US 4,603,323 AGrant
Method for Extending the Exponent Range of an IBM 370-Type Floating Point Processor
Issue Date:1986-07-29
•3 Claims
•2 Drawing Sheets
Abstract
A method for extending the exponent range in an IBM System/370-type floating point arithmetic processor. Very large or very small numbers are represented by a pair of words having sign, exponent, and fraction portions thereof. If the signs of the low- and high-order words of a pair are the same, then the number is deemed to be in the ordinary processor operating range. However, if the signs differ, then the number is deemed to be out of range. In this case, a range extension is obtained by using the exponent field of the low-order word which otherwise would be transparent to the floating point processor.
Metadata
Assignee
- International Business Machines Corporation
Inventors
- Anthony Hassitt
- Alan H. Karp
- Randolph G. Scarborough
Application Information
Application Number:US 6144101
Filing Date:1984-05-25
Priority Date:1984-05-25
Art Unit:217
Classifications
IPC:
H03M 112
Field of Search:
364235340748310347 DD
Patent Drawings (2 sheets)
Description
TECHNICAL FIELD
This invention relates to digital computers employing floating point arithmetic operations and, more particularly, to the use in such computers of a new representation of floating point numbers.
BACKGROUND
As pointed out in the IBM System/370 Principles of Operation, 7th Edition, March 1980, IBM publication GA22-7000-6, in chapter 9, there is a characterization of a floating point number as consisting of a signed hexadecimal "fraction" and an unsigned 7-bit binary integer called the "characteristic". The "characteristic" represents a signed exponent and is obtained by adding 64 to the exponent value. This is termed "excess-64 notation". The range of the characteristic is 0-127, which corresponds to an exponent range of -64 to +63. The value of a floating point number is the product of its fraction and the number 16 raised to the power of the exponent which is represented by its characteristic.
The fraction of a floating point number is treated as a hexadecimal number because it is considered to be multiplied by a number which is a power of 16. The term "fraction" indicates that the radix point is assumed to be immediately to the left of the leftmost fraction digit. The fraction is represented by its absolute value and a separate sign bit. The entire number is positive or negative, depending on whether the sign bit of the fraction is 0 or 1, respectively.
When a floating point operation would cause the result exponent to exceed 63, the characteristic wraps around from 127 to 0 and an exponent-overflow condition exists. The resultant characteristic is then too small by 128. When an operation would cause the exponent to be less than -64, the characteristic wraps around from 0 to 127 and an exponent-underflow condition exists. The resultant characteristic is then too large by 128, except that a 0 characteristic is produced when a true 0 is forced.
In computers of the IBM System/370 type having a floating point arithmetic capability, floating point numbers have a 32-bit (short) format, a 64-bit (long) format, or a 128-bit (extended) format. The extended floating point number has a 28-digit fraction and consists of two long floating point numbers which are called the high-order and low-order parts. The high-order part may be any long floating point number. The fraction of the high-order part contains the leftmost 14 hexadecimal significant digits of the 28-digit fraction. The characteristic and sign of the high-order part are the characteristic and sign of the extended floating point number. If the high-order part is normalized, the extended number is considered normalized. The fraction of the low-order part contains the rightmost 14 digits of the 28-digit fraction. The sign and characteristic of the low-order part of an extended operand are ignored. Relatedly, the process of normalization consists of shifting the fraction left, one digit at a time, until the high-order hexadecimal digit is non-zero and concurrently reducing the characteristic by the number of hexadecimal digits shifted.
In the extended precision format, a floating point number of magnitude m lies in the range
In decimal terms this is approximately
Others have attempted to use multiple fields in a computer word to extend the range of floating point magnitude representations within a digital computer. Illustrative, is Morris, U.S. Pat. No. 3,742,198, issued June 26, 1973, "Apparatus for Utilizing a Three-field Word to Represent a Floating Point Number". Morris subdivides a fixed length digital computer word into three fields. A first and second field is subdivided into a variable length exponent field and a variable length fraction field. The third field serves to specify the size of the variable length exponent field. This permits a tradeoff between accuracy and range in the floating point representation of the number.
It is accordingly an object of this invention to devise a method for extending the exponent range of numbers utilized by an IBM System/370-type floating point arithmetic processor while maintaining compatibility with existing floating point number representations.
THE INVENTION
The foregoing object is attained by a method for operatively extending the exponent range in a floating point arithmetic processor of the type in which each numerical signal value is expressed as a pair of formatted words. The first word corresponds to a high-order sign, exponent, and fraction while the second word corresponds to a low-order sign, exponent, and fraction. The processor responsive to the formatted word pair does not ordinarily operate upon selective low-order fields. The method steps include (a) ascertaining the floating point range of the operand of each numerical signal value and of any resultant obtained by the arithmetic operation upon said operand or multiple operands; and (b) reformatting any operands or resultants which lie in the extended range in a sign bit-manipulated and protected field of a corresponding low-order word. Relatedly, the sign bit-manipulated and extended field of the low-order word also includes the sign and exponent fields.
The method of this invention relates to a new use of an atomic instruction. That is, namely, one of reformatting the low-order word of a word pair to contain an operand/resultant when said operand/resultant lies in the extended range. This is absent from the prior art. Indeed, the aforementioned Morris patent teaches only the conversion of a fixed field computer word to a three-field floating point word and vice versa.
BRIEF DESCRIPTION OF THE DRAWING
FIG. 1 depicts the extended precision of floating point number format used in the invention.
FIG. 2 sets out a high-level logical flow in floating point addition of two numbers invoking the inventive method.
FIG. 3 sets out the multiplication of two floating point numbers according to the invention in high-level flow.
DESCRIPTION OF THE PREFERRED EMBODIMENT AND INDUSTRIAL APPLICABILITY
The Execution Environment
As a preliminary, the method of this invention is executable upon a computing system of the type including a main store, input/output channel, control unit, direct access storage devices, local addressable clocks, and other I/O devices coupled thereto. Such a system is described in Amdahl et al, U.S. Pat. No. 3,400,371, issued Sept. 3, 1968, and entitled "Data Processing System". The Amdahl system includes as a resource, all of the facilities of either the computing system or of an operating system running thereon which are required for execution of a process including the method of this invention. Typical resources include a main store, I/O devices, the CPU, data sets, and control or processing programs. Furthermore, such systems are capable of multiprogramming. This pertains to the concurrent execution of two or more processes by a computing system, and can be managed on a computer running under an IBM System/370 operating system described in IBM publication GC28-6646, July 1973, and listed in IBM System/360 bibliography, GA22-6822.
Extended Floating Point Number Representation
Referring now to FIG. 1, there is shown a floating point number in the extended exponent, extended precision format. This format consists of a hexadecimal fraction of 112 bits, an unsigned binary integer of 14 bits termed the characteristic, a sign bit, and a flag bit for a total of 128 bits. The characteristic represents a signed exponent and is obtained by adding a bias to the exponent value. For purposes of this description, a bias may be taken to mean an offset or an adjustment.
The interpretation of the characteristic and the bias depends on the value of a flag bit. In this sense, a flag bit is a single bit used to differentiate between two states of interest. If the flag bit is the same as the sign bit, then the number has an exponent between -64 and +63. Further, the number fits into the normal extended precision number range. In the case that the signed bit is the same as the flag bit, the high-order 7 bits of a characteristic represent the exponent biased by 64 and the low-order 7 bits are ignored. If the flag bit is different from the sign bit, then the number is in the new number format and the characteristic is treated as a 14-bit exponent biased by 8192. Note that the magnitude 8192 is 2.sup.13.
When the result of an extended precision floating point operation has an exponent between -64 and +63, then irrespective of the formats of the operands of the operation, the flag bit and the sign bit are made identical and the characteristic is made a 7-bit exponent biased by 64. When the result of an extended precision floating point operation has an exponent which is greater than +63, then the flag bit is made different from the sign bit and the characteristic is made a 14-bit exponent biased by 8192. If the result exponent exceeds 8190, then an exponent-overflow condition exists, and the exponent, if it exceeds 8191, wraps around to -8192. When the result of an extended precision floating point operation has an exponent which is less than -64, then the flag bit is made different from the sign bit and the characteristic is made a 14-bit exponent biased by 8192. If the result exponent is less than -8191, then an exponent-underflow condition exists, and the exponent, if it is less than -8192, wraps around to 8191.
The fraction of a floating point number is treated the same in this invention whether the exponent is a 7-bit exponent or a 14-bit exponent. It is a number between 1/16 and 1 that is to be multiplied by a power of 16. The radix point is assumed to be immediately to the left of the leftmost fraction digit. The fraction in turn is represented by its absolute value and a separate sign bit. The entire number is positive or negative, depending on whether the sign bit of the fraction is 0 or 1, respectively.
Referring again to FIG. 1, it is apparent that the first bit (bit 0) is a sign bit. In both the 7-bit and 14-bit exponent formats, the number is negative, if and only if, this bit is set to 1. Bits 1-7 and 65-71 are respectively the high-order and low-order parts of the characteristic. Bit 64 is used as a flag in order to distinguish 7-bit exponents from 14-bit exponents.
If bit 64 is the same as bit 0, then the number has a 7-bit exponent biased by 64. In this format, the low-order 7 bits of the characteristic are ignored. However, if bit 64 is different from bit 0, then the number has a 14-bit exponent biased by 8192. In this format, bits 1-7 concatenated with bits 65-71 represent a 14-bit characteristic.
The range covered by the magnitude m of a normalized extended precision number depends on the format. In the 7-bit exponent format, reference can be made to the prior description. In the 14-bit exponent format, m lies in the range
or
In the ensuing description of the invention, several tables depicting software fragments embodying the preferred representation of the invention will be used in several examples. In this regard, Table 1, in conjunction with FIG. 2, illustrates the addition of two floating point numbers involving the extended range method. Likewise, Table 2, in conjunction with FIG. 3, shows the multiplication of two floating point numbers. Further, Table 3 sets forth software sequences invoking either the addition or multiplication operations. Lastly, Table 4 sets out the computational results of these operations.
Referring now to Table 1 and FIG. 2, there is illustrated the addition of two floating point numbers according to the invention. The first step in adding two numbers is to extract the exponents of the two operands. This is done by determining if each number is in the normal range. If it is in the normal range, a 7-bit exponent is extracted. If it is not in the normal range, a 14-bit exponent is extracted. The next step in the procedure is to scale the fraction of one of the operands based on the difference of the exponents of the two operands. The next step is to form the fraction of the result by adding the fractions of the two operands as scaled. Next, the exponent of the result is formed by adding the exponent of the unscaled operand to the exponent of the fraction of the result. If the result is in the normal range, an exponent in the 7-bit format is produced. If the result is not in the normal range, an exponent in the 14-bit format is produced. An embodiment to this algorithm is shown in Table 1 and is set out in the IBM System/370 Assembly language. ##SPC1##
Referring now to Table 2 and FIG. 3, there is illustrated the multiplication of two floating point numbers utilizing the invention. In order to multiply two floating point numbers, each exponent is extracted by determining if it is in the normal range. If it is in the normal range, a 7-bit exponent is extracted. If it is not in the normal range, a 14-bit exponent is extracted. The next step in the multiplication procedure is to multiply the fractions of the two arguments to produce the fraction of the result. The exponent of the result is formed by adding the exponents of the two input arguments and the exponent of the fraction of the result. If the result is in the normal range, a result with a 7-bit exponent is produced. If the result is not in the normal range, a result with a 14-bit exponent is produced. The software embodiment of the multiplication operation is set out in detail in Table 2. ##SPC2##
Referring now to Table 3, there is shown an interactive program fragment invoking the add and multiply floating point functions embodying the invention. The invocation is respectively defined by statements 100-130 for floating point add and 300-330 for floating point multiply. Relatedly, Table 4 shows a displayed output of the code execution of Table 3 in combination with the execution shown in Tables 1 and 2. Note that the output is that of a menu from which an operator may select one or more functions. For purposes of illustration, either the add or multiply functions may be selected, operands A and B entered, and the resultant depicted as either A+B or A.times.B.
Alterations and modifications can be made by those having skill in this art without departing from the spirit and scope of the invention. One technique might be to modify high-level language compilers such as a FORTRAN compiler. Such a compiler will either encode the extended floating point arithmetic operations or invoke calls to subroutines to execute these functions. Alternatively, the compiler could generate several so-called "no operations" (NOP) around the extended precision operations in order to allow a post-processor to scan the object code and make the needed changes.
Claims
We claim:
1. A new use of the atomic operation of reformatting at least a portion of a pair of words used to represent either a numerical operand or resultant in a floating point arithmetic processor of the IBM System/370 type, the first word of a pair corresponding to a high-order sign, exponent, and fraction, while the second word corresponds to the low-order sign, exponent, and fraction, comprising the steps of: (a) ascertaining the floating point range of a numerical operand and of any resultant obtained by the arithmetic operation upon said operand or multiple operands, said step includes comparing the signs of the high- and low-order words of a pair such that a first match condition indicates that the operand or resultant lies within a normal range while a second match condition indicates that the operand or resultant lies within the extended range; and (b) reformatting the low-order word to contain an operand or resultant when said operand or resultant lies in an extended range.
2. A method for operatively extending the exponent range in a floating point arithmetic processor of the type in which each numerical signal value is expressed as a pair of formatted words; the first word corresponding to a high-order sign, exponent, and fraction while the second word corresponding to a low-order sign, exponent, and fraction; the processor responsive to the formatted word pair ordinarily not operating upon selective low-order words or fields, comprising the steps of: (a) ascertaining the floating point range of the operand of each numerical signal value and of any resultant obtained by the arithmetic operation upon said operand or multiple operands by testing the match condition of the signs in the high- and low-order words of each pair such that a match indicates that a corresponding operand or resultant lies outside the normal range while a mismatch indicates the corresponding operand or resultant lies in the extended range; and (b) reformatting any operands or resultants which lie in the extended range in a sign bit-manipulated and protected field of the corresponding low-order word by utilizing the exponent field of the low-order word as an extension of the exponent field of the high-order word.
3. A method according to claim 2, wherein the exponent field of the low-order word is transparent to the floating point processor where the signs of the high- and low-order words of a pair match.
Patent Citations (2)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US3400371 | 1968-09-01 | Amdahl et al. | |
| US3742198 | 1973-06-01 | Morris |
Non-Patent Literature (1)
- IBM System/370 Principles of Operation, IBM Publication GA 22-7000-G, 7th Edition, Chapter 9, Mar. 1980.