Patent Application
20170293471
The present disclosure relates to data processing and more specifically to devices for adding floating point numbers, devices for multiplying floating point numbers, devices for floating-point fused multiply-add operation, devices for for performing fixed point number operation, and associated converters thereof.
In information processing systems, the representation of numbers is performed by binary strings. The bits can be arranged in digits depending on the radix or base.
The numbers may be represented in various formats. The formats mostly used are the Floating Point (FP) format and the Fixed point Format (FF). In fixed point format, which includes the integer numbers, the number of fractional and integer digits is fixed. In this representation, the negative numbers are typically represented in complement format, with respect to the base. For example in binary numbers a two's complement format is used.
In floating point, the number comprises the mantissa (Ma), the base (B) and the exponent (Ex). The value (Va) represented would thus be Va=Ma*B̂Ex. Then, only the numbers Ma and Ex need to be stored. The IEEE-754 standard format is the most extensive one. The standard defines five basic formats that are named for their numeric base and the number of bits used in their interchange encoding. The typical precision of the basic binary formats is one bit more than the width of its significand (or mantissa). The extra bit of precision comes from an implied (hidden) leading 1 bit. The typical floating point number will be normalized such that the most significant bit will be a one. If the leading bit is known to be one, then it need not be encoded in the interchange format.
Systems for performing operations between such numbers may use a plurality of functional units. These units may perform numerical transformations such as arithmetic operations, format conversions, function evaluation, etc. The format used for representing the numbers with which these circuits operate completely defines the design of these circuits and, therefore their fundamental efficiency parameters such as precision, range, speed, area and power. Consequently, the format used in these system influences enormously their efficiency.
Two basic circuits that are required in the majority of such functional units are rounding circuits and two's complement circuits.
The rounding circuits are used when it is necessary to reduce the number of significant digits, both in numbers in fixed format and in the mantissa in floating point format numbers. The circuit that performs a two's complement function is used to change the sign of the number. Any improvement in the efficiency of these two circuits directly affects the efficiency of the majority of the functional units that include them.
To perform the base complement of a number, first a complement to the base minus one is performed, an operation that is performed with all the digits in parallel. Subsequently the Unit-in-the-Last-Place (ULP) digit is added to the number. In the binary case, for a circuit to perform the two's complement of a number of n bits, n inverters and an n-bit adder would be required. In case of a subtraction operation (X−Y=X+(−Y)), which actually involves a sum with the two's complement of the subtrahend, the input carry bit of the adder is typically used to add to the ULP. However, this does not mean that every time that it is required to perform the two's complement the reason is a subtraction. Such cases are the absolute value operation or the addition/subtraction of numbers in sign-magnitude representation, a representation typically used in floating point.
With respect to rounding circuits, there are various forms of rounding used. One that demonstrates significant properties and is used most is the “rounding to nearest tie to even”. In this mode, the value that it is used as final value is the value that it is closer to the real value and, in case of a tie, the even value. Using this type of rounding, an error inferior to +−0.5ULP is achieved and there is no statistical deviation in the errors.
Given a number of d1-digits, to perform a rounding operation of d2-Digits, assuming d1>d2, d1-d2 digits need to be discarded. In order for the rounding to be to the nearest number, it is important to examine the value of the most significant digit of the ones that need to be discarded (MD) and the least significant digit of the ones that remain (LD):
Therefore, the basic circuit to implement this rounding type requires an adder to add one if necessary and a circuit to calculate the sticky bit.
The rounding and base's complement circuits are required in functional units such as adders, multipliers, dividers, FMAD units, absolute value operators, format converters or precision converters etc. The additional cost, e.g. in area or delay, that such circuits pose in the aforementioned functioning units is usually substantial, mostly because they are typically in the critical path.
Various attempts have been made in the prior art to reduce the effects of these calculations, namely the two's complement, the rounding and sticky bit calculations. In certain prior art documents it has been proposed to precalculate the sticky bit or remove these operations from the critical path or reduce the overall number of rounding operations needed or combine rounding and two's complement.
It would be desirable to have circuits and methods that reduce the cost in area, delay or power in rounding to nearest circuits and/or in base's complement circuits.
The present disclosure relates to various methods and devices for avoiding or at least partly reducing this problem.
The present disclosure is directed to fixed point operations configurations and circuits that implement techniques for encoding numbers to perform “round to nearest” and base's complement functions without the need to perform an addition. Thus systems using the proposed encoding type and requiring these operations may, simultaneously, reduce area, delay and power consumption.
To this end the present disclosure focuses on the design of more efficient (faster, lower cost, lower power consumption) digital information processing systems through the use of a new family of formats or a modification of the numerical coding formats, applicable to most current formats, which implies changes in the circuits that process said formats. These formats drastically simplify rounding to the nearest and base's complement circuits, without adversely affecting the rest of the circuit.
In a first aspect, a device for performing a desired addition or subtraction operation of at least two preprocessed floating point numbers to generate a third preprocessed floating point number is disclosed. Each number may have a preprocessed significand of m+2 digits. The device may comprise an exponent data path and a significand data path. The significand data path may comprise a first input arranged to receive at most the m+1 Most Significant Digits (MSDs) of the preprocessed significand of first number and a second input arranged to receive at most the m+1 MSDs of the preprocessed significand of the second number. The significand data path may be arranged to generate at most the m+1 MSDs of the preprocessed significand of the third number. The Least Significant Digit (LSD) of all preprocessed significands may be equal to B/2, B being the base of the numerical system. In case the numerical system is binary, then B=2 and the LSD is equal to 1.
One advantage of the device is the ability to perform the aforementioned operations without using explicitly the LSD of the significand of the floating point numbers. To achieve this, the floating point numbers need to be in a preprocessed format. The proposed format may be derived from any unprocessed format, either fixed point or floating point format. In case of fixed point numbers the preprocessed format may be obtained by adding a new digit as a Least Significant Digit (LSD). The value of said digit (KD) is equal to the representation base divided by two. In case of floating point numbers, the same process takes place for the significand of the FP number.
Therefore, in principle, the preprocessed numbers need one more digit than the unprocessed ones with the same precision. However, as this KD digit (or LSD) is a constant, it does not have to be stored or transmitted explicitly. It may only be required to represent this digit in an explicit form when there is a need to perform operations (arithmetic, conversions or other type) with those numbers. Therefore, the storage and transmission of preprocessed format numbers (implied) is equivalent to the conventional one.
Furthermore, the number of values represented exactly in the two corresponding formats (preprocessed and unprocessed) shall be the same. However, the values exactly represented in each format, shall be different. For example, in a binary fixed point format with only two fractional bits, four values are exactly represented (0, 0.25, 0.5, 0.75), and in the corresponding preprocessed format (i.e., three fractional bits), also four values are exactly represented but different ones (0.125, 0.375, 0.625, 0.875). More specifically, the values exactly represented in the preprocessed format will appear exactly at the halfway points between the exact numerical representations of the unprocessed exactly represented values in the original unprocessed format. This means that the accuracy will be equivalent in both formats, but conversion between them may not be exact.
A digital system using the preprocessed format may be implemented more efficiently if the digit KD is implicit. Said digit KD may be added at the input of a processing circuit or be introduced when an operation requires its presence. On the other hand if the number needs to include explicitly the digit KD, e.g. for a subsequent operation, then the digit KD may be added at the output of a previous operation.
Summarizing, a preprocessed fixed point format is a fixed point format wherein the LSD of all numbers exactly represented in said format is equal to B/2 (i.e. one for binary radix), and the rest are rounded to one of these numbers. Thus, said LSB may be stored, transmitted, or even operated implicitly. A preprocessed floating point format is a floating point format wherein the significand is a preprocessed fixed point number.
Using preprocessed format numbers greatly simplifies the operation of rounding to “nearest” or to “nearest tie to even”. This is the principal advantage of using this format. Given a fixed point number or the significand of a floating point number of d1-digits, the rounding operation “to nearest” to d2+1-digit preprocessed format, assuming d1 and d2 are natural numbers such that d1>d2, is performed by discarding the d1-d2 LSDs (truncation). In the case of rounding “to nearest tie to even”, before operating it is necessary to check if the d1-d2 LSDs are all zero (which is typically performed by calculating the sticky bit). If so, while eliminating the d1-d2 least significant digits, the following process would be performed on the next digit:
Using preprocessed format numbers also simplifies the operation of base's complement. Due to the specific value of the LSD, the addition of 1 ULP after complementing the number to the base minus one simply returns the value of the LSD to B/2 and no carry is produced towards the rest of the digits. For example, in binary format, after one's complementing a preprocessed binary number, the LSB is equal to zero and the addition of one ULP does not produce any carry but only sets the LSB to one again. Therefore, the implementation of the base's complement of a preprocessed number only requires complementing to the base minus one all digits but the LSD which remains equal.
Implementations according to said aspect have the advantage that there is no need for a rounding up logic. The elimination of the logic for rounding up, which usually is an independent adder (incrementer) or a compound adder (adder which returns X+Y and X+Y+1) along with other control logic is made possible because the rounding “to nearest” to obtain a preprocessed number is performed, as it is explained before, only by truncation. Furthermore, there is no need for logic for computing the sticky bit. The elimination of logic for computation of the sticky bit is possible because, if alignment is required, the sticky is always one since the last hidden digit is necessarily always B/2 (digit KD). This is advantageous for rounding and for when the effective operation is a subtraction. Finally, another advantage is that no overflow may occur after rounding, since rounding up is not performed.
In the following description of embodiments it is generally considered that the floating point format uses unsigned significands and an independent sign bit, however, one skilled in the art may apply the teaching disclosed herein also for signed significands in a straightforward manner.
In some embodiments, the exponent data path may be arranged to define the effective operation between the significands according to the desired floating point operation and the signs of the inputs. Furthermore, it may be arranged to detect the floating point number with the highest exponent and generate a first shift amount to align the input significands. It may also be arranged to compute the output exponent and the output sign. Finally, it may be arranged to detect special values of the inputs, such as zero, infinity, “not a number” or denormalize numbers, and instruct the adder to produce the result accordingly. Besides, it may be arranged to detect and resolve exceptions, such as overflow or underflow, and special values, such as the previous ones, after said effective operation.
In some embodiments said preprocessed significands may be normalized. Normalisation means that except for the number zero, a real number is represented with one integer digit with a value not equal to zero and a fractional part. In those embodiments said first and second inputs may be arranged to receive the m MSDs of the fractional part of the first and second preprocessed significands, respectively.
In some embodiments the device may further comprise a third input for receiving the LSD of said first and second preprocessed significands. Alternatively, the third input may have the value of B/2, as the LSD of preprocessed significands is equal to B/2. Therefore, the entire preprocessed significand shall be used for the subsequent operations, although it was not necessary to transmit the entire significand up to the input of the device.
In floating point addition, the operation of the significand data path is generally split into various cases. In some implementations it may be split into two cases: the close path, when it computes the effective subtractions for an exponent difference |d|≦1, and the far path when it computes all the effective additions and the effective subtractions for an exponent difference |d|>1. In some implementations said significand data path, or any part of it, may be implemented using two or more parallel paths to calculate separately the cases to achieve better performance. Each sub-path performs the computation supposing a different case and a final multiplexor selects the correct result for the actual case. In the following description of embodiments it is generally considered a unified implementation of the significand data path, however, one skilled in the art may appreciate that the various modules described herein may be used in a replicated or divided form, with minor modifications, to implement them in parallel paths. Furthermore, although the following descriptions of embodiments represent circuits designed for binary logic, the person skilled in the art may apply the teaching disclosed herein also for non-binary logic circuits in a straightforward manner.
In some embodiments, the significand data path may comprise at least one adding module arranged to receive the at most m+1 MSBs of the first and second preprocessed significand. If the number is normalised then it may receive only m LSBs of the m+1 MSBs as the MSB of a normalised number is always 1 and needs not be received. Otherwise, it may receive all the m+1 MSBs. The significand data path may be arranged to receive an instruction from the exponent data path about the significand corresponding to the number with the highest exponent, the first shift amount and the effective operation. Furthermore, the significand data path may be arranged to generate a value that corresponds either to the addition or to the subtraction operation between said preprocessed significands after aligning them.
In some embodiments said at least one adding module is further arranged to generate a value that corresponds to the absolute value of the result of the effective operation between said preprocessed significands.
In some embodiments, the adding module may comprise a first shifting module arranged to receive the at most m+1 MSBs of the preprocessed significand corresponding to the number with the lowest exponent at a first input and the first shift amount at a second input and generate an output value corresponding to the right shifting of said preprocessed significand corresponding to the number with the lowest exponent. The first shifting module may further comprise a third input having the value of 1 to aggregate explicitly the LSB to the significand before shifting it. A swapping module may be used to receive an indication of the significand corresponding to the number with the lowest exponent and provide it to the first shifting module. In the case that both exponents are equal, any of the significands may be provided as the one corresponding to the lowest exponent, with no change in the functionality. For clarity in the explanation, although both exponents were equal, we will name “the significand corresponding to the number with the lowest exponent” to refer to one of significands and the opposite to refer to the other one. The first shifting module may be arranged to selectively negate the output value. Since the significand is a preprocessed number, this negation may be implemented by only inverting all bits but the LSB, and no addition is required. In some implementations, the sign bit of the significand may be included initially as the MSB of the significand whereas in others a sing bit may be added to the left of the significand before inverting it. In other implementations, the sign bit may be added after the inversion, just before operating with the number. In an alternative implementation, the significant of the floating point format may be signed and therefore negation would not be necessary.
In some embodiments the first shifting module may comprise a right shifter coupled to a conditional bit inverter. In some implementations, the right shifter is placed before the conditional bit inverter and additional logic may be required to set to one the LSB of the output after inversion if the exponents are equal since no shifting is performed and the LSB of the significand is explicitly represented. In other implementations, the right shifter, which should be implemented with sign extension, is placed after the conditional bit inverter and no additional logic is required since the LSB of the significand is added after the inverter circuit.
In some embodiments the adding module may further comprise an integer adder having a first input coupled to the output of the first shifting module and a second input arranged to receive the at most m+1 MSBs of the preprocessed significand with the highest exponent. The integer adder may be arranged to generate a value that corresponds to the result of the effective operation between said preprocessed significands after aligning them. In some implementations the integer adder may be further arranged to generate an overflow signal as an independent output, whereas others may add an extra MSB to the output. In some implementations the integer adder may produce a negative output and a sign bit may be generate. In some implementations the sign bit may be delivered as an independent output, whereas others may add it as the MSB of the output.
In some implementations the integer adder may be arranged to incorporate explicitly the LSB of the preprocessed significand with the highest exponent, which is always one, before the effective operation is performed. In other implementations the integer adder may be arranged to take into account said LSB internally when the effective operation is performed.
In some embodiments the integer adder may be arranged to selectively negate the preprocessed significand corresponding to the number with the highest exponent. This may be used when the effective operation is subtraction, a positive result is required and the exponents are equal.
In some embodiments the integer adder may comprise a conditional bit inverter to selectively negate the preprocessed significand with the highest exponent. Again, an advantage of the proposed embodiments is that in order to negate only an inversion is necessary. In some implementations, the sign bit of the significand may be included initially as the MSB of the significand whereas in others a sign bit may be added to the left of the significand before inverting it.
In some embodiments the adding module may further comprise a control circuit arranged to receive the effective operation and selectively instruct the first shifting module or the integer adder to selectively negate. The control circuit may be different according to output requirements, for example when the output is required in absolute value format or when a negative output is allowed.
In some embodiments the device may further comprise a normalization module. The normalization module of the FP adder may have a first input coupled to the output of the adding module and a second input for receiving a second shift amount. The normalization module may be arranged to generate the at most m+1 MSBs of the third preprocessed significand by selectively left or right shifting the output of the adding module. Since the output is a preprocessed number then the rounding to nearest may be performed by a simple truncation but some bias may appear after rounding.
In some embodiments the normalization module of the FP adder may further be arranged to selectively generate a value equivalent to subtracting one from the LSB of the result of the shifting operation when a selected bit or a combination of selected bits of the output of the adding module is equal to one. This arrangement allows the normalisation module to remove the bias (tie to even) when d={1, 0} and the effective operation is a subtraction, i.e. the close path case.
In some embodiments the normalization module may further be arranged to selectively generate the one complement of the result of said shifting or said subsequent subtraction. It allows a positive output, when the integer adder provides a negative output and, furthermore, removes the bias of the rounding when d=0 and the effective operation is a subtraction.
In some embodiments the normalization module may further be arranged to selectively fill the vacant positions, after a left-shifting operation, by setting them to zero or by setting the MSB of the vacant positions to zero and the rest to one or by setting the MSB of the vacant positions to one and the rest to zero.
In some embodiments, the normalization module may be arranged to selectively fill said vacant positions randomly based on the value of a selected bit or of a combination of selected bits of the first input of the normalization module when the difference of the exponents is equal to 1. In alternative implementations, said value may be any bit or combination of bits with adequate random characteristics. In other implementations, a new input may be arranged. This allows to remove any bias in the rounding when d=1.
In some embodiments, the normalization module may further be arranged to force to zero the second LSB of the value that corresponds to the third preprocessed significand when the input operands have the same exponent, the values of the second LSB of the preprocessed significands of said operands are different, and the effective operation is addition. This allows removing the bias in the rounding for the aligned sum (tie to even).
In some embodiments the device may further comprise a circuit arranged to identify the position of the leading significant bit of the output of the adding module and calculate the second shift amount to be used by the exponent data path to compute the output exponent and by the normalization module to normalize the significand.
In a second aspect, a device for performing a multiplication operation of at least two preprocessed floating point numbers to generate a third preprocessed floating point number is disclosed. Each number has a preprocessed significand of m+2 digits. The device comprises an exponent data path and a significand data path. The significand data path may comprise a first input arranged to receive at most the m+1 Most Significant Digits (MSDs) of the preprocessed significand of first number and a second input arranged to receive at most the m+1 MSDs of the preprocessed significand of the second number. The significand data path may be arranged to generate at most the m+1 MSDs of the preprocessed significand of the third number. The Least Significant Digit (LSD) of all preprocessed significands may be equal to B/2, B being the base of the numerical system. In case the numerical system is binary, then B=2 and the LSD is equal to 1.
In some embodiments, the exponent data path may be arranged to compute the output exponent and the sign of the output. Furthermore, it may be arranged to detect special values of the inputs, such as zero, infinity, “not a number” or denormalize numbers, and instruct the multiplier to produce the result accordingly. Besides, it may be arranged to detect and resolve exceptions, such as overflow or underflow, and special values, such as the previous ones, after said operation.
In some embodiments said preprocessed significands may be normalized.
In some embodiments the device may further comprise a third input for receiving the LSD of said first and second preprocessed significands. Alternatively, the third input may have the value of B/2, as the LSD of preprocessed significands is equal to B/2. Therefore, the entire preprocessed significand shall be used for the subsequent operations, although it was not necessary to transmit the entire significand up to the input of the device.
In some embodiments, the mantissa data path may comprise a fixed point multiplying module arranged to receive, at a first and a second input, the at most m+1 MSBs of the first and second preprocessed mantissas respectively. If the numbers are normalized then it may receive only m LSBs of the m+1 MSBs as the MSB of a normalized number is always 1 and needs not be received. Otherwise, it may receive all the m+1 MSBs. The fixed point multiplying module may be arranged to generate the m+2 MSBs of the value that corresponds to the multiplication operation between said preprocessed mantissas.
Implementations according to embodiments disclosed herein have the advantage that the LSB of the significands of the operands is not required explicitly, only the m+2 MSBs of the product have to be generated and there is no need for a rounding logic, including the computation of the sticky bit. In some implementations of said fixed point multiplier module, a standard fixed point multiplier having two m+2 bit input may be used by setting the LSB of said two inputs to one and the remaining bits equal to the inputs of said multiplier module whereas, in other implementations, the implicit LSB is taken into account internally to the multiplier.
In some embodiments the fixed point multiplying module may comprise a redundant multiplier arranged to receive, at a first and a second input, the at most m+1 MSBs of the first and second preprocessed mantissas respectively and generate, in a redundant representation format, the 2*m+3 MSDs of a value corresponding to the multiplication operation between said preprocessed mantissas. Furthermore, the fixed point multiplying module may comprise a conversion module, coupled to the output of said multiplying module, arranged to receive the m+2 MSDs of the output of said redundant multiplier and a carry bit, and generate an m+2 bits output corresponding to the conversion of the received redundant value to non-redundant representation format. Furthermore, the fixed point multiplying module may comprise a carry net module arranged to receive the m+1 LSDs of the output of said redundant multiplier and generate said carry bit corresponding to the output carry of the conversion of the m+1 LSDs of the output of said redundant multiplier to a non-redundant representation.
Someone skilled in the art may appreciate that the word lengths of the intermediate values of in embodiments disclosed herein guarantee the lowest rounding error. However, if a greater error is allowed those sizes may be reduced to simplify the hardware in a straightforward manner. For example, the size of the output of the redundant multiplier may be lower than 2m+3 digits, such as the input of the conversion module remains the same whereas the input of the carry net module may be reduced accordingly.
In some embodiments the redundant multiplier may comprise a partial product generator arranged to receive, at a first and a second input, the at most m+1 MSBs of the first and second preprocessed mantissas respectively and generate their partial products at an output. Furthermore, the redundant multiplier may comprise a compressor tree, having a first input coupled to the output of the partial product generator and a second input arranged to receive the at most m+1 MSBs of the first and second preprocessed mantissas, said compressor tree arranged to generate, in a redundant representation, the 2*m+3 MSDs of a value corresponding to the multiplication operation between said preprocessed mantissas at an output.
As the LSB of the preprocessed significands is equal to 1, the partial product generator is not required to generate partial products for said LSBs and they may be considered already generated. They are directly introduced in the compressor tree (externally or internally) which results in fewer operations and logic for the partial product generator.
In some embodiments the fixed point multiplying module may comprise a third input having the value of 1.
In some embodiments the device may further comprise a normalization module having an input coupled to the output of the fixed point multiplying module, wherein the normalization module is arranged to generate the at most m+1 MSBs of the third preprocessed mantissa by selecting the m+1 LSBs of its input if the MSB is equal to zero or the m+1 MSBs if said bit is equal to one.
In a third aspect, a device for performing a floating-point fused multiply-add operation among three floating point preprocessed numbers to generate a fourth preprocessed floating-point number is disclosed. Each number has a significand of m+2 digits. The device comprises an exponent data path configured to receive the exponents of the three preprocessed numbers and generate the exponent of the result of the floating-point fused multiply-add operation, and a significand data path. The significand data path comprises a multiplication path and an adding path. The multiplication path comprises a first input arranged to receive at most the m+1 Most Significant Digits (MSDs) of the preprocessed significand of the first number and a second input arranged to receive at most the m+1 Most Significant Digits (MSDs) of the preprocessed significand of the second number. The multiplication path is configured to multiply said preprocessed significands of the first and second numbers and generate a multiplication result at an output. The adding path is configured to receive at most the m+1 Most Significant Digits (MSDs) of the preprocessed significand of the third number at a first input and the multiplication result at a second input and generate the at most m+1 MSDs of the significand of the fourth preprocessed number. The Least Significant Digit (LSD) of all preprocessed significands is equal to B/2, B being the base of the numerical system. When B=2 the digits are bits.
In some embodiments the exponent data path may be arranged to define the effective operation between the third significand and the multiplication result according the signs of the inputs; compute the output exponent; compute the output sign; and detect and resolve exceptions, such as overflow or underflow, and special values of the inputs or said operation.
In some embodiments the preprocessed significands may be normalized.
In some embodiments the device may further comprise a fourth input for receiving the LSD of said first, second, and third preprocessed significands. Alternatively, the fourth input may have the value of B/2, as the LSD of preprocessed significands is equal to B/2. Therefore, the entire preprocessed significand shall be used for the subsequent operations, although it was not necessary to transmit the entire significand up to the input of the device.
In some embodiments the adding path may comprise a first shifting module, configured to receive the at most m+1 Most Significant Bits (MSBs) of the third preprocessed significand at a first input. If the number is normalized then it may receive only m LSBs of the m+1 MSBs as the MSB of a normalized number is always 1 and needs not be received. Otherwise, it may receive all the m+1 MSBs. The first shifting module may further be arranged to receive an instruction from the exponent data path about the first shift amount and the effective operation between the third preprocessed significand and the output of the multiplication path, and align them, accordingly. The adding path may further comprise an adding module, configured to add the aligned output of the first shifting module with the output of the multiplication path. In these embodiments the LSB of the third significant is not required to obtain the aligned significand.
In some embodiments the multiplication path may comprise a multiplication module, configured to receive, at an input, the at most m+1 MSBs of the significands of the first and second floating point numbers, respectively, and generate the 2*m+3 MSBs of a value corresponding to the multiplication between said preprocessed significands at an output. If the numbers are normalized then it may receive only m LSBs of the m+1 MSBs as the MSB of a normalized number is always 1 and needs not be received. Otherwise, it may receive all the m+1 MSBs.
In some embodiments the multiplication path may comprise a redundant multiplier arranged to receive, at a first and a second input, the at most m+1 MSBs of the first and second preprocessed mantissas respectively and generate, in a redundant representation format, the 2*m+3 MSDs of a value corresponding to the multiplication operation between said preprocessed mantissas. Again If the numbers are normalized then it may receive only m LSBs of the m+1 MSBs as the MSB of a normalized number is always 1 and needs not be received. Otherwise, it may receive all the m+1 MSBs.
Not only the embodiments with a multiplication module but also the embodiments with a redundant multiplier have the advantage that the LSB of the input operands is not required explicitly, and the LSD (or LSB) of the output needs not be generated. In some implementations, a standard fixed point multiplier having two m+2 bit inputs may be used by setting the LSB of said two inputs to one and the remaining bits equal to the inputs of said multiplier module whereas, in other implementations, the implicit LSB may be taken into account internally to the multiplier. Similar argument is valid for the redundant multiplier.
In some embodiments the redundant multiplier may comprise a partial product generator and a compressor tree. The partial product generator may be arranged to receive, at a first and a second input, the at most m+1 MSBs of the first and second preprocessed mantissas and generate their partial products at an output. The compressor tree may have a first input coupled to the output of the partial product generator and a second input arranged to receive the at most m+1 MSBs of the first and second preprocessed mantissas, said compressor tree arranged to generate, in a redundant representation, the 2*m+3 MSDs of a value corresponding to the multiplication operation between said preprocessed mantissas at an output. As the LSB of the preprocessed significands is equal to 1, the partial product generator is not required to generate partial products for the LSBs and they may be considered already generated. They are directly introduced in the compressor tree which results in fewer operations and logic for the partial product generator.
In some embodiments the multiplication module may further comprise a third input having the value of 1.
In some embodiments the first shifting module may be arranged to receive the at most m+1 MSBs of the third preprocessed significand at a first input and the first shift amount at a second input and generate an output value corresponding to the right shifting of said preprocessed significand.
In some embodiments the first shifting module may be arranged to selectively negate the output value. Since the significand is a preprocessed number, this negation may be implemented by only inverting all bits but the LSB, and no addition is required. In some implementations, the sign bit of the significand may be included initially as the MSB of the significand whereas in others a sign bit may be added to the left of the significand before inverting it. In other implementations, the sign bit may be added after the inversion, just before operating with the number. In an alternative implementation, the significant of the floating point format may be signed and therefore negation would not be necessary.
In some embodiments the first shifting module may further comprise a third input having the value of one to aggregate explicitly the LSB to the significand before shifting it.
In some embodiments the first shifting module may comprise a right shifter coupled to a conditional bit inverter. In some implementations, the right shifter, which should be implemented with sign extension, is placed after the conditional bit inverter and no additional logic is required since the LSB of the significand is added after the inverter circuit. In other implementations, the right shifter is placed before the conditional bit inverter but additional logic may be required add one to the LSB of the output after inversion since said output is not a preprocessed number.
In some embodiments the adding module may comprise an adder configured to receive the output of the multiplication path at a first input and the output of the first shifting module at a second input, and generate a value corresponding to the signed addition of both values at an output.
In some embodiments, said adder may be configured to receive the 2*m+3 MSBs of the multiplication of the first and second preprocessed significands at a first input and the output of the first shifting module at a second input and generate a value corresponding to a signed addition of both values at an output. In other embodiments said adder may be configured to receive the 2*m+3 MSDs of the multiplication of the first and second preprocessed significands, in a redundant representation format, at a first input and the output of the first shifting module at a second input and generate a value corresponding to the signed addition of both values at an output. Implementations according to embodiments disclosed herein may have the advantage that the LSD (or LSB) of said multiplication result is not received explicitly. In some implementations the adder may be arranged to incorporate explicitly said LSB, which is always one, before the effective operation is performed. In other implementations the adder may be arranged to take into account said LSB internally, when the effective operation is performed.
In some embodiments, said signed addition may comprise n bits, n>m, and said adder may be configured to generate the at most n−1 MSBs of said signed addition at a first output. The LSB may be implicit when it is equal to one or not required for certain cases. In some embodiments, said adder may be further configured to generate the LSB of said signed addition at a second output. In some implementations, said n bits may be aligned with the multiplication result, i.e., the LSB of said n bits has the same weight than the LSB of the multiplication result. However, in other implementations, bits with less weight may be considered, but they do not contribute to obtain more precise final result. Similarly, in other implementations, the LSB of said n bits may have more weight than the LSB of the multiplication result, but the final result may be less precise in certain cases. In some implementation, n may be equal to 3*m+6 and a signal may be generated to detect overflow. In other implementation, n may be equal to 3*m+7, and the MSB may be the sign bit and no overflow signal is required.
In some embodiments the significand data path may further comprise a normalization module having a first input coupled to the adding module and a second input for receiving a second shift amount, wherein the normalization module is arranged to generate the at most m+1 MSBs of the fourth preprocessed significand by left shifting the output of the adding module. Since the output is a preprocessed number then the rounding to nearest may be performed by a simple truncation but some bias may appear after rounding.
In some embodiments the normalization module may further be arranged to selectively generate the value equivalent to subtracting one from the LSB of the result of the shifting operation when a selected bit or a combination of selected bits is equal to one. In some implementations, this bit or bits may be selected from the first input of the normalization module. In other implementations, a new input may be arranged. This arrangement allows the normalization module to remove the bias of the rounding.
In some embodiments the normalization module may further be arranged to selectively fill the vacant positions, after a left-shifting operation, by setting them to zero or by setting the MSB of the vacant positions to zero and the rest to one or by setting the MSB of the vacant positions to one and the rest to zero. This arrangement allows the normalization module to provide the correct result in certain cases, such as when the LSB of the addition result is implicit.
In some embodiments the normalization module may be arranged to selectively fill said vacant positions randomly based on the value of a selected bit or of a combination of selected bits with adequate random characteristic. In some implementations, this bit or bits may be selected from the first input of the normalization module. In other implementations, a new input may be arranged. Such arrangements allow the normalization module to remove the bias of the rounding.
The normalization modules arranged according to some of the embodiments described herein allow performing rounding to nearest without bias in certain cases. One such case is after an FMAD operation, when the normalization requires a left shift of more than 2*m+2 bits. Filling the vacant positions to the right with zeros produces an effective rounding up and consequently some bias. Since, in this case, the LSB of the result of the addition is always one, the normalization module may be easily arranged, as described previously, to produce randomly a rounding down which eliminates said bias. If said LSB is received explicitly, this is performed by randomly subtracting one from the LSB of the shifted value. Now, if the LSB is not received explicitly this may be achieved by setting randomly either the MSB of the vacant positions to zero and the rest to one or by setting the MSB of the vacant positions to one and the rest to zero. The same solutions may be used when the operation is a sole addition and the exponent of the third input is one greater than the exponent of the other addend. We name sole addition the case when either the first or the second input is equal to one and then the FMAD operation is effectively just an addition between the third input and the input which is not one. Similarly, another case when bias may be produce is, if after a sole addition when the exponent of the third input is one lower than the exponent of the other addend, the normalization requires a left shift of more than 2*m+2 bits. In this case, the bias may be avoided by setting randomly either the MSB of the vacant positions to zero and the rest to one or by setting the MSB of the vacant positions to one and the rest to zero, since the LSB of the result of the addition is implicit and equal to one. Finally, another case is after a sole addition when the exponent of the third input and the exponent of the other addend are equal. Since, in this case, the result of the addition may be either positive or negative and its LSB is zero, the bias may be avoided by two ways. One way is by just filling the vacant positions with zeroes. Another way is by filling with zeroes and also subtracting one from the LSB of the shifted value if a selected bit, or combination of them, of the result of the sole addition is one.
In some embodiments, the normalization module may be further arranged to force to zero the second LSB of the value that corresponds to the fourth preprocessed significand when the operation is a sole addition, the third input operand and the other addend have the same exponent and sign, and the values of the second LSB of the preprocessed significands of said operands are different. This allows removing the bias in the rounding for the aligned sole addition (tie to even).
In some embodiments the normalization module may further be arranged to selectively generate the one complement of the result of said shifting or said subsequent subtraction operation. It allows a positive output, when the adding module provides a negative preprocessed number. Since it is a preprocessed number, this negation may be implemented by only inverting all bits but the LSB, and no addition is required. The adder could provide a negative unprocessed number only when performing a sole addition of two numbers with the same exponent and different sign. In this case, the bit inversion would change the sign and also remove the bias of the rounding. In alternative implementations, the significant of the floating point format may be signed and the inversion would not be necessary.
In an alternative implementation, the exponent data path may be arranged to distinguish among a fused multiply-add operation or sole multiplication or sole addition. The sole multiplication may be recognized if the third input operand is a special value zero and the device may be instructed to produce the result of a sole multiplication. In some implementations, the sole addition may be recognized if either the first or second input operand is a special value one, whereas in others, it may be recognized by an external instruction. In some implementations the multiplication path may be instructed to generate an output corresponding either to the first or second significand, if sole addition is recognized. In some implementations, the normalization module may be instructed, if sole addition is recognized, to generate an output accordingly.
In some implementations the device may further comprise a circuit arranged to identify the position of the leading significant bit of the output of the adding module and calculate the second shift amount to be used by the exponent data path to compute the output exponent and by the normalization module to normalize the significand.
In a fourth aspect, a device configured to be connected to an arithmetic unit is disclosed. Said arithmetic unit is configured to process at least a first preprocessed floating point number to generate at least a second preprocessed floating point number. Said preprocessed floating point numbers have a significand with an LSD equal to B/2, B being the base of the numerical system. The device is configured to convert an input number to said first preprocessed floating point number or said second preprocessed floating point number to an output number.
One advantage of the device is that it allows numbers represented in unprocessed format to operate in arithmetic units for preprocessed floating-point numbers, and deliver the results also in a format different from a preprocessed one.
In the following description of embodiments, it is generally considered that the fixed point numbers, both unprocessed and preprocessed, are represented in two's complement representation, but minor modifications to the disclosed embodiments are required to support other formats.
In some embodiments the device may further comprise a preprocessed-fixed-point-to-preprocessed-floating-point numbers converter for converting a preprocessed n+2-bit fixed-point number to a preprocessed floating point number having a significand of m+2 bits. The preprocessed-fixed-point-to-preprocessed-floating-point converter may comprise a shift amount calculator, an exponent calculator having a first input for receiving a third shift amount from the shift amount calculator and an output for generating the exponent of the preprocessed floating point number, and a significand calculator. The significand calculator may comprise a normalization module having a first input for receiving the n MSBs of the n+1 LSBs of the fixed-point number and a second input for receiving the third shift amount. The normalization module may be arranged to left shift the n MSBs according to said shift amount and fill the vacant positions by setting the MSB of the vacant positions to zero and the rest to one or by setting the MSB of the vacant positions to one and the rest to zero to generate the at most m+1 MSBs of the significand. The sign of the preprocessed floating point number may correspond to the MSB of the preprocessed fixed point number. Introducing such a converter before the adding module allows a number in a preprocessed fixed format to be processed by adding devices according to embodiments described herein.
In some embodiments the normalization module of the significand calculator may be arranged to randomly fill said vacant positions based on the value of a selected bit or of a combination of selected bits. In some implementations said bit (or bits) may be selected from the fixed point number. In other implementations, a new input may be arranged.
In some embodiments the normalization module of the significand calculator may be further arranged to selectively generate the one complement of the result of said shifting.
In some embodiments the device may further comprise an unprocessed-fixed-to-preprocessed-floating-point numbers converter for converting an unprocessed fixed-point number of R bits to a preprocessed floating point number having a significand of m+2 bits. The unprocessed-fixed-to-preprocessed-floating-point numbers converter may comprise a shift amount calculator, a normalization module arranged to receive the R bits of the unprocessed fixed point number and generate the at most m+1 MSBs of the significand of the preprocessed floating point number, and an exponent calculator having a first input for receiving a fourth shift amount from the shift amount calculator and an output for generating the exponent of the preprocessed floating point number. The sign of the preprocessed floating point number may correspond to the MSB of the unprocessed fixed point number. Introducing such a converter before the adding module allows a number in an unprocessed fixed format to be processable by adding devices according to embodiments described herein.
In some embodiments the normalization module of the unprocessed-fixed-to-preprocessed-floating-point numbers converter may comprise a first input for receiving the r bits of the unprocessed fixed-point number and a second input for receiving the fourth shift amount. The normalization module may be arranged to generate a value that corresponds to the at most m+1 MSBs of the preprocessed significand by left shifting the r−2 MSBs of the r−1 LSBs of the first input followed to the right by a zero bit and by filling the vacant positions with the value of the LSB of the first input.
In some embodiments the normalization module of the unprocessed-fixed-to-preprocessed-floating-point numbers converter may be further arranged to selectively generate the one complement of said value if the input is negative.
In some embodiments the normalization module of the unprocessed-fixed-to-preprocessed-floating-point numbers converter may comprise a first input for receiving the r bits of the unprocessed fixed-point number and a second input for receiving a fourth shift amount, wherein the normalization module is arranged to generate a value that corresponds to the at most m+1 MSBs of the preprocessed significand by left shifting the r−1 LSBs of the first input.
The normalization modules according to some embodiments herein, may comprise a special barrel left shifter arranged to receive a bit for filling the vacant positions. In some embodiments, the special barrel left shifter may comprise a number of successive multiplexers that is equal to the first integer greater or equal to the base 2 logarithm of a maximum shift amount [log 2(maximum shift amount)]. Each multiplexer may be arranged to perform a left shifting operation that is equal to 2̂i places, iε[0, number of multiplexers−1] and arranged to fill the vacant positions using the value of said received bit.
Furthermore, the normalisation modules according to some embodiments herein may be further arranged to selectively generate the one complement of the result of said shifting operation.
In some embodiments the exponent calculator of the unprocessed-fixed-to-preprocessed-floating-point numbers converter may be arranged to decrement, according to the fourth shift amount, a base value to obtain the exponent.
In some embodiments the exponent calculator of the unprocessed-fixed-numbers preprocessed-floating-point converter may be further arranged to detect underflow, overflow or zero values and instruct the converter to generate the output accordingly.
In some embodiments the device may further comprise a preprocessed-floating-point-to-unprocessed-fixed-point numbers converter for converting the third preprocessed floating point number to a third unprocessed fixed-point number. When the unprocessed fixed-point number has h+1 bits, the converter comprises a preprocessed-floating-point-to-preprocessed-fixed-point numbers converter having an output of h+2 bits coupled to a rounding module.
In some embodiments, the rounding module of the preprocessed-floating-point-to-unprocessed-fixed-point numbers converter may comprise an adder. Said adder may be arranged to receive, at an input, the h+1 MSBs of the output of said preprocessed-floating-point-to-preprocessed-fixed-point numbers converter and increment said input value if the LSB of said output is equal to 1. Introducing such a converter after devices according to embodiments disclosed herein allows for the result of the operations to be used by circuits functioning in unprocessed format.
In some embodiments the device may further comprise a preprocessed-floating-point-preprocessed-floating-point numbers converter for converting an initial preprocessed floating point number having a significand of j+2 bits to a subsequent preprocessed floating point number. Said subsequent preprocessed floating point number may have at least a different size of significand. This may be useful, for example, when the two operands are provided to the adder from different sources and need to have significands of equal size to allow operations between them. Accordingly, if the result of the operation needs to be converted to a floating point number having a significand of different size so that it may be used by a subsequent circuit. Therefore, the converter may be placed either before or after the FP adder, accordingly.
When the subsequent preprocessed floating point number has a significand with j+2−p bits, p<j+1 then the converter may comprise a rounding unit for removing the p+1 LSBs of the j+2 bits of the initial preprocessed significand to generate at most j+1−p MSBs of the significand of the subsequent preprocessed floating point number. The LSB of the significand of the subsequent preprocessed floating point number is equal to 1. The converter may further comprise an exponent calculator for generating the exponent of the subsequent preprocessed floating point number.
When the subsequent preprocessed floating point number has a significand with j+2+q bits then the converter may comprise a filling module, arranged to receive the at most j+1 MSBs of the significand of the initial preprocessed floating point number and generate the at most j+q+1 MSBs of the significand of the subsequent preprocessed floating point number by setting the MSB of the q LSBs to one or zero and the remaining q−1 bits of said q LSBs to the complement of said MSB. The at most j+1 MSBs of the significand of the subsequent preprocessed floating point number may be the same as the at most j+1 MSBs of the significand of the initial preprocessed floating point number The converter may further comprise an exponent calculator for generating the exponent of the subsequent preprocessed floating point number.
In some embodiments the filling module of the preprocessed-floating-point-to-preprocessed-floating-point numbers converter may be arranged to randomly set said MSB based on the value of a selected bit or of a combination of selected bits. In some implementations, said bit (or bits) may be selected from the significand of the initial preprocessed floating point number.
In some embodiments the device may further comprise a preprocessed-floating-point-preprocessed-fixed-point numbers converter for converting a preprocessed floating point number having a significand of f+2 bits to a preprocessed fixed-point number. Introducing such a converter after devices according to embodiments disclosed herein allows for the result of the operations to be used by circuits functioning in preprocessed fixed point format.
When the preprocessed fixed-point number comprises L bits, wherein L<f+4, the preprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise a shift amount calculator receiving the exponent of the preprocessed floating point number at an input and generating a fifth shift amount at an output. The converter may further comprise a shifting module having a first input for receiving the L−1 MSBs of the significand of the preprocessed floating point number and a second input coupled to the output of the shift amount calculator and a third input for receiving the sign of said floating point number to generate the L−1 MSBs of the preprocessed fixed-point number at an output. The LSB of said preprocessed fixed point number is equal to B/2 and may be implicit.
In some embodiments the shifting module of the preprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise an arithmetic right shifter coupled to a conditional bit inverter.
When the preprocessed fixed-point number comprises f+c+3 bits, c>0, the preprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise a shift amount calculator receiving the exponent of the preprocessed number at an input and generating a fifth shift amount at an output, and an arithmetic right shifting module having a first input coupled to the output of the shift amount calculator and arranged to generate the f+c+2 MSBs of the preprocessed fixed point number by arithmetic right shifting an intermediate f+c+2 bit value. Said intermediate value may have, from left to right, the sign bit, the f+1 MSBs of the significand of the preprocessed floating point number, and the MSB of the c LSBs set to zero and the rest to one or the MSB of the c LSBs set to one and the rest to zero.
In some embodiments, the arithmetic right shifting module may be arranged to randomly set said MSB of the c LSBs of said intermediate f+c+2 bit value based on the value of a selected bit or of a combination of selected bits. In some implementations, said bit (or bits) may be selected from the preprocessed floating point number.
In some embodiments the arithmetic right shifting module may be further arranged to selectively generate the one complement of the result of said shifting operation.
In some embodiments, the device may further comprise a unprocessed-floating-point-to-preprocessed-floating-point numbers converter for converting an unprocessed floating point number having a significand of e+2 bits to a preprocessed floating point number. Introducing this converter at some stage before a device according to embodiments described herein, allows for numbers that are not in the preprocessed format to be processed by the aforementioned devices.
When the preprocessed floating point number has a significand with e+2−d bits, d<e+1 then the unprocessed-floating-point-to-preprocessed-floating-point numbers converter may comprise a rounding unit arranged to remove the d+1 LSBs of the significand of the unprocessed floating point number to generate the e+1−d MSBs of the significand of the preprocessed floating point number. The LSB of the significand of the preprocessed floating point number is equal to one. The unprocessed-floating-point-to-preprocessed-floating-point numbers converter may further comprise an exponent calculator for generating the exponent of the preprocessed floating point number.
In some embodiments, the rounding unit of the unprocessed-floating-point-to-preprocessed-floating-point numbers converter may be further arranged to selectively set to zero the second LSB of the significand of the preprocessed floating point number if all of the d+1 LSBs of the significand of the unprocessed floating point number are equal to zero.
When the preprocessed floating point number has a significand with e+2+g bits then the unprocessed-floating-point-to-preprocessed-floating-point numbers converter may comprise a filling module, arranged to receive the significand of the unprocessed floating point number and generate the e+g+1 MSBs of the significand of the preprocessed floating point number by setting the e+2 MSBs of the preprocessed floating point number equal to the value of the e+2 bits of the significand of the unprocessed floating point number and the remaining bits to zero. The LSB of the significand of the preprocessed floating point number is equal to one. The unprocessed-floating-point-to-preprocessed-floating-point may further comprise an exponent calculator arranged to generate the exponent of the preprocessed floating point number.
In some embodiments the filling module of the unprocessed-floating-point-to-preprocessed-floating-point numbers converter may be further arranged to selectively generate the value corresponding to subtracting one from the second LSB of the said generate significand when a selected bit or a combination of selected bits of the input unprocessed significand is equal to one.
In some embodiments the device may further comprise a preprocessed-floating-point-to-unprocessed-floating-point numbers converter for converting a preprocessed floating point number of u+2 bits to an unprocessed floating point number. Introducing such a converter after devices according to embodiments described herein allows for the result of the operation to be processed by common floating point circuits.
When the unprocessed floating point number has a significand with u+2−v bits, then the converter may comprise a rounding module, arranged to receive the at most u+3−v MSBs of the significand of the preprocessed floating point number and generate the at most u+2−v bits of the significand of the unprocessed floating point number, and an exponent calculator arranged to generate the exponent of the unprocessed floating point number.
In some embodiments the rounding module of the preprocessed-floating-point-to-unprocessed-floating-point numbers converter may comprise an adder. The adder may be arranged to receive, at an input, the at most u+2−v MSBs of the significand of the preprocessed floating point number and increment said input value if the u+3−vth MSB of said significand is equal to 1, and generate an instruction to the exponent calculator, if an overflow is produced.
In some embodiments, the exponent calculator may be further arranged to increment the output exponent when said instruction from the rounding module is generated.
When the unprocessed floating point number has a significand with U+2+W bits then the preprocessed-floating-point-to-unprocessed-floating-point numbers converter may comprise a filling module, arranged to receive the at most u+1 MSBs of the significand of the preprocessed floating point number and generate the u+w+2 bits of the significand of the unprocessed floating point number by setting the MSB of the w+1 LSBs to one and the remaining bits to zero, and an exponent calculator arranged to generate the exponent of the preprocessed floating point number.
In a fifth aspect, a device for performing a desired operation of at least a first preprocessed fixed point number having n+1 digits to generate at least a second preprocessed fixed point number having z+1 digits is disclosed. The device comprises at least one arithmetic unit having a first input for receiving the n MSDs of said at least first preprocessed fixed point number. The at least one arithmetic unit is arranged to generate the z MSDs of the at least second preprocessed fixed point number. The Least Significant Digit (LSD) of all preprocessed fixed point numbers is equal to B/2, B being the base of the numerical system.
In some embodiments, the at least one arithmetic unit may further comprise at least a second input for receiving the L MSDs of a third preprocessed fixed point number having L+1 digits, wherein L≧N and the LSD is equal to B/2. One skilled in the art may appreciate that if L<N, both numbers, i.e. the first and third number, may be interchanged to fulfil said condition. Said arithmetic unit may further comprise an addition module to generate a value corresponding to the second preprocessed fixed point number. Said second preprocessed fixed point number may be the result, rounded to nearest, of the addition of the first and the third preprocessed fixed point numbers. In alternative implementations, said third preprocessed fixed point number may be a constant and may not be received explicitly. In these implementations the adding module may be further optimized to perform the addition of said constant number.
In some embodiments, the addition module may comprise an adder configured to receive the n MSBs of the first and third preprocessed fixed point number, at a first and second input, respectively. In the following embodiments the LSB of the first preprocessed fixed point number is considered implicitly to perform the addition. In alternative implementations the adder may be arranged to incorporate explicitly the LSB of said number, which is always one, augmenting by one bit the size of the adder.
When z≦n, said adder may be configured to generate the z MSBs of a value equivalent to adding said two inputs plus a carry input. Said carry input may be equal to the n+1 th MSB of the third preprocessed fixed point number, since the LSB of the first preprocessed fixed point number is one. The main advantages of this configuration is that no additional circuit is required to perform rounding to nearest of the result and even the generation of the n−z LSBs is not required. Thus, one skilled in the art may appreciate that a significant part of said adder may be optimized internally, since only the last carry signal corresponding to the addition of the n−z LSBs is required.
On the other hand, when z=n=L, the LSB of the exact result of the addition is zero and thus a rounding up is always performed which produces some bias. In this case the adding module may be further configured to set to zero the second LSB of the second preprocessed fixed point number. This additional configuration avoids said bias. Besides, the adder may be simplified since said second LSB may not be generated. In alternative implementations, to avoid said rounding up, the arithmetic unit or the device may be configured to deliver the exact result of the addition which is an unprocessed number (since the LSB is zero).
When z>n, said adder may be configured to generate the n MSBs of the second preprocessed fixed point number by producing a value equivalent to adding said two inputs plus a carry input. Said carry input may be equal to the n+1th MSB of the third preprocessed fixed point number, since the LSB of the first preprocessed fixed point number is one. The adding module may be further configured to set the n+1th MSB of the second preprocessed fixed point number equal to the inverse of the n+1th MSB of the third preprocessed fixed point number, which is equivalent to adding one to it. Said adding module may be further configured to set the remaining z−n−1 LSBs of the z MSBs of the second preprocessed fixed point number equal to the z−n−1 LSBs of the z MSBs of the third preprocessed fixed point number. The LSB of the second preprocessed fixed point number is implicit and equal to one. Again, no additional circuit is required to perform rounding to nearest of the result.
In some embodiments the adding module may be further arranged to negate one of the input numbers. As stated before, said negation is performed by inverting all bits except the LSB.
In some embodiments said negation operation may be performed selectively according to a control signal.
In other implementations, the adding module may comprise more than two inputs for receiving more than two preprocessed numbers to be added, respectively. In this case, the LSB of all input preprocessed numbers may be added to the result of the addition of the remaining bits as a constant value being the result of the addition of the LSB of all input preprocessed numbers. For instance, if the adding module is configured to receive nn preprocessed input operands, all having mm+1 bits, the result of the adding module may be obtained by adding the value nn (which is the addition of the LSB of all inputs), correctly aligned, to the result of the addition of the mm MSBs of all input numbers. If the sizes of the input numbers are not the same, the weight of each LSB needs to be taken into account to generate said constant value. On the other hand, if said constant value is odd then the result of the addition is a preprocessed number. Otherwise, the second LSB of the result may be set to zero to avoid the bias due to rounding.
Although the adding modules of the embodiments disclosed herein have the output result in non-redundant format, one skilled in the art may appreciate that the extension of these embodiments to implementations having the output in a redundant format, such as carry-save or signed-digit formats, may be performed in a straightforward manner.
In some embodiments the at least one arithmetic unit may comprise a multiplication module to generate a value corresponding to the second preprocessed fixed point number.
In some embodiments the multiplication module may be a squarer. Said in a different way the multiplication module may be configured to generate said value corresponding to the second preprocessed fixed point number which may be the result, rounded to nearest, of the square of the first preprocessed fixed point number, having the LSD equal to B/2.
When the first preprocessed fixed point number is signed, the squarer may comprise a module arranged to generate the n−1 MSBs of the magnitude (i.e., the value without sign) of the first preprocessed fixed point number. In this case, an unsigned squarer may be used to compute the magnitude of the second preprocessed fixed point number whereas the sign, which is always positive, may be added later. In alternative implementations a signed squarer may be used instead of the magnitude calculator and the unsigned squarer. In other implementations, the first approach may be used to design a combined unsigned/signed squarer.
In some embodiments the multiplication module may be configured to generate said value corresponding to the second preprocessed fixed point number which may be the result, rounded to nearest, of the multiplication of the first preprocessed fixed point number and a fourth preprocessed fixed point number of t+1 digits, having the LSD equal to B/2.
When the fourth preprocessed fixed point number is a constant number, the multiplication module may be a constant multiplier. In this case, said constant number may not be received explicitly. One skilled in the art may appreciate that any optimization technique for implementation of constant multipliers may be applied to the disclosed invention in a straightforward manner.
In some embodiments the at least one arithmetic unit may further comprise at least a second input for receiving the t MSDs of the fourth preprocessed fixed point number.
In some embodiments the multiplication module may comprise a multiplier. The multiplier may be configured to generate the n+t+1 MSBs of the result of the multiplication, since the LSB of said result is always one for preprocessed input numbers. If the multiplication module is a squarer only the 2*n MSB are required to be generated since, also, the second LSB is always zero. The multiplication module may further comprise a truncation module, coupled to the output of the multiplier for receiving the n+t+1 MSBs result and generating the z MSBs of the second number by truncating said output. The LSB of the second preprocessed fixed point number is implicit and equal to one. Again, no additional circuits are required to perform rounding to nearest of the result, such as an adder for rounding up or a sticky calculator.
Since the n+t−z+2 LSBs of the exact result of the multiplication are not required to obtain a correctly rounded second preprocessed fixed point number, the multiplier module may be optimized by avoiding the explicit generation of said n+t−z+2 LSBs. Thus, in some embodiments the multiplication module may comprise a redundant multiplication module arranged to receive, at a first input, the n MSBs of the first preprocessed fixed point numbers and generate, in a redundant representation format, at most the n+t+1 MSDs of a value corresponding to the multiplication operation between said preprocessed number and the fourth preprocessed fixed point number. The LSD of the result of said multiplication is implicit and equal to one. If the multiplication module is a squarer the second LSB is also constant but equal to zero and it may not be generated. The multiplication module may further comprise a conversion module, coupled to the output of said redundant multiplication module, arranged to receive the z MSDs of the output of said redundant multiplication module and a carry bit, and generate a z-bit output corresponding to the conversion of the received redundant value to non-redundant representation format. The multiplication module may further comprise a carry net module arranged to receive the at most n+t+1−z LSDs of the output of said redundant multiplication module and generate said carry bit corresponding to the output carry of the conversion of the n+t+1−z LSDs of the output of said redundant multiplication module to a non-redundant representation.
One skilled in the art may appreciate that the word lengths of the intermediate values of embodiments disclosed herein guarantee the lowest rounding error. However, if a greater error is allowable those sizes may be reduced to simplify the hardware in a straightforward manner. For example, the size of the output of the redundant multiplier may be lower than n+t+1 digits, such that the input of the conversion module may remain the same whereas the input of the carry net module may be reduced accordingly.
One skilled on the art may appreciate that, besides the approach described above, different optimization techniques which may take advantage of the fact that the n+t−z+2 LSBs are not required explicitly, such as truncated multipliers, may be applied to the disclosed invention in a straightforward manner.
In some embodiments the redundant multiplication module may comprise a partial product generator arranged to receive, at a first input, the n MSBs of the first preprocessed number and generate, at an output, the partial products corresponding to the multiplication of said input and the t MSBs of the fourth preprocessed fixed point number. If said fourth preprocessed fixed point number is a constant, said partial product generator may be optimized to generate a reduced set of partial products corresponding to the multiplication of said first input times said constant number without receiving the constant explicitly. If it is not a constant, said partial product generator may be arranged to receive said t MSBs. The redundant multiplication module may further comprise a compressor tree, having a first input coupled to the output of the partial product generator and a second input arranged to receive the n MSBs and the t MSBs of the first and fourth preprocessed numbers, respectively. In an alternative implementation, when the fourth preprocessed number is a constant, said t MSBs may be taken into account within the compressor tree to generate a more optimized circuit. Said compressor tree may be arranged to generate, in a redundant representation, at most the n+t+1 MSDs of a value corresponding to the multiplication operation between said preprocessed numbers at an output. As the LSB of the preprocessed numbers is equal to 1, the partial product generator is not required to generate partial products for said LSBs and they may be considered already generated. They may be directly introduced in the compressor tree (externally or internally) which results in fewer operations and logic for the partial product generator. In an alternative implementation, said LSBs may be considered within the partial product generator and said values may not be introduced at said second input of the compressor tree.
In some embodiments the arithmetic unit may comprise a left shifting module configured to generate a value corresponding to the second preprocessed fixed point number. Said second preprocessed fixed point number may be the result, rounded to nearest, of the left shifting of the first preprocessed fixed point number. Although, the left shifting operation (i.e., the multiplication by a power of the base) for unprocessed fixed-point formats is an exact operation, i.e. the result does not need any rounding, this is not true for preprocessed fixed-point formats. The exact result of left shifting a preprocessed fixed-point number is not a preprocessed number, since its LSD is not equal to B/2. Thus, a rounding operation is required, which at first may not imply any additional operation. However, this rounding may produce some bias introduced by the fact that a rounding up is always performed. In alternative implementations, to avoid said rounding up, the arithmetic unit or the device may be configured to deliver the exact result of the shifting which is an unprocessed number.
In some embodiments the left shifting module may be further arranged to selectively fill the vacant positions, after the left-shifting operation, by setting the MSB of the vacant positions to zero and the rest to one, or by setting the MSB of the vacant positions to one and the rest to zero. This configuration produces a rounding down for the former and a rounding up for the latter.
In some embodiments the left shifting module may be arranged to selectively fill said vacant positions randomly based on the value of a selected bit or of a combination of selected bits. This configuration allows avoiding bias in the rounding. In some embodiments said selected bit (or bits) may be part of the input number, while in other embodiments a new input may be configured.
In some embodiments the left shifting module may be further arranged to receive the shift amount to select the number of bits to shift.
In some embodiments the left shifting module may comprise a barrel shifter arranged to receive a bit for filling the vacant positions.
In some embodiments, the barrel shifter may comprise a number of successive multiplexers that is equal to the first integer greater or equal to the base 2 logarithm of a maximum shift amount [log 2(maximum shift amount)], each multiplexer arranged to perform a left shifting operation that is equal to 2̂i places, iε[0, number of multiplexers−1] and arranged to fill the vacant positions using the value of said received bit.
In some embodiments, at least one arithmetic unit may comprise an absolute value module to generate a value corresponding to the second preprocessed fixed point number. Said second preprocessed fixed point number may be the result of the absolute value of the first preprocessed fixed point number. This operation involves the negation of the input number if it is negative. Since the input number is preprocessed, this negation may be implemented by only inverting all bits but the LSB, and no addition is required. Thus, the absolute value module may comprise a conditional bit inverter arranged to receive, at a first input, the n MSBs of the first preprocessed number. Said conditional bit inverter may generate a value corresponding to the one complement of the first input if its MSB is equal to one.
In some implementations at least one arithmetic unit may comprise an elementary function calculator module to generate a value corresponding to the second preprocessed fixed point number. Said second preprocessed fixed point number may be the result, rounded to nearest, of applying an elementary function to the first preprocessed fixed point number. Said elementary function may be any mathematical function of one variable, such as trigonometric functions, logarithm, exponential, etc. But, one skilled in the art may appreciate that an extension to multivariable functions is straightforward. The elementary function calculator module may comprise a table lookup arranged to receive, at a first input, the n MSDs of the first preprocessed numbers. Said table lookup may be further arranged to store and to deliver the z MSDs of said second preprocessed fixed point number corresponding to each possible input. The LSD of said second preprocessed fixed point number is equal to B/2 and may be implicit. One advantage of this proposal is that the LSB of the output number does not need to be stored or delivered explicitly. Another advantage is that the value stored in the table lookup is exactly rounded to any precision below z+1 Digits, just by truncation.
In some embodiments, the device may further comprise an unprocessed-to-preprocessed fixed point numbers converter coupled at an input of the arithmetic unit configured to receive an unprocessed fixed point number of e+1 bits and generate a preprocessed fixed point number. Introducing such a converter according to embodiments disclosed herein allows a number in an unprocessed fixed-point format to be operated by said arithmetic units functioning in preprocessed fixed point format.
When the preprocessed fixed point number has e+1−k1 bits, with k1<e, then the converter may comprise a rounding unit arranged to remove the k1+1 LSBs of the unprocessed fixed point number to generate the e−k1 MSBs of the preprocessed fixed point number. The LSB of said preprocessed fixed point number is equal to B/2 and is implicit.
In some embodiments the rounding unit may be further arranged to selectively set to zero the second LSB of the preprocessed fixed point number if all of the k1+1 LSBs of the unprocessed fixed point number are equal to zero. This configuration avoids the bias due to rounding.
When the preprocessed fixed point number has e+1+k2 bits then the converter may comprise a filling module, arranged to receive the unprocessed fixed point number and generate the e+k2 MSBs of the preprocessed fixed point number by setting the e+1 MSBs of the preprocessed fixed point number equal to the value of the e+1 bits of the unprocessed fixed point number and the remaining bits to zero. The LSB of the preprocessed fixed point number is equal to one and is implicit.
In some embodiments, the filling module may be further arranged to selectively generate the value corresponding to subtracting one from the second LSB of the said preprocessed fixed point number when a selected bit or a combination of selected bits of the input unprocessed number is equal to one. This configuration avoids the bias due to rounding.
In some embodiments the device may further comprise a preprocessed-to-preprocessed fixed point numbers converter coupled at an input and/or output of the arithmetic unit and configured to receive an initial preprocessed fixed point number of j+1 bits and generate a subsequent preprocessed fixed point number of different size. This may be useful at the input, for example, when an operand is provided to the arithmetic unit with more precision (or with less precision) than needed. Accordingly, if the result of the operation needs to be converted to a number of different size so that it may be used by a subsequent circuit. Therefore, the converter may be placed either before or after the arithmetic unit, accordingly.
When the subsequent preprocessed fixed point number has j+1−p1 bits, p1<J then the converter may comprise a rounding unit for removing the p1+1 LSBs of the J+1 bits of the initial preprocessed number to generate the j−p1 MSBs of the subsequent preprocessed fixed point number. The LSB of the subsequent preprocessed fixed point number is equal to B/2 and is implicit.
When the subsequent preprocessed fixed point number has j+1+p2 bits then the converter may comprise a filling module, arranged to receive the j MSBs of the initial preprocessed fixed point number and generate the j+p2 MSBs of the subsequent preprocessed fixed point number by setting the MSB of the p2 LSBs to one or zero and the remaining p2−1 bits of said p2 LSBs to the complement of said MSB. Depending of the value of said MSB, an effective rounding up or rounding down is produced. The j MSBs of the subsequent preprocessed fixed point number may be the same as the j MSBs of the initial preprocessed fixed point number and is implicit.
In some embodiments the filling module may be further arranged to randomly set said MSB based on the value of a selected bit or on a combination of selected bits. In some implementations, said bit (or bits) may be selected from the initial preprocessed fixed point number.
In some embodiments, the device may further comprise a preprocessed-to-unprocessed fixed point numbers converter, coupled at the output of an arithmetic unit and configured to receive a preprocessed fixed point number of w+1 bits and generate an unprocessed fixed point number. Introducing such a converter according to embodiments disclosed herein allows a preprocessed number generated by said arithmetic unit to be operated by common fixed-point circuits.
When the unprocessed fixed point number has w+1−v1 bits, v1<w, then the converter may comprise a rounding module, arranged to receive the w+2−v1 MSBs of the preprocessed fixed point number and generate the w+1−v1 bits of the unprocessed fixed point number.
In some embodiments, the rounding module may comprise an adder. Said adder may be arranged to receive, at an input, the w+1−v1 MSBs of the preprocessed fixed point number and increment said input value if the w+2−v1th MSB of said preprocessed number is equal to 1. The computation of the sticky bit is not required since the input is a preprocessed number and its LSB is equal to one.
When the unprocessed fixed point number has w+1+v2 bits then the converter may comprise a filling module, arranged to receive the W MSBs of the preprocessed fixed point number and generate the w+v2+1 bits of the unprocessed fixed point number by setting the MSB of the v2+1 LSBs to one and the remaining bits to zero.
In the following embodiments of converters, it is considered that the floating point numbers, both unprocessed and preprocessed, are represented by a sign bit, an exponent and a normalized significand without sign, the MSB being equal to one and explicitly included in the significand representation. However, one skilled in the art may appreciate that other formats with a different representation may be used with minor modifications in the circuits described herein.
In some embodiments the device may further comprise a preprocessed-floating-point-to-preprocessed-fixed-point numbers converter coupled at the input of an arithmetic unit and configured to receive a preprocessed floating point number having a significand of f+2 bits and to generate a preprocessed fixed-point number. Introducing such a converter before an arithmetic unit according to embodiments disclosed herein allows a number in a preprocessed floating point format to be operated by said arithmetic units functioning in preprocessed fixed point format.
When the preprocessed fixed-point number comprises g bits, wherein g<f+4, the preprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise a shift amount calculator receiving the exponent of the preprocessed floating point number at an input and generating a shift amount at an output. The converter may further comprise a shifting module having a first input for receiving the g−1 MSBs of the significand of the preprocessed floating point number and a second input coupled to the output of the shift amount calculator and a third input for receiving the sign of said floating point number to generate the g−1 MSBs of preprocessed fixed-point number at an output. The LSB of said preprocessed fixed point number is equal to B/2 and may be implicit.
In some embodiments the shifting module of the preprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise an arithmetic right shifter coupled to a conditional bit inverter. In some embodiments the inverter is before the shifting module, in others it may be in the contrary.
When the preprocessed fixed-point number comprises f+c+3 bits, c>0, the preprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise a shift amount calculator receiving the exponent of the preprocessed number at an input and generating a shift amount at an output, and an arithmetic right shifting module having a first input coupled to the output of the shift amount calculator and arranged to generate the f+c+2 MSBs of the preprocessed fixed point number by arithmetic right shifting an intermediate f+c+2 bit value. Said intermediate value may have, from left to right, the sign bit, the f+1 MSBs of the significand of the preprocessed floating point number, and the MSB of the c LSBs set to zero and the rest to one or the MSB of the c LSBs set to one and the rest to zero.
In some embodiments, the arithmetic right shifting module may be arranged to randomly set said MSB of the c LSBs of said intermediate f+c+2 bit value based on the value of a selected bit or of a combination of selected bits. In some implementations, said bit (or bits) may be selected from the preprocessed floating point number.
In some embodiments the arithmetic right shifting module may be further arranged to selectively generate the one complement of the result of said shifting operation.
In some embodiments the device may further comprise a preprocessed-fixed-point-to-preprocessed-floating-point numbers converter coupled at an output of an arithmetic unit, and configured to convert a preprocessed q+2-bit fixed-point number to a preprocessed floating point number having a significand of m+2 bits. The preprocessed-fixed-point-to-preprocessed-floating-point numbers converter may comprise a shift amount calculator, an exponent calculator having a first input for receiving a shift amount from the shift amount calculator and an output for generating the exponent of the preprocessed floating point number, and a significand calculator. The significand calculator may comprise a normalization module having a first input for receiving the q MSBs of the q+1 LSBs of the fixed-point number and a second input for receiving the third shift amount. The normalization module may be arranged to left shift the q MSBs according to said shift amount and fill the vacant positions by setting the MSB of the vacant positions to zero and the rest to one or by setting the MSB of the vacant positions to one and the rest to zero to generate the at most m+1 MSBs of the significand. The sign of the preprocessed floating point number may correspond to the MSB of the preprocessed fixed point number. Introducing such a converter after an arithmetic unit according to embodiments disclosed herein allows a number in a preprocessed fixed format generated by it, to be processed by preprocessed FP devices.
In some embodiments the normalization module of the significand calculator may be arranged to randomly fill said vacant positions based on the value of a selected bit or of a combination of selected bits. In some implementations said bit (or bits) may be selected from the fixed point number. In other implementations, a new input may be arranged.
In some embodiments the normalization module of the significand calculator may be further arranged to selectively generate the one complement of the result of said shifting.
In some embodiments the device may further comprise a preprocessed-fixed-point-to-unprocessed-floating-point numbers converter, coupled at an output of an arithmetic unit, and configured to convert a preprocessed h+2 bit fixed-point number to an unprocessed floating point number having a significand of r+1 bits.
In some embodiments said preprocessed-fixed-point-to-unprocessed-floating-point numbers converter may comprise a shift amount calculator, an exponent calculator and a significand calculator. Said exponent calculator may have a first input for receiving a shift amount from the shift amount calculator and an output for generating the exponent of the unprocessed floating point number. The significand calculator may comprise a normalization module having a first input for receiving the h MSBs of the h+1 LSBs of the fixed-point number and a second input for receiving the shift amount. Said normalization module may be arranged to generate a value corresponding to the at most r+2 MSBs of the h+1 LSBs of the fixed point number left shifted according to said shift amount. Said significand calculator may further comprise a rounding module arranged to receive the output of the normalization module and generate the at most r+1 MSBs of the significand of the unprocessed floating point number. The sign of the unprocessed floating point number may correspond to the MSB of the preprocessed fixed point number.
In some embodiments said normalization module may be further arranged to selectively generate the negation of said at most r+2 bit value.
In some embodiments the rounding module may comprise an adder. Said adder may be arrange to receive, at an input, the at most r+1 MSBs of the output of the normalization module and to increment said input value if the LSB of said output is equal to 1.
In some embodiments the device may further comprise an unprocessed-floating-point-to-preprocessed-fixed-point numbers converter, coupled at an input of an arithmetic unit, and configured to convert an unprocessed floating point number having a significand of s bits to a preprocessed fixed-point number of A+2 bits. Introducing such a converter according to embodiments disclosed herein allows a number in a unprocessed floating point format to be operated by said arithmetic units functioning in preprocessed fixed point format.
In some embodiments said unprocessed-floating-point-to-preprocessed-fixed-point numbers converter may comprise a shift amount calculator, receiving the exponent of the unprocessed floating point number at an input and generating shift amount at an output, an unprocessed-to-preprocessed fixed point numbers converter according to embodiments disclosed herein, and a shifting module. Said unprocessed-to-preprocessed fixed point numbers converter may be arranged to receive the at most s bits of the significand of the unprocessed floating point number and to generate the A MSBs of a preprocessed fixed-point number. The shifting module may have a first input for receiving the A bit output of said converter and a second input coupled to the output of the shift amount calculator and a third input for receiving the sign of said floating point number. Said shifting module may be arranged to generate the A+1 MSBs of the output preprocessed fixed-point number by right shifting, according to the second input, the first input augmented to the left with the sign bit. The LSB of said preprocessed fixed point number is equal to B/2 and may be implicit. In some implementations the MSB of the significand of the floating point number may be implicit, since it is always equal to one, and it may not be received explicitly by the converter.
In some embodiments said shifting module may be further arranged to selectively generate a value equal to the one complement of the result of said shifting.
In some embodiments the shifting module may comprise an arithmetic right shifter coupled to a conditional bit inverter. In some embodiments the inverter is before the shifting module, in others it may be in the contrary.
In some embodiments the device may further comprise a third input and/or output for receiving and/or generating the LSD of said first and/or third preprocessed fixed-point numbers. Alternatively, said third input and/or output may have the value of B/2, as the LSD of preprocessed fixed-point numbers is equal to B/2. Therefore, the entire preprocessed number shall be used for the subsequent operations, although it was not necessary to transmit the entire number up to the input and/or output of the device.
In some embodiments, the device may comprise a plurality of arithmetic units and an operation selection input for receiving a desired operation signal. Said device may be configured to select the output of an arithmetic unit from the plurality of arithmetic units based on said received desired operation signal.
Particular embodiments of the present invention will be described in the following by way of non-limiting examples, with reference to the appended drawings, in which:
b illustrate the significand data path of a floating point (FP) multiplier according to two examples;
a and 17b illustrate example implementations of a preprocessed-floating-point-to-preprocessed-fixed-point numbers converter;
a, 19b illustrate example implementations of the significand data path of a unprocessed-floating-point-to-preprocessed-floating-point numbers converter;
Accordingly, FP adder further comprises a conditional bit inverter 125 having a first input coupled to a second output of the swap module 105 for receiving the m MSBs of the significand that is not input to the right shifter 115 and a second input coupled to the second output of the comparator module. The conditional bit inverter 125 is a conventional conditional bit inverter with no special cases as the LSB of the significand is not introduced at its input. Now, the conditional bit inverter 125 generates an m bit number. When the effective operation is subtraction and d=0, the comparator module 110 compares the input significands and instructs either the conditional bit inverter 120 or the conditional bit inverter 125 to negate the lower one. If d< >0, the conditional bit inverter 120 always negates the input to perform an effective subtraction. The FP adder 100 further comprises a two's complement adding module 130 having a first input coupled to the output of the conditional bit inverter 125 and a second input coupled to the output of the special conditional bit inverter 120. The first input receives m bits while the second input receives m+1 bits. Therefore, the two's complement adding module 130 further comprises a third input coupled to 1, so that the m bits at the output of the conditional bit inverter 125 are augmented to the right by 1 bit. However, in alternative implementations, the introduction of the additional one may be performed internally to the module 130 without the need for a special input. It is merely illustrated in the example of
It should be noted that in this implementation the m MSBs of the significand include the sign bit and the integer bit. In an alternative implementation, the sign bit may be discarded after the addition, since it is always zero and, similarly, the integer bit may be discarded after normalization, since it is always one.
The special left-shifter 245, shown in
It should be noted that in this implementation the m MSBs of the significand include the sign bit and the integer bit. In an alternative implementation, the sign bit may be extracted after the addition and, similarly, the integer bit may be discarded.
The first output of two's complement adding module 320 is coupled to a first input to shifter 335 of normalization module 330. A second input of shifter 335 is coupled to the output of LZA module 325. The FP adder 300 further comprises LZA module 325 having a first and a second input coupled to the first and second output of swap module 305, respectively, and a third input coupled to the LSB of the exponent difference. Similar to the LZA module of
The first output of two's complement adding module 420 is coupled to a first input of close rounding module 425 of normalization module 430. The normalization module 430 further comprises a special shifting module 435 having a first input coupled to the first output of close rounding module 425 for receiving m+2 bits. This is a special shifter in such a way that in a left shift of the first input, the vacant positions are filled with a third input which is coupled to a second output of close rounding module 425 for receiving 1 bit. The close rounding module 425 provides the adequate values to the special shifter module 435 to obtain correctly rounded and no biased results after normalization if the effective operation is a subtraction and the difference of exponents is less or equal to one (op=1,d={0,1}, close path case).
The shifting module 435 provides an m+1-bit output corresponding to the MSBs of the first input (m+2 bits) after shifting it one bit to the right (overflow) or shifting it to the left according to the second input, which is coupled to the output of LZA module 445. The FP adder 400 further comprises LZA module 445 having a first input coupled to the second output of swap module 405 and a second input coupled to the output of right shifter 415. Similar to the LZA module of
The significand data path 400m further comprises a far rounding module 440 having an input coupled to the output of the shifting module 435. The far rounding module 440 prevents rounding with bias in the aligned sum. The far rounding module 440 provides a m-bit bus at the output from a m+1-bit at the input.
The exponent data path comprises an exponent difference module 450 having a first input for receiving the first exponent Ex and a second input for receiving the second exponent Ey and generating, at an output, a value representing the difference of the exponents d. This value includes information relevant to the sign of the difference and the magnitude of the difference. A multiplexer 455 receives the exponents at a first and second input, respectively, and the sign of the difference of the exponents at a third input. The exponent data path further comprises an exponent update module 460 having a first input receiving the output of multiplexer 455, a second input receiving the output of LZA module 445 and a third input receiving the overflow bit from two's complement adder 420. The exponent update module generates the exponent Ez of the result of the effective operation. Further, a sign module 465 receives the sign bits Sx and Sy of the operands, the sign of the difference of the exponents (sign(d)) and the sign (sign(c)) of the difference of the significands, and generates the bit indicative of the effective operation (op) and the sign bit Sz of the result of the FP operation.
FP adder 500 further comprises conditional bit inverter 510 having a first input coupled at a first output of swap module 505 for receiving the m bits of the significand with the lowest exponent. The conditional bit inverter 510 shall carry out a bit-wise inversion of the m bits if the effective operation is a subtraction. The conditional bit inverter 510 has a second input for receiving a control bit indicating the effective operation. The FP adder 500 further comprises a right shifter 515 having a first input coupled to the output of conditional bit inverter 510 and a second input for receiving the shift amount (|d|). The right shifter 515 may further comprise a third input coupled to a logical 1 to receive the LSB of the significand. The right shifter 515 shall right shift this m+1 significand according to the shift amount received and generate a right shifted m+1 bit number. The FP adder 500 also further comprises a two's complement adding module 520 having a first input coupled to the output of the right shifter 515 and a second input coupled to a second output of the swap module 505. Similarly to the two's complement adding modules of
The FP adder 500 further comprises a second right shifter 525 having a first input coupled to the output of conditional bit inverter 510. The second right shifter 525 further comprises a second input coupled to a logical 1, so that the m bits at the output of the conditional bit inverter 510 are augmented by 1 bit. The second right shifter 525 shall right shift at most one position this m+1 number generating a right shifted m+1 bit number.
The FP adder 500 further comprises a second two's complement adding module 530 having a first input coupled to the output of second right shifter 525 and a second input coupled to the second output of swap module 505. Similarly to the adding module 520, the first input receives m+1 bits while the second input receives m bits. Therefore, the second two's complement adding module 530 further comprises a third input coupled to 1, so that the m bits at the output of the swap module 505 are augmented by 1 bit. The two's complement adding module 530 performs an addition of the two signed numbers and generates an m+1 bit result at an output.
The output of two's complement adding module 530 is coupled to a first input of close rounding module 550 of normalization module 540.
The normalization module 540 further comprises a special left shifter 555. The special left shifter is equal to the one described with reference to
The m-bit output of special left shifter 555, which is the output of normalization module 540, is introduced as a first input in multiplexer 565. The second input of multiplexer 565 is coupled to the output of far rounding module 560. Far rounding unit 560 is coupled to the m+1 bit output of shifting module 545 which, in turn, has an input coupled to the output of two's complement adding module 520. The shifting module 545 produces a right or left shifting of maximum one position to normalize the result of the far path. The far rounding unit 560 is equal to the one described with reference to
The multiplexer 565 receives the effective operation and the difference of the exponents and generates the m bits of the significand Mz that is the sum or difference of significands Mx and My after aligning them. The LSB of significand Mz is implicit, in the same way discussed with reference to
It should be noted that in this implementation the m MSBs of the significand include the integer bit. In an alternative implementation, the integer bit of the output significand may be discarded after normalization, since it is always one.
Implementation examples of fixed point multipliers are commented later in the text.
The significand data path 110F comprises a multiplication module 115F for receiving the m MSBs of the significands of the preprocessed FP numbers X and Y. The significands are represented by symbols Mx and My in
The significand data path 210F comprises a multiplication module 215F for receiving the m MSBs of the significands of the preprocessed numbers X and Y. Again, the significands are represented by symbols Mx and My in
In a parallel path, the m MSBs of the significand Mz of the third preprocessed number is input to first shifting module 220F that is configured to align Mz so that it can be added to the result of the multiplication. First shifting module 220F comprises a conditional bit inverter 222F that is controlled by the bit op, and an arithmetic right shifter 224F. This bit op indicates the effective operation, which depends on the sign of the input floating point numbers (XOR of the three input signs). The m-bit output of conditional bit inverter 222F, augmented to the left with the op bit as its sign bit and to the right with the LSB of Mz, is input to the arithmetic right shifter 224F. Again, the right arithmetic shifter 224F is controlled by an output of the exponent logic 207F that indicates the difference (d) of the exponent of Z and the sum of the other two exponents. The output of the first shifting module 220F is a 3*m+3 bit number and is coupled to adding module 230F. At first, said number should have 3*m+4 bits to cover all displacement cases with minimum error. However, the sign bit (MSB of the shifted value) is omitted and its second MSB is used instead, since both bits are equal, except if no shifting is performed. In this last case, no addition is really performed, since no shifting means that the two numbers are too distant (Ez>>Ex+Ey and more specifically Ez>Ex+Ey+m+1). Thus, the sign of the result of the addition is not its MSB, but the bit to indicate the effective operation (op). In an alternative implementation, the inversion in both conditional inverters 222F and 244F may be prevented when this situation (Ez>Ex+Ey+m+1) is produced, and, concordantly, the sign may always be positive in this situation. In another alternative implementations, the sign of the result of the addition may be always its MSB and the overflow signal may be avoided if 3*m+4 bits are used for representing the aligned significant and the result of the addition.
The adding module 230F generates, in a non-redundant representation, the addition between the redundant output of the multiplication module 215F and the aligned output of the first shifting module 220F. In this particular example, since carry-save is used as redundant representation, the adding module 230F comprises a 3:2 counter 232F to add the two outputs of the multiplication module 215F and the 2*m+2 LSBs of the output of the first shifting module 220F. The 3:2 counter 232F generates two 2*m+2 bit words as a carry-save output. The adding module 230F further comprises a two's complement adder 234F couples to the output of the 3:2 counter 232F and an incrementing module 235F, having a first input for receiving the m+1 MSBs of the output of the first shifting module 220F and a second input for receiving a carry out bit from the two's complement adder 234F, to produce a significand in a non-redundant representation. In an alternative implementation, both modules may be substituted by a 3*m+3 bit two's complement adder having the m+1 MSBs of one of its inputs couple to zero, or a different circuit if the redundant representation selected is other. The m+1 bit output of the incrementing module 235F and the 2*m+2 output of the two's complement adder 234F comprise a 3*m+3 bit number that corresponds to the significand of result of the fused multiply-add operation before normalization. Said 3*m+3 bit number is input to a normalization module 240F. The incrementing module 235F further produces an overflow bit at a second output. In other implementations, the overflow information may be obtained from the output of a Leading Zero Anticipator (LZA), and this explicit output may not be required.
The significand data path 210F further comprises a Leading Zero Anticipator (LZA) 237F having a first input coupled to the output of 3:2 counter 232F and a second input for receiving the m+1 MSBs of the output of the first shifting module 220F. LZA 237F also receive an instruction (not shown in the figure) about the effective operation when no shifting is performed in the first shifting module. LZA 237F calculates the required left shift to normalize the result. In an alternative implementation, the LZA may take the inputs directly from the outputs of the multiplication module 215F and the first shifting module 220F or at a later stage from the output of the adding module 230F.
One skilled in the art may appreciate that the adding module 230F and the LZA 237F may be implemented (together or separately) in many different ways without departing from the scope of the invention.
The normalization module 240F comprises a left shifting module 242F and a conditional bit inverter 244F. The left shifting module 242F receives the 3*m+3 number from the adding module 230F at a first input and generates an m+1 bit normalized and rounded preprocessed number, having the LSB implicit and equal to one. It performs this operation based on a second shift amount received from the LZA 237F at a second input. The m MSBs of said preprocessed number is then input to conditional bit inverter 244F to negate it if its MSB is zero. This last indicates a negative result of the addition, since said MSB is the integer bit and it should be one (normalized number). One skilled in the art may appreciate that different options to detect a negative result of the addition may be used. On the other hand, in an alternative implementation, the conditional bit inverter may be before the left shifting module. The m-bit output of the conditional inverter 244F corresponds to the m MSBs of the preprocessed significand of the final result of the FMAD operation. The LSB of said preprocessed significand is implicit and it is equal to 1. It should be noted that in this implementation the m MSBs of the significand include the integer bit which is always one. Therefore, in an alternative implementation, the integer bit may be discarded after the normalization.
The FP arithmetic units described above require FP numbers that have been preprocessed according to the invention as described also above. These preprocessed numbers may be generated by circuits, such as the aforementioned FP adders, that are designed to function with preprocessed numbers or they may be generated by converters, designed to convert unprocessed numbers or preprocessed non-FP numbers to preprocessed numbers. Furthermore, the preprocessed numbers generated by the adders described above may, accordingly, require converters so that the numbers generated may be used by circuits that are not designed to process preprocessed numbers.
In the following examples, it is considered that floating point numbers, both unprocessed and preprocessed) are represented by a sign bit, an exponent, and a unsigned normalized significand such as the MSB is equal to one and it is explicitly included in the significand representation. In the same way, fixed point numbers, both unprocessed and preprocessed, are represented in two's complement representation, being the MSB equivalent to the sign bit. However, one skilled in the art may appreciate that other formats having a different representation may be used with minor modifications to the disclosed circuits. Some of these variations may be:
One category of such converters is converters for converting preprocessed integer numbers to preprocessed FP numbers.
Using this special left-shifter 610a, the input value of preprocessed left shifter 610 is augmented with an additional LSB set to any random bit (for instance, the LSB of the initial input value) and the third input of the special left-shifter is set to the inverted random value to fill both, the vacant positions required to complete the size of n if n>m+1, and the vacant positions produced after shifting. The output of preprocessed left shifter 610 comprises the n MSBs of the significand Mz of the preprocessed FP number. Said output corresponds only to the n MSBs of the shifted value if n<m. The LSB of the significand Mz is implicit and it is equal to 1.
In a parallel path, the converter 600 comprises el módulo detector de uno de cabecera (LOD) 615 having an input coupled to the output of conditional bit inverter 605 and an output for generating the shift amount for the special preprocessed left shifter 610 which is also used as input to exponent computation module 620 to generate the exponent Ez of the preprocessed FP number. Alternatively, the input of LOD module 615 may be directly coupled at the input of converter 600, but in this case it should detect the first zero, instead of the one, when the number is negative.
Compared with conventional fixed point FP converters, when M>N, there is no rounding up taking place after the shifting operation and therefore there is a reduction in components and processing. When M<N, then there is no bias produced because of the rounding with the used of the proposed converter.
Another category of converters are converters for converting unprocessed fixed point numbers to preprocessed floating point numbers.
La
Returning to the converter 700 in
Another category of converters is converters for converting preprocessed FP numbers to preprocessed FP numbers of different size of significand.
Another category of converters is converters for converting preprocessed FP numbers to preprocessed fixed point number numbers.
In other implementations of the examples in
Another category of converters is converters for converting unprocessed FP numbers to preprocessed FP numbers. In a first case, the significand of the original FP number is larger than the significand of the target FP number. The converter discussed with reference to
When the significand of the FP preprocessed number shall have more bits (n+m+1) than the significand of the unprocessed FP number (n), then:
a) in the case of biased rounding, the significand of the unprocessed number is expanded with as many zeros as necessary. This is illustrated in
b) in the case of unbiased rounding, the n−1 MSBs are the same. The nth bit is forced to zero. The m+1 bits to the right are set equal to the LSB of the unprocessed significand. This is illustrated in
Another category of converters is converters for converting preprocessed FP numbers to unprocessed FP numbers. When the significand of the preprocessed FP number has more bits (n+m+1) than the significand of the unprocessed one (n), then the circuit illustrated in
In an alternative implementation, the rounder may perform another type of rounding.
Yet another category of converters is converters for converting preprocessed FP numbers to unprocessed fixed point numbers.
If the number of bits of the input significand is lower than the number of bits of the output fixed point number, such a converter may be identical to the converter 1000a described in
On the other hand, there are two different cases when one of the input numbers is an unprocessed number. When the size of the unprocessed input number is equal or greater than the size of the preprocessed number, the exact result of the addition may be an unprocessed number. The implementation of a fixed point adding module arranged to receive the N MSBs of a first preprocessed fixed point number of n+1 bits and the n+m+1 bits of a second unprocessed fixed point number, may be similar to the circuit shown in
In the following examples of multipliers (including squarers and constant multipliers), it is considered, unless otherwise stated, that fixed point numbers are unsigned. However, one skilled in the art may appreciate that two's complement numbers may be operated instead, by making known modifications to the disclosed circuits, such as sign extension instead of zero extension for additions.
Since, only the z MSBs of the multiplication are delivered, the multiplier circuit may be optimized by avoiding the computation of the LSBs.
The carry net module 207MFJ receives the n+1 LSDs of the output, which does not include the implicit LSB of the preprocessed format, of said redundant multiplier and generates the carry bit corresponding to the conversion of said digits to a non-redundant binary representation. In this particular example, since carry-save representation is used, the carry net module 207MFJ receives the n+1 LSBs of the sum and carry words, at a first and a second input, respectively, and generates the last carry bit corresponding to the addition of both inputs.
The conversion module 209MFJ receives the m MSDs of the output of the redundant multiplier 205MFJ and the carry bit from the carry net module 207MFJ, and generates the m bits corresponding to the m MSBs of the value of the multiplication of the input fixed point numbers in a non-redundant representation. In this particular example, since carry-save representation is used, the conversion module 209MFJ receives the m MSBs of the sum and carry words at a first and a second input, respectively, and the carry bit at a third input and generates a value corresponding to the addition of both input words and the carry bit. Besides, in this particular example, the size of the output and the first input are equal, but in an alternative implementation the size of the output may be z+1 bits, being z<n+m. In this case, the carry net module 207MFJ may receive the n+m−z+1 LSDs of the output of the redundant multiplier, and the conversion module 209MFJ, the z MSDs.
The redundant multiplier for preprocessed numbers represented in
The compressor tree 330MFJ receives the output of the partial product generator module 325MFJ and a copy of the two inputs of the partial product generator module 325MFJ and generates a m+n+1 digit redundant output corresponding to the addition of all its inputs correctly aligned. It should be noted that said copies are aligned, such that the second LSB is aligned with the LSB of the least significant partial product (the one corresponding to the LSB of the second input). In this particular example, since carry-save representation is used, two m+n+1 bit numbers are produced corresponding to the sum and carry words. In an alternative implementation, a different redundant representation format may be used. In other implementations if non redundant output is desired, a conversion module may be used to transform the output of the compressor tree 330MFJ to a non-redundant m+n+1 bit number corresponding to the m+n+1 MSBs of the product of the initial preprocessed numbers.
The preprocessed redundant multiplier represented in
The architectures shown with reference to
Alternatively, preprocessed fixed point squaring module 100bSQFJ of
Since, only the z MSBs of the squaring are delivered, the squarer circuit may be optimized by avoiding the computation of the LSBs.
The carry net module 307SQFJ receives the 2*m−z LSDs of the output of said redundant squaring module 305SQFJ, and generates the carry bit corresponding to the conversion of said digits to a non-redundant binary representation. In this particular example, since carry-save representation is used, the carry net module 307SQFJ receives the 2*m−z LSBs of the sum and carry words, at a first and a second input, respectively, and generates the last carry bit corresponding to the addition of both inputs.
The conversion module 309SQFJ receives the z MSDs of the output of the redundant squaring module 305SQFJ and the carry bit from the carry net module 307SQFJ, and generates the z bits corresponding to the z MSBs of the value of the squaring of the input fixed point number in a non-redundant representation. In this particular example, since carry-save representation is used, the conversion module 309SQFJ receives the z MSBs of the sum and carry words, at a first and a second input, respectively, and the carry bit at a third input and generates a value corresponding to the addition of both input words and the carry bit.
The compressor tree 430SQFJ receives the output of the partial product generator module 425SQFJ and a copy of the m MSBs of the preprocessed input number and generates a 2*m digit redundant output corresponding to the addition of all its inputs correctly aligned. We should note that said m MSBs are aligned in such a way, that, its LSB is aligned with the LSB of the least significant partial product. In an alternative implementation said m MSBs may be introduced either within the compressor tree 430SQFJ or the partial product generator module 425SQFJ. In this particular example, since carry-save representation is used, two 2*m bit numbers are produced corresponding to the sum and carry words. In an alternative implementation, a different redundant representation format may be used. In other implementations if non redundant output is desired, a conversion module may be used to transform the output of the compressor tree 430SQFJ to a non-redundant 2*m bit number corresponding to the 2*m MSBs of the squaring of the initial preprocessed number.
In the examples shown in
The examples shown in figures from
Alternatively, preprocessed fixed point constant multiplication module 200MCFJ of
In alternative implementations of the examples of
Since, only the z MSBs of the multiplication are delivered, the multiplier circuit may be optimized by avoiding the computation of the LSBs.
The carry net module 307MCFJ receives the n+1 LSDs of the output of said redundant constant multiplication module 305MCFJ, which does not include the implicit LSB of the preprocessed format, and generates the carry bit corresponding to the conversion of said digits to a non-redundant binary representation. In this particular example, since carry-save representation is used, the carry net module 307MCFJ receives the n+1 LSBs of the sum and carry words, at a first and a second input, respectively, and generates the last carry bit corresponding to the addition of both inputs.
The conversion module 309MCFJ receives the m MSDs of the output of the redundant constant multiplication module 305MCFJ and the carry bit from the carry net module 307MCFJ, and generates the m bits corresponding to the m MSBs of the value of the multiplication of the input fixed point number and the constant number in a non-redundant representation. In this particular example, since carry-save representation is used, the conversion module 309MCFJ receives the m MSBs of the sum and carry words at a first and a second input, respectively, and the carry bit at a third input and generates a value corresponding to the addition of both input words and the carry bit. Besides, in this particular example, the size of the output and the first input are equal, but in an alternative implementation the size of the output may be z+1 bits, being z<n+m+1. In this case, the carry net module 307MCFJ may receive the n+m−z+1 LSDs of the output of the redundant multiplier, and the conversion module 309MCFJ, the z MSDs.
The compressor tree 430MCFJ receives the output of the partial product generator module 425MCFJ, a copy of the m-bit input and the n MSBs of the constant preprocessed number, and generates a m+n+1 digit redundant output corresponding to the addition of all its inputs correctly aligned. We should note that said copy and said n MSBs are aligned, in such a way that their second LSB is aligned with the LSB of the least significant partial product. In an alternative implementation said copy and said n MSBs of the constant number may be introduced either within the compressor tree 430MCFJ or the partial product generator module 425MCFJ. In this particular example, since carry-save representation is used, two m+n+1 bit numbers are produced corresponding to the sum and carry words. In an alternative implementation, a different redundant representation format may be used. In other implementations if non redundant output is desired, a conversion module may be used to transform the output of the compressor tree 4300MCFJ to a non-redundant m+n+1 bit number corresponding to the m+n+1 MSBs of the product of the initial preprocessed number and the constant number.
The architectures shown with reference from
The implementation of a preprocessed left shifter is described
Other category of converters is converters for converting preprocessed fixed point numbers to preprocessed Fixed Point numbers of different size.
Another category of converters is converters for converting preprocessed fixed point numbers to unprocessed fixed point numbers
Another category of converters is converters for converting preprocessed FP numbers to preprocessed fixed-point numbers (
Another category of converters is converters for converting preprocessed fixed-point numbers to preprocessed FP numbers.
In this example of
The input value of preprocessed left shifter 610FJ is augmented with an additional LSB sets to any bit with a random value (for instance, the LSB of the initial input value) and both, the vacant positions required to complete the size of n, if n>m+1, and the vacant positions produced after shifting are set to the invert of the random value. The output of preprocessed left shifter 610FJ comprises the n MSBs of the significand Mz of the preprocessed FP number. Said output corresponds only to the n MSBs of the shifted value if n<m. The LSB of the significand Mz is implicit and it is equal to 1.
In a parallel path, the converter 600FJ comprises LOD module 615FJ having an input coupled to the output of conditional bit inverter 605FJ and an output for generating the shift amount for the preprocessed left shifter 610FJ which is also used as input to exponent computation module 620FJ to generate the exponent Ez of the preprocessed FP number. Alternatively, the input of LOD module 615FJ may be directly coupled at the input of converter 600FJ but, in this case, it should detect the first zero instead of the one when the number is negative.
Compared to conventional fixed-point-to-FP numbers converters, when m>n, there is no rounding up taking place after the shifting operation and therefore there is a reduction in components and processing. When m<n, then there is no bias produced with the used of the proposed converter.
Another category of such converters is converters for converting preprocessed fixed-point numbers to unprocessed FP numbers.
In a parallel path, the converter 1500FJ comprises LOD module 1515FJ having an input coupled to the output of conditional bit inverter 1505FJ and an output for generating the shift amount for the left shifter 1510FJ which is also used, along with the overflow signal, as input to exponent computation module 1520FJ to generate the exponent Ez of the preprocessed FP number. Alternatively, the input of LOD module 1515FJ may be directly coupled at the input of converter 1500FJ. The converter shown in this example may produce some bias when n<m and the input number is in such a way that the LSB of the output of the left shifter 1510FJ coincides with the LSB of said input number. This bias may be avoided by applying standard techniques when this situation occurs; such as only perform the rounding up if the second LSB of the number is also one. In some implementations, said situation may be detected by checking the shift amount whereas in others it may be detected by computing the sticky bit over the m-n LSBs of the shifted value.
Another category of converters is converters for converting unprocessed FP numbers to preprocessed fixed point numbers.
The sign of the unprocessed FP number is used to control the conditional bit inverter 1605FJ. The output of the conditional bit inverter 1605FJ along with the sign (sign_x) is input to right shifter 1610FJ. Right shifter 1610FJ has another input for receiving the shift amount from shift amount calculator 1615FJ. Shift amount calculator 1615FJ receives the exponent of the unprocessed FP number and generates the shift amount. The output of the right shifter 1610FJ is the n+1 MSBs of the preprocessed fixed point number. The LSB is similarly equal to 1 and is neither generated nor depicted. In an alternative implementation the conditional bit inverter may be placed after the right shifter.
Although only a number of particular embodiments and examples of the invention have been disclosed herein, it will be understood by those skilled in the art that other alternative embodiments and/or uses of the invention and obvious modifications and equivalents thereof are possible. Furthermore, the present invention covers all possible combinations of the particular embodiments described. Thus, the scope of the present invention should not be limited by particular embodiments, but should be determined only by a fair reading of the claims that follow.
Furthermore, the described embodiments of the invention with reference to the drawings comprise computer systems and processes performed in computer systems, characterized functionally, and independent of the support or technology used for implementation. This support means may be, for example, an application specific integrated circuit (ASIC, acronym) circuit, a programmable logic circuit (FPGA or CPLD, acronym in English) including a memory, or any other device, such circuits being adapted or configured to perform, or for use in performing, the relevant processes.
Although the embodiments described comprise computing devices, the invention also extends to computer programs, more particularly to computer programs in a carrier means adapted to carry out the invention. The computer program may be in source code, object code or an intermediate code between source code and object code such as in partially compiled form, or in any other form suitable for use in the implementation of the processes according to the invention. The carrier may be any entity or device capable of carrying the program.
For example, the carrier may comprise a storage medium such as a ROM, for example a CD ROM or semiconductor ROM, or magnetic recording medium, for example a floppy disc or hard disk. In addition, the carrier may be a transmissible carrier medium such as an electrical or optical signal that can be transmitted via electrical or optical cable or by radio or other means.
When the computer program is contained in a signal that can be transmitted directly via cable or other device or means, the carrier may be constituted by such cable or other device or means.
| Number | Date | Country | Kind |
|---|---|---|---|
| P201430451 | Mar 2014 | ES | national |
| P201430453 | Mar 2014 | ES | national |
| P201430454 | Mar 2014 | ES | national |
| P201430455 | Mar 2014 | ES | national |
| P201430456 | Mar 2014 | ES | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/ES2015/000050 | 3/27/2015 | WO | 00 |
