Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!rutgers!sun-barr!texsun!pitstop!sundc!seismo!uunet!mcvax!inria!crcge1!lri!eb
From: eb@lri.lri.fr (Erick Bizouarn)
Newsgroups: comp.sys.transputer
Subject: Transputer timing
Message-ID: <412@lri.lri.fr>
Date: 10 May 89 16:51:51 GMT
Sender: eb@lri.lri.fr
Lines: 85

We have some questions on the notion of time of the transputer. It has a 
strange behavior. Consider the following example where we measures the
execution time of the ldc (load constant) instruction. The time does not
always increase with the number of the instruction in the measured loop.

We have obtained the following results on a T414:

	-----------------------------------------
	| number of ldc inst |  execution time	|
	-----------------------------------------
	|	0	     |  T = 16832 us.	|
	-----------------------------------------
	|	1	     |  T - 960 us.	|
	-----------------------------------------
	|	2	     |  T - 1984 us.	|
	-----------------------------------------
	|	3	     |  T - 512 us.	|
	-----------------------------------------
	|	4	     |  T + 2048 us.	|
	-----------------------------------------


-------------------------------------------------------------------------------

	.text
_mess:
	.ascii	"time = %d us.\12\0"
.globl _main
_main:
	adjw	-3
	ldc	0
	sttimer
	ldtimer
	stl	1
	ldc	10000
	stl	2
	.align 2	; alignment on a divisible by 4 boundary.
L0:
###
ldc	0
.....	(n times)
ldc	0
###
	ldl	2
	adc	-1
	stl	2

	ldl	2
	eqc	0
	cjmp	L0

	ldtimer
	ldl	1
	sub
	ldc	64	; for low priority timer
	mul
	adc	-16832	; time for a loop without any ldc 0 instruction.
	ldc	_mess-L1
	ldpi
L1:
	ldl	4
	call	_printf
	adjw	3
	ret

--------------------------------------------------------------------------------

Where is the mistake ???

May a longer program be executed faster ? Perhaps it will be explained by
instruction prefetch, but we would like an detailled explanation.

Furthermore is there a means to write out of the memory of the INMOS B004 
transputer card without halting the transputer.

Thanks in advance for any hint.

********************************************************************************

Bizouarn Erick
L.R.I. (Universite d'Orsay) France

E-mail: eb@lri.lri.fr

********************************************************************************