courseware high-level synthesis an introduction prof. jan madsen informatics and mathematical...

courseware

High-Level Synthesisan introduction

Prof. Jan Madsen

Informatics and Mathematical ModellingTechnical University of Denmark

Richard Petersens Plads, Building 321DK2800 Lyngby, Denmark

M-1 High-Level Synthesis 2SoC-MOBINET courseware

Hardware synthesis

P1P2

P3

CPU ASIC

P1 P2 & P3

Starts from an abstract behavioral description

Generates an RTL description

Need to restrict the target hardware – otherwise search space is too large


Hardware synthesis

P1P2

P3

CPU ASIC

P1 P2 & P3

How is the behavior specified? Natural languages C/C++ VHDL/Verilog

What is the target architecture of the ASIC?


Hardware model - components

Most synthesis systems are targeted towards synchronous hardware

Functional units: Can perform one or more computations Addition, multiplication, comparison, ALU, etc.

Registers: Store inputs, intermediate results and outputs May be organized as a register file


Hardware model - interconnection

Multiplexers: Select one output from several inputs

Busses: Connection shared between several components Only one component can write data at a specific time Exclusive writing may be controlled by tri-state drivers


Hardware model – parameters

Clocking strategy Single or multiple phase clocks

Interconnect Allowing or disallowing busses

Clocking of functional units Multicycle operations Chaining Pipelined units


Hardware model – example


Hardware concepts

Data path Network of functional units, registers, multiplexers and

buses

Control Takes care of having the data present at the right

place at a specific time Takes care of presenting the right instructions to a

programmable unit

Often high-level synthsis concentrates on data path synthesis


Methodology

implementationdesign specification

Physical domain

Mathematical domain

specification

create a model of the physical problem

synthesis

create an alogorithm to solve the problem

implementation

Transform the optimized model back to the physical domain


Input format

Input Behavior described in textual form Conventional programming language Hardware description language (HDL)

Has to be parsed and transformed into an internal representation

Conventional compiler techniques are used


Internal representation

Data-flow graph (DFG) Used by most systems May or may not contain information on control flow

vertex (node): represent computation

edge: represent precedence relations


Data flow

x := a * b;

y := c + d;

z := x + y;

a b c d

x

*

y

+

z

+


DFG semantics

a b c d

x

*

y

+

z

+


Exercise 1: data flow graph of DiffEq

Solve the second order differential equation y´´ + 3zy´+ 3y = 0

Iterative solution

While (z<a) {

z1 := z + dz;

u1 := u – (3*z*u*dz) – (3*y*dz);

y1 := y + (u*dz);

z := z1; u := u1; y := y1;

}


Exercise 1 - result

+

u1

-

*

-

*

*

* *

* + <

u dz 3 z 3 y udz z dz

y1 ctrl


High-level synthesis

a b c d

x y

+

z

+

*



Scheduling Determine for each operation the time at which it

should be performed such that no precedence contraint is violated

Allocation Specify the hardware resources that will be necessary

Assignment Provide a mapping from each operation to a specific

functional unit and from each variable to a register



Scheduling, allocation and assignment are strongly interrelated

But are often solved separately! Scheduling is NP-complete – heuristics have to

be used!


Scheduling

Input DFG G(V, E) Library of ressource types R

Mapping : V R, (vi ) = r

a given operation may be mapped to different ressource type, e.g. + may be performed by an adder or an ALU

execution delay:(vi ) = di

ressource type cost: (r)


Scheduling

Start time of operations T = { ti : i = 0, 1, …, n }

Scheduling is the task of determining the start times subject to the precedence constraints of the DFG : V Z+

vi ) = ti such that ti tj + dj, i, j : (vj, vi ) E

Latency: = tn – t0

Cost of schedule: r R(r) Nr()]


Scheduling

implementation specification

Physical domain

Mathematical domain

specification synthesis implementation

C program

DFG

Scheduling algorithm

Scheduled DFG


Scheduling – ASAP

Map operations to their earliest possible start time not violating the precedence constraints

Easy and fast to compute Find longest path in a directed acyclic graph No attemp to optimize ressource cost

Gives the fastest possible schedule if unlimited amount of resources are available

Gives an upper bound on execution speed


ASAP algorithm

For each node vi V do

if pred(vi) = Ø then

Ei = 1;

V = V – { vi };

else

Ei = 0;

endif

endfor

While V ≠ Ø do

for each node vi V do

if all_sched(pred(vi),E) then

Ei = max(pred(vi),E) + 1;

V = V – { vi };

endif

endfor

endwhile


DiffEq

+

u1

-

*

-

*

*

* *

* + <

u dz 3 z 3 y z dz

y1 ctrl

udz


Exercise 2 – latency and resources

Assume: cycle time = 25 ns d*, d+, d-, d< = 25 ns

What is the latency of the schedule? How many resources are needed? How many resources are needed, if we introduce an ALU

(+,-,<) What is the latency if we have only 1 multiplier? What is the latency if

d* = 25ns and dALU = 12ns


Exercise 2 – result

What is the latency of the schedule? 4*25ns = 100ns

How many resources are needed? 4*, 1+, 1-, 1<

How many resources are needed, if we introduce an ALU (+,-,<) 4*, 2ALU

What is the latency if we have only 1 multiplier? 7*25ns = 175ns

What is the latency if d* = 25ns and dALU = 12ns 3*25ns = 75ns (operator chaining)


Scheduling - ALAP

+

u1

-

*

-

*

* *

**

+ <

u dz 3 z 3 y z dz

y1 ctrl

udz


Scheduling – ALAP

Map operations to their latest possible start time not violating the precedence constraints

Needs a latency constraint Easy and fast to compute

Find longest path in a directed acyclic graph No attemp to optimize ressource cost


Scheduling – ASAP / ALAP

Are ASAP and ALAP useful? ASAPvi ) = Ei

ALAPvi ) = Li

Operator flexibility = Li – Ei

Also known as mobility

Mobility = 0 operator has to be scheduled at Ei

otherwise latency constraint is violated

Mobility > 0 gives scheduling freedom


Scheduling – list based

Generalization of ASAP Priority-list of ready nodes A ready node is an operator that has all

predecessors already scheduled The priority-list is always sorted with respect to a

priority function


List scheduling algorithm

ins_ready_ops(V,PListr1, PListr2

,…, PListrm);

Cstep = 0;

While ((PListr1 Ø) or … or ((PListrm

Ø)) do

Cstep = Cstep + 1;

for k = 1 to m do

for funit = 1 to Nk do

if PListrk Ø then

schdule_op(first(Plistrk),Cstep);

Plistrk = delete(Plistrk

,first(Plistrk));

endif

endfor

endfor

ins_ready_ops(V,PListr1, PListr2

,…, PListrm);

endwhile


DiffEq

+

u1

-

*

-

*

*

* *

* + <

u dz 3 z 3 y z dz

y1 ctrl

udz

[h,0]

[g,0]

[a,0] [b,0]

[e,0] [f,1]

[c,1] [d,2]

[i,2] [k,2]

[j,2] Plist*:

Plist+:

Plist-:

Plist<:

a,b,c,d

j

Ø

Ø

* * +

- <

* * c,d e,c,d *

Ø

k

*

*

*

Ø

f,d

g


List scheduling

Priority may be based on other measures than mobility

Length of longest path to a node with no immediate successor

Number of immediate successor nodes High number means high priority

courseware high-level synthesis an introduction prof. jan madsen informatics and mathematical...

Documents