4 ms·

It's neat that the author was able to completely offload the work to the PIO. I'm not sure I really understand what's so poor about this display's design, thoug

It's neat that the author was able to completely offload the work to the PIO. I'm not sure I really understand what's so poor about this display's design, though? It sounds like it's just a framebuffer on a SPI bus. Even if you were stuck with a more limited DMA peripheral on a different MCU, you could poll the touch screen and interrupt on every scan line and do 30fps with fewer than 10khz interrupts, and have plenty of cpu cycles left over.

dmitrygr 4y ago
The fact that touch and display are on the same SPI bus making it impossible to just set up a repeating DMA to the display. On a non-RP2040 it wouldn’t be possible to drive it easily. Most DMA units wouldn’t allow you to feed a scan line at a time and then poll touch. They aren’t that flexible. You’d be stuck with a lot of IRQs
- sgtnoodle 4y ago
  The last time I worked with DMA was on an Atmel cortex-M7. Definitely a higher end MCU, though. It supported arbitrarily complex DMA chaining via linked lists in RAM. That being said, doing ~10Khz or so IRQs just to kick off a DMA transfer wouldn't have been much CPU in the grand scheme of things. For the touch input, is it necessary to truly sample at a high rate continuously, or is it more a need to oversample and take an average? If it's a need for oversampling, you could sample the touch screen between frames rather than interleaved between scanlines. Likewise, you could read from the SD card between frames as well. Writing to an SD card would be pretty terrible, given that cards often block for hundreds of milliseconds. Once again, awesome work with the PIO. I imagine you're running the display at its theoretical update limit unless you try overclocking?
  - dmitrygr 4y ago
    Between frames would be too rare for touch. You really do not want touch at just 45Hz. More samples are for averaging for smoothing. The touch controller here is noisy and the board layout waveshare did … does not follow the data sheet recommendations for lowering noise. Sd between refreshes is a poor idea for the reason you mention, yes. And yes screen interface is run as fast as it can as per data sheet. Overclocking the MCU wont help as we are not cycle bound there. Overclocking the SPI bus will work but not far. Past 70MHz it gets glitchy.
    - sgtnoodle 4y ago
      I specifically meant rapidly sampling the touch screen N times between frames rather than 1 time between frames. Depending on the characteristics of the EMI, it seems like that could work just as well.
      - dmitrygr 4y ago
        But in terms of temporal resolution you still have 45Hz which is too low for good inking.
- crest 4y ago
  You could use the DMA engine to reconfigure the function routing using a pair of DMA channels.
  - dmitrygr 4y ago
    Sadly in many chips this is not possible. Eg STM32H7 has 3 types of DMA units and each has a complex table of where they can read and write from. Two kinds of DMA cannot write to anywhere that contains a DMA controller. Rp2040 is rather unique in how powerful its DMA units are due to their reach. But it also has faults. Its DMA units seemingly cannot access GPIO, since that is bolted directly to the core (via the single cycle IO port). The only way to affect pins from DMA is via PIO as far as I can tell.
    - sgtnoodle 4y ago
      Reminds me of a memorable Ethernet peripheral. The peripheral would use the MCU's RAM for its own buffering. There were frame headers that both the peripheral and CPU would need to manipulate. Being an M7 the CPU had a data cache, though, and so the CPU couldn't safely manipulate the headers without first completely disabling the cache. I ended up using a DMA channel to indirectly manipulate the headers, bypassing the data cache. It was pretty silly but worked, and was way more efficient than disabling the data cache.
      - dmitrygr 4y ago
        On an m7, you can use the MPU to assign non-cacheable attribute to a memory area. As small as 32bytes and as large as 4GB. Any size representable as 2^n - k * 2^(n - 3) for k = 0..7 and n >= 8
        sgtnoodle 4y ago
        The particular driver was being retrofitted into a large enough variety of existing firmwares that we really wanted to avoid introducing any sort of MPU or link time configuration. The DMA channel approach was the least impactful path.
- sgtnoodle 4y ago
  Just for the fun of it, I've been reading a bit about the STM32H7 since you mentioned it. One half baked idea I have to achieve only one interrupt per scan line would be to use multiple DMA channels to drive the same SPI peripheral, one for the display SS and the other for the touch SS. Have a scanline ISR start both transfers at the same time, and rely on the well defined DMA channel priorities to deconflict them. It seems like this could possibly work given that the high level DMA documentation says it can control SS. What I don't know at the moment is how exactly the DMA channel actually controls SS. Does the channel itself have knowledge of the signal, or does it just poke the SPI peripheral at the right time? Said another way, are SS pins assigned to DMA channels, or are they assigned to the SPI peripheral? There's also clock frequency and data mode that would need to be associated with the DMA channel. My guess is it's the latter and my idea wouldn't work. Both options seem like they would have been reasonable from a design perspective. Being able to associate a specific slave device with a DMA channel and letting the hardware arbitrate the bus would certainly be a beneficial feature. The implementation would be more complicated than just letting the SPI peripheral do all the configuration and state management, and simply enabling DMA to shovel bytes in and out of the FIFO seems the simplest. My guess is that it's the absolute simplest implementation, and the SPI peripheral just has a setting to automatically de-assert SS when its TX FIFO and shift register is empty. Once again practically, though, I don't think it's too crazy to drive this particular display and touch screen with per-scanline interrupts. Assuming 30Hz 240 lines, that's a baseline of 7.2Khz IRQs. If you want 400Hz touch samples, that's only 400Hz more IRQs. Go crazy and sample the touch screen at 2.8Khz as long as there's enough bus bandwidth. While it feels a bit inelegant to not have the scanlines be harmonic with the touch updates, the resulting jitter is at a sufficiently short timescale that it doesn't matter. Also, for this class of hardware, you're devoting a significant chunk of RAM to any sort of framebuffer. Expecting to budget a modest amount of CPU to push the framebuffer to the display seems completely reasonable. As you mentioned, it seems like the original intent of the designers were for folk to treat the display hardware as a SPI accessible frame buffer. They were probably targeting 8-bit AVRs popularly used in 3D printers? Update rate was probably the last of their requirements. They probably didn't even consider someone trying to drive the display with DMA. The fact that you're able to drive 50fps is quite remarkable, on top of your achievement of fully offloading it on the rp2040.

dmitrygr 4y ago

The fact that touch and display are on the same SPI bus making it impossible to just set up a repeating DMA to the display. On a non-RP2040 it wouldn’t be possible to drive it easily. Most DMA units wouldn’t allow you to feed a scan line at a time and then poll touch. They aren’t that flexible. You’d be stuck with a lot of IRQs

sgtnoodle 4y ago
The last time I worked with DMA was on an Atmel cortex-M7. Definitely a higher end MCU, though. It supported arbitrarily complex DMA chaining via linked lists in RAM. That being said, doing ~10Khz or so IRQs just to kick off a DMA transfer wouldn't have been much CPU in the grand scheme of things. For the touch input, is it necessary to truly sample at a high rate continuously, or is it more a need to oversample and take an average? If it's a need for oversampling, you could sample the touch screen between frames rather than interleaved between scanlines. Likewise, you could read from the SD card between frames as well. Writing to an SD card would be pretty terrible, given that cards often block for hundreds of milliseconds. Once again, awesome work with the PIO. I imagine you're running the display at its theoretical update limit unless you try overclocking?
- dmitrygr 4y ago
  Between frames would be too rare for touch. You really do not want touch at just 45Hz. More samples are for averaging for smoothing. The touch controller here is noisy and the board layout waveshare did … does not follow the data sheet recommendations for lowering noise. Sd between refreshes is a poor idea for the reason you mention, yes. And yes screen interface is run as fast as it can as per data sheet. Overclocking the MCU wont help as we are not cycle bound there. Overclocking the SPI bus will work but not far. Past 70MHz it gets glitchy.
  - sgtnoodle 4y ago
    I specifically meant rapidly sampling the touch screen N times between frames rather than 1 time between frames. Depending on the characteristics of the EMI, it seems like that could work just as well.
    - dmitrygr 4y ago
      But in terms of temporal resolution you still have 45Hz which is too low for good inking.
crest 4y ago
You could use the DMA engine to reconfigure the function routing using a pair of DMA channels.
- dmitrygr 4y ago
  Sadly in many chips this is not possible. Eg STM32H7 has 3 types of DMA units and each has a complex table of where they can read and write from. Two kinds of DMA cannot write to anywhere that contains a DMA controller. Rp2040 is rather unique in how powerful its DMA units are due to their reach. But it also has faults. Its DMA units seemingly cannot access GPIO, since that is bolted directly to the core (via the single cycle IO port). The only way to affect pins from DMA is via PIO as far as I can tell.
  - sgtnoodle 4y ago
    Reminds me of a memorable Ethernet peripheral. The peripheral would use the MCU's RAM for its own buffering. There were frame headers that both the peripheral and CPU would need to manipulate. Being an M7 the CPU had a data cache, though, and so the CPU couldn't safely manipulate the headers without first completely disabling the cache. I ended up using a DMA channel to indirectly manipulate the headers, bypassing the data cache. It was pretty silly but worked, and was way more efficient than disabling the data cache.
    - dmitrygr 4y ago
      On an m7, you can use the MPU to assign non-cacheable attribute to a memory area. As small as 32bytes and as large as 4GB. Any size representable as 2^n - k * 2^(n - 3) for k = 0..7 and n >= 8
      - sgtnoodle 4y ago
        The particular driver was being retrofitted into a large enough variety of existing firmwares that we really wanted to avoid introducing any sort of MPU or link time configuration. The DMA channel approach was the least impactful path.
sgtnoodle 4y ago
Just for the fun of it, I've been reading a bit about the STM32H7 since you mentioned it. One half baked idea I have to achieve only one interrupt per scan line would be to use multiple DMA channels to drive the same SPI peripheral, one for the display SS and the other for the touch SS. Have a scanline ISR start both transfers at the same time, and rely on the well defined DMA channel priorities to deconflict them. It seems like this could possibly work given that the high level DMA documentation says it can control SS. What I don't know at the moment is how exactly the DMA channel actually controls SS. Does the channel itself have knowledge of the signal, or does it just poke the SPI peripheral at the right time? Said another way, are SS pins assigned to DMA channels, or are they assigned to the SPI peripheral? There's also clock frequency and data mode that would need to be associated with the DMA channel. My guess is it's the latter and my idea wouldn't work. Both options seem like they would have been reasonable from a design perspective. Being able to associate a specific slave device with a DMA channel and letting the hardware arbitrate the bus would certainly be a beneficial feature. The implementation would be more complicated than just letting the SPI peripheral do all the configuration and state management, and simply enabling DMA to shovel bytes in and out of the FIFO seems the simplest. My guess is that it's the absolute simplest implementation, and the SPI peripheral just has a setting to automatically de-assert SS when its TX FIFO and shift register is empty. Once again practically, though, I don't think it's too crazy to drive this particular display and touch screen with per-scanline interrupts. Assuming 30Hz 240 lines, that's a baseline of 7.2Khz IRQs. If you want 400Hz touch samples, that's only 400Hz more IRQs. Go crazy and sample the touch screen at 2.8Khz as long as there's enough bus bandwidth. While it feels a bit inelegant to not have the scanlines be harmonic with the touch updates, the resulting jitter is at a sufficiently short timescale that it doesn't matter. Also, for this class of hardware, you're devoting a significant chunk of RAM to any sort of framebuffer. Expecting to budget a modest amount of CPU to push the framebuffer to the display seems completely reasonable. As you mentioned, it seems like the original intent of the designers were for folk to treat the display hardware as a SPI accessible frame buffer. They were probably targeting 8-bit AVRs popularly used in 3D printers? Update rate was probably the last of their requirements. They probably didn't even consider someone trying to drive the display with DMA. The fact that you're able to drive 50fps is quite remarkable, on top of your achievement of fully offloading it on the rp2040.