Skip to content

sampling.md

Latest commit

 

History

History
80 lines (66 loc) · 3.93 KB

sampling.md

File metadata and controls

80 lines (66 loc) · 3.93 KB
Raw

RSSI sampling rate

How fast the timer reads the receiver, what sets that rate, and a faster option kept for later. Measured on an ESP32 dev board with an RX5808, powered from USB.

Today (v1.1.0)

The timing core reads the RX5808's RSSI with Arduino's analogRead() in a loop and feeds every reading through a Kalman filter. One analogRead() takes about 90 µs: most of it is setup repeated on every call (pin mode, attenuation, locks) and code run from flash through the ESP32's cache. So the timer takes 6 500-10 600 readings per second, and which depends on the build: where the linker places the code decides how often the cache misses.

Same timer, same conditions (/api/debug/load → samplesPerSec):

Build Samples/s
the build before the transmit-gain fix (ae181de, 2026-09-30 morning) 8 400-8 750
v1.1.0 as released (4ee4de8) 6 390-6 540
the same code with a few lines of diagnostics added 10 600
that build with 2 000 unused instructions added (padding) 9 420

Not the cause: heat (the same rate right after power-up and after running a while), the WiFi mode, or the hotspot's transmit gain (switching the power loop on and off while running changed nothing).

It is enough: 6 500/s is still 6-7 readings per millisecond, and a pass lasts tens of milliseconds. The filter smooths over about 70 readings, so its time constant is ~7 ms at 10 600/s and ~11 ms at 6 500/s; a pass is timed at the middle of its peak either way. Compare firmware versions only A/B on the same timer.

Kept for later: fast reads, fixed filter step

On branch perf/fast-adc (commit 6dc7bd9), bench-tested, not flown. Not merged because 6 500-10 000 samples/s is enough. Worth it if the filter's timing should be the same in every build, or the RSSI should be less noisy.

  1. RX5808::readRssiAdc() starts the conversion through the ADC's registers, from IRAM (classic ESP32, ADC1 pins): ~13 µs, ~78 000 readings/s. init() sets the pin, attenuation and width up with one analogRead() and keeps the ADC powered (adc_power_acquire()).
  2. LapTimer::scan() reads until the next filter step is due and filters the average: one step every 100 µs (10 000/s) in every build, each the average of ~8 readings.
// RX5808.cpp, classic ESP32 (#include "driver/adc.h" and "soc/sens_struct.h")
uint16_t IRAM_ATTR RX5808::readRssiAdc() {
    SENS.sar_read_ctrl.sar1_dig_force = 0;       // RTC controller, started by software
    SENS.sar_meas_start1.meas1_start_force = 1;
    SENS.sar_meas_start1.sar1_en_pad_force = 1;
    SENS.sar_meas_start1.sar1_en_pad = 1 << adcChannel;
    SENS.sar_meas_start1.meas1_start_sar = 0;
    SENS.sar_meas_start1.meas1_start_sar = 1;
    while (!SENS.sar_meas_start1.meas1_done_sar) {
    }
    return SENS.sar_meas_start1.meas1_data_sar;
}

// LapTimer::scan(): one Kalman step every RSSI_STEP_US (100)
uint32_t sum = 0, n = 0;
do {
    uint16_t raw = rx->readRssiAdc();
    sum += raw > 2047 ? 2047 : raw;
    n++;
} while ((int32_t)(ESP.getCycleCount() - stepDueCycles) < 0);
uint32_t now = ESP.getCycleCount();
stepDueCycles += stepCycles;                    // RSSI_STEP_US * CPU MHz
if ((int32_t)(now - stepDueCycles) >= 0)
    stepDueCycles = now + stepCycles;           // after a pause: start again from now
sample(round(filter.filter((sum / n) >> 3, 0)), nowMs);

Bench results: 10 000 steps/s and 77 800-77 950 readings/s with 0, 1 000 or 2 000 instructions of padding; the RSSI on the same scale as with analogRead() (noise floor 50); spectrum scan and step test work; device test 22/22; 10/10 boots; hotspot unchanged (-59 to -63 dBm); all five targets build. C3/S3 keep analogRead() with the same 100 µs step: their ADCs differ, and they haven't been tested.

Before using it: add a timeout to the conversion wait (fall back to analogRead(), which sets the ADC up again); fly passes and compare lap detection with the current build (tools/rssi_log.py); check Enter/Exit, as the RSSI is less noisy.