AVR-Chat

On Feb 13, 2006, at 11:31 AM, kernels_nz wrote:

> 2. if you had :
>
>    mydata[0] = PIND;
>    mydata[1] = PIND;
>    mydata[2] = PIND;
>           .
>           .
>           .
>    mydata[255] = PIND;
>
> all written out the hard way like that, it would be slightly faster,
> because the "for" loop condition would not be checked every time, it
> would of course occupy much more code space, and be considered
> terrible programming.

Yes, but there's still a middle way, which is partial unrolling:

#define ATONCE 8
uint8_t mydata[256];

void myloop(void)
{
     uint8_t *p = mydata;
     for (uint8_t counter = sizeof(mydata) / ATONCE; counter--; p +=  
ATONCE)
     {
         p[0] = PIND;
         p[1] = PIND;
         p[2] = PIND;
         p[3] = PIND;
         p[4] = PIND;
         p[5] = PIND;
         p[6] = PIND;
         p[7] = PIND;
     }
}


which results in (with avr-gcc 4.0 and -O2):

    9:test3.c       **** void myloop(void)
   10:test3.c       **** {
   74               	.LM0:
   75               	/* prologue: frame size=0 */
   76               	/* prologue end (size=0) */
   77 0000 E0E0      		ldi r30,lo8(mydata)
   78 0002 F0E0      		ldi r31,hi8(mydata)
   79               	.L2:
   80               	.LBB2:
   11:test3.c       ****     uint8_t *p = mydata;
   12:test3.c       ****     for (uint8_t counter = sizeof(mydata) /  
ATONCE; counter--; p += ATONCE)
   13:test3.c       ****     {
   14:test3.c       ****         p[0] = PIND;
   82               	.LM1:
   83 0004 80B3      		in r24,48-0x20
   84 0006 8083      		st Z,r24
   15:test3.c       ****         p[1] = PIND;
   86               	.LM2:
   87 0008 80B3      		in r24,48-0x20
   88 000a 8183      		std Z+1,r24
   16:test3.c       ****         p[2] = PIND;
   90               	.LM3:
   91 000c 80B3      		in r24,48-0x20
   92 000e 8283      		std Z+2,r24

... etc ...

   21:test3.c       ****         p[7] = PIND;
110               	.LM8:
111 0020 80B3      		in r24,48-0x20
112 0022 8783      		std Z+7,r24
114               	.LM9:
115 0024 3896      		adiw r30,8  ; 2
116 0026 80E0      		ldi r24,hi8(mydata+256)  ; 1
117 0028 E030      		cpi r30,lo8(mydata+256)  ; 1
118 002a F807      		cpc r31,r24 ; 1
119 002c 59F7      		brne .L2 ; 2

As a result, you amortize the cost of lines 115-119 (7 cycles) over 8  
bytes of transfer.
For a cost/byte transferred of 3+7/8 cycles.

-- 
Ned Konz
ned@bike-nomad.com