The document discusses optimizing SPI performance for a multi-LED project involving RGB LEDs controlled by shift registers. It highlights various techniques, such as staggered SPI output and loop unrolling, to reduce CPU usage and improve refresh rates while minimizing bit-banging, which is noted to be significantly slower. The author also mentions various experiences and humorous anecdotes related to moose to provide a light-hearted context.

1. Staggering SPI
Performance
Or how I learned to stop bit-banging and love SPDR
--Dg

2. Project background
• Multi-led project
• Shift register based controllers
• Software PWM
• 40-60% cpu use target
Wi nøt tei a høliday in Sweden this yer?

3. Software PWM
• 32 rgb leds, 96 individual leds
• 12 8-bit shift registers to turn leds on/off
• 12 bytes output per round
• 128 levels of color desired
• ~100Hz refresh rate wanted
• 12,800 rounds of sending 12 bytes/second
See the løveli lakes

4. NO BITBANGING
• Bit twiddling may be fun, but it is slow
• Order of magnitude slower than hardware
spi - numbers later
The wonderful telephøne system

5. Arduino Hardware SPI
• Suggestions for use found online suggest:
while(morestuff) {
data = morestuff;
SPDR = data;
while(!(SPSR & (1<<SPIF)));
}
And mani interesting furry animals

6. Waiting for ?
• 63.2µs / 12 byte round of output
• 12,500 rounds/second - 80% cpu used
• 28% of time spent waiting for SPI to be
ready
• waiting sucks
Including the majestic møøse

7. Staggered SPI output
• Do work while waiting:
data = stuff;
SPDR=data;
while(morestuff) {
data = morestuff;
while(!(SPSR & (1<<SPIF)));
SPDR=data;
}
A Møøse once bit my sister ...

8. Staggered numbers
• 49.1µs/round - 61% cpu spent doing PWM
• 7% of that still spent spinning for SPSR
• More to peel out?
No realli! She was Karving her initials on the møøse with the sharpened end of an interspace tøøthbrush given her by Svenge - her brother-in-law - an Oslo dentist and the star of many Norwegian møvies: “The Høt Hands of an Oslo Dentist”, “Fillings
of Passion”, “The Huge Mølars of Horst Nordfink”

9. Squeezing more
• Loop unrolling buys more time
• 42.4µs/round - 53% cpu usage
SPDR = data; data=morestuff; SPSRWait;
SPDR = data; data=morestuff; SPSRWait;
SPDR = data; data=morestuff; SPSRWait;
SPDR = data; data=morestuff; SPSRWait;
....
Mynd you, møøse bites Kan be pretty nasti...

10. Bitbanging?
• 586.6µs/round
• 73% cpu usage
• only 1000 rounds - 10 levels of color at
100Hz or 20 at 50Hz
• any higher flickers
Did you kno that Møøse like bit banging?

11. More techniques
• Smoother PWM by spreading out work
• Hiding SPSR waiting in interrupts
• Clock counting, hand assembler
• Take advantage of avr opcodes *(--pData)
becomes a single opcode (dec and
derefence)
They dø! Dirty, dirty bit banging møøses...

12. Fin
Daniel Garcia
dgarcia@dgarcia.net
--Dg
Møøse trained by TUTTE HERMSGERVORDENBROTBORDA