Implement high frequency reads for ADC and Digital In #421
Conversation
src/edgepi/adc/adc_multiplexers.py
Outdated
| @@ -27,8 +26,7 @@ def _format_mux_values(mux_p: CH, mux_n: CH): | |||
|
|
|||
| return mux_p_val, mux_n_val, mask | |||
|
|
|||
|
|
|||
| def generate_mux_opcodes(mux_updates: dict): | |||
| def generate_mux_opcodes(adc1_reg, adc1_mux, adc2_reg, adc2_mux): | |||
There was a problem hiding this comment.
pack() is a surprisingly slow function, so replacing it improved performance a lot. The logging and dictionary operations also slowed this function down a reasonable amount, which is why I removed them. I tried to avoid removing logging statements where possible, however they did have an effect on the performance.
src/edgepi/adc/adc_state.py
Outdated
| f"query_state: query_property='{adc_property}'," | ||
| " adc_property_bits={hex(adc_property_bits)}," | ||
| f" adc_property_value='{adc_property_value}'" | ||
| "query_state: query_property='{}',".format(adc_property) + |
There was a problem hiding this comment.
format strings f"{...}" are eager, and so slow things down more. I opted to switch to lazy formatting where possible
src/edgepi/peripherals/gpio.py
Outdated
| # pylint: disable=protected-access | ||
| gpio._line = pin_num | ||
| # pylint: disable=protected-access | ||
| gpio._reopen(pin_dir, edge="none", bias=pin_bias, drive="default", inverted=False) |
There was a problem hiding this comment.
this change enables the batching of gpio reads, which minimizes the number of device open calls (I/O), which improves performance quite a bit. One tradeoff is that we touch private variables & functions, which could change if the library is updated. It seems the file these variables are in hasn't been modified for a year, but still something to keep in mind.
| # NOTE: disabling register update validation because there's no reason for us to suspect the other | ||
| # values would change? I'm not sure why this was done before. | ||
| # This is also quite slow, and affects the performance significantly | ||
| #__validate_register_updates(original_regs, register_values) |
There was a problem hiding this comment.
this validation function, combined with the deepcopy above is quite slow, and I can't figure out if it actually does anything. If it does, I'll probably want to make a function unsafe_apply_opcodes() which removes it, so that the rest of the codebase isn't adversely affected.
| @@ -980,6 +975,127 @@ def set_config( | |||
| [None]) | |||
| self.__config(**args) | |||
|
|
|||
| def adc1_config_and_read_samples_batch( | |||
There was a problem hiding this comment.
This function is massive because it is a combination of set_config and read_sample, but also most of the child functions. One issue with this codebase is that get_state is called a large number of times, even when only a few properties are needed. The construction of the ADCState object, and sometimes reading from the hardware is quite slow. By combining all the functions, we can read the register state from the hardware just once, then reuse those values & update our local copy after we complete all the hardware writes.
| command_tup, register_values = create_commands(-1, register_values, mux_p, mux_n) | ||
| command_tup_list += [command_tup] | ||
|
|
||
| data_list = self.spi_apply_adc_commands(command_tup_list) |
There was a problem hiding this comment.
we apply all the adc config. start, and read commands at once
|
The goal of this PR was to improve the performance of the SDK's ADC reads to be higher than 100hz. All 8 ADC inputs can consistently be read in less than 10ms on average (with a somewhat surprisingly high variance, likely due to the OS), while all 8 digital inputs can be read in less than 1ms on average.
It's possible to improve the performance more than 100 hz for all 8 ADC, however 1/3 of all time is being spend sleeping, while another 1/3 is spent transferring data over SPI. I think it's unlikely to get faster than 200 hz.
The speed improvements are due to bit optimizations, minimizing data sent over the channel, python optimizations, and batching of requests (minimize system calls opening & closing devices), all of which were found via profiling. For more info, see this (private) page in the wiki.