Subject:
|
Re: RCX Firmware speed versus legOS
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Thu, 11 Apr 2002 23:47:19 GMT
|
Reply-To:
|
Dick Swan <DICKSWA@SBCGLOBAL.NETstopspam>
|
Viewed:
|
537 times
|
| |
|
|
"Michael Obenland" <obenland@t-online.de> wrote in message
news:GuEtHr.M39@lugnet.com...
> <<snip>>
> ----------------------------------------------------------
>
> It took 0.211 sec. running empty, i.e. without reading the sensor • but only
> doing calculations, and 1.113 sec. with reading the sensor. So I can • assume
> that reading 10000 light values takes about 1 second.
I was surprized at the wide disparity between 'empty loop' and loop
with 'reading the sensor'. I suspect there is significant compile time
code optimization at play here. I use Hitachi's H8 tool chain and not
GCC. Hitachi's compiler did lots of compile time optimization on the
empty loop in recognizing that many of the variables have constant
value.
I compiled three different versions of the program to see how many
instructions were included in the main loop. Results as follows:
[1] 9 7 'empty loop'
[2] 13 11 add sensor read to loop.
[3] 22 20 eliminated compile time code optimization.
Case [1] had nine instructions in the loop. 7 of these instructions
are executed; the other two are part of 'if..then..' code that is
never executed. In case [3] need to add the time taken in the call to
'divide' subroutine. Case [3] generated inline code for the multiply
by 3 code.
Note that a smarter compiler could generate even more efficient code
and get the loop code down to two instructions by eliminating the code
between labels L541-l544 in the snippet below.
The remainder of this note contains the actual code generated by
compiler for any 'hard core' readers.
The code generated by Hitachi's H8 compiler for [1] is next. I
manually inserted '//' comments.
max = 0;
SUB.W R4,R4 // compiler used R4 a register variable
min = 5000;
MOV.W #5000:16,R3 // compiler used R3a register variable
mean = 0; // compiler generated no code.
value = 0; // compiler generated no code.
for (index2 = 1; index2 <= kNumberOfLoops; ++index2)
MOV.W #1:16,R0 // compiler used R0a register variable
BRA L542
{
//value = sensor2.sensorRaw();
if( max < value ) max = value;
L541: MOV.W R4,R4
BGE L543
SUB.W R4,R4
L543:
if( min > value ) min = value;
MOV.W R3,R3
BLE L544
SUB.W R3,R3
L544: ADDS.W #1,R0
L542: CMP.W R1,R0
BLE L541
mean = ( mean * 3 - mean + value ) / 3; // compiler generated no
code.
}
I recompiled with the sensor read instruction included, i.e. [2]
above. The same optimizations occurred as shown below.
max = 0;
SUB.W R4,R4
min = 5000;
MOV.W #5000:16,R0
MOV.W R0,@SP
mean = 0;
value = 0;
for (index2 = 1; index2 <= kNumberOfLoops; ++index2)
MOV.W #1:16,R5
BRA L547
{
value = sensorRead(2);
L546: MOV.W #2:16,R0
BSR _sensorRead
MOV.W R0,R6
if( max < value ) max = value;
CMP.W R0,R4
BGE L548
MOV.W R0,R4
if( min > value ) min = value;
L548: MOV.W @SP,R0
CMP.W R6,R0
BLE L549
MOV.W R6,@SP
L549: ADDS.W #1,R5
L547: CMP.W R3,R5
BLE L546
mean = ( mean * 3 - mean + value ) / 3;
}
Finally, I forced all code to be generating by adding, outside the
loop, the statement 'extVariable = mean + value', i.e. case [3]. Lots
more code generated.
max = 0;
SUB.W R4,R4
min = 5000;
MOV.W #5000:16,R0
MOV.W R0,@(2:16,SP)
mean = 0;
value = 0;
SUB.W R6,R6
MOV.W R6,R5
for (index2 = 1; index2 <= kNumberOfLoops; ++index2)
MOV.W #1:16,R3
BRA L548
{
value = sensorRead(2);
L547: MOV.W #2:16,R0
BSR _sensorRead
MOV.W R0,R6
if( max < value ) max = value;
CMP.W R0,R4
BGE L549
MOV.W R0,R4
if( min > value ) min = value;
L549: MOV.W @(2:16,SP),R0
CMP.W R6,R0
BLE L550
MOV.W R6,@(2:16,SP)
L550:
mean = ( mean * 3 - mean + value ) / 3;
MOV.W R5,R0
MOV.W R0,R1
ADD.W R0,R0
ADD.W R1,R0 // inline code for '*3'
SUB.W R5,R0
ADD.W R6,R0
MOV.W #3:16,R1
JSR @$DIVI$3:16
MOV.W R0,R5
ADDS.W #1,R3
L548:MOV.W @SP,R0
CMP.W R0,R3
BLE L547
}
extVariable = mean + value; // to 'disable' compile time code
optimization
ADD.W R6,R5
MOV.W R5,@_extVariable:16
|
|
Message has 1 Reply:
 | | Re: RCX Firmware speed versus legOS
|
| (...) Of course you are right. <shame on me> (...) Hm? You don't use the gcc toolchain for the H8 but another compiler? Can you tell a bit more about your compiler? I revised my testings without optimizing out the calculation and the empty loop took (...) (22 years ago, 12-Apr-02, to lugnet.robotics)
|
Message is in Reply To:
| | RCX Firmware speed versus legOS
|
| The discussion about the Firmware speed is intresting indeed. So I did a fast, by no means representative, reserch. The following program is a legOS program reading the light sensor, calculating some values and printing the elapsed time: ---...--- (...) (22 years ago, 11-Apr-02, to lugnet.robotics)
|
3 Messages in This Thread:
 
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
This Message and its Replies on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
