Canonical signed digit multiplier

Abstract

A multiplier is able to multiply an input data value by a selected constant value in CSD form. The selected constant value has a plurality of pairs of bits, and the multiplier includes multiplexers, each controlled by a respective pair of bits of the selected constant value. Each of the multiplexers has a plurality of inputs, and is connected to receive the input data value, the inverse of the input data value, and all zeros on said inputs, and it is controlled such that it outputs either the input data value, the inverse of the input data value, or all zeros, depending on the values of the respective pair of bits of the selected constant value. Variable shift blocks are each connected to receive an input from a respective one of said multiplexers, and are each adapted to shift their received input by a first bit shift value or a second bit shift value, depending on the values of the respective pair of bits of the selected constant value, wherein the first bit shift value and the second bit shift value differ by 1. The multiplier also includes combination circuitry, for receiving the outputs from the plurality of shift blocks, and for combining the outputs from the plurality of shift blocks and applying further bit shifts, to form an output value equal to the result of multiplying the input data value by the selected constant value.

Description

This invention relates to a multiplier, and in particular to a multiplier for multiplying input data by multiple constants.

In many digital signal processing applications, it is necessary to multiply input data by multiple constants. For example, in the case of a finite impulse response (FIR) filter, or a fast fourier transform (FFT) processor, it is necessary to multiply input data by constant coefficients. FIR filters and FFT processors are used in many digital signal processing applications, for example in wireless communications applications.

Since it is computationally expensive to perform multiplications, it is known to perform the required calculations by representing the coefficients by means of canonical signed digit (CSD) multipliers. This allows the required multiplication to be performed by means of a sequence of shift operations and add operations. A number is said to be in CSD format if it is in the form b₀b₁b₂ . . . b_N−1, where each b_i, for i=0, 1, 2, . . . , N−1, takes one of the values 0, +1, −1, and where no two consecutive b_i are non-zero.

The document “Implementation of Orthogonal Frequency Division Multiplexing Modem Using Radix-N Pipeline Fast Fourier Transform (FFT) Processor”, Oh, et al, Jpn. J. Appl. Phys. Vol. 42 (2003) pp. 1-6 discloses a hardware structure for implementation of a CSD-encoded multiplier. An input signal is applied to a switching network through a set of right shift elements in parallel. The switching network then operates to apply the right-shifted inputs to a set of output multiplexers, based on the required multiplication constants. The output multiplexers are controlled by a select signal to allow the processed data to be output sequentially.

This structure becomes complex when the number of required multiplication constants increases. Specifically, the design of the switching network becomes complex as the number of required multiplication constants increases, while the size of each of the output multiplexers also increases as the number of required multiplication constants increases.

This structure also has the disadvantage that, since the low order bits are discarded before the output values are accumulated, the structure leads to relatively large truncation errors.

According to the present invention, there is provided a multiplier, for multiplying an input data value by a selected constant value in CSD form, wherein the selected constant value is made up of pairs of bits, the multiplier comprising:

a plurality of multiplexers, wherein each of said multiplexers is controlled by a respective pair of bits of the selected constant value, and wherein each of said multiplexers has a plurality of inputs, and is connected to receive the input data value, the inverse of the input data value, and all zeros on said inputs, and being controlled such that it outputs either the input data value, the inverse of the input data value, or all zeros, depending on the values of the respective pair of bits of the selected constant value;

a plurality of variable shift blocks, each connected to receive an input from a respective one of said multiplexers, and each being adapted to shift its received input by a first bit shift value or a second bit shift value, depending on the values of the respective pair of bits of the selected constant value, wherein the first bit shift value and the second bit shift value differ by 1; and

combination circuitry, for receiving the outputs from the plurality of shift blocks, and for combining the outputs from the plurality of shift blocks and applying further bit shifts, to form an output value equal to the result of multiplying the input data value by the selected constant value.

This has the advantage that it allows implementation of a CSD-encoded multiplier efficiently in hardware, with an acceptable level of truncation error.

FIG. 1 is a block schematic diagram showing the general form of a multiplier in accordance with the present invention.

FIG. 2 is a block schematic diagram showing in more detail the general form of a multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

FIG. 3 is a block schematic diagram showing in more detail the general form of a multiplier in accordance with the present invention, for multiplying data by 12-bit constants.

FIG. 4 is a block schematic diagram showing a first multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

FIG. 5 is a block schematic diagram showing a second multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

FIG. 6 is a block schematic diagram showing a third multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

FIG. 7 is a block schematic diagram showing a fourth multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

FIG. 1 shows the general form of a multiplier 10 in accordance with the present invention. As is well known, the function of a canonical signed digit (CSD) multiplier is to multiply successive input data values by one or more predetermined constant values, where the input data values are in binary format, and the constant values are in CSD format. That is, the input data values are in the form d₀d₁d₂ . . . d_M−1, where each d_i, for i=0, 1, 2, . . . , M−1, takes one of the values 0 or +1, while the constant values are in the form b₀b₁b₂ . . . b_N−1, where each b_i, for i=0, 1, 2, . . . , N−1, takes one of the values 0, +1, −1, and where no two consecutive b_i are non-zero.

Thus, the CSD multiplier must be able to receive an input data value, and multiply it by a selected one of the predetermined constant values, and then supply the result as an output value.

The result of the multiplication is d₀d₁d₂ . . . d_M−1×b₀b₁b₂ . . . b_N−1, and, as is conventional, the multiplier 10 calculates this result by means of appropriate additions and shift operations.

In FIG. 1, the input data values are received on a data line 12. Where the input data values are m-bit values, the data line 12 is an m-bit line. The input data values received on the data line 12 are applied to respective first inputs of each of n multiplexers 14₀, 14₁, . . . , 14_n−1.

A second input line 16 carries m zeros in parallel, and these are applied to respective second inputs of the multiplexers 14₀, 14₁, . . . , 14_n−1.

The input data values received on the data line 12 are also applied to an inverter 18, to form the twos-complement of the input data values, and the resulting twos-complemented version of the input data values are applied on the inverse line 20 to respective third inputs of the multiplexers 14₀, 14₁, . . . , 14_n−1.

The multiplexers 14₀, 14₁, . . . , 14_n−1 are controlled by a control signal on a first control line 22. As mentioned above, there are n multiplexers 14₀, 14₁, . . . , 14_n−1, and this is the case when the constant values in CSD format each have 2n bits. That is, the constant values are in the form b₀b₁b₂ . . . b_2n−2b_2n−1, where each b_i, for i=0, 1, 2, . . . , 2n−1, takes one of the values 0, +1, −1, and where no two consecutive b_i are non-zero.

Thus, the first multiplexer 14₀ is controlled by the first two bits b₀b₁, of the constant value which is in use at this moment, the second multiplexer 14₁, is controlled by the third and fourth bits b₂b₃ of the constant value which is in use at this moment, and so on, while the final multiplexer 14_n−1 is controlled by the penultimate and final bits b_2n−2b_2n−1 of the constant value which is in use at this moment.

In the constant value which is in use at this moment, it can be taken that no two consecutive bits are non-zero, because this is a condition of the CSD format. Therefore, taking as an example the first multiplexer 14₀, this is controlled by the combination of the two bits b₀b₁, and this combination has five possible values in the CSD representation, namely: 0 0, 0 1, 1 0, −1 0 and 0 −1.

If b₀b₁ equals 0 0, the first multiplexer 14₀ is controlled such that it passes to its output the value on its second input, namely all zeros. If b₀b₁ equals 0 1 or 1 0, the first multiplexer 14₀ is controlled such that it passes to its output the value on its first input, namely the input data value. If b₀b₁ equals 0 −1 or −1 0, the first multiplexer 14₀ is controlled such that it passes to its output the value on its third input, namely the twos-complemented version of the input data value.

The other multiplexers are controlled by their respective pairs of bits from the constant value which is in use at this moment, in a corresponding way.

The outputs from the multiplexers 14₀, 14₁, . . . , 14_n−1 are passed to respective variable shift blocks 24₀, 24₁, . . . , 24_n−1, and the variable shift blocks 24₀, 24₁, . . . , 24_n−1 are controlled by a control signal on a second control line 26.

Specifically, the first variable shift block 24₀ is controlled on the basis of the first two bits b₀b₁ of the constant value which is in use at this moment, the second variable shift block 24₁ is controlled on the basis of the third and fourth bits b₂b₃ of the constant value which is in use at this moment, and so on, while the final variable shift block 24_n−1 is controlled on the basis of the penultimate and final bits.

The variable shift blocks 24₀, 24₁, . . . , 24_n−1, can be right shifters, or left shifters, or a combination of right shifters and left shifters. In any event, the input to each of the variable shift blocks 24₀, 24₁, . . . , 24_n−1, is the output from the respective one of the multiplexers 14₀, 14₁, . . . , 14_n−1. In each case, it is desired that the output of the variable shift block should be proportional to the result of multiplying the input data value by the appropriate pair of bits of the constant value.

The sign of this output is determined by the control of the multiplexer, while the magnitude of the output is determined by the control of the shift that is applied by the variable shift block 24. Each of the variable shift blocks 24₀, 24₁, . . . , 24_n−1, is controlled such that it applies a respective shift value of S₀, S₁, . . . , S_n−1, and each of these shift values can take either a first or a second value, which differ by one, depending on whether the non-zero bit in the appropriate pair of bits of the constant value is the more significant bit or the less significant bit.

Thus, considering the variable shift block 24₀, this shifts the input data based on the value of the bit pair b₀b₁. If the more significant bit b₀ is nonzero, it does not perform any right shift, but simply passes through its input. If the less significant bit b₁ is nonzero, it performs a right shift by 1 bit position. If b₀ and b₁ are both zero, it is irrelevant whether the shifter performs a right shift by 1 bit position or just passes through its input because, as described above, that input will consist entirely of zeros.

It should be noted that although, as described here, the variable shift block 24₀ does not perform any right shift if the more significant bit b₀ is nonzero, and performs a right shift by 1 bit position if the less significant bit b₁ is nonzero, it is the relative sizes of these two possible shift values that is important, rather than their absolute values. Thus, for example, the variable shift block 24₀ could perform a right shift by 2 bit positions if the more significant bit b₀ is nonzero, and could then perform a right shift by 3 bit positions if the less significant bit b₁ is nonzero. Alternatively, the variable shift block 24₀ could perform a left shift by 3 bit positions if the more significant bit b₀ is nonzero, and could then perform a left shift by 2 bit positions if the less significant bit b₁ is nonzero. In each case, the bit shift if the more significant bit b₀ is nonzero is one bit more to the left (or one bit less to the right) than the bit shift if the less significant bit b₁ is nonzero.

Exactly the same applies to the other variable shift blocks 24₁, . . . , 24_n−1.

The outputs of the variable shift blocks 24₀, 24₁, . . . , 24_n−1 are then passed to combination circuitry 28, where they are combined, and then to an output line 30. The function of the combination circuitry 28 is to add together the outputs of the variable shift blocks 24₀, 24₁, . . . , 24_n−1, while applying appropriate further bit shifts so that the correct final result is obtained.

The variable shift block 24₀ is providing the result of multiplying the input data by the two most significant bits of the constant value, while the variable shift block 24₁ is providing the result of multiplying the input data by the next two most significant bits of the constant value, and so on, and the variable shift block 24_n−1 is providing the result of multiplying the input data by the two least significant bits of the constant value. Therefore, the output of the variable shift block 24₀ must be shifted to the left by two bits more (or to the right by two bits less) in total than the output of the variable shift block 24₁, and so on, until the output of the variable shift block 24₀ is shifted to the left by 2(n−2) bits more (or to the right by 2(n−2) bits less) in total than the output of the variable shift block 24_n−1.

This can be achieved by setting different possible shift values for the variable shift blocks 24₀, 24₁, . . . , 24_n−1, or by applying different shift values to the outputs of the variable shift blocks, or by a combination of these two elements.

The result is thus a multiplier which uses relatively little hardware, and can be controlled as desired to produce the required result.

FIG. 2 is a block schematic diagram showing the general form of a multiplier in accordance with the present invention, for multiplying data by 8-bit constants, and illustrating how the combination circuitry can operate.

In FIG. 2, the input data values to the multiplexer 210 are received on a data line 212. The constants are each of eight bits, and so the input data values received on the data line 212 are applied to respective first inputs of each of four multiplexers 214₀, 214₁, . . . 214₃.

A second input line 216 carries all zeros, and these are applied to respective second inputs of the multiplexers 214₀, 214₁, . . . , 214₃. The input data values received on the data line 212 are also applied to an inverter 218, to form the twos-complement of the input data values, and the resulting twos-complemented version of the input data values are applied on the inverse line 220 to respective third inputs of the multiplexers 214₀, 214₁, . . . , 214₃.

The multiplexers 214₀, 214₁, . . . , 214₃ are controlled by a control signal on a control line (not shown), as described with reference to FIG. 1. Thus, the first multiplexer 214₀ is controlled by the first two bits b₀b₁ of the constant value which is in use at this moment, the second multiplexer 214₁ is controlled by the third and fourth bits b₂b₃ of the constant value which is in use at this moment, the third multiplexer 214₂ is controlled by the fifth and sixth bits b₄b₅ of the constant value which is in use at this moment, and the fourth multiplexer 214₃ is controlled by the seventh and eighth bits b₆b₇ of the constant value which is in use at this moment.

If b₀b₁ equals 0 0, the first multiplexer 214₀ is controlled such that it passes to its output the value on its second input, namely all zeros. If b₀b₁ equals 0 1 or 1 0, the first multiplexer 214₀ is controlled such that it passes to its output the value on its first input, namely the input data value. If b₀b₁ equals 0 −1 or −1 0, the first multiplexer 214₀ is controlled such that it passes to its output the value on its third input, namely the twos-complemented version of the input data value.

The other multiplexers are controlled by their respective pairs of bits from the constant value which is in use at this moment, in a corresponding way.

The outputs from the multiplexers 214₀, 214₁, . . . , 214₃ are passed to respective variable shift blocks 224₀, 224₁, . . . , 224₃, and the variable shift blocks 224₀, 224₁, . . . , 224₃ are controlled by a control signal on a control line (not shown) as described above with reference to FIG. 1.

Specifically, the first variable shift block 224₀ is controlled on the basis of the first two bits b₀b₁ of the constant value which is in use at this moment, the second variable shift block 224₁ is controlled on the basis of the third and fourth bits b₂b₃ of the constant value which is in use at this moment, and so on.

The variable shift blocks 224₀, 224₁, . . . , 224₃ can be right shifters, or left shifters, or a combination of right shifters and left shifters, as described above.

The outputs of the first and second variable shift blocks 224₀ and 224₁ are passed to a first adder 232, while the outputs of the third and fourth variable shift blocks 224₂ and 224₃ are passed to a second adder 234.

The output of the first adder 232 is passed to a first fixed shift block 236, which applies a first fixed shift value. The output of the second adder 234 is passed to a second fixed shift block 238, which applies a second fixed shift value. The outputs of the first fixed shift block 236 and the second fixed shift block 238 are passed to a third adder 240, and the output of the third adder 240 is supplied as a data output on the output line 242.

The shift values applied by the four variable shift blocks 224₀, 224₁, . . . , 224₃ and the two fixed shift blocks 236, 238 are then set such that the output of the first variable shift block 224₀ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the second variable shift block 224₁, the output of the second variable shift block 224₁ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the third variable shift block 224₂, and the output of the third variable shift block 224₂ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the fourth variable shift block 224₃.

FIG. 3 is a block schematic diagram showing the general form of a multiplier in accordance with the present invention, for multiplying data by 12-bit constants, and illustrating how the combination circuitry can operate in this case.

In FIG. 3, the input data values to the multiplier 310 are received on a data line 312. The constants are each of twelve bits, and so the input data values received on the data line 312 are applied to respective first inputs of each of six multiplexers 314₀, 314₁, . . . , 314₅.

A second input line 316 carries all zeros, and these are applied to respective second inputs of the multiplexers 314₀, 314₁, . . . , 314₅. The input data values received on the data line 312 are also applied to an inverter 318, to form the twos-complement of the input data values, and the resulting twos-complemented version of the input data values are applied on the inverse line 320 to respective third inputs of the multiplexers 314₀, 314₁, . . . , 314₅.

The multiplexers 314₀, 314₁, . . . , 314₅ are controlled by a control signal on a control line (not shown), as described with reference to FIG. 1. Thus, the first multiplexer 314₀ is controlled by the first two bits b₀b₁ of the constant value which is in use at this moment, the second multiplexer 314₁ is controlled by the third and fourth bits b₂b₃ of the constant value which is in use at this moment, the third multiplexer 314₂ is controlled by the fifth and sixth bits b₄b₅ of the constant value which is in use at this moment, the fourth multiplexer 314₃ is controlled by the seventh and eighth bits b₆b₇ of the constant value which is in use at this moment, the fifth multiplexer 314₄ is controlled by the ninth and tenth bits b₈b₉ of the constant value which is in use at this moment, and the sixth multiplexer 314₅ is controlled by the eleventh and twelfth bits b₁₀b₁₁ of the constant value which is in use at this moment.

If b₀b₁ equals 0 0, the first multiplexer 314₀ is controlled such that it passes to its output the value on its second input, namely all zeros. If b₀b₁ equals 0 1 or 1 0, the first multiplexer 314₀ is controlled such that it passes to its output the value on its first input, namely the input data value. If b₀b₁ equals 0 −1 or −1 0, the first multiplexer 314₀ is controlled such that it passes to its output the value on its third input, namely the twos-complemented version of the input data value.

The other multiplexers are controlled by their respective pairs of bits from the constant value which is in use at this moment, in a corresponding way.

The outputs from the multiplexers 314₀, 314₁, . . . , 314₅ are passed to respective variable shift blocks 324₀, 324₁, . . . , 324₅, and the variable shift blocks 324₀, 324₁, . . . , 324₅ are controlled by a control signal on a control line (not shown) as described above with reference to FIG. 1.

Specifically, the first variable shift block 324₀ is controlled on the basis of the first two bits b₀b₁ of the constant value which is in use at this moment, the second variable shift block 324₁ is controlled on the basis of the third and fourth bits b₂b₃ of the constant value which is in use at this moment, and so on.

The variable shift blocks 324₀, 324₁, . . . , 324₅ can be right shifters, or left shifters, or a combination of right shifters and left shifters, as described above.

The outputs of the first and second variable shift blocks 324₀ and 324₁ are passed to a first adder 332, while the outputs of the third and fourth variable shift blocks 324₂ and 324₃ are passed to a second adder 334, and the outputs of the fifth and sixth variable shift blocks 324₄ and 324₅ are passed to a third adder 336.

The output of the first adder 332 is passed to a first fixed shift block 338, which applies a first fixed shift value. The output of the second adder 334 is passed to a second fixed shift block 340, which applies a second fixed shift value. The output of the third adder 336 is passed to a third fixed shift block 342, which applies a third fixed shift value.

The outputs of the first fixed shift block 336 and the second fixed shift block 340 are passed to a fourth adder 344. The outputs of the third fixed shift block 342 and of the fourth adder 344 are passed to a fifth adder 346, and the output of the fifth adder 346 is supplied as a data output on the output line 348.

The shift values applied by the six variable shift blocks 324₀, 324₁, . . . , 324₅ and the three fixed shift blocks 338, 340, 342 are then set such that the output of the first variable shift block 324₀ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the second variable shift block 324₁, the output of the second variable shift block 324₁ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the third variable shift block 324₂, the output of the third variable shift block 324₂ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the fourth variable shift block 324₃, the output of the fourth variable shift block 324₃ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the fifth variable shift block 324₄, and the output of the fifth variable shift block 324₄ is shifted to the left by two bits more (or to the right by two bits less) in total than the output of the sixth variable shift block 324₅.

FIG. 4 is a block schematic diagram showing a specific form of a multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

In FIG. 4, the input data values to the multiplexer 410 are received on a data line 412. The constants are each of eight bits, and so the input data values received on the data line 412 are applied to respective first inputs of each of four multiplexers 414₀, 414₁, . . . , 414₃.

A second input line 416 carries all zeros, and these are applied to respective second inputs of the multiplexers 414₀, 414₁, . . . , 414₃. The input data values received on the data line 412 are also applied to an inverter 418, to form the twos-complement of the input data values, and the resulting twos-complemented version of the input data values are applied on the inverse line 420 to respective third inputs of the multiplexers 414₀, 414₁, . . . , 414₃.

The multiplexers 414₀, 414₁, . . . , 414₃ are controlled by a control signal on a control line (not shown), as described with reference to FIG. 1. Thus, the first multiplexer 414₀ is controlled by the first two bits b₀b₁ of the constant value which is in use at this moment, the second multiplexer 414₁ is controlled by the third and fourth bits b₂b₃ of the constant value which is in use at this moment, the third multiplexer 414₂ is controlled by the fifth and sixth bits b₄b₅ of the constant value which is in use at this moment, and the fourth multiplexer 414₃ is controlled by the seventh and eighth bits b₆b₇ of the constant value which is in use at this moment.

If b₀b₁ equals 0 0, the first multiplexer 414₀ is controlled such that it passes to its output the value on its second input, namely all zeros. If b₀b₁ equals 0 1 or 1 0, the first multiplexer 414₀ is controlled such that it passes to its output the value on its first input, namely the input data value. If b₀b₁ equals 0 −1 or −1 0, the first multiplexer 414₀ is controlled such that it passes to its output the value on its third input, namely the twos-complemented version of the input data value.

The other multiplexers are controlled by their respective pairs of bits from the constant value which is in use at this moment, in a corresponding way.

The outputs from the multiplexers 414₀, 414₁, . . . , 414₃ are passed to respective variable shift blocks 424₀, 424₁, . . . , 424₃, and the variable shift blocks 424₀, 424₁, . . . , 424₃ are controlled by a control signal on a control line (not shown) as described above with reference to FIG. 1.

Specifically, the first variable shift block 424₀ is controlled on the basis of the first two bits b₀b₁ of the constant value which is in use at this moment, the second variable shift block 424₁ is controlled on the basis of the third and fourth bits b₂b₃ of the constant value which is in use at this moment, and so on.

The first variable shift block 224₀ can either apply no shift, or can shift the data to the right by one bit, the second variable shift block 224₁ can either shift the data to the right by two bits or by three bits, the third variable shift block 224₂ can either apply no shift, or can shift the data to the right by one bit, and the fourth variable shift block 224₃ can either shift the data to the right by two bits or by three bits.

The outputs of the first and second variable shift blocks 424₀ and 424₁ are passed to a first adder 432, while the outputs of the third and fourth variable shift blocks 424₂ and 424₃ are passed to a second adder 434.

No fixed shift is applied to the output of the first adder 232. The output of the second adder 434 is passed to a fixed shift block 438, which applies a fixed shift of four bits to the right. The outputs of the first adder 432 and the fixed shift block 438 are passed to a third adder 440, and the output of the third adder 440 is supplied as a data output on the output line 442.

Thus, as required, the output of the first variable shift block 424₀ is shifted to the right by 0 bits or 1 bit; the output of the second variable shift block 424₁ is shifted to the right by 2 bits or 3 bits; the output of the third variable shift block 424₂ is shifted to the right by 4 bits or 5 bits (by means of the variable shift block 424₂ and the fixed shift block 438); and the output of the fourth variable shift block 424₃ is shifted to the right by 6 bits or 7 bits.

This structure has a lower truncation error than some conventional devices. However, in accordance with the invention, the truncation error can be reduced further.

FIG. 5 is therefore a block schematic diagram showing a specific form of a multiplier in accordance with the present invention, for multiplying data by 8-bit constants.

In FIG. 5, the input data values to the multiplexer 510 are received on a data line 512. The input data values received on the data line 512 are applied to respective first inputs of each of four multiplexers 514₀, 514₁, . . . , 514₃.

A second input line 516 carries all zeros, and the input data values received on the data line 512 are also applied to an inverter 518, with the inverter output being applied on the inverse line 520 to respective third inputs of the multiplexers 514₀, 514₁, . . . , 514₃.

The operation of the multiplexers 514₀, 514₁, . . . , 514₃ is exactly the same as that of the multiplexers 414₀, 414₁, . . . , 414₃ in FIG. 4, and this will not be described again.

The outputs from the multiplexers 514₀, 514₁, . . . , 514₃ are passed to respective variable shift blocks 524₀, 524₁, . . . , 524₃, and the principle of operation of the variable shift blocks 524₀, 524₁, . . . , 524₃ is exactly the same as that of the variable shift blocks 424₀, 424₁, . . . , 424₃ in FIG. 4, and this will not be described again.

However, in this case, the first variable shift block 524₀ can either shift the data to the left by three bits or by two bits, without truncation, the second variable shift block 524₁ can either shift the data to the left by one bit, without truncation, or can apply no bit shift, the third variable shift block 524₂ can either shift the data to the right by one bit or by two bits, and the fourth variable shift block 224₃ can either shift the data to the right by three bits or by four bits.

The fact that the first variable shift block 524₀ and the second variable shift block 524₁ can shift the data to the left, without truncation, improves the precision of the device, albeit at the cost of slightly more complex hardware.

The outputs of the first and second variable shift blocks 524₀ and 524₁ are passed to a first adder 532, while the outputs of the third and fourth variable shift blocks 524₂ and 524₃ are passed to a second adder 534.

No fixed shift is applied to the output of either the first adder 532 or the second adder 534. The outputs of the first adder 532 and the second adder 534 are passed to a third adder 540, and the output of the third adder 540 is supplied as a data output on the output line 542.

Thus, as required, the output of the first variable shift block 524₀ is shifted to the left by 3 bits or 2 bits; the output of the second variable shift block 524₁ is shifted to the left by 1 bits or 0 bits; the output of the third variable shift block 524₂ is shifted to the right by 1 bit or 2 bits; and the output of the fourth variable shift block 524₃ is shifted to the right by 3 bits or 4 bits.

FIG. 6 is a block schematic diagram showing a further specific form of a multiplier in accordance with the present invention, for multiplying data by 8-bit constants, again achieving smaller truncation errors than the embodiment shown in FIG. 4.

In FIG. 6, the input data values to the multiplier 610 are received on a data line 612. The input data values received on the data line 612 are applied to respective first inputs of each of four multiplexers 614₀, 614₁, . . . , 614₃.

A second input line 616 carries all zeros, and the input data values received on the data line 612 are also applied to an inverter 618, with the inverter output being applied on the inverse line 620 to respective third inputs of the multiplexers 614₀, 614₁, . . . , 614₃.

The operation of the multiplexers 614₀, 614₁, . . . , 614₃ is exactly the same as that of the multiplexers 414₀, 414₁, . . . , 414₃ in FIG. 4, and this will not be described again.

The outputs from the multiplexers 614₀, 614₁, . . . , 614₃ are passed to respective variable shift blocks 624₀, 624₁, . . . , 624₃, and the principle of operation of the variable shift blocks 624₀, 624₁, . . . , 624₃ is exactly the same as that of the variable shift blocks 424₀, 424₁, . . . , 424₃ in FIG. 4, and this will not be described again.

However, in this case, the first variable shift block 624₀ can either shift the data to the left by three bits or by two bits, without truncation, the second variable shift block 624₁ can either shift the data to the left by one bit, without truncation, or can apply no bit shift, the third variable shift block 624₂ can either shift the data to the left by three bits or by two bits, without truncation, and the fourth variable shift block 624₃ can either shift the data to the left by one bit, without truncation, or can apply no bit shift.

Again, the fact that the first variable shift blocks 624₀-624₃ can shift the data to the left, without truncation, improves the precision of the device, albeit at the cost of slightly more complex hardware.

The outputs of the first and second variable shift blocks 624₀ and 624₁ are passed to a first adder 632, while the outputs of the third and fourth variable shift blocks 624₂ and 624₃ are passed to a second adder 634.

No fixed shift is applied to the output of the first adder 632, while the output of the second adder 634 is applied to a fixed shift block 638, which applies a shift of 4 bits to the right. The outputs of the first adder 632 and the fixed shift block 638 are passed to a third adder 640, and the output of the third adder 640 is supplied as a data output on the output line 642.

Thus, as required, the output of the first variable shift block 624₀ is shifted to the left by 3 bits or 2 bits; the output of the second variable shift block 624₁ is shifted to the left by 1 bits or 0 bits; the output of the third variable shift block 624₂ is shifted to the right by 1 bit or 2 bits (considering the effect of the third variable shift block 624₂ and the fixed shift block 638 together); and the output of the fourth variable shift block 624₃ is shifted to the right by 3 bits or 4 bits (considering the effect of the fourth variable shift block 624₃ and the fixed shift block 638 together).

It is also possible to use a structure which suffers from no truncation errors, again at a cost of slightly further increased hardware complexity, because the bit precision of the outputs must be increased. FIG. 7 is a block schematic diagram showing a further specific form of a multiplier in accordance with the present invention, for multiplying data by 8-bit constants, without truncation errors.

In FIG. 7, the input data values to the multiplier 710 are received on a data line 712. The input data values received on the data line 712 are applied to respective first inputs of each of four multiplexers 714₀, 714₁, . . . , 714₃.

A second input line 716 carries all zeros, and the input data values received on the data line 712 are also applied to an inverter 718, with the inverter output being applied on the inverse line 720 to respective third inputs of the multiplexers 714₀, 714₁, . . . , 714₃.

The operation of the multiplexers 714₀, 714₁, . . . , 714₃ is exactly the same as that of the multiplexers 414₀, 414₁, . . . , 414₃ in FIG. 4, and this will not be described again.

The outputs from the multiplexers 714₀, 714₁, . . . , 714₃ are passed to respective variable shift blocks 724₀, 724₁, . . . , 724₃, and the principle of operation of the variable shift blocks 724₀, 724₁, . . . , 724₃ is exactly the same as that of the variable shift blocks 424₀, 424₁, . . . , 424₃ in FIG. 4, and this will not be described again.

However, in this case, the first variable shift block 724₀ can either shift the data to the left by three bits or by two bits, without truncation, the second variable shift block 724₁ can either shift the data to the left by one bit, without truncation, or can apply no bit shift, the third variable shift block 724₂ can either shift the data to the left by three bits or by two bits, without truncation, and the fourth variable shift block 724₃ can either shift the data to the left by one bit, without truncation, or can apply no bit shift.

The outputs of the first and second variable shift blocks 724₀ and 724₁ are passed to a first adder 732, while the outputs of the third and fourth variable shift blocks 724₂ and 724₃ are passed to a second adder 734.

No fixed shift is applied to the output of the second adder 734, while the output of the first adder 732 is applied to a fixed shift block 736, which applies a shift of 4 bits to the left.

Again, the fact that the variable shift blocks 724₀-724₃ and the fixed shift block 736 can shift the data to the left, without truncation, improves the precision of the device, albeit at the cost of slightly more complex hardware.

The outputs of the second adder 734 and the fixed shift block 736 are passed to a third adder 740, and the output of the third adder 740 is supplied as a data output on the output line 742.

Thus, as required, the output of the first variable shift block 724₀ is shifted to the left by 7 bits or 6 bits (considering the effect of the first variable shift block 724₀ and the fixed shift block 736 together); the output of the second variable shift block 624₁ is shifted to the left by 5 bits or 4 bits (considering the effect of the second variable shift block 724₁ and the fixed shift block 736 together); the output of the third variable shift block 724₂ is shifted to the left by 3 bits or 2 bits; and the output of the fourth variable shift block 724₃ is shifted to the left by 1 bit or has no bit shift applied.

There are thus described specific embodiments for multiplying data by any set of 8-bit constants, and it will be appreciated that the same principle can be applied to constants of any length.

When the multiplier is to be used with a set of constants that is known in advance, some simplifications can be made, since the multiplier may not need to be able to operate with some of the possible values.

Specifically, each pair of two consecutive bits (i.e. b_2nb_2n+1) should be checked, for the entire set of constants to be supported by the CSD multiplier. For each value of n (i.e. n=0, 1, 2, 3 in the case of 8-bit constants), all of the values of b_2nb_2n+1 can be listed. Then, again for each value of n, the listed values are examined, to determine whether any simplification of the corresponding multiplexer, shifter and adder can be made.

With regard to the multiplexer, it was mentioned above that there are five possible pairs of values for two consecutive bits. These can be divided into three subsets: subset0={0 0}; subset1={1 0, 0 1}; subset2={−1 0, 0 −1}. If the listed values of b_2nb_2n+1 come from just one of the above subsets, the respective multiplexer can be removed, as the required multiplexer output is always zeros, or the data input, or the twos-complement of the data input, depending on the subset. If the listed values come from two of the above subsets, the corresponding multiplexer is reduced from 3:1 to 2:1, since one of the possible multiplexer inputs is never selected.

With regard to the variable bit shifter, the five possible pairs of values for two consecutive bits can again be divided into three subsets, though these are different from the subsets mentioned above, namely subset0={0 0}; subset1*={1 0, −1 0}; subset2*={0 1, 0 −1}. If only values in one of subset1* and subset2* are contained in the listed values, the shifter controlled by the bit pair b_2nb_2n+1 is optimized to a fixed shifter, since only one shift value is ever applied. If the listed values only come from subset0, the shifter can be removed.

With regard to the adder, then, if a shifter is removed as described above, the corresponding adder, that would have received one of its inputs from that shifter, can also be removed.

There is thus described a multiplier that can provide outputs at a required degree of precision, without excessive hardware requirements.

Claims

1. A multiplier, for multiplying an input data value by a selected constant value in CSD form, the selected constant value comprising a plurality of pairs of bits, the multiplier comprising: a plurality of multiplexers, wherein each of said multiplexers is controlled by a respective pair of bits of the selected constant value, and wherein each of said multiplexers has a plurality of inputs, and is connected to receive the input data value, the inverse of the input data value, and all zeros on said inputs, and being controlled such that it outputs either the input data value, the inverse of the input data value, or all zeros, depending on the values of the respective pair of bits of the selected constant value; a plurality of variable shift blocks, each connected to receive an input from a respective one of said multiplexers, and each being adapted to shift its received input by a first bit shift value or a second bit shift value, depending on the values of the respective pair of bits of the selected constant value, wherein the first bit shift value and the second bit shift value differ by 1; and combination circuitry, for receiving the outputs from the plurality of variable shift blocks, and for combining the outputs from the plurality of variable shift blocks and applying further bit shifts, to form an output value equal to the result of multiplying the input data value by the selected constant value.

2. A multiplier as claimed in claim 1, wherein said combination circuitry comprises an adder, connected to receive the outputs from two of said plurality of variable shift blocks, and a fixed shift block, adapted to receive an output from the adder as its input, and to shift its received input by a fixed bit shift value.

3. A multiplier as claimed in claim 2, wherein the fixed shift block comprises a left shifter, for shifting its input by one or more bit positions to the left, without truncation.

4. A multiplier as claimed in claim 2, further comprising at least one further fixed shift block, connected to receive an input from a respective one of said multiplexers, and being adapted to shift its received input by a respective fixed bit shift value.

5. A multiplier as claimed in claim 1, wherein there is one said multiplexer for each said pair of bits of the selected constant value.

6. A multiplier as claimed in claim 1, further comprising at least one multiplexer, controlled by a respective pair of bits of the selected constant value, wherein said multiplexer has at least one input, and is connected to receive one or more of the input data value, the inverse of the input data value, and all zeros on said at least one input, and being controlled such that it outputs one of said received values, depending on the values of the respective pair of bits of the selected constant value.

7. A multiplier as claimed in claim 1, wherein at least one of said variable shift blocks comprises a left shifter, for shifting its input by one or more bit positions to the left, without truncation.

8. A method of multiplying an input data value by a selected constant value in CSD form, the method comprising: dividing the selected constant value into a plurality of pairs of bits; controlling each of a plurality of multiplexers on the basis of bit values of a respective one of said pairs of bits of the selected constant value; outputting from each of said plurality of multiplexers either the input data value, the inverse of the input data value, or all zeros, depending on the bit values of the respective one of said pairs of bits of the selected constant value; applying the outputs from said plurality of multiplexers to respective variable shift blocks; controlling each of the variable shift blocks on the basis of the bit values of the respective one of said pairs of bits of the selected constant value, such that said variable shift block shifts its received input by a respective first bit shift value or a respective second bit shift value, wherein the first bit shift value and the second bit shift value differ by 1; and combining the outputs from the plurality of shift blocks and applying further bit shifts, to form an output value equal to the result of multiplying the input data value by the selected constant value.

9. A method as claimed in claim 8, wherein said combining step includes a step of adding the outputs from two of said plurality of variable shift blocks, and applying a fixed bit shift to a result of said addition.