ADC unrealiable in ATmega32?

[Psi]

Nope, there's no sequence point within the quoted expression, so the compiler is free to evaluate ADCL and ( ADCH << 8 ) in whichever order it feels like today.
...
Ah, now, that's subtly different -- when you're doing that you use the short-circuiting boolean operators && and ||, which are explicitly defined to evaluate the left operand first and only evaluate the right operand if required to compute the final result...

Just tested this with PIC C and guess what?
...
It starts by reading the high value 

Thanks guys, learn something every day


Safe way to combine low and high is going to be

Code: [Select]

uint16_t adc_value;
adc_value = ADCL;
adc_value = adcvalue | (ADCH << 8) ;Or just use ADCW

[free_electron]

Thats the technique i described above : sliding window integration.

No, it's not. Do read again. Do try it. Compare results. Repeat.

Their difference is that they shift down only 3 positions where i do 4. So they claim to get 'additional bits, which is nonsense. You can't get 16 bit out of a 10 bit converter no matter how hard you try.

The second trick they employ is injecting noise in the input. They are essentially making a sigma delta on top of their a/d. But this again is nonsense... I wouldn't trust that. First of all your 'random' generator is not going to be random at all. It'll be based on a prbs polynomal. If you try to convert a perfectly stable static input it will still bounce all over the place..

[aluck]

No, it's not. Do read again. Do try it. Compare results. Repeat.

The second trick they employ is injecting noise in the input. They are essentially making a sigma delta on top of their a/d. But this again is nonsense... I wouldn't trust that. First of all your 'random' generator is not going to be random at all. It'll be based on a prbs polynomal.

You missed 2 points:
 1. It is not just "a noise". If you superimpose 10Vpp noise over 250 mV signal - it wouldn't make any sense. You need artificially produced noise "with a ripple peak-to-peak value of a few LSB" (page 9).
 2. It is not designed to replace 16-bit ADC with el-cheapo 8-bit. You are losing sampling rate at exponentional progression. For each additional 1 bit you have to double the number of measurements, so for getting additional 8 bit you will need 2^8 oversampling.

So far, you are just theorizing. Just give it a try and let us know of your results. If you still not satisfied - let's summon Atmel people here and question them a little bit Torquemada style.