US20010027461A1 - Seed rom for reciprocal computation - Google Patents
Seed rom for reciprocal computation Download PDFInfo
- Publication number
- US20010027461A1 US20010027461A1 US09/867,851 US86785101A US2001027461A1 US 20010027461 A1 US20010027461 A1 US 20010027461A1 US 86785101 A US86785101 A US 86785101A US 2001027461 A1 US2001027461 A1 US 2001027461A1
- Authority
- US
- United States
- Prior art keywords
- reciprocal
- estimated
- lookup table
- term
- memory
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
- G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
- G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
- G06F7/52—Multiplying; Dividing
- G06F7/535—Dividing only
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2207/00—Indexing scheme relating to methods or arrangements for processing data by operating upon the order or content of the data handled
- G06F2207/535—Indexing scheme relating to groups G06F7/535 - G06F7/5375
- G06F2207/5356—Via reciprocal, i.e. calculate reciprocal only, or calculate reciprocal first and then the quotient from the reciprocal and the numerator
Definitions
- Math functions including addition and subtraction might be accomplished by performing the steps of 1) loading registers with the operands; 2) executing the operational code (“opcode”) for the desired function; and then, 3) retrieving the result from one or more registers.
- the opcode is simply a command to perform the function implemented in hardware.
- the steps for computing a reciprocal might include 1) scaling or normalizing an operand; 2) looking up a reciprocal value corresponding to the normalized value in a lookup “seed” table; and 3) scaling or denormalizing the looked up value to reflect the reciprocal of the denormalized operand.
- the lookup table is typically stored in a read only memory (ROM). The error associated with this technique is controlled to a great extent by the “height” and “width” or the precision of the entry in the lookup table.
- One disadvantage of the conventional lookup approach is that the area required to store the lookup table on the integrated circuit may prevent the lookup approach from being a cost effective technique for given precision requirements.
- a cost effective way of implementing mathematical functions such as reciprocal or division to within a consistent error margin in a give time frame is a desirable feature of a computer system.
- a reciprocal of an operand is determined by looking up an estimated reciprocal term in a first lookup table stored in a first computer memory wherein the estimated reciprocal term corresponds to at least a portion of the operand.
- An error term is looked up in a second lookup table stored in a second computer memory. The error term corresponds to at least a portion of the operand.
- the reciprocal is generated from the estimated reciprocal term and the error term.
- a method of performing a divide operation in a computer includes the step of looking up an estimated reciprocal term in a first lookup table stored in a first computer memory.
- the estimated reciprocal term corresponds to at least a portion of a given divisor.
- a reciprocal error term is looked up in a second lookup table stored in a second computer memory.
- the reciprocal error term corresponds to at least a portion of the divisor.
- a reciprocal of the divisor is generated from the estimated reciprocal term and the error term.
- a dividend is multiplied by the reciprocal of the divisor to generate a quotient.
- An apparatus for computing a reciprocal of an operand includes a first memory containing estimated reciprocal terms, and a second memory containing reciprocal error terms.
- the apparatus further includes an adder for adding a selected estimated reciprocal term from the first memory and a selected reciprocal error term from the second memory to provide the reciprocal, wherein the selected estimated reciprocal term and the selected reciprocal error term correspond to at least a portion of the operand.
- An apparatus for performing a divide operation includes a first memory containing estimated reciprocal terms, and a second memory containing reciprocal error terms.
- the apparatus also includes an adder for adding a selected estimated reciprocal term from the first memory and a selected reciprocal error term from the second memory to provide the reciprocal.
- the selected estimated reciprocal term and the selected reciprocal error term correspond to at least a portion of a divisor.
- a multiplier multiplies a dividend by the reciprocal to generate a quotient.
- FIG. 1 illustrates prior art circuitry for computing a reciprocal.
- FIG. 2 illustrates circuitry for computing a reciprocal.
- apparatus 100 generates a 15 bit reciprocal from an 11 bit input value. At least a portion of the 11 bit input value is used to lookup a value stored in read only memory (ROM) 110 .
- ROM 110 may be referred to as a “seed” ROM. A portion of the 11 bit input value is used as an address or a lookup key to ROM 110 . When read, the ROM returns a 13 bit lookup value associated with that lookup key or address. Another portion of the 11 bit operand is used to generate the 2 most significant bits of the result.
- the most significant bits (msbs) generator circuitry ( 130 ) is typically combinatorial logic. The 13 bits and the generated most significant 2 bits are then combined to provide the 15 bit reciprocal result which may be called the seed.
- the input to the seed ROM is x′ instead of x, where x represents the operand for which a reciprocal is to be determined and x′ represents a normalized value for x.
- the seed may need to be scaled or denormalized in order to reflect the reciprocal of the operand x as opposed to x′.
- FIG. 2 illustrates one embodiment of a dual lookup table implementation that accomplishes the reciprocal function with the same precision and approximately the same accuracy as the circuitry of FIG. 1.
- FIG. 2 illustrates circuitry for determining the reciprocal of a 12 bit normalized operand. The result is a 15 bit normalized reciprocal result.
- First ROM 210 contains a lookup table of estimated reciprocal terms.
- Second ROM 220 contains a lookup table of reciprocal error terms.
- one ROM might be used instead of two.
- one or both tables are realized in a logic gate implementation.
- the mechanism for storing the tables will be generally referred to as computer memory and can include nonvolatile memory such as read only memory or a logic gate implementation.
- the most significant 8 bits of the 12 bit normalized operand are used by first ROM 210 to lookup an estimated reciprocal term.
- the most significant 8 bits of the 12 bit operand are also used by msbs generator 230 to determine the 2 most significant bits of the estimated reciprocal term.
- the least 8 significant bits are used by second ROM 220 to lookup a reciprocal error term.
- the reciprocal error term is a 6 bit expression.
- Adder 240 adds the 6 bit reciprocal error term to the 15 bit estimated reciprocal term.
- the 6 bit reciprocal error term is aligned with the least significant 6 bits of the estimated reciprocal term for purposes of accomplishing the addition.
- the result is a 15 bit normalized reciprocal result.
- Msbs generator 230 examines the most significant bit of the 12 bit normalized input value. If this most significant bit is a “1” then msbs generator 230 will output a “01”. If this most significant bit is a “0” then msbs generator 230 will output a “10”. Msbs generator 230 can be realized using combinatorial logic. The 2 bits generated become the 2 most significant bits of the 15 bit normalized reciprocal result.
- circuitry of FIG. 2 has the same number of bits of precision and approximately the same level of accuracy as the circuitry of FIG. 1. In order to achieve this level of accuracy a 2048 ⁇ 13 bit (26,624 bit) ROM was required in FIG. 1.
- the embodiment in FIG. 2 uses a 256 ⁇ 13 bit estimated reciprocal ROM and a 256 ⁇ 6 bit error ROM. This equates to a total of 256 ⁇ (13+6), or 4864 bits of storage.
- the lookup tables of FIG. 2 require considerably less storage than the lookup table in FIG. 1. This equates to a substantial savings in area required for implementation on an integrated circuit.
- the embodiment shown in FIG. 2 utilizes unsigned integers.
- the entries in ROMs 210 and 220 are treated as positive numbers. Because the normalized reciprocal result is generated by adding the estimated reciprocal and error terms, the reciprocal will be approached from the lower side. In other words, the estimated reciprocal term is less than or equal to the actual reciprocal. The correct value is approximated by adding a positive error term to the estimated reciprocal term.
- Groups or ranges of normalized operands will share the same error term because only a portion of the normalized operand's 12 bits is used for addressing or looking up entries in the error ROM. Because only 8 of the bits are used, all 12 bit normalized operands having the same 4 most significant bits and the same least significant 4 bits will share the same error term. Thus groups of 16 input values will share the same error term. Due to the address decoding structure, this also means that 16 consecutive entries in seed ROM 210 will be associated with the same error term.
- the data stored in the estimated reciprocal lookup table and in the reciprocal error lookup table is dependent to a great extent upon the normalization routine used.
- 32 bit data word 250 is used for the input operand.
- the 32 bit data word has bit positions ranging from bit 0 to bit 31 where bit 31 represents the most significant bit position.
- bit position 31 contains a “1”.
- the input operand is left-shifted until the most significant “1” bit of the original input operand is in bit position 31 .
- Shifting is well known in the art and is not illustrated in FIG. 2.
- the result of 32 minus the bit position of the most significant “1” in the original operand is stored in a scale register and is referred to as the scale factor. This value is equivalent to the number of positions shifted plus one in this embodiment.
- Bits 30 to 19 are the normalized 12 bits ( 260 ) provided as an input to the reciprocal computation.
- the normalized reciprocal must be denormalized so that it corresponds to the reciprocal of the original input operand as opposed to the reciprocal of the normalized operand. This is accomplished by right-shifting the normalized reciprocal.
- the number of positions to right-shift depends upon the position of the radix point of the input operand and the scale factor. For example, if the input radix point is between bits 16 and 15 , the reciprocal result is determined by right-shifting the normalized reciprocal by (scale factor ⁇ 2) positions.
- the shifting circuitry is well known in the art and is not illustrated in FIG. 1. In other words, the normalized reciprocal result will be shifted m bit positions to the right, where m is determined from the scale factor and the location of the radix of the input operand.
- the data stored in the estimated reciprocal lookup table is based upon this normalization scheme and the observation that 16 consecutive input operand values will share the same entry in the estimated reciprocal lookup table.
- To determine the estimated reciprocal lookup table entries first the 2 12 possible input operand values are treated as groups of 16 consecutive values. The input operand values within each group are treated as floating point numbers. Next the input operand values are inverted (i.e., actual reciprocals are computed). The 13 bit estimated reciprocal value is chosen as the value which is less than or equal to the smallest actual reciprocal result. This process is repeated for all 28 groups.
- the data stored in the reciprocal error lookup table is based upon the same normalization scheme. For each set of 16 consecutive seed ROM entries, an error term is calculated. The error term is determined by selecting the smallest error term such that the estimated reciprocal value plus the error term is less than or equal to the actual reciprocal corresponding to each of the 16 consecutive seed ROM entries.
- Appendix 1 contains a listing of a program for generating lookup table data for the seed ROM and the error term ROM.
- Subroutine “main” is used to generate the estimated reciprocal table and to print out the results (if desired).
- Subroutine “main” also contains provisions to print out the error terms.
- the program was written in Gnu C++ for execution on a Sun Microsystems (Mountainview, Calif.) workstation using the Sun Solaris operating system.
- Appendix 2 contains a listing of a program for emulating the hardware seed ROM. The data that would be placed into a ROM is loaded into an array. The program prints the seed value, scale factor, and the actual inverse for a series of input values.
- a divide operation can be accomplished by multiplying the intended dividend by the reciprocal of the divisor, as calculated above.
- the divisor is normalized.
- a normalized reciprocal is then calculated.
- the normalized reciprocal is denormalized by using the scale factor to generate the result.
- Another operand and the reciprocal are then sent to the hardware multipliers (not shown).
Abstract
Description
- This pertains to the field of computers. In particular, this pertains to methods and apparatus for computing reciprocals and performing divide operations.
- Computers are well known for their ability to perform mathematical functions at a high rate of speed. In order to achieve such speed, computers use various techniques to perform mathematical functions on given operands. Limitations in the representation of numbers and in numerical calculations within the computer lead to computed results that are only accurate to within some known or estimated error term. The ability to perform basic functions such as addition, subtraction, multiplication, and division to a consistent accuracy in a timely manner is a consideration in computational performance of the computer system.
- One technique for implementing a mathematical function is to implement the function hardware. Math functions including addition and subtraction might be accomplished by performing the steps of 1) loading registers with the operands; 2) executing the operational code (“opcode”) for the desired function; and then, 3) retrieving the result from one or more registers. The opcode is simply a command to perform the function implemented in hardware.
- One disadvantage of this prior art technique is that some functions, such as division, typically require several cycles for completion which in turn tends to adversely affect system performance.
-
- often involves looking the result up in a table. For example, the steps for computing a reciprocal might include 1) scaling or normalizing an operand; 2) looking up a reciprocal value corresponding to the normalized value in a lookup “seed” table; and 3) scaling or denormalizing the looked up value to reflect the reciprocal of the denormalized operand. The lookup table is typically stored in a read only memory (ROM). The error associated with this technique is controlled to a great extent by the “height” and “width” or the precision of the entry in the lookup table.
- One disadvantage of the conventional lookup approach is that the area required to store the lookup table on the integrated circuit may prevent the lookup approach from being a cost effective technique for given precision requirements.
- A cost effective way of implementing mathematical functions such as reciprocal or division to within a consistent error margin in a give time frame is a desirable feature of a computer system.
- Methods and apparatus for computing reciprocals and performing divide operations in a computer are described.
- A reciprocal of an operand is determined by looking up an estimated reciprocal term in a first lookup table stored in a first computer memory wherein the estimated reciprocal term corresponds to at least a portion of the operand. An error term is looked up in a second lookup table stored in a second computer memory. The error term corresponds to at least a portion of the operand. The reciprocal is generated from the estimated reciprocal term and the error term.
- A method of performing a divide operation in a computer includes the step of looking up an estimated reciprocal term in a first lookup table stored in a first computer memory. The estimated reciprocal term corresponds to at least a portion of a given divisor. A reciprocal error term is looked up in a second lookup table stored in a second computer memory. The reciprocal error term corresponds to at least a portion of the divisor. A reciprocal of the divisor is generated from the estimated reciprocal term and the error term. A dividend is multiplied by the reciprocal of the divisor to generate a quotient.
- An apparatus for computing a reciprocal of an operand includes a first memory containing estimated reciprocal terms, and a second memory containing reciprocal error terms. The apparatus further includes an adder for adding a selected estimated reciprocal term from the first memory and a selected reciprocal error term from the second memory to provide the reciprocal, wherein the selected estimated reciprocal term and the selected reciprocal error term correspond to at least a portion of the operand.
- An apparatus for performing a divide operation includes a first memory containing estimated reciprocal terms, and a second memory containing reciprocal error terms. The apparatus also includes an adder for adding a selected estimated reciprocal term from the first memory and a selected reciprocal error term from the second memory to provide the reciprocal. The selected estimated reciprocal term and the selected reciprocal error term correspond to at least a portion of a divisor. A multiplier multiplies a dividend by the reciprocal to generate a quotient.
- Other objects, features, and advantages of the present invention will be apparent from the accompanying drawings and from the detailed description that follows below.
- The present invention is illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:
- FIG. 1 illustrates prior art circuitry for computing a reciprocal.
- FIG. 2 illustrates circuitry for computing a reciprocal.
- One method of accomplishing a division such as f(x,y)={fraction (y/x)} is to take the reciprocal of x and multiply it by y. This requires an ability to multiply and means for determining the reciprocal of x. Hardware multipliers, however, tend to be more prevalent and complete an operation faster than hardware dividers.
- FIG. 1 illustrates one embodiment of prior art circuitry for a lookup table implementation of a reciprocal function (e.g., f(x)={fraction (1/x)}). As shown,
apparatus 100 generates a 15 bit reciprocal from an 11 bit input value. At least a portion of the 11 bit input value is used to lookup a value stored in read only memory (ROM) 110.ROM 110 may be referred to as a “seed” ROM. A portion of the 11 bit input value is used as an address or a lookup key toROM 110. When read, the ROM returns a 13 bit lookup value associated with that lookup key or address. Another portion of the 11 bit operand is used to generate the 2 most significant bits of the result. The most significant bits (msbs) generator circuitry (130) is typically combinatorial logic. The 13 bits and the generated most significant 2 bits are then combined to provide the 15 bit reciprocal result which may be called the seed. - Although the normalization function is not illustrated in FIG. 1, typically, the input to the seed ROM is x′ instead of x, where x represents the operand for which a reciprocal is to be determined and x′ represents a normalized value for x. The seed (reciprocal result) may need to be scaled or denormalized in order to reflect the reciprocal of the operand x as opposed to x′.
- One disadvantage of this technique is that the table, as implemented in read only memory, can take up a relatively large amount of silicon. In the embodiment illustrated, a 2048 by 13 bit ROM is necessary to calculate a 15 bit reciprocal that is within approximately 2−10.5 of the true or actual reciprocal.
- FIG. 2 illustrates one embodiment of a dual lookup table implementation that accomplishes the reciprocal function with the same precision and approximately the same accuracy as the circuitry of FIG. 1.
- FIG. 2 illustrates circuitry for determining the reciprocal of a 12 bit normalized operand. The result is a 15 bit normalized reciprocal result.
First ROM 210 contains a lookup table of estimated reciprocal terms.Second ROM 220 contains a lookup table of reciprocal error terms. In an alternative embodiment, one ROM might be used instead of two. In other embodiments, one or both tables are realized in a logic gate implementation. The mechanism for storing the tables will be generally referred to as computer memory and can include nonvolatile memory such as read only memory or a logic gate implementation. - The most significant 8 bits of the 12 bit normalized operand are used by
first ROM 210 to lookup an estimated reciprocal term. The most significant 8 bits of the 12 bit operand are also used bymsbs generator 230 to determine the 2 most significant bits of the estimated reciprocal term. The least 8 significant bits are used bysecond ROM 220 to lookup a reciprocal error term. In this embodiment the reciprocal error term is a 6 bit expression.Adder 240 adds the 6 bit reciprocal error term to the 15 bit estimated reciprocal term. The 6 bit reciprocal error term is aligned with the least significant 6 bits of the estimated reciprocal term for purposes of accomplishing the addition. The result is a 15 bit normalized reciprocal result. -
Msbs generator 230 examines the most significant bit of the 12 bit normalized input value. If this most significant bit is a “1” then msbsgenerator 230 will output a “01”. If this most significant bit is a “0” then msbsgenerator 230 will output a “10”.Msbs generator 230 can be realized using combinatorial logic. The 2 bits generated become the 2 most significant bits of the 15 bit normalized reciprocal result. - One advantage of the circuitry of FIG. 2 is the difference in the size of the lookup tables. In this embodiment, the circuitry of FIG. 2 has the same number of bits of precision and approximately the same level of accuracy as the circuitry of FIG. 1. In order to achieve this level of accuracy a 2048×13 bit (26,624 bit) ROM was required in FIG. 1. The embodiment in FIG. 2, however, uses a 256×13 bit estimated reciprocal ROM and a 256×6 bit error ROM. This equates to a total of 256×(13+6), or 4864 bits of storage. Thus the lookup tables of FIG. 2 require considerably less storage than the lookup table in FIG. 1. This equates to a substantial savings in area required for implementation on an integrated circuit.
- The embodiment shown in FIG. 2 utilizes unsigned integers. In other words, the entries in
ROMs - Although a 12 bit operand (normalized) is available to the estimate and error ROMs, only a portion of the 12 bits is used for addressing or looking up entries in estimated reciprocal and error lookup tables. This indicates that groups or ranges of normalized operands will share the same estimated reciprocal entry. In this embodiment, all12 bit input normalized operands having the same 8 most significant bits will share the same estimated reciprocal table entry or “seed”. This means that the least significant 4 bits play no role in selecting the seed, so 16 consecutive input values share the same seed value.
- Groups or ranges of normalized operands will share the same error term because only a portion of the normalized operand's 12 bits is used for addressing or looking up entries in the error ROM. Because only 8 of the bits are used, all 12 bit normalized operands having the same 4 most significant bits and the same least significant 4 bits will share the same error term. Thus groups of 16 input values will share the same error term. Due to the address decoding structure, this also means that16 consecutive entries in
seed ROM 210 will be associated with the same error term. - Referring to FIG. 2, the data stored in the estimated reciprocal lookup table and in the reciprocal error lookup table is dependent to a great extent upon the normalization routine used.
- In this embodiment 32
bit data word 250 is used for the input operand. The 32 bit data word has bit positions ranging from bit 0 to bit 31 wherebit 31 represents the most significant bit position. - To normalize the input operand, the operand is left-shifted until
bit position 31 contains a “1”. In other words, the input operand is left-shifted until the most significant “1” bit of the original input operand is inbit position 31. Shifting is well known in the art and is not illustrated in FIG. 2. The result of 32 minus the bit position of the most significant “1” in the original operand is stored in a scale register and is referred to as the scale factor. This value is equivalent to the number of positions shifted plus one in this embodiment.Bits 30 to 19 are the normalized 12 bits (260) provided as an input to the reciprocal computation. - After the reciprocal has been calculated, the normalized reciprocal must be denormalized so that it corresponds to the reciprocal of the original input operand as opposed to the reciprocal of the normalized operand. This is accomplished by right-shifting the normalized reciprocal. The number of positions to right-shift depends upon the position of the radix point of the input operand and the scale factor. For example, if the input radix point is between
bits - The data stored in the estimated reciprocal lookup table is based upon this normalization scheme and the observation that 16 consecutive input operand values will share the same entry in the estimated reciprocal lookup table. To determine the estimated reciprocal lookup table entries, first the 212 possible input operand values are treated as groups of 16 consecutive values. The input operand values within each group are treated as floating point numbers. Next the input operand values are inverted (i.e., actual reciprocals are computed). The 13 bit estimated reciprocal value is chosen as the value which is less than or equal to the smallest actual reciprocal result. This process is repeated for all 28 groups.
- The data stored in the reciprocal error lookup table is based upon the same normalization scheme. For each set of16 consecutive seed ROM entries, an error term is calculated. The error term is determined by selecting the smallest error term such that the estimated reciprocal value plus the error term is less than or equal to the actual reciprocal corresponding to each of the 16 consecutive seed ROM entries.
-
Appendix 1 contains a listing of a program for generating lookup table data for the seed ROM and the error term ROM. Subroutine “main” is used to generate the estimated reciprocal table and to print out the results (if desired). Subroutine “main” also contains provisions to print out the error terms. The program was written in Gnu C++ for execution on a Sun Microsystems (Mountainview, Calif.) workstation using the Sun Solaris operating system. -
Appendix 2 contains a listing of a program for emulating the hardware seed ROM. The data that would be placed into a ROM is loaded into an array. The program prints the seed value, scale factor, and the actual inverse for a series of input values. - Although a 32 bit data word has been used in examples presented above, other data word sizes can be used. Examples of common data word sizes include 8, 16, 32, 64, and 128 bit data words. As discussed above, the “height” and “width” of the tables control the precision and accuracy of the resulting computed reciprocal. For the 32 bit example presented above, only 12 bits were used for accessing the lookup tables. (Inherent in the table data, however, was the assumption of a 13th bit that had a value of “1”—in this
case bit 31.) Only 8 bits were used to access the ROMs, and therefore only 28 (i.e., 256) unique addresses could be looked up. For greater accuracy, more bits from the data word could be used. This would permit accessing larger (i.e., “longer” or “higher” data tables). In other words, the tables could have more entries in them. The number of bits (i.e., “width”) for each entry in the tables can also be varied. A greater number of bits for each entry will increase the precision of the result. The tradeoff, however, is that a greater number of entries or an increase in the number of bits for each entry will require more storage space and therefore will require additional area on an integrated circuit. - Assuming that multipliers are available, a divide operation can be accomplished by multiplying the intended dividend by the reciprocal of the divisor, as calculated above. Referring to FIG. 2, the divisor is normalized. A normalized reciprocal is then calculated. The normalized reciprocal is denormalized by using the scale factor to generate the result. Another operand and the reciprocal are then sent to the hardware multipliers (not shown).
- In the preceding detailed description, the invention is described with reference to specific exemplary embodiments thereof. It will, however, be evident that various modifications and changes may be made thereto without departing from the broader spirit and scope of the invention as set forth in the claims. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense.
Claims (26)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US09/867,851 US6446106B2 (en) | 1995-08-22 | 2001-05-29 | Seed ROM for reciprocal computation |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US08/517,902 US6240338B1 (en) | 1995-08-22 | 1995-08-22 | Seed ROM for reciprocal computation |
US09/867,851 US6446106B2 (en) | 1995-08-22 | 2001-05-29 | Seed ROM for reciprocal computation |
Related Parent Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US08/517,902 Continuation US6240338B1 (en) | 1995-08-22 | 1995-08-22 | Seed ROM for reciprocal computation |
Publications (2)
Publication Number | Publication Date |
---|---|
US20010027461A1 true US20010027461A1 (en) | 2001-10-04 |
US6446106B2 US6446106B2 (en) | 2002-09-03 |
Family
ID=24061709
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US08/517,902 Expired - Lifetime US6240338B1 (en) | 1995-08-22 | 1995-08-22 | Seed ROM for reciprocal computation |
US09/867,851 Expired - Lifetime US6446106B2 (en) | 1995-08-22 | 2001-05-29 | Seed ROM for reciprocal computation |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US08/517,902 Expired - Lifetime US6240338B1 (en) | 1995-08-22 | 1995-08-22 | Seed ROM for reciprocal computation |
Country Status (3)
Country | Link |
---|---|
US (2) | US6240338B1 (en) |
AU (1) | AU6770796A (en) |
WO (1) | WO1997008615A1 (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20020055962A1 (en) * | 1999-11-12 | 2002-05-09 | Richard Schroeppel | Automatically solving equations in finite fields |
US20040186873A1 (en) * | 2003-03-21 | 2004-09-23 | Intel Corporation | Narrow data path for very high radix division |
US7346642B1 (en) * | 2003-11-14 | 2008-03-18 | Advanced Micro Devices, Inc. | Arithmetic processor utilizing multi-table look up to obtain reciprocal operands |
US20080091754A1 (en) * | 2006-10-10 | 2008-04-17 | Olympus Corporation | Reciprocal calculation unit and reciprocal calculation method |
US20080201396A1 (en) * | 2007-02-15 | 2008-08-21 | Fujitsu Limited | Signal processing apparatus and the correcting method |
US20090164543A1 (en) * | 2007-12-23 | 2009-06-25 | Vinodh Gopal | Apparatus and method to compute reciprocal approximations |
Families Citing this family (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6240338B1 (en) * | 1995-08-22 | 2001-05-29 | Micron Technology, Inc. | Seed ROM for reciprocal computation |
US6223198B1 (en) * | 1998-08-14 | 2001-04-24 | Advanced Micro Devices, Inc. | Method and apparatus for multi-function arithmetic |
US6735610B1 (en) * | 1999-04-29 | 2004-05-11 | Walter E. Pelton | Apparatus, methods, and computer program products for determining the coefficients of a function with decreased latency |
JP3447614B2 (en) * | 1999-05-31 | 2003-09-16 | 株式会社東芝 | Fraction calculator, graphic setup engine, fraction calculation method, and machine-readable storage medium |
US6922712B2 (en) | 2000-02-26 | 2005-07-26 | Walter E. Pelton | Apparatus, methods, and computer program products for accurately determining the coefficients of a function |
AU2001275444A1 (en) * | 2000-06-09 | 2001-12-17 | K. Walt Herridge | Apparatus, methods and computer program products for performing high speed division calculations |
WO2001095142A2 (en) * | 2000-06-09 | 2001-12-13 | Pelton Walter E | Methods for reducing the number of computations in a discrete fourier transform |
US6732134B1 (en) * | 2000-09-11 | 2004-05-04 | Apple Computer, Inc. | Handler for floating-point denormalized numbers |
US7007058B1 (en) * | 2001-07-06 | 2006-02-28 | Mercury Computer Systems, Inc. | Methods and apparatus for binary division using look-up table |
US7065546B2 (en) * | 2002-04-09 | 2006-06-20 | Sony Electronics Inc. | Method of performing quantization within a multimedia bitstream utilizing division-free instructions |
US20070083586A1 (en) * | 2005-10-12 | 2007-04-12 | Jianjun Luo | System and method for optimized reciprocal operations |
US10133553B2 (en) * | 2015-02-22 | 2018-11-20 | The Regents Of The University Of Michigan | Reciprocal unit |
US20200311554A1 (en) * | 2019-03-27 | 2020-10-01 | International Business Machines Corporation | Permutation-invariant optimization metrics for neural networks |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4413326A (en) * | 1978-10-18 | 1983-11-01 | Honeywell Inc. | Floating point division control |
US4823301A (en) * | 1987-10-22 | 1989-04-18 | Tektronix, Inc. | Method and circuit for computing reciprocals |
JPH02156328A (en) * | 1988-12-08 | 1990-06-15 | Toshiba Corp | Reciprocal circuit |
US5249149A (en) * | 1989-01-13 | 1993-09-28 | International Business Machines Corporation | Method and apparatus for performining floating point division |
US5068816A (en) * | 1990-02-16 | 1991-11-26 | Noetzel Andrew S | Interplating memory function evaluation |
US5079716A (en) * | 1990-05-01 | 1992-01-07 | Globe-Union, Inc. | Method and apparatus for estimating a battery temperature |
US5157624A (en) * | 1990-12-13 | 1992-10-20 | Micron Technology, Inc. | Machine method to perform newton iterations for reciprocal square roots |
US5220524A (en) * | 1990-12-13 | 1993-06-15 | Micron Technology, Inc. | Machine method to perform newton iterations for reciprocals |
US5206823A (en) * | 1990-12-13 | 1993-04-27 | Micron Technology, Inc. | Apparatus to perform Newton iterations for reciprocal and reciprocal square root |
US5305248A (en) * | 1993-04-23 | 1994-04-19 | International Business Machines Corporation | Fast IEEE double precision reciprocals and square roots |
US5341321A (en) * | 1993-05-05 | 1994-08-23 | Hewlett-Packard Company | Floating point arithmetic unit using modified Newton-Raphson technique for division and square root |
US5499272A (en) * | 1994-05-31 | 1996-03-12 | Ericsson Ge Mobile Communications Inc. | Diversity receiver for signals with multipath time dispersion |
US5729481A (en) * | 1995-03-31 | 1998-03-17 | International Business Machines Corporation | Method and system of rounding for quadratically converging division or square root |
US6240338B1 (en) * | 1995-08-22 | 2001-05-29 | Micron Technology, Inc. | Seed ROM for reciprocal computation |
-
1995
- 1995-08-22 US US08/517,902 patent/US6240338B1/en not_active Expired - Lifetime
-
1996
- 1996-08-08 WO PCT/US1996/012999 patent/WO1997008615A1/en active Application Filing
- 1996-08-08 AU AU67707/96A patent/AU6770796A/en not_active Abandoned
-
2001
- 2001-05-29 US US09/867,851 patent/US6446106B2/en not_active Expired - Lifetime
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20020055962A1 (en) * | 1999-11-12 | 2002-05-09 | Richard Schroeppel | Automatically solving equations in finite fields |
US20040186873A1 (en) * | 2003-03-21 | 2004-09-23 | Intel Corporation | Narrow data path for very high radix division |
US7167891B2 (en) * | 2003-03-21 | 2007-01-23 | Intel Corporation | Narrow data path for very high radix division |
US7346642B1 (en) * | 2003-11-14 | 2008-03-18 | Advanced Micro Devices, Inc. | Arithmetic processor utilizing multi-table look up to obtain reciprocal operands |
US20080091754A1 (en) * | 2006-10-10 | 2008-04-17 | Olympus Corporation | Reciprocal calculation unit and reciprocal calculation method |
US20080201396A1 (en) * | 2007-02-15 | 2008-08-21 | Fujitsu Limited | Signal processing apparatus and the correcting method |
US20090164543A1 (en) * | 2007-12-23 | 2009-06-25 | Vinodh Gopal | Apparatus and method to compute reciprocal approximations |
US8301680B2 (en) * | 2007-12-23 | 2012-10-30 | Intel Corporation | Apparatus and method to compute reciprocal approximations |
Also Published As
Publication number | Publication date |
---|---|
WO1997008615A1 (en) | 1997-03-06 |
AU6770796A (en) | 1997-03-19 |
US6240338B1 (en) | 2001-05-29 |
US6446106B2 (en) | 2002-09-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US6240338B1 (en) | Seed ROM for reciprocal computation | |
US8028015B2 (en) | Method and system for large number multiplication | |
US5570310A (en) | Method and data processor for finding a logarithm of a number | |
US6360241B1 (en) | Computer method and apparatus for division and square root operations using signed digit | |
US5515308A (en) | Floating point arithmetic unit using modified Newton-Raphson technique for division and square root | |
Weber | The accelerated integer GCD algorithm | |
US20140067889A1 (en) | Datapath circuit for digital signal processors | |
US3828175A (en) | Method and apparatus for division employing table-lookup and functional iteration | |
EP0938042A2 (en) | High accuracy estimates of elementary functions | |
US5307303A (en) | Method and apparatus for performing division using a rectangular aspect ratio multiplier | |
US5463574A (en) | Apparatus for argument reduction in exponential computations of IEEE standard floating-point numbers | |
US20080071850A1 (en) | Methods and apparatus for extracting integer remainders | |
US20130185345A1 (en) | Algebraic processor | |
US9170776B2 (en) | Digital signal processor having instruction set with a logarithm function using reduced look-up table | |
JPH11353158A (en) | Division and/or extraction of square root arithmetic circuit | |
US7412476B2 (en) | Decimal multiplication for superscaler processors | |
EP1223505B1 (en) | Partial match partial output cache for computer arithmetic operations | |
US6711596B1 (en) | Method and apparatus for determining the approximate valve of a logarithmic function | |
Buell et al. | A multiprecise integer arithmetic package | |
US20020116431A1 (en) | System and method for improving the accuracy of reciprocal operations performed by a floating-point unit | |
US5648924A (en) | Method and apparatus for finding arctangents | |
US20070124357A1 (en) | Methods and apparatus fro performing calculations using reduced-width data | |
US5768171A (en) | Method and apparatus for improving the precision or area of a memory table used in floating-point computations | |
US20040015882A1 (en) | Branch-free software methodology for transcendental functions | |
US7167885B2 (en) | Emod a fast modulus calculation for computer systems |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: MICRON TECHNOLOGY, INC., IDAHO Free format text: MERGER;ASSIGNOR:RENDITION, INC.;REEL/FRAME:013019/0458 Effective date: 19980911 |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |
|
FPAY | Fee payment |
Year of fee payment: 4 |
|
FPAY | Fee payment |
Year of fee payment: 8 |
|
FPAY | Fee payment |
Year of fee payment: 12 |
|
AS | Assignment |
Owner name: U.S. BANK NATIONAL ASSOCIATION, AS COLLATERAL AGENT, CALIFORNIA Free format text: SECURITY INTEREST;ASSIGNOR:MICRON TECHNOLOGY, INC.;REEL/FRAME:038669/0001 Effective date: 20160426 Owner name: U.S. BANK NATIONAL ASSOCIATION, AS COLLATERAL AGEN Free format text: SECURITY INTEREST;ASSIGNOR:MICRON TECHNOLOGY, INC.;REEL/FRAME:038669/0001 Effective date: 20160426 |
|
AS | Assignment |
Owner name: MORGAN STANLEY SENIOR FUNDING, INC., AS COLLATERAL AGENT, MARYLAND Free format text: PATENT SECURITY AGREEMENT;ASSIGNOR:MICRON TECHNOLOGY, INC.;REEL/FRAME:038954/0001 Effective date: 20160426 Owner name: MORGAN STANLEY SENIOR FUNDING, INC., AS COLLATERAL Free format text: PATENT SECURITY AGREEMENT;ASSIGNOR:MICRON TECHNOLOGY, INC.;REEL/FRAME:038954/0001 Effective date: 20160426 |
|
AS | Assignment |
Owner name: U.S. BANK NATIONAL ASSOCIATION, AS COLLATERAL AGENT, CALIFORNIA Free format text: CORRECTIVE ASSIGNMENT TO CORRECT THE REPLACE ERRONEOUSLY FILED PATENT #7358718 WITH THE CORRECT PATENT #7358178 PREVIOUSLY RECORDED ON REEL 038669 FRAME 0001. ASSIGNOR(S) HEREBY CONFIRMS THE SECURITY INTEREST;ASSIGNOR:MICRON TECHNOLOGY, INC.;REEL/FRAME:043079/0001 Effective date: 20160426 Owner name: U.S. BANK NATIONAL ASSOCIATION, AS COLLATERAL AGEN Free format text: CORRECTIVE ASSIGNMENT TO CORRECT THE REPLACE ERRONEOUSLY FILED PATENT #7358718 WITH THE CORRECT PATENT #7358178 PREVIOUSLY RECORDED ON REEL 038669 FRAME 0001. ASSIGNOR(S) HEREBY CONFIRMS THE SECURITY INTEREST;ASSIGNOR:MICRON TECHNOLOGY, INC.;REEL/FRAME:043079/0001 Effective date: 20160426 |
|
AS | Assignment |
Owner name: JPMORGAN CHASE BANK, N.A., AS COLLATERAL AGENT, ILLINOIS Free format text: SECURITY INTEREST;ASSIGNORS:MICRON TECHNOLOGY, INC.;MICRON SEMICONDUCTOR PRODUCTS, INC.;REEL/FRAME:047540/0001 Effective date: 20180703 Owner name: JPMORGAN CHASE BANK, N.A., AS COLLATERAL AGENT, IL Free format text: SECURITY INTEREST;ASSIGNORS:MICRON TECHNOLOGY, INC.;MICRON SEMICONDUCTOR PRODUCTS, INC.;REEL/FRAME:047540/0001 Effective date: 20180703 |
|
AS | Assignment |
Owner name: MICRON TECHNOLOGY, INC., IDAHO Free format text: RELEASE BY SECURED PARTY;ASSIGNOR:U.S. BANK NATIONAL ASSOCIATION, AS COLLATERAL AGENT;REEL/FRAME:047243/0001 Effective date: 20180629 |
|
AS | Assignment |
Owner name: MICRON TECHNOLOGY, INC., IDAHO Free format text: RELEASE BY SECURED PARTY;ASSIGNOR:MORGAN STANLEY SENIOR FUNDING, INC., AS COLLATERAL AGENT;REEL/FRAME:050937/0001 Effective date: 20190731 |
|
AS | Assignment |
Owner name: MICRON TECHNOLOGY, INC., IDAHO Free format text: RELEASE BY SECURED PARTY;ASSIGNOR:JPMORGAN CHASE BANK, N.A., AS COLLATERAL AGENT;REEL/FRAME:051028/0001 Effective date: 20190731 Owner name: MICRON SEMICONDUCTOR PRODUCTS, INC., IDAHO Free format text: RELEASE BY SECURED PARTY;ASSIGNOR:JPMORGAN CHASE BANK, N.A., AS COLLATERAL AGENT;REEL/FRAME:051028/0001 Effective date: 20190731 |