High Level Synthesis: from Algorithm to Digital Circuit- P29 ppt

The number of bits executed in every cycle coincides with the width of the addition fragment scheduled in that cycle Unlike additions, the algorithm must select the exact portion of the

14.3.3 Bound Update

Once an operation (or a fragment) has been scheduled in a cycle c, it is checked if the

distribution defined by the actual value of the bound is still reachable Otherwise the value of the bound is updated with the next most uniform distribution still reachable This occurs when:

• The sum of the computational costs of operations scheduled in cycle c does not reach the bound and there are not new operations left that could be scheduled in

it, either because they are already scheduled, or their mobilities have changed

(CCS(τ,c) < bound)∧ (UOPτc=φ),

where

UOPτ

c: set of unscheduled operations of type τ whose mobility makes their

scheduling possible in cycle c.

The new bound value is the previous one plus the value needed to reach the

bound in cycle c divided by the number of open cycles (included in the mobility

of the unscheduled operations)

NewBound = bound + bound − CCS(τ,c)

OC where, OC = {c ∈ N|UOPτc φ}.

• The sum of the computational costs of the operations scheduled in cycle c equals

the bound and there exists at least one unscheduled operation whose mobility

includes cycle c, but even fragmented cannot be scheduled in its mobility cycles.

(CCS(τ,c) = bound)∧



∃op ∈ UOPτc | c ∈∑μop (bound−CCS(τ,c)) < width(op)



,

where

μop: set of cycles included in the mobility of operation op.

The new bound value is the old one plus, for every operation satisfying the above condition, the computational cost of the operation fragment that cannot be scheduled divided by the number of cycles of its mobility

NewBound = bound +

COST(op) − ∑

c ∈ μop (bound−CCS(τ,c))

μop .

14.3.4 Operation Fragmentation

In order to schedule an addition fragment in a certain cycle, it is not necessary to define the portion of the addition to be calculated in that cycle It will be fixed once

the operation has been completely scheduled, i.e when all the addition fragments have been scheduled Then the algorithm selects the LSB of the operation to be exe-cuted in the earliest of its execution cycles, and so on until the MSB are calculated

in the last cycle Due to carry propagations among addition fragments, any other arrangement of the addition bits would require more computations to produce the correct result The number of bits executed in every cycle coincides with the width

of the addition fragment scheduled in that cycle

Unlike additions, the algorithm must select the exact portion of the multiplica-tion that will be executed in the selected cycle To do so, it transforms the operamultiplica-tion into a set of smaller multiplications and additions One of these new multiplications corresponds to the fragment to be scheduled there, and the other fragments continue unscheduled The selection of every fragment type and width is required to calculate the mobility of the unscheduled part of the multiplication, and of the predecessors and successors of the original operation as well Thus, it must be done immedi-ately after scheduling a multiplication fragment in order to avoid reductions in the mobility of all the affected operations

Many different ways can be found to transform one multiplication into several multiplications and additions However, it is not always possible to obtain a multi-plication fragment of a certain computational cost In these cases, the multimulti-plication

is transformed in order to obtain several multiplication fragments whose sum of computational costs equals the desired cost

In order to avoid reductions in the mobility of the successors and predecessors

of fragmented operations, these must be fragmented too In the case of additions, every predecessor and successor is fragmented into two new operations, one of them as wide as the scheduled fragment The mobility of each immediate prede-cessor ends just before where the addition fragment is scheduled, and the mobility

of each immediate successor begins in the next cycle The remaining fragments of its predecessors and successors inherit the mobility of their original operations These fragmentations divide the computational path into two new independent ones, where the two fragments of a same operation have different mobility

In the case of multiplications, their immediate successors and predecessors may not become immediate successors and predecessors of the new operations Data dependencies among operations are not directly inherited during the fragmenta-tion Instead, the immediate predecessors and successors of every fragment must

be calculated after each fragmentation

14.4 Applications to Allocation Algorithms

The proposed techniques to reduce the HW waste during the allocation phase can be easily implemented in most algorithms This chapter presents a heuristic algorithm that includes most of the proposed techniques [2] First it calculates the mini-mum set of functional, storage, and routing units needed to allocate the operations

of the given schedule, and afterwards, it successively transforms the specification

operations to allocate them to the set of FUs The set of datapath resources can also

be modified in the allocation to avoid the HW waste These modifications consist basically on the substitution of functional, storage, or routing resources for several smaller ones, but do not represent an increment of the datapath area

This algorithm also exploits the proposed allocation techniques to guarantee the maximum bit-level reuse of storage and routing units In order to minimize the stor-age area, some variables may be stored simultaneously in the same register (wider than or equal to the sum of the variables widths), and some variables may be frag-mented and every fragment stored in a different register (the sum of the registers widths must be greater than or equal to the sum of the variables widths) And

to achieve the minimal routing area, some variables may be transmitted through the same multiplexer, and some variables may be fragmented and every fragment transmitted through a different multiplexer

The proposed algorithm takes as input one scheduled behavioural specification and outputs one controller and one datapath formed by a set of adders, a set of multipliers, a set of other types of FUs, some glue logic needed to execute additive and multiplicative operations over adders and multipliers, a set of registers, and a set of multiplexers The algorithm is executed in two phases:

(1) Multiplier selection and binding A set of multipliers is selected and some

specification multiplications are bound to them Some other multiplications are transformed into smaller multiplications and some additions in order to increase the multipliers reuse, and the remaining ones are converted into additions to be allocated during the next phase

(2) Adder selection and binding A set of adders is selected and every addition

bound to it These additions may come from the original specification, the trans-formation of additive operations, or the transtrans-formation of multiplications into smaller ones or directly into additions

The next sections explain the central phases of the algorithm proposed, but first some concepts are introduced to ease their understanding

14.4.1 Definitions

• Internal Wastage (IW) of a FU in a cycle: percentage of bits discarded from the

result in that cycle (due to the execution of one operation over a wider FU)

• Maximum Internal Wastage Allowed (MIWA): Maximum average IW of every

multiplier in the datapath allowed by the designer A MIWA value of 0% means that no HW waste is permitted (i.e every multiplier in the datapath must execute one operation of its same width in every cycle)

• Multiplication order: One multiplication of width m × n (being m ≥ n) is bigger than other one of width k × l (being k ≥ l) if either (m > k) or (m = k and n > l).

• Occurrence of width n in cycle c: number of operations of width n scheduled in cycle c.

• Candidate: set of operations of the same type which satisfy the following

conditions:

– all of them are scheduled in different cycles

– (m ≥ n) for every width n of the candidate operations, where m is the width of

the biggest operation of the candidate

There exist many different bit alignments of the operations comprised in a can-didate In order to reduce the algorithm complexity, only those candidates with the LSB and the MSB aligned are considered Thus, if one operation is executed over a wider FU the MSB or the LSB of the result produced are discarded

• Interconnection saving of candidate C (IS): sum of the number of bits of the operands of C candidate operations that may come from the same sources, and the number of bits of the results of C candidate operations which may be stored

in the same registers

IS(C) = BitsOpe(C) + BitsRes(C),

where

BitsOpe(C): number of bits of the left and right operands that may come from

the same sources

BitsRes(C): number of bits of the C candidate results that may be stored in the

same set of storage units

• Maximum Computed Additions Allowed per Cycle (MCAAC): maximum number

of addition bits computed per cycle This parameter is calculated once there are not unallocated multiplications left, and it is obtained as the maximal sum of the addition widths in every cycle

14.4.2 Multiplier Selection and Binding

In order to avoid excessive multiplication transformations, and thus obtain more structured datapaths, this algorithm allows some HW waste in the instanced multi-pliers The maximum HW waste allowed by the designer in every circuit is defined

by the MIWA parameter This phase is divided into the following four steps, and fin-ishes when either there are not remaining unallocated multiplications left, or when

it is not possible to instance a new multiplier without exceeding MIWA (due to the given scheduling) This check is performed after the completion of every step The steps 1–3 are executed until it is not possible to instance a new multiplier with a valid MIWA Then, step 4 is executed followed by the adder selection and binding phase

14.4.2.1 Instantiation and Binding of Multipliers Without IW

For every different width m × n of multiplications, the algorithm instances as many

multipliers of that width as the minimum occurrence of multiplications of that width

per cycle Next, the algorithm allocates operations to them For every instanced

mul-tiplier of width m ×n, it calculates the candidates formed by as many multiplications

of the selected width as the circuit latency, and the IS of every candidate The algo-rithm allocates to every multiplier the operations of the candidate with the highest

IS Multipliers instanced in this step execute one operation of its same width per cycle, and therefore their IW is zero in all cycles

14.4.2.2 Instantiation and Binding of Multipliers with Some IW

The set of multiplications considered in this step may come from either the orig-inal specification, or the transformation of multiplications (performed in the next

step) For every different width m × n of multiplications, and from the biggest, the algorithm checks if it is possible to instance one m × n multiplier without

exceed-ing MIWA It considers in every cycle the operation (able to be executed over an

m × n multiplier) that produces the lowest IW of an m × n multiplier After every

successful check the algorithm instances one multiplier of the checked width, and allocates operations to it Now the candidates are formed by as many operations as the number of cycles in which at least there is one operation that may be executed

over one m × n multiplier The width of the candidate operation scheduled in cycle

c equals the width of the operation used in cycle c to perform the check, such that

each candidate has the same number of operations of equal width Once all can-didates have been calculated, the algorithm computes their corresponding IS, and allocates the operations of the candidate with the highest IS Multipliers instanced

in this step may be unused during several cycles, and may also be used to execute narrower operations (being the IW average of these multipliers in compliance with MIWA)

14.4.2.3 Transformation of Multiplications into Several Smaller

Multiplications

This step is only performed when it is not possible to instance a new multiplier

of the same width as any of the yet unallocated multiplications without exceeding MIWA It transforms some multiplications to obtain one multiplication fragment of

width k × l from each of them These transformations increase the number of k × l

multiplications, which may result in the final instance of a multiplier of that width (during previous steps) First the algorithm selects both the width of the operations

to be transformed and the fragment width, and afterwards a set of multiplications of the selected width, which are finally fragmented

The following criteria are used to select the multiplication and fragment widths:

(1) The algorithm selects as m × n (width of the operations to be transformed) and

k × l (fragment width), the widths of the two biggest multiplications that satisfy

the following two conditions:

• There is at least one k × l multiplication, being k × l < m × n, that can be executed over one m × n multiplier (i.e m ≥ k and n ≥ l).

• At least in one cycle there is one m × n multiplication scheduled and there are not k × l multiplications scheduled.

(2) The algorithm selects two different widths as the widths of the operations to

be fragmented, and a fragment width independent of the remaining unallocated

multiplications The widths selected of the operations to be fragmented m ×n and k ×l, are those of the biggest multiplications that satisfy the following

conditions:

• At least in one cycle there is one m × n multiplication scheduled and there are not k × l multiplications scheduled.

• At least in one cycle there is one k × l multiplication scheduled and there are not m × n multiplications scheduled.

In this case the fragment width equals the maximum common multiplicative kernel

of m × n and k × l multiplications, i.e min(m,k) × min(n,l).

Next the algorithm selects the set of operations to be fragmented In the first case

it is formed by one m × n multiplication per every cycle where there are not k × l multiplications scheduled And in the second one, it is formed by either one m × n

or one k × l multiplication per cycle In the cycles where there exist operations of

both widths scheduled, only one multiplication of the largest width is selected Once the set of operations to be fragmented and the desired fragment width are selected, the algorithm decides which one out of the eight different possible fragmentations

is selected, according to the following criteria:

• The best fragmentations are the ones that obtain, in addition to one multiplication

fragment of the desired width, other multiplication fragments of the same width

as any of the yet unallocated multiplications

• Among the fragmentations with identical multiplication fragments, the one that

requires the lowest cost in adders is preferable

Figure 14.7 illustrates the eight different fragmentations of one m ×n multiplica-tion explored by the algorithm to obtain one k × p multiplication fragment.

Fig 14.7 Multiplication fragmentations explored by the algorithm

14.4.2.4 Transformation of Multiplications into Additions

Due to the given schedule it is not always possible to instance a new multiplier without exceeding MIWA Therefore, unallocated multiplications are transformed into several additions

14.4.3 Adder Selection and Binding

14.4.3.1 Instantiation and Binding of Adders Without IW

The set of additions considered here may come from the original specification, the transformation of multiplications (performed in the previous phase), or the

transfor-mation of additions (step 4.3.3) For every different width n of unallocated additions,

the algorithm instances as many adders of that width as the minimum occurrence of additions of that width per cycle Next, operations are allocated to them For every

instanced adder of width n, it calculates the candidates formed by as many additions

of the selected width as the circuit latency, and the IS of every candidate The algo-rithm allocates to every adder the operations of the candidate with the highest IS The IW of the adders instanced here is zero in all the cycles

14.4.3.2 Instantiation and Binding of Adders with Some IW

For every different width n of unallocated additions, and from the biggest, the algo-rithm checks if it is possible to instance one n adder without exceeding MCAAC It considers in every cycle the operation (able to be executed over an n adder) that pro-duces the lowest IW of an n bits adder After every successful check, the algorithm

instances one adder of the checked width, and allocates operations to it Now the candidates are formed by as many operations as the number of cycles where there is

at least one operation that may be executed over one n bits adder The width of the candidate operation scheduled in cycle c equals the width of the operation used in cycle c to perform the check Once all candidates are calculated, their corresponding

IS are computed, and the additions of the candidate with the highest IS allocated Adders instanced in this step may be unused during several cycles, and may also be used to execute narrower operations (being the IW of these adders in compliance with MCAAC)

14.4.3.3 Transformation of Additions

This step is only performed when it is not possible to instance a new adder of the same width as any of the yet unallocated additions without exceeding MCAAC

Some additions are transformed to obtain one addition fragment of width m from

each of them These transformations increase the number of m bits additions, which

may result in the final instance of an adder of that width (during previous steps) First the algorithm selects both the set of the operations to be transformed and the fragment width, and afterwards it performs the fragmentation of the selected addi-tions The fragment size is the minimum width of the widest unallocated operation scheduled in every cycle A maximum of one operation per cycle is fragmented each time, but only in cycles without unallocated operations of the selected width The set of fragmented operations is formed by the widest unallocated addition scheduled

in every cycle without operations of the selected width Every selected addition is decomposed into two smaller ones, being one of fragments of the desired width These fragmentations produce the allocation of at least one new adder of the selected width during the execution of the previous steps, and may also contribute to the allocation of additional adders

14.5 Analysis of the Implementations Synthesized Using

the Proposed Techniques

This section presents some of the synthesis results obtained by the algorithms described previously which include some of the bit level design techniques pro-posed in this chapter These results have been compared to those obtained by a HLS

commercial tool, Synopsys Behavioral Compiler (BC) version 2001.08, to evaluate

the quality of the proposed methods and their implementations in HLS algorithms The area of the implementations synthesized is measured in number of inverters, and includes the area of the FUs, storage and routing units, glue logic, and controller The clock cycle length is measured in nanoseconds The RT-level implementations

produced have been translated into VHDL descriptions to be processed by Synopsys Design Compiler (DC) to obtain the area and time reports The design library used

in all the experiments is VTVTLIB25 by Virginia Tech based on 0 25μm TSMC

technology

14.5.1 Implementation Quality: Influential Factors

The main difference between conventional synthesis algorithms and our approach is the number of factors that influence the quality of the implementations obtained The implementations proposed by conventional algorithms depend on the specification size, the operation mobility, and the specification heterogeneity, measured as the number of different triplets (type, data format, width) present in the original specifi-cation divided by the number of operations Otherwise, our algorithms minimize the influence of data dependencies and get implementations totally independent from the specification heterogeneity, i.e from the number, type, data format, and width

of the operations used to describe behaviours

Just to illustrate these influences we have synthesized different descriptions of the same behaviour, shown in Table 14.3, first with the proposed algorithms, and after-wards with BC These descriptions have been created by progressively transforming

(from circuit A to G) some of the specification operations into several smaller ones,

in order to increase the number of operations of every different type and width, such that the specification heterogeneities of these descriptions have been progressively

reduced from A to G Circuit A is the original specification formed by 30

opera-tions with six different operation types (MAC, multiplication, addition, subtraction, comparison, and maximum), two different data formats (unsigned and two’s com-plement), and eight different operation widths (4, 8, 12, 16, 24, 32, 48, and 64 bits)

And circuit G is the specification obtained after the last transformation It consists

of 86 unsigned multiplications and additions of 26 different widths, thus being the description with the smallest heterogeneity Table 14.3 shows the number of opera-tions (# Operaopera-tions), the number of different operation types (# Types), the number

of different data formats (# Formats), and the number of different operation widths (# Widths) present in each synthesized specification The latency in all cases is equal

to ten cycles The amount of area saved by the algorithm grows, in general, with the specification heterogeneity In the circuits synthesized, the homogeneous distribu-tion of the computadistribu-tional costs among cycles achieved by our algorithm has also resulted in substantial clock cycle length reductions Figure 14.8 shows in this set

of examples the amount of area and clock cycle length saved by our approach in function of the heterogeneity

Table 14.3 Features of the synthesized descriptions

Circuit # Operations # Types # Formats # Widths Heterogeneity

Fig 14.8 Experimental area and execution time of different descriptions of one specification

Table 14.4 Area results of the synthesis of some modules of the ADPCM decoder algorithm

ADPCM decoder Datapath Commercial tool Fragmentation techniques

OPFC + SCA FUs 928 478

OPFC + SCA Controller 62 66

OPFC + SCA Multiplexers 418 470

OPFC + SCA Registers 461 475

OPFC + SCA Total area 1,873 1,503 gates (19.7% saved)

Total all modules 3,897 3,032 (22.2% saved)

14.5.2 An Application Example

As an example of a real circuit, we have synthesized the following modules of the ADPCM decoding algorithm described in the Recommendation G.721 of CCITT:

• Inverse Adaptative Quantizer (IAQ)

• Output PCM Format Conversion (OPFC)

• Synchronous Coding Adjustment (SCA)

• Tone and Transition Detector (TTD)

Table 14.4 compares the area of the modules synthesized by our approach and BC for a fixed value of the circuit latency The amount of area saved by our algorithm averages 22%

OPFC and SCA modules have been synthesized together, and IAQ and TTD inde-pendently Better results could be obtained if all modules were synthesized together, because it would increase the number of operations that could be executed over the same FU, and the number of variables that could be stored in the same register

14.5.3 Synthesis of Non Heterogeneous Specifications

As shown in previous sections, our algorithm substantially reduces the area of cir-cuits synthesized from heterogeneous specifications, as compared with commercial tools and previous known approaches But the application method is not just lim-ited to heterogeneous specifications Important area reductions can also be achieved when specifications formed by operations with the same data formats and widths are