ESE535:
Electronic Design Automation
Day 13: February 27, 2013
Dataflow
Penn ESE535 Spring 2013 -- DeHon
1
Previously
A1
A3
A5
A2
A4
A6
A7
A8
A9 A10 A11 A12 A13
B2
B1
B3
B10
B11
B4
B5
B6
B7
• Scheduling of
concurrent operations
B8
B9
Penn ESE535 Spring 2013 -- DeHon
2
Want to See
• Abstract compute model
– natural for parallelism and hardware
• Describe computation abstracted from
implementation
– Defines correctness
Penn ESE535 Spring 2013 -- DeHon
3
Behavioral
(C, MATLAB, …)
Today
Arch. Select
Schedule
RTL
• Dataflow
• SDF
– Single rate
– Multirate
• Dynamic Dataflow
• Expression
FSM assign
Two-level,
Multilevel opt.
Covering
Retiming
Gate Netlist
Placement
Routing
Layout
Masks
Penn ESE535 Spring 2013 -- DeHon
4
Parallelism Motivation
Penn ESE535 Spring 2013 -- DeHon
5
Producer-Consumer Parallelism
Stock
predictions
encrypt
• Can run concurrently
• Just let consumer know when producer
sending data
Penn ESE535 Spring 2013 -- DeHon
6
Pipeline Parallelism
ME
DCT
VQ
code
• Can potentially all run in parallel
• Like physical pipeline
• Useful to think about stream of data
between operators
Penn ESE535 Spring 2013 -- DeHon
7
DAG Parallelism
Decode
video
Check/decode
block
synchronize
Decode
audio
• Doesn’t need to be linear pipeline
• Synchronize inputs
Penn ESE535 Spring 2013 -- DeHon
8
Graphs with Feedback
+
+
×k
×k
×k
×k
• In general may hold state
• Very natural for many tasks
Penn ESE535 Spring 2013 -- DeHon
9
Definitions
Penn ESE535 Spring 2013 -- DeHon
10
Dataflow / Control Flow
Dataflow
• Program is a graph
of operators
• Operator consumes
tokens and
produces tokens
• All operators run
concurrently
Control flow (e.g. C)
• Program is a
sequence of
operations
• Operator reads
inputs and writes
outputs into
common store
• One operator runs
at a time
– defines successor
Penn ESE535 Spring 2013 -- DeHon
11
Token
• Data value with presence indication
– May be conceptual
• Only exist in high-level model
• Not kept around at runtime
– Or may be physically represented
• One bit represents presence/absence of data
Penn ESE535 Spring 2013 -- DeHon
12
Token Examples?
• What are familiar cases where data may
come with presence tokens?
– Network packets
– Memory references from processor
• Variable latency depending on cache presence
– Start bit on serial communication
Penn ESE535 Spring 2013 -- DeHon
13
Operator
• Takes in one or more inputs
• Computes on the inputs
• Produces results
• Logically self-timed
– “Fires” only when input set present
– Signals availability of output
Penn ESE535 Spring 2013 -- DeHon
14
Penn ESE535 Spring 2013 -- DeHon
15
Dataflow Graph
• Represents
– computation sub-blocks
– linkage
• Abstractly
– controlled by data presence
Penn ESE535 Spring 2013 -- DeHon
16
Dataflow Graph Example
Penn ESE535 Spring 2013 -- DeHon
17
In-Class Dataflow Example
Penn ESE535 Spring 2013 -- DeHon
18
Stream
• Logical abstraction of a persistent pointto-point communication link
– Has a (single) source and sink
– Carries data presence / flow control
– Provides in-order (FIFO) delivery of data
from source to sink
stream
Penn ESE535 Spring 2013 -- DeHon
19
Streams
• Captures communications structure
– Explicit producerconsumer link up
• Abstract communications
– Physical resources or implementation
– Delay from source to sink
Penn ESE535 Spring 2013 -- DeHon
20
Register Transfer Level
(RTL)
Behavioral
(C, MATLAB, …)
• Describe computation as logic
and registers
• Equations (logic) define values
to be clocked into next register
• Typically what you right in
VHDL, Verilog
Arch. Select
Schedule
RTL
FSM assign
Two-level,
Multilevel opt.
Covering
Retiming
Gate Netlist
Placement
Routing
Layout
Masks
Penn ESE535 Spring 2013 -- DeHon
21
Dataflow Abstracts Timing
• Doesn’t say
– on which cycle calculation occurs [contrast RTL]
• Does say
– What order operations occur in
– How data interacts
• i.e. which inputs get mixed together
• Permits
– Scheduling on different # of resources
– Operators with variable delay [examples?]
– Variable delay in interconnect [examples?]
Penn ESE535 Spring 2013 -- DeHon
22
Examples
• Operators with Variable Delay
– Cached memory or computation
– Shift-and-add multiply
– Iterative divide or square-root
• Variable delay interconnect
– Shared bus
– Distance changes
• Wireless, longer/shorter cables
– Computation placed on different cores?
Penn ESE535 Spring 2013 -- DeHon
23
Difference:
Dataflow Graph/Pipeline
Penn ESE535 Spring 2013 -- DeHon
24
Clock Independent Semantics
Interconnect
Takes n-clocks
Penn ESE535 Spring 2013 -- DeHon
25
Semantics
• Need to implement semantics
– i.e. get same result as if computed as
indicated
• But can implement any way we want
– That preserves the semantics
– Exploit freedom of implementation
Penn ESE535 Spring 2013 -- DeHon
26
Dataflow Variants
Penn ESE535 Spring 2013 -- DeHon
27
Synchronous Dataflow (SDF)
• Particular, restricted form of dataflow
• Each operator
– Consumes a fixed number of input tokens
– Produces a fixed number of output tokens
– When full set of inputs are available
• Can produce output
– Can fire any (all) operators with inputs
available at any point in time
Penn ESE535 Spring 2013 -- DeHon
28
Synchronous Dataflow
+
+
×k
×k
×k
×k
Penn ESE535 Spring 2013 -- DeHon
29
SDF: Execution Semantics
while (true)
Pick up any operator
If operator has full set of inputs
Compute operator
Produce outputs
Send outputs to consumers
Penn ESE535 Spring 2013 -- DeHon
30
Multirate Synchronous Dataflow
• Rates can be different
– Allow lower frequency operations
– Communicates rates to CAD
• Something not clear in RTL
• Use in scheduling, provisioning
– Rates must be constant
• Data independent
2
decimate
1
Penn ESE535 Spring 2013 -- DeHon
31
SDF
• Can validate flows to check legal
– Like KCL  token flow must be conserved
– No node should
• be starved of tokens
• Collect tokens
• Schedule onto processing elements
– Provisioning of operators
• Provide real-time guarantees
• Simulink is SDF model
Penn ESE535 Spring 2013 -- DeHon
32
SDF: good/bad graphs
1
1
2
1
1
1
1
Penn ESE535 Spring 2013 -- DeHon
1
1
1
1
1
33
SDF: good/bad graphs
1
1
1
2
1
2
1
1
1
1
Penn ESE535 Spring 2013 -- DeHon
2
2
1
1
1
1
1
1
34
Dynamic Rates?
• When might static rates be limiting?
– Compress/decompress
• Lossless
• Even Run-Length-Encoding
– Filtering
• Discard all packets from geraldo
– Anything data dependent
Penn ESE535 Spring 2013 -- DeHon
35
Data Dependence
• Add Two Operators
– Switch
– Select
Penn ESE535 Spring 2013 -- DeHon
36
Switch
Penn ESE535 Spring 2013 -- DeHon
37
Filtering Example
Geraldo?
discard
dup
switch
Penn ESE535 Spring 2013 -- DeHon
38
Select
Penn ESE535 Spring 2013 -- DeHon
39
Constructing
If-Then-Else
Penn ESE535 Spring 2013 -- DeHon
40
Select Example
In-Order
Merge of
Streams
(smallest
to largest)
select
select
>
switch
switch
select
Penn ESE535 Spring 2013 -- DeHon
41
Looping
• for (i=0;i<Limit;i++)
Penn ESE535 Spring 2013 -- DeHon
42
Universal
• Once we add switch and select,
the dataflow model is as power as any
other
– E.g. can do anything we could do in C
– “Turing Complete” in formal CS terms
Penn ESE535 Spring 2013 -- DeHon
43
Dynamic Challenges
• In general, cannot say
– If a graph is well formed
• Will not deadlock
– How many tokens may have to buffer in
stream
– Right proportion of operators for
computation
Penn ESE535 Spring 2013 -- DeHon
44
Feels a bit tacked-on
…not flow that well with rest.
Might think about how to do better.
Expression
How would we capture this in a
Programming Language?
Penn ESE535 Spring 2013 -- DeHon
45
Expression
• Could express operators in C/Java
– Each is own thread
• Link together with Streams
• E.g. SystemC
Penn ESE535 Spring 2013 -- DeHon
46
C Example
while (!(eos(stream_a) && !(eos(stream_b))
A=stream_a.read();
B=stream_b.read();
Out=(a+b)*(a-b);
stream_out.write(Out);
Penn ESE535 Spring 2013 -- DeHon
47
Connecting up Dataflow
stream stream1=new stream();
operator prod=new stock(stream1);
operator cons=new encrypt(stream1);
Stock
predictions
Penn ESE535 Spring 2013 -- DeHon
encrypt
48
What’s the Point?
• Seen repeatedly: exploit freedom
– To reduce costs
• A1: How do we capture freedom that exists
in computational?
– Higher-level than an implementation
– Perhaps as a useful intermediate
• A2: How do we allow freedom for
implementations (or instances) to take
variable time?
Penn ESE535 Spring 2013 -- DeHon
49
Summary
• Dataflow Models
– Simple pipelines
– DAGs
– SDF (single, multi)-rate
– Dynamic Dataflow
• Allow
– express parallelism
– freedom of implementation
Penn ESE535 Spring 2013 -- DeHon
50
Big Ideas:
• Dataflow
– Natural model for capturing computations
– Communicates useful information for
optimization
• Linkage, operator usage rates
• Abstract representations
– Leave freedom to implementation
Penn ESE535 Spring 2013 -- DeHon
51
Admin
• Homework 4 Due Today
• Spring Break next week
• Back on Monday 3/11
– Reading on Blackboard
Penn ESE535 Spring 2013 -- DeHon
52