Efficient Software Performance Estimation Methods for Hardware/Software Codesign

Kei SuzukiAlberto Sangiovanni-

Vincentelli

Present: Yanmei Li

10/29/2002 EE249 Discussion Session

Introduction One of the most important purposes of hw/sw

codesign is to find the optimum hw/sw partition of a system level specification under particular criteria

Criteria Performance(speed, or the number of clock cycles) Cost(number of components, die size, or code size)

Estimation At a lower abstraction level

easy and accurate, but long design iteration time At a higher abstraction level

reduce the exploring time Play an important role in the synthesis and optimization

10/29/2002 EE249 Discussion Session

Software Performance Estimation

Cost of a mixed hw/sw system based on a standard micro-processor depends on the hw size

Solution: Implement a given functionality with a program on the microprocessor

Problem: Software implementation often fails to meet the performance requirement

Tradeoff: To implement the critical portion in the program

with hardware Software performance estimation is the key

10/29/2002 EE249 Discussion Session

POLIS System CFSMs (Codesign Finite State Machines)

Does not discriminate between hw and sw Estimation provides preliminary timing information

and also a measure for hw/sw partitioning A partitioning process takes place to identify the

candidate components for sw implementation S-Graph (Software graph)

To optimize the trade-off between the performance and the code size of the final implementation

Estimation is helpful for s-graph optimization and sw module scheduling

10/29/2002 EE249 Discussion Session

Related Work Software performance depends on the structure

of the software program as well as on the components of the target system

The structure of the software program is more difficult to estimate as the abstraction level rises

Most of the results are from the object code level which is the lowest level of abstraction, and are concerned with software that has a limited structure

A number of approaches have been proposed A simple prediction method Statistical methods ……

10/29/2002 EE249 Discussion Session

Abstraction Models in POLIS

CFSM HW: be mapped into

an abstract hardware description format, and synthesized into a combinational circuit and a set of latches

SW: be is translated into a data structure called s-graph

10/29/2002 EE249 Discussion Session

Abstraction Models in POLIS

S-Graph: A DAG(directed

acyclic graph) with one source node and one sink node

Represent the control flow of a given behavior

Four types of node: BEGIN, END, TEST, ASSIGN

10/29/2002 EE249 Discussion Session

S-Graph Semantics:

Start with the BEGIN node Traverse each node along its edge, until reaching the END

node At a TEST node, select one corresponding child with the

value of the associated predicate P(V) At an ASSIGN node, assign the value of the associated

function A(V) to the output variable z Translate an s-graph into a C program

Traverse the graph in a depth-first manner TEST: if (or switch) statement ASSIGN: assignment statement

The resulting C program has the same structure

10/29/2002 EE249 Discussion Session

Performance Estimation Methods

Modeling the target system The structure of C code generated by POLIS:

Function() ……(1)

{

Initialization of local variable(assignment statements); ……(2)

Structure of mixed if or switch statements and assignment statements; ……(3)

Return; ……(4)

}

10/29/2002 EE249 Discussion Session

Modeling the Target System

Execution timeT=Tpp +k Tinit +Tstruct

Code sizeS=Spp +k Sinit +Sstruct

Tpp (Spp) :for entering and exiting the function (1)+(4) Tinit ( Sinit):for initializing local variables(2).

k is the number of local variables. Tstruct (Sstruct):for the structure of mixed conditional

statements generated from TEST nodes and assignment statements generated from ASSIGN nodes(3).

10/29/2002 EE249 Discussion Session

Modeling the Target System(cont.)

Tpp, Spp , Tinit , Sinit are constant which can be determined beforehand

Tstruct =ΣPi Ct (node_type_of(i), variable_type_of(i)) Sstruct =ΣCs (node_type_of(i), variable_type_of(i))

Pi =1 if node i is on a path, otherwise Pi =0 Ct and Cs can be obtained by using simple

benchmark programs containing a mix of the C statement that appears in the generated C programs and analyzing the execution time and code size of the programs on the target compiler and the target CPU

10/29/2002 EE249 Discussion Session

Benchmark Model Four attributes to characterize a system

Name of the parameter set, a name for a unit of execution time, a name for a unit of code size, and the size of an integer variable

seventeen cost parameters to model the execution time, and fifteen cost parameters to model the code size

A TEST node with an event-type variable/multi-valued variable with a bit mask/multi-valued variable

An ASSIGN node with an event-type variable/which assigns a constant to a variable/which assigns one variable to another one

Pre-processing and post-processing A branch operation Initialization of a local variable Average execution time and size for pre-defined software library

functions The size of pointers The size of integer variables

10/29/2002 EE249 Discussion Session

S-graph Level Estimation Property

Property 1. Each node in an S-graph has a one-to-one correspondence with only a few statements in the synthesized C code

Property 2. The form of each statement is determined by the type of corresponding node

Property 3. The S-graph is a DAG, hence it does not include loops in its structure

Each node/edge is weighted according to pre-calculated cost parameters in the pre-process

10/29/2002 EE249 Discussion Session

S-graph Level Estimation Algorithm: SGtrace(sgi)

If (sgi==NULL) return (C(0,∞,0));If(sgi has been visited)

return (pre-calculated Ci(*,*,0) associated with sgi);Ci=initialize (max_time=0; min_time=∞; code_size=0);For each child sgj of sgi{

Cij=SGtrace(sgj)+edge cost for edge eij

If(Cij.max_time> Ci.max_time)Ci.max_time= Cij.max_time;

If(Cij.min_time< Ci.min_time)Ci.min_time= Cij.min_time;

Ci.code_size+= Cij.code_size;}Ci+= node cost for node sgi;Return(Ci);

10/29/2002 EE249 Discussion Session

S-graph Level Estimation The computational complexity: O(E) Average execution time:

Cave =ΣPij (Ct (node_type_of(i), variable_type_of(i))+ Ce (i,j))

Pij is the possibility of executing node i and going to node j

Ce (i,j) is the edge cost for edge eij

10/29/2002 EE249 Discussion Session

CFSM Level Estimation Is much more difficult since a CFSM model does

not closely reflect the code structure MDDs are used to represent the transition

relation function of a CFSM (a node represents a multi-valued variable; ordering is important)

The estimation algorithm of the MDD is based on the assumption that the maximum(minimum) cost path in an MDD is usually the maximum (minimum) cost path in the s-graph that is generated from the MDD

Also based on recursive DFS traversing algorithm There is no relation between the code size of the

number of the MDD nodes

10/29/2002 EE249 Discussion Session

Experimental Results(1)

10/29/2002 EE249 Discussion Session

Experimental Results(2)

10/29/2002 EE249 Discussion Session

Experimental Results(3) Compared to an assembly-level analysis:

S-graph(Table 1): The differences in the maximum execution time

are within (-10%, +10%) The differences in the minimum execution time

are within (-20%,+20%) The differences in code size are within (-20%,

+20%) CFSM(Table 2):

The differences in the maximum execution time are within (-10%,+25%)

The differences in the minimum execution time are within (-20%,+20%)

10/29/2002 EE249 Discussion Session

Conclusions S-graph level method

provides an accurate estimation for all analysis: the maximum and minimum execution time, and code size.

It is a useful technique for optimization in software synthesis because of its accuracy.

CFSM level method is less accurate than the s-graph estimation, but it is

still accurate enough when estimating the maximum and minimum execution time.

is important for automatic partitioning of CFSMs into hardware and software parts, and also for scheduler generation.

10/29/2002 EE249 Discussion Session

Conclusions Two software performance estimation

methods for use with the POLIS hardware/software codesign system are proposed in this paper. S-graph level method CFSM level method

The experimental results showed that the accuracy of both proposed methods is high enough for use in the POLIS system.