Unable to use DMA with ST7789 + ESP32-S3

[TheWolf95]

I only recently come to know about this section of issues/discusson for TFT_eSPI library😂 sorry for my blindess. Getting to the problem (repost of #3414)

IDE: PlatformIO
Library: bodmer/TFT_eSPI@^2.5.43
MCU: ESP32-S3-WROOM-1 N8R8
TFT Driver: ST7789 - SPI (SCK, SDA, RES, DC)

I am trying to use the pushImageDMA() method to achieve faster refreshrate on a ST7789 display, without MISO and CS pin (link to aliexpress). DMA initialization is ok, but any update to the screen is not sent (no fillScreen, no pushImageDMA, not print, ecc), just the last frame is showed.

Wiring: (it is a custom board from another project, I connected spare pins exposed)

TFT_MOSI 18
TFT_SCLK 17
TFT_DC 2
TFT_RST 1

TFT config:

#define ST7789_2_DRIVER

#define TFT_RGB_ORDER TFT_RGB

 #define TFT_WIDTH  240
 #define TFT_HEIGHT  240

#define SPI_FREQUENCY 80*1000000
#define SPI_READ_FREQUENCY  20000000
#define SPI_TOUCH_FREQUENCY  2500000

#define USE_HSPI_PORT

Setup before DMA push:

if(tft.initDMA()) Serial.println("[DMA] Init ok.");
DMA_BUFFER = (uint16_t *) heap_caps_malloc(256 * sizeof(uint16_t), MALLOC_CAP_DMA);  // These buffers could be 240 only
SCREENBUFFER = (uint16_t *) heap_caps_malloc(256 * sizeof(uint16_t), MALLOC_CAP_DMA);

Push in loop function: (this function runs on a parallel thread, priority 2)

extern "C" int lcd_write_frame(const uint16_t x, const uint16_t y, const uint16_t width, const uint16_t height) {
...
  for (i = 0; i < height; i++) {
      for (X_ = 0; X_ < width; X_++) { 
          
          SCREENBUFFER[X_] = gbc_16bit[(0x3F & ((uint8_t *)SCREENMEMORY[i])[X_])];

      }

     tft.startWrite();  // I know this should be called to keep the SPI interface reserved through CS pin (not my case), but who knows
     tft.pushImageDMA(0, i, 240, 1, (uint16_t *) SCREENBUFFER, (uint16_t *) DMA_BUFFER);
     tft.endWrite();

  }
...
}

I get a refresh times of around 80ms (12 FPS) for 240x240 16bit images, I was aiming on improving them (the code is a fork of GBCCAT, ported to PlatformIO + TFT_eSPI library)

UPDATE:
With some code tinkering I improved the rendering times to 45ms (~22 FPS) witout DMA. The code is similar to the one above tho.
I should add that in the meantime I tried some examples from DMA folder and got no results still. I'm waiting for a devkit with ESP32-WROOM-1 MCU to arrive, to test wether the pin configuration is off

UPDATE 2:
I got hands on a ESP32-S3-DevKit-C1 like board to further test on pin mapping

#define TFT_MISO 13
#define TFT_MOSI 11
#define TFT_SCLK 12
#define TFT_CS   10 
#define TFT_DC   38 
#define TFT_RST  2 
#define TFT_BL 9

using this setup, with no interface specified (should fall back to VSPI which should be FSPI in S3). I still get the screen freezing to last pushed frame

[TheWolf95]

Using the setup as per UPDATE2, I'm able to send a 240x240 buffer via DMA in a simple code

#include <TFT_eSPI.h> // Include the graphics library

TFT_eSPI tft = TFT_eSPI();  // Create an instance of the library
uint16_t *dmaBuffer;

void setup() {
  Serial.begin(115200);
  tft.init();
  tft.initDMA();
  tft.setRotation(5);

  tft.setSwapBytes(true);

  // Allocate the DMA buffer
  dmaBuffer = (uint16_t*)heap_caps_malloc(TFT_WIDTH * 240 * sizeof(uint16_t), MALLOC_CAP_DMA | MALLOC_CAP_32BIT);
  if (!dmaBuffer) {
    Serial.println("Failed to allocate DMA buffer!");
    while (1);
  }

  // Fill the DMA buffer with initial data
  for (int i = 0; i < TFT_WIDTH * 240; i++) {
    dmaBuffer[i] = TFT_BLACK; // Initial fill color
  }

  // Perform an initial DMA transfer
  tft.startWrite();
  tft.pushImageDMA(0, 0, TFT_WIDTH, 240, dmaBuffer);
  tft.endWrite();

}

void loop() {
  // Update the DMA buffer with new frame data
  for (int i = 0; i < TFT_WIDTH * 240; i++) {
    dmaBuffer[i] = (i % 3 == 0) ? TFT_RED : ((i % 2 == 0) ? TFT_GREEN : TFT_BLUE); // Example pattern
  }
  Serial.printf("Showing color: 0x%04x\n", dmaBuffer[0]);

  // Perform the DMA transfer
  tft.startWrite();
  tft.pushImageDMA(0, 0, TFT_WIDTH, 240, dmaBuffer);
  tft.endWrite();

  delay(1000);

  tft.fillScreen(TFT_BLACK);

  delay(500);
}

MISO (alternatively define #define TFT_SDA_READ) and CS must be defined (even randomly) and not -1 and the malloc must be done into a DMA and 32bit accessible part of internal RAM, via the flags above.

I now face another problem in the main code:

Free heap: 0, internal: 81328, spiram: 6054475
Guru Meditation Error: Core  0 panic'ed (LoadProhibited). Exception was unhandled.


Core  0 register dump:

PC      : 0x4201b67c  PS      : 0x00060a30  A0      : 0x8201c1bf  A1      : 0x3fcee4e0  
A2      : 0x3fcf7488  A3      : 0x3fc9adc0  A4      : 0x00000000  A5      : 0x00000000  
A6      : 0x60024000  A7      : 0x6002409c  A8      : 0x00000000  A9      : 0x00000000  
A10     : 0x00000000  A11     : 0x00000000  A12     : 0x8037e292  A13     : 0x3fc97400  
A14     : 0x00000000  A15     : 0x00000000  SAR     : 0x0000001d  EXCCAUSE: 0x0000001c  
EXCVADDR: 0x00000088  LBEG    : 0x400570e8  LEND    : 0x400570f3  LCOUNT  : 0x00000000 


Backtrace: 0x4201b679:0x3fcee4e0 0x4201c1bc:0x3fcee510 0x4200cac3:0x3fcee550 0x4200881c:0x3fcee590 0x42006eb5:0x3fcee5c0


  #0  0x4201b679:0x3fcee4e0 in check_trans_valid at /Users/ficeto/Desktop/ESP32/ESP32S2/esp-idf-public/components/driver/spi_master.c:686 (discriminator 2)
  #1  0x4201c1bc:0x3fcee510 in spi_device_queue_trans at /Users/ficeto/Desktop/ESP32/ESP32S2/esp-idf-public/components/driver/spi_master.c:803
  #2  0x4200cac3:0x3fcee550 in TFT_eSPI::pushImageDMA(int, int, int, int, unsigned short*, unsigned short*) at .pio/libdeps/esp32-s3-devkitc-1/TFT_eSPI/Processors/TFT_eSPI_ESP32_S3.c:797
  #3  0x4200881c:0x3fcee590 in lcd_write_frame at src/main.cpp:365
  #4  0x42006eb5:0x3fcee5c0 in fb_draw_task at src/GBCCAT/gnuboy.c:256

the EXCVADRR value suggests a NULL pointer in a struct, which I suspect being static spi_transaction_t trans's tx_buffer pointer in TFT_eSPI class.
Does anyone have an idea on why the pointer I pass to the TFT_eSPI::pushImageDMA(int, int, int, int, unsigned short*, unsigned short*) method is NULL even after I check the malloc beforehand?

UPDATE:
I correct myself, I misread documentation: the error states that the pointer to the struct itself is NULL. So lesser clues than before on what could it be

[TheWolf95]

The solution to this problem was within the fact I used xTask methods to create tasks to take care of video rendering. The stack size (which still now remains a mistery how to determine its size), was too small, causing the methods executed to not being able to read memory properly (this is what I understood, probably the thruth is similar, but not exact).

TL;DR: are you using xTaskCreate? Try increasing the stack size.

I've now to face tearing issues, or better: random pixels are appearing on the first lines of the frame, varying in thickness over time. Curious fact: decreasing the SPI frequency worsen the effect

[rsiemens77]

using this setup, with no interface specified (should fall back to VSPI which should be FSPI in S3). I still get the screen freezing to last pushed frame

Using the setup as per UPDATE2, I'm able to send a 240x240 buffer via DMA in a simple code

Can you clarify? So you are not specifying HSPI? Everything I read says you need to do this, and indeed, unless I do, I get the commonly reported crash. How did you get it to work without defining HSPI?

Looking at the code, if you do

#define USE_HSPI_PORT

It results in using

#define SPI_PORT 3 //HSPI is port 3 on ESP32 S2

Is this what you are using?

I can't get my S3 devkit to do DMA, although the non-DMA call works fine. I used the same pins as you and as listed in the Espressif S3 documentation. I still get the single frame and then freeze.

Pin Name | GPIO Number (SPI2)

CS0 1 | 10
SCLK | 12
MISO | 13
MOSI | 11
QUADWP | 14
QUADHD | 9

[TheWolf95]

Yes, I didn't define HSPI (so it would fall back to VSPI) plus all the other settings and mallocs.
Maybe share your UserSetup and code you are trying to reproduce, so I or someone else can help you further!

[rsiemens77]

Thanks, I was following the discussion here #2233 where it was reported working with HSPI. Let me try your way without HSPI, and then I'll follow up. Thanks for quick response!

[rsiemens77]

I started this thread with my detailed trials and results if you have any insights, I'd appreciate it.

#3473