Arithmetic Module Generator (AMG) supports various hardware algorithms for two-operand adders and multi-operand adders. These hardware algorithms are also used to generate multipliers, constant-coefficient multipliers and multiply accumulators. In the following, we briefly describe the hardware algorithms that can be handled by AMG.
The most straightforward implementation of a final stage adder for two n-bit operands is a ripple carry adder, which requires n full adders (FAs). The carry-out of the ith FA is connected to the carry-in of the (i+1)th FA. Figure 1 shows a ripple carry adder for n-bit operands, producing n-bit sum outputs and a carry out.
The main idea behind carry look-ahead addition is an attempt to generate all
incoming carries in parallel and avoid waiting until the correct carry
propagates from the stage (FA) of the adder where it has been generated.
The carry, Ci+1, produced at the ith stage is given as follows:
.
The equation can be interpreted as stating
that there is a carry either if one is generated at that stage or if
one is propagated from the preceding stage. In other words, a carry
is generated if both operand bits are 1, and an incoming carry is
propagated if one of the operand bits is 1 and the other is 0.
Therefore, let Gi and Pi denote the generation and propagation at
the ith stage, we have:
for operand bit xi and yi and carry-in ci. These expressions allow us to
calculate all the carries in parallel from the operands.
For example, the carries for a 4-bit adder are given as
A 4-bit carry look-ahead adder is designed so as to realize the above expressions.
The idea of the ripple-block carry look-ahead addition is to lessen the fan-in and fan-out difficulties inherent in carry look-ahead adders. A ripple-block carry look-ahead adder (RCLA) consists of N m-bit blocks arranged in such a way that carries within blocks are generated by carry look-ahead but carries between blocks are rippled.
The block size m is fixed to 4 in the generator. The RCLA design is obtained by using multiple levels of carry look-ahead. If there are five or more blocks in a RCLA, 4 blocks are grouped into a single superblock, with the second level of look-ahead applied to the superblocks.
Figure 2 shows the parallel prefix graph of a 64-bit RCLA, where the symbol (solid circle) indicates an extension of the fundamental carry operator described at Parallel prefix adders.
Another way to design a practical carry look-ahead adder is to reverse the basic design principle of the RCLA, that is, to ripple carries within blocks but to generate carries between blocks by look-ahead. A block carry look-ahead adder (BCLA) is based on the above idea.
Figure 3 shows the parallel prefix graph of a 64-bit BCLA, where the symbol (solid circle) indicates an extension of the fundamental carry operator described at Parallel prefix adders.
Parallel prefix adders are constructed out of fundamental carry
operators denoted by ¢ as follows:
(G'', P'') ¢ (G', P') = (G''+G'·P'', P'·P''),
where P'' and P' indicate the propagations, G'' and G' indicate the generations.
The fundamental carry operator is represented as Figure 4.
A parallel prefix adder can be represented as a parallel prefix graph consisting of carry operator nodes.
Figure 5 is the parallel prefix graph of a Ladner-Fischer adder. This adder structure has minimum logic depth, but has large fan-out requirement up to n/2.
Figure 6 is the parallel prefix graph of a Kogge-Stone adder. This adder structure has minimum logic depth, and full binary tree with minimum fun-out, resulting in a fast adder but with a large area.
Figure 7 is the parallel prefix graph of a Brent-Kung adder. This adder is the extreme case of maximum logic depth and minimum area.
Figure 8 is the parallel prefix graph of a Han-Carlson adder. This adder has a hybrid design combining stages from the Brent-Kung and Kogge-Stone adder.
The basic idea in the conditional sum adder is to generate two sets of outputs for a given group of operand bits, say, k bits. Each set includes k sum bits and an outgoing carry. One set assumes that the eventual incoming carry will be zero, while the other assumes that it will be one. Once the incoming carry is known, we need only to select the correct set of outputs (out of the two sets) without waiting for the carry to further propagate through the k positions.
In this generator, we divide the given n-bit operands into two groups of size n/2 bits each. Each of these can be further divided into two groups of n/4 bits each. This process can, in principle, be continued until a group of size 1 is reached. The above idea is applied to each of groups separately.
Figure 9 depicts a conditional sum adder for 4-bit operands.
The underlying strategy of the carry-select adder is similar to that of the conditional-sum adder. Each group generates two sets of sum bits and an outgoing carry. One set assumes that the incoming carry into the group is 0, the other assumes that it is 1. When the incoming carry into the group is assigned, its final value is selected out of the two sets. Unlike the conditional-sum adder, the sizes of the kth group is chosen so as to equalize the delay of the ripple-carry within the group and the delay of the carry-select chain from group 1 to group k. In this generator, the group lengths follow the simple arithmetic progression 1, 1, 2, 3, ...
Figure 10 shows a 16-bit carry select adder.
A carry-skip adder reduces the carry-propagation time by skipping over groups of consecutive adder stages. The carry-skip adder is usually comparable in speed to the carry look-ahead technique, but it requires less chip area and consumes less power.
In the carry-skip adder, any adder stage can be skipped for which Pm = xm exor ym = 1, where Pm indicates the m th carry propagate. The adder structure is divided into blocks of consecutive stages with a simple ripple-carry scheme. Every block also generates a block-carry-propagate signal that equals 1 if all stages internal to the block satisfy Pm = 1. This signal can be used to allow an incoming carry to skip all the stages within the block and generate a block-carry-out. Figure 11 shows an example block consisting of k bit positions j, j+1, ..., j+k-1.
Figure 12 shows an 8-bit carry-skip adder
consisting of four fixed-size blocks, each of size 2. The fixed block
size should be selected so that the time for the longest
carry-propagation chain can be minimized.
The optimal block size k_opt
follows:
.
Figure 13 shows a 16-bit carry-skip adder consisting of seven variable-size blocks. This optimal organization of block size includes L blocks with sizes k1, k2, ..., kL = 1, 2, 3, ..., 3, 2, 1. This reduces the ripple-carry delay through these blocks.
Please note that the delay information of carry-skip adders
in Reference data
page
is simply estimated by using false
paths instead of true paths.
Figure 14 compares the delay information of true paths and
that of false paths in the case of Hitachi 0.18 um process.
Table 1 shows hardware algorithms that can be selected for multi-operand adders in AMG, where the bit-level optimized design indicates that the matrix of partial product bits is reorganized to optimize the number of basic components.
Array is a straightforward way to accumulate partial products
using a number of adders.
The n-operand array consists of n-2 carry-save adder. Figure 15 shows an array
for 18-operand, producing 2 outputs, where CSA indicates a carry-save adder having three multi-bit inputs and two multi-bit outputs.
Wallace tree is known for their optimal computation time, when adding multiple operands to two outputs using carry-save adders. The Wallace tree guarantees the lowest overall delay but requires the largest number of wiring tracks (vertical feedthroughs between adjacent bit-slices). The number of wiring tracks is a measure of wiring complexity. Figure 16 shows an 18-operand Wallace tree, where CSA indicates a carry-save adder having three multi-bit inputs and two multi-bit outputs.
Balanced delay tree requires the smallest number of wiring tracks but has
the highest overall delay compared with the Wallace tree and
the overturned-stairs tree.
Figure 17 shows an 18-operand balanced delay tree, where CSA indicates a carry-save adder having three multi-bit inputs and two multi-bit outputs.
Overturned-stairs tree requires smaller number of wiring tracks
compared with the Wallace tree and has lower overall delay compared with the balanced
delay tree.
Figure 18 shows an 18-operand overturned-stairs tree, where CSA indicates a carry-save adder having three multi-bit inputs and two multi-bit outputs.
(4;2) compressor tree has a more regular structure than an
ordinary CSA tree made of (3,2) counters because the partial products
are added up in the form of a binary tree.
Figure 19 shows an 18-operand (4;2) compressor tree, where (4;2)
indicates a carry-save adder having four multi-bit inputs and two
multi-bit outputs.
Dadda tree is based on (3,2) counters. To reduce the hardware complexity, we allow the use of (2,2) counters in addition to (3,2) counters. Given the matrix of partial product bits, the number of bits in each column is reduced to minimize the number of (3,2) and (2,2) counters.
A (7,3) counter tree is based on (7,3) counters. To reduce the hardware complexity, we allow the use of (6,3), (5,3), (4,3), (3,2), and (2,2) counters in addition to (7,3) counters. We employ Dadda's strategy for constructing (7,3) counter trees.
Redundant binary (RB) addition tree has a more regular structure than an ordinary CSA tree made of (3,2) counters because the RB partial products are added up in the binary tree form by RB adders. The RB addition tree is closely related to (4;2) compressor tree. Figure 20 shows a 9-operand RB addition tree, where RB indicates a carry-save adder having two RB inputs and one RB output. Note here that the RB number should be encoded into a vector of binary digit in the standard binary-logic implementation. In this generator, we employ a minimum length encoding based on positive-negative representation.
Partial products result from the logical AND of the multiplicand with the multiplier.
Partial products are generated with Radix-4 modified Booth recoding. The Booth recoding of the multiplier reduces the number of partial products (and hence has a possibility of reducing the amount of hardware involved and the execution time).
AMG provides parallel multipliers consisting of Partial Product
Generator (PPG), Partial Product Accumulator (PPA), and Final Stage
Adder (FSA) as shown in Figure 21. The PPG stage first generates
partial products from the multiplicand and multiplier in parallel.
The PPA stage then performs multi-operand addition for all the
generated partial products and produces their sum in carry-save form.
Finally, the carry-save form is converted to the corresponding binary
output at FSA.
> generate multipliers
AMG provides constant-coefficient multipliers in the form: P=R*x, where R is an integer coefficient, and X and P are the integer input and output. The hardware algorithms for constant-coefficient multiplication are based on multi-input 1-output addition algorithms (i.e., combinations of PPAs and FSAs). There are many possible choices for the multiplier structure for a specific coefficient R. The complexity of multiplier structures significantly varies with the coefficient value R.
We consider here the use of special number representation
called Signed-Weight (SW) number system, which is useful for
constructing compact PPAs. At present, the combination of
CSD (Canonic Signed-Digit) coefficient encoding technique
with the SW-based PPAs seems to
provide the practical hardware implementation of fast
constant-coefficient multipliers. As a result, AMG supports
such hardware algorithms for constant-coefficient
multiplication, where the range of R is from -2^{31} to
2^{31}-1.
A constant-coefficient multiplier is given as a part of MACs as follow.
> generate constant-coefficient multiplications
AMG provides multiply accumulators in the form: P = ∑
X_{i} * Y_{i}, where X_{i} indicates an
integer variable or constant, Y_{i} indicates an integer
variable. Figure 22 shows a n-term multiply accumulator. A multiply
accumulator is generated by a combination of hardware algorithms for
multipliers and constant-coefficient multipliers. All the partial
products from PPGs are accumulated in carry-save form by a single PPA.
The carry-save form is converted to the corresponding binary output by
an FSA. Figure 23 show a simple multiply accumulators with function P = X*Y+Z,
which is frequently used in DSP systems.
> generate generalized MAC
> generate simple MAC
Figure 22. Generalized MAC | Figure 23. Simple MAC |
Figure 24 shows how the MAC is employed in actual DSP applications. The structure (a) illustrates a typical situation, where the MAC is used to perform a multiply-add operation in an iterative fashion. On the other hand, the structure (b) shows a faster design, where two product terms are computed simultaneously in a single iteration. You can further increase the number of product terms computed in a single cycle depending on your target applications.
(a) | (b) | |
Figure 24. Application of MAC |