Pico2 DMA Channel Interference

[samneggs]

I have two WAV players, DMA0 feeding PIO0 and DMA1 feeding PIO1. The PIOs act as a PWM. Either works fine until the other is triggered. Then the second DMA seems to occasionally start without being triggered. I looked at the CH0_READ register and verified it was changing when only CH1 was active. Also, eventually I'll get a DMA CH0 read error. Maybe my pwm1 and pwm2 are not independent?
Sam

# wav player
# https://antirez.com/news/143
# ffmpeg -i ohno.wav -ar 24000 -acodec pcm_u8 -f u8 def_fire.raw


from machine import Pin, PWM, mem32, mem8
from time import sleep_us, sleep
from rp2 import PIO, StateMachine, asm_pio
from uctypes import addressof
from rp2 import DMA
import gc, array

PIO0_BASE = const(0x50200000)
PIO1_BASE = const(0x50300000)
TXF0 = const(0x010) 

DMA_BASE        = const(0x50000000)
CH0_READ        = const(0x00)
CH0_WRITE       = const(0x04)
CH0_TRANS_COUNT = const(0x08)
CH0_CTRL_TRIG   = const(0x0c)
CH1_READ        = const(0x40)
CH1_WRITE       = const(0x44)
CH1_TRANS_COUNT = const(0x48)
CH1_CTRL_TRIG   = const(0x4c)

DREQ_PIO0_TX0 = const(0)
DREQ_PIO1_TX0 = const(8)

@asm_pio(sideset_init=PIO.OUT_LOW)
def pwm_prog():
    pull(block) .side(0)
    out(x,8)
    mov(y, isr) # ISR must be preloaded with PWM count max
    label("pwmloop")
    jmp(x_not_y, "skip")
    nop()         .side(1)
    label("skip")
    jmp(y_dec, "pwmloop")


class PIOPWM:
    def __init__(self, sm_id, pin , max_count, count_freq):
        self._sm = StateMachine(sm_id, pwm_prog, freq=2 * count_freq, sideset_base=Pin(pin))
        # Use exec() to load max count into ISR
        self.sm_id = sm_id
        self._sm.put(max_count)
        self._sm.exec("pull()")
        self._sm.exec("mov(isr, osr)")
        self._sm.active(1)
        self._max_count = max_count
    def set(self, value):
        # Minimum value is -1 (completely turn off), 0 actually still produces narrow pulse
        value = max(value, -1)
        value = min(value, self._max_count)
        self._sm.put(value)
    @micropython.viper
    def effect(self,datainptr:ptr32,length:int):
        dma_base = ptr32(DMA_BASE)
        print(self.sm_id,datainptr,length)
        if int(self.sm_id) == 0:
            print(dma_base[CH0_TRANS_COUNT//4],dma_base[CH1_TRANS_COUNT//4])
            dma_base[CH0_READ//4] =  int(datainptr)
            dma_base[CH0_WRITE//4] = int(PIO0_BASE +  TXF0 )
            dma_base[CH0_TRANS_COUNT//4]   = length
            #                                    EN  DATA_SIZE   INC_READ   INC_WRITE   DREQ                     
            dma_base[CH0_CTRL_TRIG//4]          = 1 | 0<<2     | 1<<4     | 0<<6       | DREQ_PIO0_TX0<<17
        else:
            print(dma_base[CH0_TRANS_COUNT//4],dma_base[CH1_TRANS_COUNT//4])
            dma_base[CH1_READ//4] =  int(datainptr)
            dma_base[CH1_WRITE//4] = int(PIO1_BASE +  TXF0 )
            dma_base[CH1_TRANS_COUNT//4]   = length
            #                                    EN  DATA_SIZE   INC_READ   INC_WRITE   DREQ                     
            dma_base[CH1_CTRL_TRIG//4]          = 1 | 0<<2     | 1<<4     | 0<<6       | DREQ_PIO1_TX0<<17
        

if __name__=='__main__': 
    pwm1 = PIOPWM(0, 20, max_count=(1 << 10) - 1, count_freq=20_000_000)
    pwm2 = PIOPWM(4, 21, max_count=(1 << 10) - 1, count_freq=20_000_000)
    FIRE_BUTTON = Pin(14, Pin.IN, Pin.PULL_UP)
    BOMB_BUTTON = Pin(15, Pin.IN, Pin.PULL_UP)
    f = open("def_fire.raw","rb")
    buf1 = f.read()
    f.close()
    f = open("deflanddie.raw","rb")
    buf2 = f.read()
    f.close()

    #machine.freq(220_000_000) #220
    #machine.mem32[0x40010048] = 1<<11 # enable peri_ctrl clock
    length1 = len(buf1)
    length2 = len(buf2)
    print('buf1 ', hex(addressof(buf1)))
    print('buf2 ', hex(addressof(buf2)))
    print(mem32[DMA_BASE + CH0_CTRL_TRIG] & (1<<29))     # write error?
    #mem32[DMA_BASE + CH0_CTRL_TRIG] = (1<<30)  # clear error
    print(mem32[DMA_BASE + CH0_CTRL_TRIG] & (1<<30),hex(mem32[DMA_BASE + CH0_READ]))     # read error?
    print(mem32[DMA_BASE + CH1_CTRL_TRIG] & (1<<29))     # write error?
    print(mem32[DMA_BASE + CH1_CTRL_TRIG] & (1<<30))     # read error?
    while True:
        if FIRE_BUTTON.value() == 0:    
            pwm1.effect(buf1,length1)
        if BOMB_BUTTON.value() == 0:    
            pwm2.effect(buf2,length2)
        sleep_us(150000)

[rkompass]

Hello Sam,

your problem description looks like the DMA channels are triggering each other.

From a short inspection of the code I note that

#                                        EN  DATA_SIZE   INC_READ   INC_WRITE   DREQ                     
     dma_base[CH0_CTRL_TRIG//4]        = 1   | 0<<2     | 1<<4     | 0<<6     | DREQ_PIO0_TX0<<17

AFAIU contains errors: INC_WRITE should be 0<<5 and DREQ should be DREQ_PIO0_TX0<<15. Same with CH1_CTRL_TRIG.
With CH1 this leads to DREQ_PIO1_TX0 (=8) becoming 32=DREQ_I2C_TX.
But that does not explain a triggering.
I didn't inspect the last lines of code yet, where you set the triggers.

In order to make DMA programming simpler I devised a class where the DMA configures basically itself, depending on which objects are given it as source and destination.
This predated the pull of the rp2.DMA code. Actually I should rewrite it to use that code, complete it to include SPI and I2C, among others and offer it as something like SimpleDMA.
There are examples (which I have tested) where the PIO is destination. Also, example 5 produces sound via PWM, in 11 bit quality. This happens cyclically without interruption by using a second DMA for reprogramming the first.

If you don't have an idea what to do next to find the problem, you might try that out.

[samneggs]

Hi rkompass,
Thanks for taking the time to look into my problem. The INC_WRITE and other bits have moved around with the RP2350 to add new features so I believe they are correct. I'll take a look at your DMA code for some ideas.
Sam

[samneggs]

AAARRRGG!
I’m basically describing DMA chaining as a problem. The default configuration is zero so all channels chain back to channel zero unless written to chain to themselves. DMA chaining is an opt out feature!
Surprised this isn’t a more common occurrence.

[rkompass]

Hi samneggs,

thanks for the clarification that the ctrl/trig config bits are different in Pico2. I could not imagine that naively.
So it seems you have found the problem. And we can expect more problem reports of this sort :-)