LARD simulation system: LARD to C translation

1
LARD simulation system LARD to C translation
Behavioralsimulation
Balsa
LARD
synthesis
circuits
2
LARD example
Divide_and_apply( minval(Ttype), maxval(T) )
expr(void) ( if (max-mingtMAX_diff) then
( middlevar(T). middle
(minmax)/2 Divide_and_apply( min,
middle) Divide_and_apply( middle, max
) ) else f( middlevar(T)
(minmax)/2. fct(middle) ) )
3
1) Block declarations2) Threads3) Types
( ( ( ) )
forever ( ) )
4
1) Block declarations2) Threads some figures3)
Types

• LARD code simulation
• Dhrystone benchmark (15000 instructions)
• 495,404 thread creations
• 34,767,587 thread switching
• BALSA code simulation
• hello world with spamulet (130 instructions)
• 51,480 thread creations
• 25,563,604 thread switching

Jabez 2 min
Jabez 1 min
5
1) Block declarations2) Threads thread
switching3) Types
F( ) sub_1b( ) ... ...
forever( sub_1b ) ... forever( expr )
while (1) expr( )
Method 1 switch stacks and
registers
6
1) Block declarations2) Threads thread
switching3) Types
Method 2
F( ) sub_1b( ) ... ...
forever( sub_1b ) ... forever( expr )
while (1) expr( )
Yield( )
Save sub_1b position
Save F position
Save expr, forever position
7
1) Block declarations2) Threads thread
switching3) Types

• Method 2 saverestore sub-function
  states
• Typical sub-routines chain
• lt ? while ? function ? channel_send ? yield
  gt
• switching cost
• 4 saves 4 loads 4 function returns 4
  function calls

8
1) Block declarations2) Threads methods
comparison3) Types

• Method 1 Method 2
• stack switch sub-fonction state
• saverestore

? switch speed 5x faster ?
? on the stack local vars difficult ?
already done ?
not optimal speed ?
? difficult getting thread info easy ?
9
1) Block declarations2) Threads Switching
method 23) Types
( ( ( ) )
forever ( ... Yield ) )

• Analyse each fct and sub-fct
• Need to saverestore thread?
• Need debugging information?
• Optimise argument passing

10
1) Block declarations2) Threads Results3) Types
Old
( Total 1545 s )
Lard2c
( Total 100 s )
( Total 61 s )
11
1) Block declarations2) Threads Conclusions3)
Types

• Is it possible to switch threads faster ?
• How to get debugging information ?
• Easy variable control OR speed ?

Method 1 seems promising but
12
1) Block declarations2) Threads3) Types

• Types represented by their size
• string and mpint pointers to external
  structures
• need to be explicitly initialised and destroyed
• cant use generic functions

Divide_and_apply( minval(Ttype), maxval(T) )
expr(void) ( if (max-mingtMAX_diff) then
( middlevar(T). middle
(minmax)/2 Divide_and_apply( min,
middle) Divide_and_apply( middle, max
) ) else f( middlevar(T)
(minmax)/2. fct(middle) ) )
gt Need a new representation for types
13
Conclusions

• Thread creation and switching can still be
  improved
• Need new system for handling types
• To get
• Faster simulation
• More robust (Types)
• Saverestore simulations