the accuracy of the output result, simplify the post-processing of the intermediate result, and reduce the overall processing time are highly desirable.

SUMMARY OF THE INVENTION The invention provides a floating-point processors capable of performing multiply-add (Madd) operations and having improved accuracy, reduced circuit complexity, and possibly enhanced operating speed. Improved performance is achieved by processing the intermediate result from the multiplier unit, when operating in a specified operating mode, in a particular manner. Specifically, the intermediate result is rounded but not normalized or denormalized, as described in detail below.

An embodiment of the invention provides a floating-point unit (FPU) configurable to perform multiply-add (Madd) operations. The FPU includes a multiplier unit coupled to an adder unit. The multiplier unit receives and multiplies the mantissas for the first and second operands to generate a multiplier output mantissa. The multiplier unit is configurable to operate in a first operating mode defined by the multiplier output mantissa being rounded and having a pseudo-normalized format. The adder unit couples to the multiplier unit and receives and combines the multiplier output mantissa and a mantissa for a third operand to generate a FPU output mantissa.

In an embodiment, the pseudo-normalized format is characterized by at least one bit to the left of a binary point having a value of one. The multiplier output mantissa can be rounded at a bit position defined by IEEE standard, and can also be rounded without regard to its associated exponent.

The multiplier unit can be designed to selectively operate in one of a number of operating modes. Besides the first operating mode described above, a second operating mode can be defined by the multiplier output mantissa being in conformance with the IEEE standard. In this mode, the multiplier output mantissa is normalized or denormalized, as necessary, and rounded in accordance with the IEEE standard. A third operating mode can be defined by the multiplier output mantissa being flushed to zero or other predefined values when a denormalized multiplier output is detected. The operating mode can be selected by a value stored in a control register, a control signal, or some other mechanisms.

The FPU can provided improved performance when configured to execute a set of operations designed to approximate a reciprocal of a number or a reciprocal square root of a number. These approximation can be performed in accordance with a Newton-Raphson algorithm.

Yet another embodiment of the invention provides a floating-point processor configurable to perform Madd operations. The floating-point processor includes a multiplier unit coupled to an adder unit. The multiplier unit includes a multiplier array operatively coupled to a first rounding unit. The multiplier array receives and multiplies mantissas for two operands. The first rounding unit is configurable to round an output from the multiplier array to generate a rounded multiplier output mantissa having a pseudo-normalized format. The adder unit includes a carry propagation adder (CPA), a second rounding unit, and a normalization unit. The CPA receives and combines the multiplier output mantissa and a mantissa for a third operand.

The second rounding unit couples to the CPA and receives and rounds the mantissa from the CPA. The normalization unit couples to the second rounding unit and receives and normalizes the rounded mantissa. Within the multiplier unit, another CPA can be coupled between the multiplier array and the first rounding unit, to receive and combine a sum output and a carry output from the multiplier array.

The FPU and floating-point processor described above typically include additional units to process the exponents for the operands. The FPU and floating-point processor can be incorporated within a microprocessor or other hardware structure, and can also be described and/or implemented using hardware design languages (e. g., Verilog).

Yet another embodiment of the invention provides a method for performing a floating-point Madd operation. In accordance with the method, the mantissas for two operands are multiplied to generate a third mantissa, which is then rounded to generate a fourth mantissa. The fourth mantissa has a pseudo-normalized format and a range greater than a normalized mantissa. The fourth mantissa is combined with a mantissa for a third operand to generate an output mantissa. The output mantissa can further be rounded and normalized to generate a representation that conforms to the IEEE standard.

The invention also provides computer program products that implement the embodiments described above.

The foregoing, together with other aspects of this invention, will become more apparent when referring to the following specification, claims, and accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 shows a simplified diagram of an embodiment of a processor that incorporates a floating-point unit (FPU) of the invention; Fig. 2 shows a block diagram of an embodiment of a FPU capable of performing add, multiply, and multiply-add operations in accordance with the invention; Fig. 3A shows a representation of a floating-point number; Fig. 3B shows representations for single and double-precision floating- point numbers as defined by the IEEE standard; Fig. 4A shows a diagram of a line that graphically represents all real numbers;

Fig. 4B shows a diagram of an exponent representation for single- precision numbers in accordance with the IEEE standard; Fig. 5A shows the representations for some normalized numbers; Fig. 5B shows the representations for some denormalized numbers and for zero (0); Fig. 6A shows two mantissa representations that can result from a multiply operation; Fig. 6B shows a representation of a normalized but unrounded mantissa; Fig. 6C shows a representation of a normalized mantissa that conforms to the IEEE standard; Fig. 7A shows two mantissa representations for the mantissa from a carry propagation adder (CPA) within a multiplier unit of the FPU in Fig. 2; Fig. 7B shows two mantissa representations for the mantissa from the rounding unit within the multiplier unit; and Fig. 8 shows a representation of a floating-point control status register ("FCSR") configured to store bits identifying a particular operating mode.

DESCRIPTION OF THE SPECIFIC EMBODIMENTS Fig. 1 shows a simplified diagram of an embodiment of a processor 100 that incorporates a floating-point unit (FPU) 110 of the invention. As shown in the specific embodiment in Fig. 1, processor 100 further includes an instruction dispatch unit (IDU) 120, a load store unit (LSU) 130, and an integer execution unit (IXU) 140. IDU 120 decodes a sequence of instructions, dispatches floating-point instructions to FPU 110, and keeps track of the state of each dispatched floating-point instruction, resource and register dependencies, and the possibility of bypassing a resultant operand to the next FPU instruction. FPU 110 performs floating-point computations, as directed by IDU 120.

LSU 130 interfaces with other elements (i. e., internal or external to processor 100) and provides data to, and receives data from FPU 110. For example, operands are loaded from LSU 130 to FPU 110 and results are provided from FPU 110 to LSU 130. IXU 140 performs integer computations, and is able to transfer data to, and receive data from FPU 110.

Fig. 1 also shows a block diagram of an embodiment of FPU 110. FPU 110 includes a floating-point register file (FPR) 152 that interfaces with LSU 130. FPR 152 includes a number of read ports (i. e., for reading up to three operands for each

arithmetic operation and one operand for a store operation) and a number of write ports (i. e., for writing one operand for each arithmetic and load operation).

A floating-point pipe file (PIP) 154 couples to FPR 152 and further interfaces with LSU 130 and IXU 140. For each instruction, PIP 154 selects and receives operands from FPR 152, a load pipe file, a result pipe file, or a ROM. PIP 154 then unpacks the received operands (i. e., from an IEEE-compliant format) into an internal data format recognized by the processing units within FPU 110. PIP 154 also packs the results from FPU 110 into a data format (i. e., IEEE-compliant format) required by the external circuitry and provides the packed results to FPR 152.

A floating-point multiplier (MUL) 156 couples to PIP 154 and executes floating-point multiply instructions as well as the multiply portion of compound instructions such as a multiply-add (MADD) instruction. MUL 156 receives the mantissas of two operands from PIP 154. In an embodiment, MUL 156 is implemented with a multiplier array that generates a set of sum and carry outputs having full precision.

The sum and carry are provided to, and combined in a carry-propagate adder (CPA) to generate a precise, unrounded resultant mantissa. The lower bits of the resultant mantissa are logically combined to form a"sticky"bit that, in combination with the round bit and the current rounding mode, are used to generate rounding information to be sent to a subsequent rounding unit within MUL 156. MUL 156 is described in further detail below.

A floating-point adder (ADD) 158 couples to PIP 154 and MUL 156.

ADD 158 executes floating-point add and subtract instructions as well as the add portion of compound instructions such as MADD. ADD 158 receives two operands and performs floating-point magnitude addition/subtraction using, for example, a prescale adder (PSA) operated in parallel with a massive cancellation adder (MCA). The final output is selected from one of the adders and provided to PIP 154 for storage. In one embodiment, adder selection is based upon the exponent difference of the subject operands and the first few bits of the mantissas of such operands (i. e., the integer bit (s) and the most significant bit of the fractional portion) so that: (1) the PSA is always used for addition and selectively used for subtraction when the result is guaranteed to be no less than 0.1000000 before rounding and normalization and (2) the MCA is selectively used for subtraction when the result is guaranteed to be less than 1.0 and may have a large number of leading zeros before rounding and normalization. Operands that satisfy both conditions may be processed by either the PSA or MCA. Often, this selection criteria results in selection of

the PSA for additions and subtractions when the difference between operand exponents is greater than two, and the selection of the MCA for subtractions when the difference between operand exponents is less than or equal to two.

A floating-point exponent unit (EXP) 160 couples to PIP 154 and ADD 158. EXP 160 receives the exponents of the operands from PIP 154, estimates an output exponent based on the operation being performed and the exponents of the operands, and adjusts the estimated output exponent based on the rounding and normalization performed on the mantissa (e. g., by MUL 156 and ADD 158). EXP 160 also performs other functions, such as overflow/underflow prediction and checks.

A floating-point divide/square-root unit (DSQ) 162 couples to MUL 156 and operates on the mantissas for divide and square-root instructions. DSQ 162 is designed to implement a particular algorithm (e. g., a radix-2 SRT or a radix-4 SRT algorithm).

A floating-point control unit (CTL) 164 is the control unit for FPU 110.

CTL 164 receives floating-point instructions and directs the operation of various units within FPU 110 to generate the proper output.

Fig. 2 shows a block diagram of an embodiment of a floating-point unit (FPU) 200 capable of performing add, multiply, and multiply-add (Madd) operations.

FPU 200 represents a portion of FPU 100 shown in Fig. 1. FPU 200 includes a multiplier unit coupled to an adder unit. Support circuitry, such as that shown in Fig. 1, is not shown in Fig. 2 for simplicity. The multiplier unit includes a multiplier array 210, a carry-propagation adder (CPA) 212, a rounding unit 216, and an exponent combine unit 220. The adder unit includes the remaining units shown in Fig. 2. FPU 200 includes several features that increase accuracy, simplify the hardware design, and improve operational performance, as described below.

At any given moment, FPU 200 can be configured to perform one of at least three different operations including addition, multiplication, and Madd. These operations are expressed by the following: Fd = Fs Ft, Fd =Fs*Ft and Fd = ( (Fs-Ft) Fr)

where Fd is the resultant output and Fr, Fs, and Ft are three input operands. The Madd operation can be further decomposed into the following operations and their corresponding names: Fd = (Fs Ft) + Fr, MADD Fd = (Fs Ft)-Fr, MSUB Fd =- ((Fs Ft) + Fr), N MADD Fd =- ( Ft)-Fr), N MSUB, As shown in Fig. 2, the mantissas, Mant S and Mant T, for two operands are provided to multiplier array 210. Array 210 can implement, for example, a Booth or modified Booth algorithm, and can include partial product generation logic and a number of carry-save adders. The partial product generation logic produces partial products based on the mantissas. The carry-save adders add a number of partial products together and send the outputs to other carry-save adders in a tree-like fashion until only two numbers are left, the final sum and carry. In a specific implementation, the carry-save adders take in four terms and combine them into two, but other configurations are possible.

Array 210 thus multiplies the two operands and provides the product in sum-and-carry format to CPA 212. CPA 212 combines the sum and carry and provides the resultant mantissa to rounding unit 216 that processes the mantissa based on a specified operating mode of the FPU. The operation of rounding unit 216 is further described below. The processed mantissa comprises the output mantissa from the multiplier unit.

The exponents, Exp S and Exp T, of the two operands are provided to exponent combination unit 220 that combines the exponents for a multiply operation.

The combined exponent from unit 220 comprises the output exponent from the multiplier unit.

In an embodiment, for improved performance (i. e., faster operating speed), the adder unit includes a prescale adder (PSA) and a massive cancellation adder (MCA) operated in parallel. Depending on the characteristics of the operands, the output from either the PSA or MCA is selected.

To perform a floating-point addition, the mantissas of the two operands are typically aligned by shifting one of the mantissa and adjusting its exponent until the

exponents of the two operands are equal. The mantissas are then combined (e. g., added or subtracted), and the resultant mantissa is normalized. The number of shifts prior to the combination can be large (e. g., when adding a large number with a small number), and the number of shifts after the combination can also be large (e. g., when subtracting two operands having similar magnitudes). The PSA and MCA are designed to efficiently process most input conditions, including these two extreme conditions.

For the PSA, the mantissa from rounding unit 216 is provided to MUXes 232a and 232b. The mantissas for operands R and T are provided to a MUX 230 that selects one of the mantissas, based on the operation to be performed, and provides the selected mantissa to MUXes 232a and 232b. MUX 232a selects the mantissa of the smaller operand and provides the selected mantissa to a right-shift unit 234. MUX 232b selects the mantissa of the larger operand and provides the selected mantissa to a CPA 236.

The exponents of operands R and T are provided to a MUX 226 that selects one of the exponents based on the operation to be performed. The selected exponent from MUX 226 and the combined exponent from unit 220 are provided to an exponent calculation unit 252 that determines the difference between the two exponents and a preliminary result exponent. Determination of the preliminary result exponent is dependent upon the arithmetic equation being performed, and is further described in U. S.

Patent Application Serial No. 09/363,638, filed July 30,1999, assigned to the assignee of the invention, and incorporated herein by reference. The preliminary result exponent (e. g., the larger exponent when performing an add operation) is provided to an exponent adjustment unit 256 and the exponent difference is provided to right-shift unit 234 that shifts the mantissa of the smaller operand to the right by the indicated exponent difference. The shifted mantissa is provided to CPA 236 that combines the two mantissas and provides a combined mantissa to a rounding unit 242. Rounding unit 242 rounds the combined mantissa and provides the rounded mantissa to a normalization unit 244.

The mantissa from CPA 236 can be in the Ol. xxx--xxxx, Ix. xxx--xxxx, or 0.1xxx--xxxx (from subtraction) format. Normalization unit 244 normalizes the result to the Ol. xxx--xx format by performing a 1-bit right-shift or left-shift, if necessary. The exponent is adjusted by exponent adjustment unit 256, as necessary, based on the normalization performed by normalization unit 244.

The MCA portion of FPU 200 includes a CPA 262, a leading zero anticipator (LZA) 264, and a left-shift unit 266. For the MCA, the mantissas from

rounding unit 216 and MUX 230 are provided to MUXes 232c and 232d. MUXes 232 facilitate a small shift of one of the mantissas, based on the exponent difference, to align the mantissas. MUXes 232 are also used to select one of the mantissas for inversion in subtraction operations (the inverter is not shown in Fig. 2 for simplicity). The outputs from MUXes 232c and 232d are provided to CPA 262 and LZA 264. CPA 262 combines the two mantissas and provides a combined mantissa to left-shift unit 266. LZA 264 anticipates the number of leading zeros in the resultant mantissa, based on the input operands. The output from LZA 264 is encoded into a control signal that defines the number of shifts for left-shift unit 266. The control signal is also provided to exponent adjust 256 to adjust the exponent.

The outputs from normalization unit 244 and left-shift unit 266 are provided to a MUX 246 that selects the output from the PSA or MCA as the output mantissa from FPU 200. The adjusted exponent from unit 256 comprises the output exponent from FPU 200. The operation of FPU 200 is further described in U. S. Patent Application Serial No. 09/364,514, filed July 20,1999, assigned to the assignee of the invention, and incorporated herein by reference.

Fig. 2 shows a Madd architecture in which two rounding operations can be performed, one after the multiply operation and the other after the add operation. This Madd architecture can generate Madd results that fulfill the IEEE rounding requirements, as if the multiply and add were executed separately.

Fig. 3A shows a representation of a normalized floating-point number.

The representation includes a sign bit 310, a mantissa 312, and an exponent 314. A normalized floating-point number is represented by a mantissa having a one (1) to the left of the binary point (i. e., the integer portion of the mantissa) and a l. xxx--xx format, where each"x"represents one bit that is either a one or a zero. As defined by the IEEE standard, the fractional portion"xxx--xx"represents 23 bits after the binary point for normalized single precision numbers and 52 bits for normalized double precision numbers. The normalized mantissa has a range of between 1.0 and 2.0 (i. e., 1.0 < mantissa < 2.0).

The IEEE standard defines the representation for floating-point numbers.

For normalized numbers, the IEEE standard mandates storing only the fractional portion of the mantissa (i. e., the"xxx--xx"portion in Fig. 3A). The leading one (1) to the left of the binary point is implicit and not stored.

Fig. 3B shows representations for single and double-precision floating- point numbers as defined by the IEEE standard. The IEEE standard defines the number of bits to be used for the exponent and for the fractional portion of the mantissa, for single (32-bit) and double (64-bit) precision floating-point numbers. As shown in Fig. 3B, the single-precision representation includes three components: a single sign bit (sign), an 8- bit exponent (exp), and a 23-bit mantissa (mant). The double-precision representation includes a single sign bit (sign), an 11-bit exponent (exp), and a 52-bit mantissa (mant).

A floating-point number (y) can be expressed as : y = (-1)"0 * mant 9 2"P. Eq. (1) Fig. 4A shows a diagram of a line that graphically represents all real numbers. It should be noted that this line representation is not drawn to scale. Generally, real numbers range from negative infinity (-oc) to positive infinity (+oc). In the line representation shown in Fig. 4A and for the floating-point expression shown in equation (1), numbers greater than 0.0 (i. e., y > 0.0) are represented by a positive sign bit (i. e., sign = 0) and numbers less than 0.0 (i. e., y < 0.0) are represented by a negative sign bit (i. e., sign = 1). Numbers having an absolute value greater than 1.0 (i. e., IYI > 1.0) are represented by a positive exponent (i. e., exp > 0) and numbers having an absolute value less than 1.0 (i. e., IYI < 1.0) are represented by a negative exponent (i. e., exp < 0).

For floating-point representations having finite resolution (such as those shown in Fig. 3B with 32 bits for single-precision and 64 bits for double-precision), only numbers within a particular range of values can be represented as normalized number using the expression shown in equation (1). This range is defined by a maximum normalized value ana,, and a minimum normalized value a", n. For single-precision numbers, amax is slightly smaller than 2+lys and Bmin = 2-. Some numbers between zero and the minimum normalized value (i. e., a" ; n > y > 0.0) can be represented as denormalized numbers, which are described below. Special representations are used for zero (0.0) and for some numbers greater than the maximum normalized value (e. g., infinity).

Fig. 4B shows a diagram of an exponent representation for single- precision numbers in accordance with the IEEE standard. To obtain both positive and negative exponent values with a binary representation, the represented exponent is offset with a bias value. For single-precision numbers, the IEEE standard defines the exponent bias value to be 127. Thus, the actual (i. e., unbiased) exponent value is equal to the

represented (i. e., biased) exponent value (i. e., as stored in the exponent portion shown in Fig. 3B) minus the exponent bias value of 127. For example, the biased exponent values of 254,127, and 1 correspond to the unbiased exponent values of +127,0, and-126, respectively. In accordance with IEEE standard, exponent values of all ones (e. g., 255) and all zeros (e. g., 0) are used to represent special cases, as described below. Thus, normalized numbers that can be represented have exponents within the range of +127 to- 126.

Fig. 5A shows the representations for some normalized numbers. As noted above, a normalized number has a mantissa that conforms to the l. xxx--xx format, where the bit to the left of the binary point is a one (1) and each"x"to the right of the binary point represents a single bit that can be either a one (1) or a zero (0). Generally, a normalized number has a biased exponent that is within the defined range (e. g., between 254 to 1 for single-precision, which corresponds to a biased exponents of+127 to-126, <BR> <BR> respectively). The biased exponent value of 0 is reserved for representing zero (i. e. , y = 0.0) and denormalized numbers, as described below.

The largest normalized number (i. e., amas) has a biased exponent of 111... 110 (corresponding to an unbiased exponent of +127 for single-precision) and a mantissa of all ones (corresponding to a mantissa value of 1.111... 11, since a"1"to the left of the binary point is implicit for a normalized number). It should be noted that the all ones exponent is reserved for special representations. The next largest number that can be represented has the same exponent value (e. g., unbiased exponent of +127 for single-precision) and a mantissa value that is one least significant bit (LSB) less than that of the largest normalized number (i. e., 1.111... 10). Generally, for decreasingly smaller numbers, the fractional part of the mantissa is decremented by one from all ones to all zeros. When the fractional part is all zeros, the next smaller number is expressed by resetting the fractional part back to all ones and decrementing the exponent value by one.

This process continues until the smallest value is reached, which is represented by an exponent of 000... 001 and a fractional part of all zeros.

Fig. 5B shows the representations for some denormalized numbers and for zero. As noted above, a zero real number (i. e., y = 0.0) is represented by an all-zero exponent and an all-zero mantissa. To extend the range of numbers that can be represented, the IEEE standard defines the representation of denormalized numbers using the all-zero exponent and non-zero mantissas. The denormalized mantissa has a format of O. xxx--xx, where the bit to the left of the binary point is a zero (0) and each"x"to the

right of the binary point is either a one (1) or a zero (0). The largest denormalized number has a mantissa value of 0.111... 11 (corresponding to a number slightly smaller than 2nez6 for single-precision) and the smallest denormalized number has a mantissa value of 0.000... 01 (i. e., corresponding to 2-l49, for single-precision). Thus, the denormalized numbers cover a range between the minimum normalized number ann and zero (i. e., a range of 2-126 to 2-l49 for single-precision).

Denormalized numbers extend the range of representable numbers, as indicated in the line representation in Fig. 4A, but are generally more difficult to handle in the FPU because extra processing is required to manipulate the exponent and mantissa.

To facilitate handling and to simplify processing of denormalized numbers, the FPU is typically designed with additional resolution to allow for representation of IEEE- compliant denormalized numbers using internal normalized representations. For example, by increasing the exponent by one additional bit, the resolution is greatly increased and denormalized numbers can be normalized using the internal representation prior to processing. As a specific example, for single-precision, a 9-bit exponent having an unbiased exponent range of +255 to-254 can easily present the smallest IEEE denormalized number of 2-l49.

Typically, numbers are stored in memory or storage in accordance with the IEEE format. As such, these numbers can be normalized numbers, denormalized numbers, or special numbers (e. g., zero and infinity). For many FPU architectures, numbers are retrieved from storage and, prior to processing by the FPU, are"unpacked" into internal normalized representations. The resultant output from the FPU may be "packed"back into the IEEE format prior to being transferred back to storage.

The IEEE standard defines guidelines to be followed to generate unique results for floating-point operations such as addition and multiplication. The IEEE standard does not define the processing for a Madd operation, which is basically a concatenation of a multiplication followed by an addition. To generate a Madd result that is IEEE-compliant, the intermediate result from the multiply portion of the Madd operation needs to be processed to generate an IEEE-compliant intermediate result that is identical to a result from an IEEE multiply operation.

Fig. 6A shows two mantissa representations that can result from a multiply operation. When performing multiplication of two normalized mantissas (i. e., using internal representations), each being within the range of 1.0 and 2.0 (i. e., corresponding to the 1. xxx--xx format) and having N bits of precision, the resultant mantissa can range

from 1.0 to 4.0 and have a Ol. xxx--xxxx or lx. xxx--xxxx format. The fractional portion "xxx--xxxx"represents up to 2N bits of precision for the unrounded multiplier result (or more than 23 bits for single precision and more than 52 bits for double precision numbers). Two bits to the left of the binary point are used to represent the range of 1.0 to 4.0. Normalization (e. g., a right-shift of one bit position) can be performed whenever the resultant operand is 2.0 or greater to maintain the mantissa within the range of 1.0 and 2.0. The exponent is adjusted accordingly whenever a shift is performed (i. e., by incrementing the exponent by one for each right shift by one bit position).

Fig. 6B shows a representation of a normalized but unrounded mantissa.

Since the resultant mantissa from the multiplier array can have up to 2N bits of precision, rounding can be performed to provide a mantissa having N bits of precision, the same precision as that of the input mantissas. The IEEE standard defines the position of the bit to be rounded as well as the available rounding modes. Essentially, the mantissa is truncated to the right of a defined bit position indicated by an arrow 620, and possibly incremented at this bit position. The increment bit is generated based on a round bit, a "sticky"bit, and the current rounding mode. The round bit is the bit to the right of arrow 620, and the sticky bit is the OR of all bits to the right of the round bit. The rounding may generate a mantissa that requires re-normalization. When this occurs, a second normalization is performed.

Fig. 6C shows a representation of a normalized mantissa that conforms to the IEEE standard. The normalized mantissa has a range of 1.0 to 2.0 and N bits of precision.

The formats shown in Figs. 6A through 6C can be deemed as variants of a "pseudo-normalized"format. As used herein, the pseudo-normalized format is defined by the presence of at least one bit having a value of one (1) located to the left of the binary point. Thus, the pseudo-normalized format includes O1. xxx--xx, l0. xxx--xx, 1 l. xxx--xx, and other formats. The pseudo-normalized format covers any number of binary digits to the right of the binary point.

As noted above, multiplication of two N-bit numbers results in a product having up to 2N bits of precision. Depending on the input operands, the product may exceed the maximum normalized value a, l, aX or may fall within the IEEE denormalized range. As part of the post-processing after the multiply operation, to generate an IEEE- compliant multiply result, a determination is made whether the intermediate result is a normalized or denormalized number. If the intermediate result is deemed to be a

denormalized number, a denormalization followed by a rounding is performed in order to generate IEEE-compliant multiply result, which is the prerequisite for generating an IEEE-compliant Madd result.

Denormalization of the intermediate result can be performed as follows.

First, the intermediate exponent associated with the intermediate result is checked to determine whether it is less than the minimum normalized exponent (e. g., less than-126 for single precision). This situation can occur when multiplying two small numbers, and the additional precision of the internal exponent representation can capture the small result. If the intermediate exponent is less than the minimum normalized exponent, the mantissa is shifted to the right and the exponent incremented by one with each bit of right-shift. The mantissa right-shift and exponent increment continues until the updated exponent is equal to the minimum normalized exponent. The mantissa is then rounded at the bit position indicated by the IEEE standard. Denormalization is necessary so that the rounding can be performed at the correct bit position as defined by the IEEE standard.

Denormalization of the intermediate result to generate an IEEE-compliant multiply result generally degrades accuracy because bits are thrown out during the denormalization process. Moreover, denormalization is a time consuming process that requires additional hardware to implement. Various approaches have been implemented to simplify the processing of denormalized number and to reduce the hardware requirement. Two of these approaches are described below.

In one approach, the FPU detects a denormalized number but the actual denormalization is performed by some other mechanisms (e. g., software). Detection of a denormalized number can be achieved using conventional techniques based upon exponent and mantissa values. Upon detection of a denormalized number, an exception occurs (i. e., a flag is raised) and corrective measures are carried out by software. The hardware detection and software processing approach can be applied to input operands as well as the intermediate result from the multiplier unit. This approach for handling denormalized numbers to generate an IEEE-compliant multiply result typically sacrifices performance (i. e., slower operating speed) for a more simplified hardware design.

In another approach, the FPU is designed with the capability to flush a denormalized number to zero. In one specific implementation of this"flush-to-zero" approach, if an operand is determined to be within the range of positive and negative minimum normalized number (i. e., between +am ; n and-am ; n) it is set to zero (0.0). Other variants of the flush-to-zero approach can also be implemented. For example, in a round-

to-positive-infinity rounding mode, the operand is flushed to the minimum normalized number a,, if it is within the range of zero (0.0) and am n, and flushed to zero (0.0) if it is within the range of-anin and zero (0.0). In a round-to-negative-infinity rounding mode, the operand is flushed to-anin if it is within the range of-amin and zero (0.0), and flushed to zero (0.0) if it is within the range of zero (0.0) and amis. The flush-to-zero operation can be performed on denormalized input, intermediate, and output operands. This approach provides improved operating speed and requires minimal additional hardware to flush the intermediate result to zero. However, the Madd result is not IEEE-compliant and accuracy is degraded since bits are thrown out when the mantissa is flushed to zero or some other values.

Many applications do not require IEEE-compliant results from Madd operations. For these applications, the intermediate result from the multiply portion of a Madd operation can be flushed to zero, as described above. However, flushing the intermediate result to zero degrades the accuracy of the associated Madd result.

In accordance with the invention, a new approach is provided that causes rounding, but not normalization or denormalization, of the intermediate result from the multiply portion of the Madd operation. In this"Madd-flush-override"approach, the intermediate result is maintained in an internal normalized format to improve accuracy.

This approach may be used exclusively or configured as one of several operating modes of a computer system. Each such operating mode may be identified, for example, by one or more bits held in a control register.

Fig. 8 shows a representation of a floating-point control status register ("FCSR") configured to store bits identifying a particular operating mode. The FCSR contains bit values that control the floating-point operations of the FPU. In a specific embodiment, the FCSR includes a flush ("FS") bit and a Madd-flush-override ("FO") bit located at bit positions 24 and 22, respectively, to specify IEEE-compliant, flush-to-zero, and Madd-flush-override modes. The flush-to-zero mode is activated when the FS bit is one (i. e., logic high), and the IEEE-compliant mode is activated when the FS bit is zero (i. e., logic low). When the FO bit is one, the Madd-flush-override mode is activated and a denormalized intermediate result of a Madd operation is not flushed nor denormalized according to the FS bit. The FCSR is further described in U. S. Patent Application Serial No. 09/364,787, filed July 20,1999, assigned to the assignee of the invention, and incorporated herein by reference.

Table 1 summarizes the operating modes and a designation for the FS and FO bits.

Table 1-Mode Definition

MODE FO FS Madd Intermediate Result Processing IEEE-0 0 Process the intermediate result from the multiplier compliant unit to generate IEEE-compliant intermediate output. This processing includes normalization/ denormalization and rounding of the intermediate result. Flush-to-0 1 Flush the intermediate result to zero or some other zero predefined value if it is a denormalized number. Madd-1 Round Madd intermediate result in internal flush-normalized format (also referred to herein as override"pseudo-normalized"format); i. e., round but do not normalize/denormalize the intermediate result.

Generally, the FPU can be designed to support any combination of the modes shown in Table 1 and other modes not listed. For example, the FPU can be designed to support only Madd-flush-override mode, IEEE-compliant and Madd-flush- override modes, all three modes listed in Table 1, or other mode combinations. In IEEE- compliant mode, the processing to generate the IEEE-compliant intermediate result can be performed in hardware, software, or a combination of both. The modes listed in Table 1 are described below.

In IEEE-compliant mode, the intermediate result from the multiplier unit in the FPU is normalized or denormalized, as necessary by hardware or software, and rounded. Referring to Fig. 6A, the intermediate result from CPA 212 can be in either the Ol. xxx--xxxx or lx. xxx--xxxx format. Initially, the exponent for the intermediate result is checked to determine whether the intermediate result is within the range of normalized or denormalized number. This can be achieved by comparing the exponent against the maximum exponent value expmax and the minimum exponent value expnMn (e. g., expnux = +127 and expni, =-126 for single-precision). If the intermediate result is determined to be a normalized number, the mantissa is normalized to the 01. xxx--xxxx format.

Alternatively, if the intermediate result is determined to be a denormalized number, the mantissa is denormalized in the manner described above. The normalized or

denormalized mantissa is then rounded at the position indicated by arrow 620 in Fig. 6B.

The normalized/denormalized and rounded intermediate result is provided to the adder unit.

In Flush-to-zero mode, the mantissa from the multiplier unit in the FPU is flushed to zero if the intermediate result is a denormalized number. Again, this determination can be achieved by checking the exponent for the intermediate result and the mantissa value. Specifically, if the exponent is less than expmjn and the mantissa is not equal to zero (e. g., mant W 0), the intermediate result is deemed to be a denormalized number and flushed to zero.

In Madd-flush-override mode, the intermediate result from the multiplier unit is rounded but not normalized or denormalized. In an embodiment, the rounding is performed without checking the exponent associated with the intermediate result. In an embodiment, the mantissa from rounding unit 216 is rounded regardless of whether it is a denormalized number or a number within the range defined by amax and amin.

Fig. 7A shows two mantissa representations for the mantissa from CPA 212 in the multiplier unit. The output from multiplier array 210 can have up to 2N bits of precision to the right of the binary point and two bits of precision to the left of the binary point. Rounding unit 216 initially determines whether the bit in the most significant bit (MSB) position 708 is a zero (0) or a one (1). The position of the bit to be rounded differs by one bit position depending on whether the mantissa has the 01. xxx--xxxx or Ix. xxx-xxxx format. For a mantissa having the 01. xxx--xxxx format, rounding unit 216 rounds the bit at the position indicated by an arrow 710a. Alternatively, for a mantissa having the 1 x. xxx--xxxx format, rounding unit 216 rounds the bit at the position indicated by an arrow 710b.

Fig. 7B shows two mantissa representations for the mantissa from rounding unit 216. The rounded mantissa ranges between 1.0 and 4.0, which is approximately twice the range of a normalized mantissa. An additional bit is used to represent the additional range in the mantissa. The subsequent units (e. g., right-shift unit 234, CPA 236, rounding unit 242, and normalization unit 244 in the PSA, and CPA 262 and left-shift unit 266 in the MCA) are designed to properly process the additional bit in the mantissa.

Several advantages are obtained by performing rounding, but not normalization or denormalization, in accordance with MADD-flush-override mode. First,

accuracy is improved since denormalization, which throws out bits, is not performed. In MADD-flush-override mode, the intermediate result is kept in internal normalized format to improve accuracy Second, some or all of the hardware can be designed with less than 2N bits of precision, since some of the precision is likely to be discarded in the rounding within the multiplier unit. For example, multiplier array 210 and CPA 212 can be implemented with less than 2N bits of precision. The subsequent units in the adder unit can also be designed to operate on number having less than 2N bits of precision, which again simplify the hardware design.

Third, implementation of these units within less precision can improve the operating speed of the FPU. As noted above, the multiplier array is typically implemented with a set of carry-save adders operated sequentially. Less processing delay, and therefore faster operating speed, may be obtain by calculating less than the full 2N bits result.

As noted above, floating-point processors capable of performing Madd operations are desirable in many applications, such as digital signal processing and graphics. Madd operations are also commonly used in arithmetic algorithms. For example, the Newton-Raphson algorithm is a convergent algorithm used for approximating reciprocals and reciprocal square roots. This algorithm performs many multiply and Madd operations. The performance of the algorithm (e. g., its speed of convergent and accuracy) is dependent on the accuracy of the Madd operation.

Moreover, since the algorithm typically includes many Madd operations, the speed of the algorithm is also dependent on the speed of the Madd operation. A floating-point processor incorporating the invention can provide improved performance (i. e., accuracy and speed) when used to implement the Newton-Raphson algorithm.

The Newton-Raphson reciprocal algorithm for approximating the reciprocal of a number R is defined as: Xi+1=Xi#(2-R#Xi),1/R# where i is an integer greater than one (i. e., i = 1,2,3,...), Xi is an approximation from the i-th iteration, and X ; +t is a more accurate approximation at the (i+1)-th iteration.

The Newton-Raphson reciprocal square root algorithm for approximating the reciprocal square root of a number (R) is defined as:

(3-R#Xi#Xi)#Xi<BR> 1/#R # Xi+1 =<BR> <BR> <BR> 2 where Xi is an approximation of the reciprocal square root of the number R from the i-th iteration, and Xi, l is a more accurate approximation at the (i+l)-th iteration.

The Newton-Raphson algorithm and its implementation using a floating- point processor is further described in U. S. Patent Application Serial No. 09/363,637, filed July 20,1999, assigned to the assignee of the invention, and incorporated herein by reference.

A floating-point instruction set architecture (ISA) for FPU 110 includes instructions for implementing the Newton-Raphson reciprocal and reciprocal square root algorithms. In one embodiment, the ISA includes three instructions for implementing an iteration of the Newton-Raphson reciprocal algorithm. The three instruction are: RECIPl (operandl), RECIP2 (operandl, operand2), and MADD (operandl, operand2, operand3).

The RECIP1 (operand 1) instruction enables FPU 110 to produce a result that is an estimate of the reciprocal of operandl. There are a variety of ways to produce the estimate. In one embodiment, a lookup table is used. The RECIP2 (operandl, operand2) instruction enables FPU 110 to produce a result equal to (1- operandloperand2). The MADD instruction is a multiply-add instruction that enables FPU 110 to produce a result equal to (operandl + operand2*operand3). The RECIP1 and RECIP2 instructions are further described in the aforementioned U. S. Patent Application No. 09/364,787.

The Newton-Raphson algorithm for approximating the reciprocal of a number (R) can be implemented by configuring FPU 110 to execute the above-described instructions in the following sequence: (1) Xs = RECIP1 (R); (2) IR = RECIP2 (R, Xi); and (3) Xi, l = MADD (XI, Xi, IR).

After FPU 110 executes the above three instructions in the above given sequence, the following quantity is obtained: Xi+1Xi#IR=Xi+Xi(1-R#Xi)=2Xi-R#Xi#Xi=Xi(2-R#Xi),+

which is the Newton-Raphson approximation for the reciprocal of the number R.

In an embodiment, the floating-point ISA includes four instructions for performing an iteration of the Newton-Raphson algorithm for approximating reciprocal square roots. The four instruction are: RSQRTl (operandl), RSQRT2 (operandl, operand2), MUL (operandl, operand2), and MADD (operandl, operand2, operand3).

The RSQRT1 (operand 1) instruction enables FPU 110 to produce a result that is an estimate of the reciprocal square root of the operand. There are a variety of ways to produce the estimate. In one embodiment, a lookup table is used. The RSQRT2 (operandl, operand2) instruction enables FPU 110 to produce a result equal to: (1-operandloperand2)/2. The MADD (operandl, operand2, operand3) instruction is described above. The MUL (operandl, operand2) instruction is a multiply instruction that enables FPU 110 to produce a result equal to: operandl'operand2. The RSQRT1 and RSQRT2 instructions are further described in the aforementioned U. S. Patent Application No. 09/364,787.

The Newton-Raphson algorithm for approximating the reciprocal square root of a number (R) can be implemented by configuring FPU 110 to execute the above- described instructions in the following sequence: (1) Xi = RSQRT1 (R); (2) IR1 = MUL (R, Xi); (3) IR2 = RSQRT2 (IR1, Xi); and (4) Xo+l = MADD (X,, X., IR2).

After FPU 110 executes the above four instructions in the above given sequence, Xj+) = (3 -RXoXs) Xj/2, which is the Newton-Raphson approximation for the reciprocal square root of the number R.

Each set of instructions shown above includes a MADD instruction. The MADD instruction has improved accuracy when executed using the MADD-flush- override mode described above. The IR value produced by the RECIP2 (R, X ;) instruction and the IR2 value produced by the RSQRT2 (IR1, Xj) instruction are usually small numbers. When these numbers are internally represented in the pseudo-normalized format as shown in Fig. 6A, a great deal of precision is achieved, thereby increasing the accuracy of the approximation. Additionally, when Xi and IR (or IR2) are small numbers,

the multiplication of Xi with IR (or IR2) during execution of a MADD instruction (as provided in the foregoing sequences) can result in a denormalized number. However, when these values are maintained in the pseudo-normalized format, accuracy is maintained since denormalization processing is typically avoided.

For clarity, the invention has described in the context of single and double- precision floating-point representations that conform to the IEEE formats. However, the invention can be adopted for used with other floating-point representations, and this is within the scope of the invention.

FPU 200 can be implemented in various manners. For example, FPU 200 can be implemented in hardware within a digital signal processor, an application specific integrated circuit (ASIC), a microprocessor, and other hardware structures.

In addition to implementations of the invention using hardware, the invention can also be embodied in an article of manufacture comprised of a computer usable medium configured to store a computer-readable program code. The program code causes the enablement of the functions or fabrication, or both, of the hardware disclosed in this specification. For example, this might be accomplished through the use of general programming languages (e. g., C, C++, and so on), hardware description language (HDL), register transfer language (RTL), Verilog HDL, VHDL, AHDL (Altera hardware description language), or other programming and/or circuit (i. e., schematic) capture tools available in the art. As a specific example, the Verilog simulator"VCS v. 4.1.1" by Synopsys, Inc. was used to simulate the invention. A book entitled"A Verilog HDL Primer"by J. Bhasker, Star Galaxy Pr., 1997 provides greater detail on Verilog HDL, and is incorporated herein by reference in its entirety for all purposes. In the program code implementation, Fig. 2 can serve as an embodiment of a flow diagram.

It is understood that the functions accomplished by the invention as described above can be represented in a core that can be utilized in programming code and transformed to hardware as part of the production of integrated circuits. Also, other implementations of the invention (e. g., FPU 200) using a combination of hardware and software are also possible. Therefore, the embodiments expressed above are within the scope of the invention and should also be considered protected by this patent.

The foregoing description of the preferred embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without the use of the inventive faculty. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.