Lookup() - number of CPU cycles to execute seems high at 40

guyadams · Joined: 22 Nov 2006 Posts: 26

Hi,

I am using Lookup() to read values stored in an array in Flash. Using the simulator it reports that it takes 40 cycles to execute the instruction. Is this correct and is there a faster method?

Just edited to make clearer Smile

Thanks
Guy

albertsm · Joined: 09 Apr 2004 Posts: 5922 Location: Holland

you should include the used processor. best way is $regfile with the value.
also, it is best to shrink the code to the essential part.

what is important : the used processor. the data type of the used variables.
in general : a small as possible program that can be compiled but demos the problem or question.

40 is right when i assume some things. it can be made quicker but that depends on the used processor.

guyadams · Joined: 22 Nov 2006 Posts: 26

Thanks Mark - have tidied the code example so it can be cut & pasted.

I have not thought about the operations needed - 40 seemed high and there are 8 Lookup() in the real code loop which equates to 320 cpu cycles...... I could migrate to an ATMega328PB with enough SRAM to copy the table into at startup if this is going to be faster?

Cheers
Guy

albertsm · Joined: 09 Apr 2004 Posts: 5922 Location: Holland

i explain how lookup works. for your micro that is.
- the address is needed for the start of the table. either in flash or ram. that requires 2 LDI instructions.
- an offset need to be added so in any case, 2 add instructions are required
- this offset need to be loaded from the used variable in order to save it to the regs used for adding
- with this info the lookup is called which will use LPM to restore the data from flash
- the data is stored to the variable of your choice.

you use 2 look ups. it would be simpler to use 1 lookup and change the table.
when you store the low and high value after each other, you can lookup a word and this safes the overhead of addressing.
when you use overlay you also have this info in byte form

dim bL as byte, bH as byte
dim W as word at BL overlay

the table data must have the form ;
data &Habcd% the % for indicating integer/word data

so this will safe a lot of cycles already.
when cycle_count increases, and is not random, you could go a step further.
that would require some asm to store the pointers.
this has the benefit that you do not need to add the offset each time.

when you want to go that route you best show (again simple) how the other lookup calls come into play.

MWS · Joined: 22 Aug 2009 Posts: 2262

One quart of the flash usage is sufficient, as a complete sinus can be mirrored from one quadrant.
This applies also for using SRam.

To split the table in upper and lower word makes no sense and only slows the code down.

For having speed, use ASM and a global read pointer, R8/R9 is a good choice, as it is used by the data pointer for Read()
Advance the read pointer by 32 bytes each and you have 8 support points.

If you don't like ASM, you can try with Read(), as it keeps track of its current read-position. andn does not need to calculate every time again an absolute read address.
By reading a word with Read(), it advances the read-pointer one word = 2 bytes. For addressing the next support point, you need to advance (lptr) R8 & (hptr) R9 by 14.
Only for the first quadrant of course, for the other three quadrants some calculations are required.

EDC · Joined: 26 Mar 2014 Posts: 971

In the simple words: take the first of your Sineh data &H82 and then join it with the first Sinel data &HD3 so it becomes a &H82D3%
The procent sign describe it as two byte variable.
Do it same till the end of those 200B two tables means to the value &H7FFF%

guyadams · Joined: 22 Nov 2006 Posts: 26

Hi Mark,

Thanks - tried it, but it seems to only save two cycles 80 -> 78. Maybe I have misunderstood something.....

Cheers
Guy

guyadams · Joined: 22 Nov 2006 Posts: 26

EDC · Joined: 26 Mar 2014 Posts: 971

This is how it can be measured. Simply insert into your code the "nop" instruction before you want to measure something. Then run Simulator and "Run to this place" then zeros the counter by double click on it. Only then run to the next nop.

I dont think that 61 instructions at 16_000_000 Hz is a critical value here.

guyadams · Joined: 22 Nov 2006 Posts: 26

EDC · Joined: 26 Mar 2014 Posts: 971

0.0038ms I round up into 0.004ms. It means 4us.
If this slows down your code then Im curious how fast is rest of it because it means 250 KHz!

MWS · Joined: 22 Aug 2009 Posts: 2262

guyadams · Joined: 22 Nov 2006 Posts: 26

EDC · Joined: 26 Mar 2014 Posts: 971

Now we're talking Very Happy

This is how it looks like with SRAM

[/code]

guyadams · Joined: 22 Nov 2006 Posts: 26