dma: don't swap buffers

* This uses a new closure-based method to the DMA HAL implementation which
  gives access to the inactive buffer directly.
* It removes changing addresses, the third buffer for DBM, the inactive
  address poisoning, and allows the cancellation of the redundant repeat
  memory barriers and compiler fences.
* This is now around 20 instructions per buffer down from about 100 cycles
  before.
* Also introduces a new `SampleBuffer` type alias.
* The required unpacking of the resources structure is a bit annoying
  but could probably abstraced away.

TODO:

* Test
* Adapt `lockin`
This commit is contained in:
Robert Jördens 2021-05-24 20:06:31 +02:00
parent d8cc3c74d9
commit 3165c680d6
6 changed files with 107 additions and 110 deletions

3
Cargo.lock generated
View File

@ -802,8 +802,7 @@ dependencies = [
[[package]]
name = "stm32h7xx-hal"
version = "0.9.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "67034b80041bc33a48df1c1c435b6ae3bb18c35e42aa7e702ce8363b96793398"
source = "git+https://github.com/quartiq/stm32h7xx-hal.git?rev=cca4ecc#cca4ecc3e0cc8cb2f7a9652c4099d50b44977493"
dependencies = [
"bare-metal 1.0.0",
"cast",

View File

@ -56,7 +56,9 @@ rev = "523d71d"
[dependencies.stm32h7xx-hal]
features = ["stm32h743v", "rt", "unproven", "ethernet", "quadspi"]
version = "0.9.0"
# version = "0.9.0"
git = "https://github.com/quartiq/stm32h7xx-hal.git"
rev = "cca4ecc"
[patch.crates-io.miniconf]
git = "https://github.com/quartiq/miniconf.git"

View File

@ -2,6 +2,8 @@
#![no_std]
#![no_main]
use core::sync::atomic::{fence, Ordering};
use stabilizer::{hardware, net};
use miniconf::Miniconf;
@ -120,51 +122,73 @@ const APP: () = {
/// Because the ADC and DAC operate at the same rate, these two constraints actually implement
/// the same time bounds, meeting one also means the other is also met.
#[task(binds=DMA1_STR4, resources=[adcs, digital_inputs, dacs, iir_state, settings, telemetry], priority=2)]
fn process(c: process::Context) {
let adc_samples = [
c.resources.adcs.0.acquire_buffer(),
c.resources.adcs.1.acquire_buffer(),
];
fn process(mut c: process::Context) {
let adc0 = &mut c.resources.adcs.0;
let adc1 = &mut c.resources.adcs.1;
let dac0 = &mut c.resources.dacs.0;
let dac1 = &mut c.resources.dacs.1;
let di = &c.resources.digital_inputs;
let settings = &c.resources.settings;
let iir_state = &mut c.resources.iir_state;
let telemetry = &mut c.resources.telemetry;
adc0.with_buffer(|a0| {
adc1.with_buffer(|a1| {
dac0.with_buffer(|d0| {
dac1.with_buffer(|d1| {
let adc_samples = [a0, a1];
let dac_samples = [d0, d1];
let dac_samples = [
c.resources.dacs.0.acquire_buffer(),
c.resources.dacs.1.acquire_buffer(),
];
let digital_inputs =
[di.0.is_high().unwrap(), di.1.is_high().unwrap()];
let digital_inputs = [
c.resources.digital_inputs.0.is_high().unwrap(),
c.resources.digital_inputs.1.is_high().unwrap(),
];
let hold = settings.force_hold
|| (digital_inputs[1] && settings.allow_hold);
let hold = c.resources.settings.force_hold
|| (digital_inputs[1] && c.resources.settings.allow_hold);
// Preserve instruction and data ordering w.r.t. DMA flag access.
fence(Ordering::SeqCst);
for channel in 0..adc_samples.len() {
for sample in 0..adc_samples[0].len() {
let mut y = f32::from(adc_samples[channel][sample] as i16);
for i in 0..c.resources.iir_state[channel].len() {
y = c.resources.settings.iir_ch[channel][i].update(
&mut c.resources.iir_state[channel][i],
y,
hold,
);
}
// Note(unsafe): The filter limits ensure that the value is in range.
// The truncation introduces 1/2 LSB distortion.
let y = unsafe { y.to_int_unchecked::<i16>() };
// Convert to DAC code
dac_samples[channel][sample] = DacCode::from(y).0;
}
}
for channel in 0..adc_samples.len() {
adc_samples[channel]
.iter()
.zip(dac_samples[channel].iter_mut())
.map(|(ai, di)| {
let x = f32::from(*ai as i16);
let y = settings.iir_ch[channel]
.iter()
.zip(iir_state[channel].iter_mut())
.fold(x, |yi, (ch, state)| {
ch.update(state, yi, hold)
});
// Note(unsafe): The filter limits must ensure that
// the value is in range.
// The truncation introduces 1/2 LSB distortion.
let y: i16 =
unsafe { y.to_int_unchecked() };
// Convert to DAC code
*di = DacCode::from(y).0;
})
.last();
}
// Update telemetry measurements.
c.resources.telemetry.adcs =
[AdcCode(adc_samples[0][0]), AdcCode(adc_samples[1][0])];
// Update telemetry measurements.
telemetry.adcs = [
AdcCode(adc_samples[0][0]),
AdcCode(adc_samples[1][0]),
];
c.resources.telemetry.dacs =
[DacCode(dac_samples[0][0]), DacCode(dac_samples[1][0])];
telemetry.dacs = [
DacCode(dac_samples[0][0]),
DacCode(dac_samples[1][0]),
];
c.resources.telemetry.digital_inputs = digital_inputs;
telemetry.digital_inputs = digital_inputs;
// Preserve instruction and data ordering w.r.t. DMA flag access.
fence(Ordering::SeqCst);
})
})
})
});
}
#[idle(resources=[network], spawn=[settings_update])]

View File

@ -29,15 +29,9 @@
///! available. When enough samples have been collected, a transfer-complete interrupt is generated
///! and the ADC samples are available for processing.
///!
///! The SPI peripheral internally has an 8- or 16-byte TX and RX FIFO, which corresponds to a 4- or
///! 8-sample buffer for incoming ADC samples. During the handling of the DMA transfer completion,
///! there is a small window where buffers are swapped over where it's possible that a sample could
///! be lost. In order to avoid this, the SPI RX FIFO is effectively used as a "sample overflow"
///! region and can buffer a number of samples until the next DMA transfer is configured. If a DMA
///! transfer is still not set in time, the SPI peripheral will generate an input-overrun interrupt.
///! This interrupt then serves as a means of detecting if samples have been lost, which will occur
///! whenever data processing takes longer than the collection period.
///!
///! After a complete transfer of a batch of samples, the inactive buffer is available to the
///! user for processing. The processing must complete before the DMA transfer of the next batch
///! completes.
///!
///! ## Starting Data Collection
///!
@ -68,13 +62,12 @@
///! sample DMA requests, which can be completed by setting e.g. ADC0's comparison to a counter
///! value of 0 and ADC1's comparison to a counter value of 1.
///!
///! In this implementation, single buffer mode DMA transfers are used because the SPI RX FIFO can
///! be used as a means to both detect and buffer ADC samples during the buffer swap-over. Because
///! of this, double-buffered mode does not offer any advantages over single-buffered mode (unless
///! double-buffered mode offers less overhead due to the DMA disable/enable procedure).
///! In this implementation, double buffer mode DMA transfers are used because the SPI RX FIFOs
///! have finite depth, FIFO access is slower than AXISRAM access, and because the single
///! buffer mode DMA disable/enable and buffer update sequence is slow.
use stm32h7xx_hal as hal;
use super::design_parameters::SAMPLE_BUFFER_SIZE;
use super::design_parameters::{SampleBuffer, SAMPLE_BUFFER_SIZE};
use super::timers;
use hal::dma::{
@ -119,8 +112,7 @@ static mut SPI_EOT_CLEAR: [u32; 1] = [0x00];
// processed). Note that the contents of AXI SRAM is uninitialized, so the buffer contents on
// startup are undefined. The dimensions are `ADC_BUF[adc_index][ping_pong_index][sample_index]`.
#[link_section = ".axisram.buffers"]
static mut ADC_BUF: [[[u16; SAMPLE_BUFFER_SIZE]; 2]; 2] =
[[[0; SAMPLE_BUFFER_SIZE]; 2]; 2];
static mut ADC_BUF: [[SampleBuffer; 2]; 2] = [[[0; SAMPLE_BUFFER_SIZE]; 2]; 2];
macro_rules! adc_input {
($name:ident, $index:literal, $trigger_stream:ident, $data_stream:ident, $clear_stream:ident,
@ -192,12 +184,11 @@ macro_rules! adc_input {
/// Represents data associated with ADC.
pub struct $name {
next_buffer: Option<&'static mut [u16; SAMPLE_BUFFER_SIZE]>,
transfer: Transfer<
hal::dma::dma::$data_stream<hal::stm32::DMA1>,
hal::spi::Spi<hal::stm32::$spi, hal::spi::Disabled, u16>,
PeripheralToMemory,
&'static mut [u16; SAMPLE_BUFFER_SIZE],
&'static mut SampleBuffer,
hal::dma::DBTransfer,
>,
trigger_transfer: Transfer<
@ -316,6 +307,7 @@ macro_rules! adc_input {
// data stream is used to trigger a transfer completion interrupt.
let data_config = DmaConfig::default()
.memory_increment(true)
.double_buffer(true)
.transfer_complete_interrupt($index == 1)
.priority(Priority::VeryHigh);
@ -333,17 +325,14 @@ macro_rules! adc_input {
Transfer::init(
data_stream,
spi,
// Note(unsafe): The ADC_BUF[$index][0] is "owned" by this peripheral.
// Note(unsafe): The ADC_BUF[$index] is "owned" by this peripheral.
// It shall not be used anywhere else in the module.
unsafe { &mut ADC_BUF[$index][0] },
None,
unsafe { Some(&mut ADC_BUF[$index][1]) },
data_config,
);
Self {
// Note(unsafe): The ADC_BUF[$index][1] is "owned" by this peripheral. It
// shall not be used anywhere else in the module.
next_buffer: unsafe { Some(&mut ADC_BUF[$index][1]) },
transfer: data_transfer,
trigger_transfer,
clear_transfer,
@ -364,27 +353,17 @@ macro_rules! adc_input {
}
/// Obtain a buffer filled with ADC samples.
/// Wait for the transfer of the currently active buffer to complete,
/// then call a function on the now inactive buffer and acknowledge the
/// transfer complete flag.
///
/// # Returns
/// A reference to the underlying buffer that has been filled with ADC samples.
pub fn acquire_buffer(&mut self) -> &[u16; SAMPLE_BUFFER_SIZE] {
// Wait for the transfer to fully complete before continuing. Note: If a device
// hangs up, check that this conditional is passing correctly, as there is no
// time-out checks here in the interest of execution speed.
while !self.transfer.get_transfer_complete_flag() {}
let next_buffer = self.next_buffer.take().unwrap();
// Start the next transfer.
self.transfer.clear_interrupts();
let (prev_buffer, _, _) =
self.transfer.next_transfer(next_buffer).unwrap();
// .unwrap_none() https://github.com/rust-lang/rust/issues/62633
self.next_buffer.replace(prev_buffer);
self.next_buffer.as_ref().unwrap()
/// NOTE(unsafe): Memory safety and access ordering is not guaranteed
/// (see the HAL DMA docs).
pub fn with_buffer<F, R>(&mut self, f: F) -> R
where
F: FnOnce(&mut SampleBuffer) -> R,
{
unsafe { self.transfer.next_dbm_transfer_with(|buf, _current| f(buf)) }
}
}
}

View File

@ -52,7 +52,7 @@
///! served promptly after the transfer completes.
use stm32h7xx_hal as hal;
use super::design_parameters::SAMPLE_BUFFER_SIZE;
use super::design_parameters::{SampleBuffer, SAMPLE_BUFFER_SIZE};
use super::timers;
use hal::dma::{
@ -66,8 +66,7 @@ use hal::dma::{
// processed). Note that the contents of AXI SRAM is uninitialized, so the buffer contents on
// startup are undefined. The dimensions are `ADC_BUF[adc_index][ping_pong_index][sample_index]`.
#[link_section = ".axisram.buffers"]
static mut DAC_BUF: [[[u16; SAMPLE_BUFFER_SIZE]; 3]; 2] =
[[[0; SAMPLE_BUFFER_SIZE]; 3]; 2];
static mut DAC_BUF: [[SampleBuffer; 2]; 2] = [[[0; SAMPLE_BUFFER_SIZE]; 2]; 2];
/// Custom type for referencing DAC output codes.
/// The internal integer is the raw code written to the DAC output register.
@ -137,13 +136,12 @@ macro_rules! dac_output {
/// Represents data associated with DAC.
pub struct $name {
next_buffer: Option<&'static mut [u16; SAMPLE_BUFFER_SIZE]>,
// Note: SPI TX functionality may not be used from this structure to ensure safety with DMA.
transfer: Transfer<
hal::dma::dma::$data_stream<hal::stm32::DMA1>,
$spi,
MemoryToPeripheral,
&'static mut [u16; SAMPLE_BUFFER_SIZE],
&'static mut SampleBuffer,
hal::dma::DBTransfer,
>,
}
@ -198,33 +196,26 @@ macro_rules! dac_output {
trigger_config,
);
Self {
transfer,
// Note(unsafe): This buffer is only used once and provided for the next DMA transfer.
next_buffer: unsafe { Some(&mut DAC_BUF[$index][2]) },
}
Self { transfer }
}
pub fn start(&mut self) {
self.transfer.start(|spi| spi.start_dma());
}
/// Acquire the next output buffer to populate it with DAC codes.
pub fn acquire_buffer(&mut self) -> &mut [u16; SAMPLE_BUFFER_SIZE] {
// Note: If a device hangs up, check that this conditional is passing correctly, as
// there is no time-out checks here in the interest of execution speed.
while !self.transfer.get_transfer_complete_flag() {}
let next_buffer = self.next_buffer.take().unwrap();
// Start the next transfer.
let (prev_buffer, _, _) =
self.transfer.next_transfer(next_buffer).unwrap();
// .unwrap_none() https://github.com/rust-lang/rust/issues/62633
self.next_buffer.replace(prev_buffer);
self.next_buffer.as_mut().unwrap()
/// Wait for the transfer of the currently active buffer to complete,
/// then call a function on the now inactive buffer and acknowledge the
/// transfer complete flag.
///
/// NOTE(unsafe): Memory safety and access ordering is not guaranteed
/// (see the HAL DMA docs).
pub fn with_buffer<F, R>(&mut self, f: F) -> R
where
F: FnOnce(&mut SampleBuffer) -> R,
{
unsafe {
self.transfer.next_dbm_transfer_with(|buf, _current| f(buf))
}
}
}
};

View File

@ -50,5 +50,7 @@ pub const ADC_SAMPLE_TICKS: u16 = 1 << ADC_SAMPLE_TICKS_LOG2;
pub const SAMPLE_BUFFER_SIZE_LOG2: u8 = 3;
pub const SAMPLE_BUFFER_SIZE: usize = 1 << SAMPLE_BUFFER_SIZE_LOG2;
pub type SampleBuffer = [u16; SAMPLE_BUFFER_SIZE];
// The MQTT broker IPv4 address
pub const MQTT_BROKER: [u8; 4] = [10, 34, 16, 10];