Gedae Signal Flow Graph Language

1
Gedae Signal Flow Graph Language

• James W. SteedGedae, Inc.Telephone 856 231
  4458Fax [phone]Email
  gedae_at_gedae.comWebsite www.gedae.com

2
Motivation

• Middleware cannot solve every problem
• The E library only has so many vector routines
• Middleware induces overhead to be more general
  purpose
• The E library must be written for several
  different contingencies (strides,
  in-place/out-of-place, etc.)
• Source code ties the developer to a processor
• The Gedae primitive library ties the developer to
  processors with a C cross compiler (FPGAs?)
• Need a language to program any processor that
  still allows assembly-level optimization

3
Overview

• Introduction to signal flow graphs
• Delay registers
• Port characteristics
• Clock derivation
• Introduction to single sample primitives
• Coding language
• Core interface
• Integration with data flow graphs

4
Introduction to Signal Flow Graphs

• Each graph is a synchronous circuit
• Streams represent buffers in memory
• Circuit is enabled by a clock
• Example out(i) in(i) in(i-1)

Memory buffers
Delay register
Enable signal
5
Delay Registers

• Delay registers create a delay of 1
• Delay registers are the only way to store state
  information in a SFG graph

in(i) in(i-1)
in(i)
in(i-1)
6
Port Characteristics

• Specify port characteristics using same context
  as primitives
• Array dimensions input stream float inN
• Parameters input float K
• Interpolation/decimation input stream float
  in(N)
• Dynamic output dynamic stream float out

7
Data Types

• In addition to the standard Gedae data types, SFG
  supports
• clock only used to specify Gedae clocks on
  top-level
• boolean true or false
• intltkgt k-bit integer
• e.g., int9, int21, int32
• uintltkgt k-bit unsigned
• e.g., uint9, uint21, uint32
• untypedltkgt k-bit untyped
• e.g., untyped9, untyped21, untyped32

8
Clocks

• Clocks are special local data items on the
  top-level of the SFG
• Gedae turns the clocks into enable signals that
  coordinate movement of data from inputs to outputs

9
Number of Clock Pulses

• If it takes more than one clock pulse to process
  a token, then this rate can be specified in the
  clocks declaration
• Example vector add by scalar parameter

Each vector has N elements. We need N pulses to
process N elements.
10
Clock Derivation

• Gedae allows the developer to create new clocks
  which are derived from other clocks
• / Divide specifies a static decimation
• clock2 clock/2
• clock 1 1 1 1 1 1 1 1 1 1 1 1 1
• clock2 1 0 1 0 1 0 1 0 1 0 1 0 1
• Addition specifies a delay
• clock3 clock/2 1
• clock 1 1 1 1 1 1 1 1 1 1 1 1 1
• clock3 0 1 0 1 0 1 0 1 0 1 0 1 0

11
Example Vector Sum
Decimating copy on every N-th sample
Vector size N
Use derived clock to clear on N-th sample
12
Single Sample Primitives

• Input and Output sections list ports
• At this level all data is a single sample
• No streams or parameters
• No dimensions
• Local sections are allowed
• But only registers can store state!
• Function method contains code

13
Coding Language

• Primitives are coded using the following
  interface
• Register R(input, output, enable)
• Delay register copies input to output after
  getting an enable signal
• Assignment A(expression, variable)
• Evaluate expression and assign to variable
• Function F(i0,i1,)(o0,o1,) expressions
• Evaluate a set of expressions
• Note Fs are often tied to a target language by
  necessity

14
Coding Language (contd)

• Primitives are coded using the following
  interface
• Conditional IF(test,expr1,expr2,output)
• Set output to expr1 if test, else set output to
  expr2
• Decimate D(variable, enable)
• The variable is only evaluated when the enable
  signal
• Clock C(variable, enable)
• Get the enable signal tied to the variable

15
Coding Language (contd)

• Primitives are coded using the following
  interface
• Memory M(output, length, sizeof)
• Create a local memory buffer
• Memory Read MR(addr, output)
• Read the output from the address
• Memory Write MW(input, addr)
• Write the input to the address

16
Core Interface

• Cores are efficient, pre-built implementations of
  common algorithms
• More generally, cores are target specific code
• Function method uses generic code
• Core method allows inclusion of target-specific
  code

Target specifies 8-bit 0 differently
17
Core Database

• CoreName allows developer to list cores that are
  used
• E.g., CoreName fifo_512x32,pcix_if
• LSP can use these names to look up information in
  a database
• E.g., library, source, or pre-synthesized file
  that the core is stored in

18
Integration with Data Flow

• Each data flow primitive can be specified by a
  SFG or code

19
Code Generation

• Gedae code generates two files from a SFG
• Standard Gedae primitive (Ansi-C)
• Target code (via the Language Support Package)
• The Gedae primitive is made from Function methods
  
• The target code is made from both Core methods
  and Function methods
• For each primitive, use the Function method if
  and only if there is no Core method

20
Alternate Implementations

• Gedae uses the generated C primitive in the DF
  graph

21
Alternate Implementations (contd)

• When mapped to the target
• Gedae switches in the target code
• Gedae inserts necessary comm primitives

Target name
Path to target code
22
(No Transcript)