ASM 32bit divide help!

Javalin

So I want to divide·long (32bit)·numbers, i.e. 30_000_000 / 4000, or 757_002 / 80, etc so I wrote this as a quick divide routine:

asmLongDivide           ' takes X / Y = R
                        mov     r,#0
:loop                   cmpsub  x, y            wc
        if_c            add     r,#1                         
        if_c            jmp     #:loop
asmLongDivide_ret       ret

It appears to work in my test case (110 / 10 = 11) - but when I start using it for real math it seems to "glitch" - i.e. suddenly from data where I should get 2,300 I get lumps of 4,700 or whatever.· I've confirmed its this code, if I let SPIN do the math it works perfectly.

Any ideas?· Its so simple I can't see how it can't not work.

Interestingly de-siliva's example has the WZ specified - but can't see how that would work either, as you'd want to still sub if X=Y?

asmLongDivide           ' takes X / Y = R
                        mov     r,#0
:loop                   cmpsub  x, y            wc, wz
        if_c            add     r,#1                         
        if_c_and_nz     jmp     #:loop
asmLongDivide_ret       ret

Cheers,

James

kuroneko

Javalin said...
It appears to work in my test case (110 / 10 = 11) - but when I start using it for real math it seems to "glitch" - i.e. suddenly from data where I should get 2,300 I get lumps of 4,700 or whatever. I've confirmed its this code, if I let SPIN do the math it works perfectly.

... example has the WZ specified - but can't see how that would work either, as you'd want to still sub if X=Y?

Works for me (your example that is). What numbers did you try and how did you call your routine?

As for wz, if x == y you have C (and Z) set and the sub is done already. Not looking at Z makes you go around the loop one more time for nothing (x < y || x == 0). Call it an early exit if you want.

Javalin

Hiya

Its called like this:

                        mov     x, rawvalue                                     ' convert to MS (X / 80_000)
                        mov     y, eighty_thousand
                        call    #asmLongDivide
                        mov     pulseMS, r

for example.

>Works for me (your example that is).
hmmmm.· Interesting.··OK will have to dig yet more.

Cheers!

James

BR

Javalin,

Are you trying to divide signed integers with that routine?

Javalin

Nope - should be positive integer only.

James

Javalin

ok, so i've done some more testing and its not that code.· Putting a loop of the range of values into the code instead of picking up from hardware proved the math.· Doh, but thanks for the code verification!

I am sampling the pulse length of a pin in ASM - high-low-high time like this:

readPulseLength         waitpeq RPMPULSEPIN,RPMPULSEPIN              ' Wait until pin goes HIGH
                        mov     starttime,cnt                        '
                        waitpeq zero,RPMPULSEPIN                     ' Wait until pin goes LOW
                        waitpeq RPMPULSEPIN,RPMPULSEPIN              ' Wait until pin goes HIGH                        
                        mov     stoptime,cnt                         '
                        sub     stoptime,starttime                   '
                        mov     rawvalue, stoptime                   ' leave the value in rawvalue                              
readPulseLength_ret     ret

the problem seems to be when the pulse length is ~ 2ms then the rawvalue times jump arround.· I know that that spin executes instructions at ~50us so I would imagine that ASM can see a 2ms pulse without breaking sweat!

I thought about using the counters, but then I still have to waitpeq/waitpne for signal edges, so why would that improve the timing? (apart from avoiding cnt roleover issues)

Am I making a mistake here?· Is there a better way to do this, preferably non-blocking so I can include other functions in the cog too?

Thanks again,

James

Post Edited (Javalin) : 6/27/2010 2:50:52 PM GMT

kuroneko

Javalin said...
the problem seems to be when the pulse length is ~ 2ms then the rawvalue times jump arround. I know that that spin executes instructions at ~50us so I would imagine that ASM can see a 2ms pulse without breaking sweat!

You can measure quite a lot with that piece of code [noparse]:)[/noparse] Imagine a 2ms pulse width but the low part is only e.g. 25% of that (.5ms). So depending on when your code is called you will get results ranging from 2ms (best case) downto .5ms (worst case). So I'd suggest catching two edges (LH) instead of level and edge.

Javalin

Hiya

Ah, I gettit. I'll check it out.

thanks,

James

Javalin

Ah.· Got it.· One line sorted it.

readPulseLength         [i]waitpeq zero,RPMPULSEPIN[/i]                      ' Wait until pin goes LOW
                        waitpeq RPMPULSEPIN,RPMPULSEPIN               ' Wait until pin goes HIGH
                        mov     starttime,cnt                         '
                        waitpeq zero,RPMPULSEPIN                      ' Wait until pin goes LOW
                        waitpeq RPMPULSEPIN,RPMPULSEPIN               ' Wait until pin goes HIGH                        
                        mov     stoptime,cnt                          '
                        sub     stoptime,starttime                    '
                        mov     rawvalue, stoptime                    ' leave the value in rawvalue                              
readPulseLength_ret     ret

Now it waits for a complete cycle and is still accurate.

Cheers all,

James

Post Edited (Javalin) : 6/28/2010 7:36:09 PM GMT