Conference rusure::math

236.0. "Matrix Multiply with SISAL" by HARE::STAN () Wed Mar 13 1985 19:40

From:	BABEL::WILL         13-MAR-1985 16:21
To:	ALGOL::STAN
Subj:	Talk on SISAL code generation

Wim Bohm will be giving an update on the status of the Machester Dataflow
Machine, this Friday afternoon in Babbage as a ZK technical seminar.  In
addition to that talk, he will also be giving a seminar on SISAL code
generation which we are inviting the languages groups to.  It will be
this Friday, 10-12, in the Euler conference room.

Wim Bohm is a Research Assistant at the University of Manchester, England.
His work on the Manchester Dataflow Project has included code generation,
support for separate compilation, and streamlining of the compilation tools.

 
OUTLINE OF SISAL CODE GENERATION TALK:
======================================
1 Very short introduction to SISAL using matrix multiply as example.
2 Relevant details of the Manchester Datflow Machine ( Proliferate, Structure
  Store orders Select and Fetch-Stream-from-Structure-Store, Array
  representation, Array access )
3 Colouring strategy applied to Call, Forall and Forinit.
4 Where did the MIPS go in the old compiler.
5 Methods to improve the MIPS/MFLOPS ratio and their effect on matrix multiply
  program (together improving this ratio by a factor of 10), including:
	Simple (not recoloured) loops 
	Constant lower bounds
	Rewriting the program (knowing more about the code generated)
	Order code improvement (example Reduction)
 
  and also (but not influencing matrix multiply):
 
	Simple (not recoloured) calls in loops
	Literal triggering
6 Other things we will do:
	Inlining calls ( making more loops simple )
	Multiple representations of data structures ( for records and streams )
	Thinking about avoiding array copying
	Throttling
 

Posted:	Wed 13-Mar-1985 14:37 EST
To:	RDVAX::CURTIN,ERLANG::FRANKEL,SYZYGY::SOPKA,SYZYGY::DUFFY,@[WELCH]TECHLANG.DIS,@[]ALL.DIS