Merge branch 'feature/digital-input-stamp' into feature/pounder-timestamping

master
Ryan Summers 2021-01-06 13:34:55 +01:00
commit 37595405c3
6 changed files with 127 additions and 171 deletions

View File

@ -18,6 +18,9 @@ load
# tbreak cortex_m_rt::reset_handler
monitor reset halt
source ../../PyCortexMDebug/cmdebug/svd_gdb.py
svd_load ~/Downloads/STM32H743x.svd
# cycle counter delta tool, place two bkpts around the section
set var $cc=0xe0001004
define qq

View File

@ -1,12 +1,29 @@
use super::hal::time::MegaHertz;
/// The ADC setup time is the number of seconds after the CSn line goes low before the serial clock
/// may begin. This is used for performing the internal ADC conversion.
pub const ADC_SETUP_TIME: f32 = 220e-9;
/// The maximum DAC/ADC serial clock line frequency. This is a hardware limit.
pub const ADC_DAC_SCK_MHZ_MAX: u32 = 50;
pub const ADC_DAC_SCK_MAX: MegaHertz = MegaHertz(50);
/// The optimal counting frequency of the hardware timers used for timestamping and sampling.
pub const TIMER_FREQUENCY_MHZ: u32 = 100;
pub const TIMER_FREQUENCY: MegaHertz = MegaHertz(100);
/// The QSPI frequency for communicating with the pounder DDS.
pub const POUNDER_QSPI_FREQUENCY: MegaHertz = MegaHertz(40);
/// The delay after initiating a QSPI transfer before asserting the IO_Update for the pounder DDS.
// Pounder Profile writes are always 16 bytes, with 2 cycles required per byte, coming out to a
// total of 32 QSPI clock cycles. The QSPI is configured for 40MHz, so this comes out to an offset
// of 800nS. We use 900ns to be safe.
pub const POUNDER_IO_UPDATE_DELAY: f32 = 900_e-9;
/// The duration to assert IO_Update for the pounder DDS.
// IO_Update should be latched for 4 SYNC_CLK cycles after the QSPI profile write. With pounder
// SYNC_CLK running at 100MHz (1/4 of the pounder reference clock of 400MHz), this corresponds to
// 40ns. To accomodate rounding errors, we use 50ns instead.
pub const POUNDER_IO_UPDATE_DURATION: f32 = 50_e-9;
/// The DDS reference clock frequency in MHz.
pub const DDS_REF_CLK_MHZ: u32 = 100;

View File

@ -3,9 +3,6 @@
///! This module provides a means of timestamping the rising edges of an external reference clock on
///! the DI0 with a timer value from TIM5.
///!
///! This module only supports input clocks on DI0 and may or may not utilize DMA to collect
///! timestamps.
///!
///! # Design
///! An input capture channel is configured on DI0 and fed into TIM5's capture channel 4. TIM5 is
///! then run in a free-running mode with a configured tick rate (PSC) and maximum count value
@ -13,12 +10,6 @@
///! recorded as a timestamp. This timestamp can be either directly read from the timer channel or
///! can be collected asynchronously via DMA collection.
///!
///! When DMA is used for timestamp collection, a DMA transfer is configured to collect as many
///! timestamps as there are samples, but it is intended that this DMA transfer should never
///! complete. Instead, when all samples are collected, the module pauses the DMA transfer and
///! checks to see how many timestamps were collected. These collected timestamps are then returned
///! for further processing.
///!
///! To prevent silently discarding timestamps, the TIM5 input capture over-capture flag is
///! continually checked. Any over-capture event (which indicates an overwritten timestamp) then
///! triggers a panic to indicate the dropped timestamp so that design parameters can be adjusted.
@ -27,35 +18,53 @@
///! It appears that DMA transfers can take a significant amount of time to disable (400ns) if they
///! are being prematurely stopped (such is the case here). As such, for a sample batch size of 1,
///! this can take up a significant amount of the total available processing time for the samples.
///! To avoid this, the module does not use DMA when the sample batch size is one. Instead, the
///! module manually checks for any captured timestamps from the timer capture channel manually. In
///! this mode, the maximum input clock frequency supported is equal to the configured sample rate.
///! This module checks for any captured timestamps from the timer capture channel manually. In
///! this mode, the maximum input clock frequency supported is dependant on the sampling rate and
///! batch size.
///!
///! There is a small window while the DMA buffers are swapped where a timestamp could potentially
///! be lost. To prevent this, the `acuire_buffer()` method should not be pre-empted. Any lost
///! timestamp will trigger an over-capture interrupt.
use super::{
hal, timers, DmaConfig, PeripheralToMemory, Transfer, SAMPLE_BUFFER_SIZE,
};
///! This module only supports DI0 for timestamping due to trigger constraints on the DIx pins. If
///! timestamping is desired in DI1, a separate timer + capture channel will be necessary.
use super::{hal, timers, ADC_SAMPLE_TICKS, SAMPLE_BUFFER_SIZE};
// The DMA buffers must exist in a location where DMA can access. By default, RAM uses DTCM, which
// is off-limits to the normal DMA peripheral. Instead, we use AXISRAM.
#[link_section = ".axisram.buffers"]
static mut BUF: [[u32; SAMPLE_BUFFER_SIZE]; 2] = [[0; SAMPLE_BUFFER_SIZE]; 2];
/// Calculate the period of the digital input timestampe timer.
///
/// # Note
/// The period returned will be 1 less than the required period in timer ticks. The value returned
/// can be immediately programmed into a hardware timer period register.
///
/// The period is calcualted to be some power-of-two multiple of the batch size, such that N batches
/// will occur between each timestamp timer overflow.
///
/// # Returns
/// A 32-bit value that can be programmed into a hardware timer period register.
pub fn calculate_timestamp_timer_period() -> u32 {
// Calculate how long a single batch requires in timer ticks.
let batch_duration_ticks: u64 =
SAMPLE_BUFFER_SIZE as u64 * ADC_SAMPLE_TICKS as u64;
// Calculate the largest power-of-two that is less than or equal to
// `batches_per_overflow`. This is completed by eliminating the least significant
// bits of the value until only the msb remains, which is always a power of two.
let batches_per_overflow: u64 =
(1u64 + u32::MAX as u64) / batch_duration_ticks;
let mut j = batches_per_overflow;
while (j & (j - 1)) != 0 {
j = j & (j - 1);
}
// Once the number of batches per timestamp overflow is calculated, we can figure out the final
// period of the timestamp timer. The period is always 1 larger than the value configured in the
// register.
let period: u64 = batch_duration_ticks * j - 1u64;
assert!(period <= u32::MAX as u64);
period as u32
}
/// The timestamper for DI0 reference clock inputs.
pub struct InputStamper {
_di0_trigger: hal::gpio::gpioa::PA3<hal::gpio::Alternate<hal::gpio::AF2>>,
next_buffer: Option<&'static mut [u32; SAMPLE_BUFFER_SIZE]>,
transfer: Option<
Transfer<
hal::dma::dma::Stream6<hal::stm32::DMA1>,
timers::tim5::Channel4InputCapture,
PeripheralToMemory,
&'static mut [u32; SAMPLE_BUFFER_SIZE],
>,
>,
capture_channel: Option<timers::tim5::Channel4InputCapture>,
capture_channel: timers::tim5::Channel4InputCapture,
}
impl InputStamper {
@ -63,100 +72,38 @@ impl InputStamper {
///
/// # Args
/// * `trigger` - The capture trigger input pin.
/// * `stream` - The DMA stream to use for collecting timestamps.
/// * `timer_channel - The timer channel used for capturing timestamps.
/// * `batch_size` - The number of samples collected per processing batch.
pub fn new(
trigger: hal::gpio::gpioa::PA3<hal::gpio::Alternate<hal::gpio::AF2>>,
stream: hal::dma::dma::Stream6<hal::stm32::DMA1>,
timer_channel: timers::tim5::Channel4,
batch_size: usize,
) -> Self {
// Utilize the TIM5 CH4 as an input capture channel - use TI4 (the DI0 input trigger) as the
// capture source.
let input_capture =
timer_channel.to_input_capture(timers::CaptureTrigger::Input24);
// For small batch sizes, the overhead of DMA can become burdensome to the point where
// timing is not met. The DMA requires 500ns overhead, whereas a direct register read only
// requires ~80ns. When batches of 2-or-greater are used, use a DMA-based approach.
let (transfer, input_capture) = if batch_size >= 2 {
input_capture.listen_dma();
// Set up the DMA transfer.
let dma_config = DmaConfig::default().memory_increment(true);
let timestamp_transfer: Transfer<_, _, PeripheralToMemory, _> =
Transfer::init(
stream,
input_capture,
unsafe { &mut BUF[0] },
None,
dma_config,
);
(Some(timestamp_transfer), None)
} else {
(None, Some(input_capture))
};
timer_channel.into_input_capture(timers::CaptureTrigger::Input24);
Self {
next_buffer: unsafe { Some(&mut BUF[1]) },
transfer,
capture_channel: input_capture,
_di0_trigger: trigger,
}
}
/// Start capture timestamps on DI0.
/// Start to capture timestamps on DI0.
pub fn start(&mut self) {
if let Some(transfer) = &mut self.transfer {
transfer.start(|capture_channel| {
capture_channel.enable();
});
} else {
self.capture_channel.as_mut().unwrap().enable();
}
self.capture_channel.enable();
}
/// Get all of the timestamps that have occurred during the last processing cycle.
pub fn acquire_buffer(&mut self) -> &[u32] {
// If we are using DMA, finish the transfer and swap over buffers.
if self.transfer.is_some() {
let next_buffer = self.next_buffer.take().unwrap();
self.transfer.as_mut().unwrap().pause(|channel| {
if channel.check_overcapture() {
panic!("DI0 timestamp overrun");
}
});
let (prev_buffer, _, remaining_transfers) = self
.transfer
.as_mut()
.unwrap()
.next_transfer(next_buffer)
.unwrap();
let valid_count = prev_buffer.len() - remaining_transfers;
self.next_buffer.replace(prev_buffer);
// Note that we likely didn't finish the transfer, so only return the number of
// timestamps actually collected.
&self.next_buffer.as_ref().unwrap()[..valid_count]
} else {
if self.capture_channel.as_ref().unwrap().check_overcapture() {
panic!("DI0 timestamp overrun");
}
// If we aren't using DMA, just manually check the input capture channel for a
// timestamp.
match self.capture_channel.as_mut().unwrap().latest_capture() {
Some(stamp) => {
self.next_buffer.as_mut().unwrap()[0] = stamp;
&self.next_buffer.as_ref().unwrap()[..1]
}
None => &[],
}
}
/// Get the latest timestamp that has occurred.
///
/// # Note
/// This function must be called sufficiently often. If an over-capture event occurs, this
/// function will panic, as this indicates a timestamp was inadvertently dropped.
///
/// To prevent timestamp loss, the batch size and sampling rate must be adjusted such that at
/// most one timestamp will occur in each data processing cycle.
pub fn latest_timestamp(&mut self) -> Option<u32> {
self.capture_channel
.latest_capture()
.expect("DI0 timestamp overrun")
}
}

View File

@ -30,8 +30,6 @@ extern crate panic_halt;
#[macro_use]
extern crate log;
use core::convert::TryInto;
// use core::sync::atomic::{AtomicU32, AtomicBool, Ordering};
use cortex_m_rt::exception;
use rtic::cyccnt::{Instant, U32Ext};
@ -58,7 +56,8 @@ use heapless::{consts::*, String};
// The number of ticks in the ADC sampling timer. The timer runs at 100MHz, so the step size is
// equal to 10ns per tick.
const ADC_SAMPLE_TICKS: u32 = 128;
// Currently, the sample rate is equal to: Fsample = 100/256 MHz = 390.625 KHz
const ADC_SAMPLE_TICKS: u32 = 256;
// The desired ADC sample processing buffer size.
const SAMPLE_BUFFER_SIZE: usize = 8;
@ -296,10 +295,10 @@ const APP: () = {
// Configure the timer to count at the designed tick rate. We will manually set the
// period below.
timer2.pause();
timer2.set_tick_freq(design_parameters::TIMER_FREQUENCY_MHZ.mhz());
timer2.set_tick_freq(design_parameters::TIMER_FREQUENCY);
let mut sampling_timer = timers::SamplingTimer::new(timer2);
sampling_timer.set_period(ADC_SAMPLE_TICKS - 1);
sampling_timer.set_period_ticks(ADC_SAMPLE_TICKS - 1);
sampling_timer
};
@ -315,32 +314,16 @@ const APP: () = {
// Configure the timer to count at the designed tick rate. We will manually set the
// period below.
timer5.pause();
timer5.set_tick_freq(design_parameters::TIMER_FREQUENCY_MHZ.mhz());
timer5.set_tick_freq(design_parameters::TIMER_FREQUENCY);
// The time stamp timer must run at exactly a multiple of the sample timer based on the
// batch size. To accomodate this, we manually set the period identical to the sample
// timer, but use a prescaler that is `BATCH_SIZE` longer.
// batch size. To accomodate this, we manually set the prescaler identical to the sample
// timer, but use a period that is longer.
let mut timer = timers::TimestampTimer::new(timer5);
let period: u32 = {
let batch_duration: u64 =
SAMPLE_BUFFER_SIZE as u64 * ADC_SAMPLE_TICKS as u64;
let batches_per_overflow: u64 =
(1u64 + u32::MAX as u64) / batch_duration;
// Calculate the largest power-of-two that is less than `batches_per_overflow`.
// This is completed by eliminating the least significant bits of the value until
// only the msb remains, which is always a power of two.
let mut j = batches_per_overflow;
while (j & (j - 1)) != 0 {
j = j & (j - 1);
}
let period: u64 = batch_duration * j - 1u64;
period.try_into().unwrap()
};
timer.set_period(period);
let period =
digital_input_stamper::calculate_timestamp_timer_period();
timer.set_period_ticks(period);
timer
};
@ -374,7 +357,7 @@ const APP: () = {
let spi: hal::spi::Spi<_, _, u16> = dp.SPI2.spi(
(spi_sck, spi_miso, hal::spi::NoMosi),
config,
design_parameters::ADC_DAC_SCK_MHZ_MAX.mhz(),
design_parameters::ADC_DAC_SCK_MAX,
ccdr.peripheral.SPI2,
&ccdr.clocks,
);
@ -412,7 +395,7 @@ const APP: () = {
let spi: hal::spi::Spi<_, _, u16> = dp.SPI3.spi(
(spi_sck, spi_miso, hal::spi::NoMosi),
config,
design_parameters::ADC_DAC_SCK_MHZ_MAX.mhz(),
design_parameters::ADC_DAC_SCK_MAX,
ccdr.peripheral.SPI3,
&ccdr.clocks,
);
@ -462,7 +445,7 @@ const APP: () = {
dp.SPI4.spi(
(spi_sck, spi_miso, hal::spi::NoMosi),
config,
design_parameters::ADC_DAC_SCK_MHZ_MAX.mhz(),
design_parameters::ADC_DAC_SCK_MAX,
ccdr.peripheral.SPI4,
&ccdr.clocks,
)
@ -494,7 +477,7 @@ const APP: () = {
dp.SPI5.spi(
(spi_sck, spi_miso, hal::spi::NoMosi),
config,
design_parameters::ADC_DAC_SCK_MHZ_MAX.mhz(),
design_parameters::ADC_DAC_SCK_MAX,
ccdr.peripheral.SPI5,
&ccdr.clocks,
)
@ -564,7 +547,7 @@ const APP: () = {
let qspi = hal::qspi::Qspi::bank2(
dp.QUADSPI,
qspi_pins,
40.mhz(),
design_parameters::POUNDER_QSPI_FREQUENCY,
&ccdr.clocks,
ccdr.peripheral.QSPI,
);
@ -689,30 +672,27 @@ const APP: () = {
ccdr.peripheral.HRTIM,
);
// IO_Update should be latched for 4 SYNC_CLK cycles after the QSPI profile
// write. With pounder SYNC_CLK running at 125MHz (1/4 of the pounder reference
// clock of 500MHz), this corresponds to 32ns. To accomodate rounding errors, we
// use 50ns instead.
//
// Profile writes are always 16 bytes, with 2 cycles required per byte, coming
// out to a total of 32 QSPI clock cycles. The QSPI is configured for 40MHz, so
// this comes out to an offset of 800nS. We use 900ns to be safe - note that the
// timer is triggered after the QSPI write, which can take approximately 120nS,
// so there is additional margin.
// IO_Update occurs after a fixed delay from the QSPI write. Note that the timer
// is triggered after the QSPI write, which can take approximately 120nS, so
// there is additional margin.
hrtimer.configure_single_shot(
hrtimer::Channel::Two,
50_e-9,
900_e-9,
design_parameters::POUNDER_IO_UPDATE_DURATION,
design_parameters::POUNDER_IO_UPDATE_DELAY,
);
// Ensure that we have enough time for an IO-update every sample.
let sample_frequency =
(design_parameters::TIMER_FREQUENCY_MHZ as f32
* 1_000_000.0)
/ ADC_SAMPLE_TICKS as f32;
let sample_frequency = {
let timer_frequency: hal::time::Hertz =
design_parameters::TIMER_FREQUENCY.into();
timer_frequency.0 as f32 / ADC_SAMPLE_TICKS as f32
};
let sample_period = 1.0 / sample_frequency;
assert!(sample_period > 900_e-9);
assert!(
sample_period
> design_parameters::POUNDER_IO_UPDATE_DELAY
);
hrtimer
};
@ -849,9 +829,7 @@ const APP: () = {
let trigger = gpioa.pa3.into_alternate_af2();
digital_input_stamper::InputStamper::new(
trigger,
dma_streams.6,
timestamp_timer_channels.ch4,
SAMPLE_BUFFER_SIZE,
)
};
@ -933,7 +911,7 @@ const APP: () = {
c.resources.dacs.1.acquire_buffer(),
];
let _timestamps = c.resources.input_stamper.acquire_buffer();
let _timestamp = c.resources.input_stamper.latest_timestamp();
for channel in 0..adc_samples.len() {
for sample in 0..adc_samples[0].len() {

View File

@ -43,7 +43,7 @@ impl Timestamper {
// The capture channel should capture whenever the trigger input occurs.
let input_capture = capture_channel
.to_input_capture(timers::CaptureTrigger::TriggerInput);
.into_input_capture(timers::CaptureTrigger::TriggerInput);
input_capture.listen_dma();
// The data transfer is always a transfer of data from the peripheral to a RAM buffer.
@ -68,7 +68,7 @@ impl Timestamper {
}
pub fn update_period(&mut self, period: u16) {
self.timer.set_period(period);
self.timer.set_period_ticks(period);
}
/// Obtain a buffer filled with timestamps.

View File

@ -85,7 +85,7 @@ macro_rules! timer_channels {
/// Manually set the period of the timer.
#[allow(dead_code)]
pub fn set_period(&mut self, period: $size) {
pub fn set_period_ticks(&mut self, period: $size) {
let regs = unsafe { &*hal::stm32::$TY::ptr() };
regs.arr.write(|w| w.arr().bits(period));
}
@ -241,7 +241,7 @@ macro_rules! timer_channels {
/// # Args
/// * `input` - The input source for the input capture event.
#[allow(dead_code)]
pub fn to_input_capture(self, input: super::CaptureTrigger) -> [< Channel $index InputCapture >]{
pub fn into_input_capture(self, input: super::CaptureTrigger) -> [< Channel $index InputCapture >]{
let regs = unsafe { &*<$TY>::ptr() };
// Note(unsafe): The bit configuration is guaranteed to be valid by the
@ -255,17 +255,28 @@ macro_rules! timer_channels {
impl [< Channel $index InputCapture >] {
/// Get the latest capture from the channel.
#[allow(dead_code)]
pub fn latest_capture(&mut self) -> Option<$size> {
pub fn latest_capture(&mut self) -> Result<Option<$size>, ()> {
// Note(unsafe): This channel owns all access to the specific timer channel.
// Only atomic operations on completed on the timer registers.
let regs = unsafe { &*<$TY>::ptr() };
let sr = regs.sr.read();
let ccx = regs.[< ccr $index >].read();
if sr.[< cc $index if >]().bit_is_set() {
regs.sr.modify(|_, w| w.[< cc $index if >]().clear_bit());
let result = if sr.[< cc $index if >]().bit_is_set() {
// Read the capture value. Reading the captured value clears the flag in the
// status register automatically.
let ccx = regs.[< ccr $index >].read();
Some(ccx.ccr().bits())
} else {
None
};
// Read SR again to check for a potential over-capture. If there is an
// overcapture, return an error.
if regs.sr.read().[< cc $index of >]().bit_is_clear() {
Ok(result)
} else {
regs.sr.modify(|_, w| w.[< cc $index of >]().clear_bit());
Err(())
}
}