Parcourir la source

Add DMAChannel (work in progress)

teensy4-core
PaulStoffregen il y a 10 ans
Parent
révision
7b040b4be0
2 fichiers modifiés avec 265 ajouts et 0 suppressions
  1. +31
    -0
      teensy3/DMAChannel.cpp
  2. +234
    -0
      teensy3/DMAChannel.h

+ 31
- 0
teensy3/DMAChannel.cpp Voir le fichier

@@ -0,0 +1,31 @@
#include "DMAChannel.h"

// The channel allocation bitmask is accessible from "C" namespace,
// so C-only code can reserve DMA channels
uint16_t dma_channel_allocated_mask = 0;

DMAChannel::DMAChannel(uint8_t channelRequest) : TCD(*(TCD_t *)0), channel(16)
{
uint8_t next, ch=channelRequest;

__disable_irq();
while (1) {
if (!(dma_channel_allocated_mask & (1 << ch))) {
dma_channel_allocated_mask |= (1 << ch);
__enable_irq();
break;
}
next = (ch + 1) & 15;
if (next == channelRequest) {
__enable_irq();
return; // no more channels available
// attempts to use this object will hardfault
}
}
channel = ch;
TCD = *(TCD_t *)(0x40009000 + ch * 32);
SIM_SCGC7 |= SIM_SCGC7_DMA;
SIM_SCGC6 |= SIM_SCGC6_DMAMUX;
DMA_CR = 0;
}


+ 234
- 0
teensy3/DMAChannel.h Voir le fichier

@@ -0,0 +1,234 @@
#ifndef DMAChannel_h_
#define DMAChannel_h_

#include "mk20dx128.h"

// This code is a work-in-progress. It's incomplete and not usable yet...

#ifdef __cplusplus

class DMAChannel {
typedef struct __attribute__((packed)) {
volatile const void * volatile SADDR;
int16_t SOFF;
union { uint16_t ATTR; struct { uint8_t ATTR_DST; uint8_t ATTR_SRC; }; };
uint32_t NBYTES;
int32_t SLAST;
volatile void * volatile DADDR;
int16_t DOFF;
volatile uint16_t CITER;
int32_t DLASTSGA;
volatile uint16_t CSR;
volatile uint16_t BITER;
} TCD_t;
public:
// Constructor - allocates which DMA channel each object actually uses
DMAChannel(uint8_t channelRequest=0);

// Triggers cause the DMA channel to actually move data.
//
// Use a hardware trigger to make the DMA channel run
void attachTrigger(uint8_t source) {
volatile uint8_t *mux;
mux = (volatile uint8_t *)&(DMAMUX0_CHCFG0) + channel;
*mux = 0;
*mux = source | DMAMUX_ENABLE;
}
// Use another DMA channel as the trigger, causing this
// channel to trigger every time it triggers. This
// effectively makes the 2 channels run in parallel.
void attachTrigger(DMAChannel &channel) {

}
// Use another DMA channel as the trigger, causing this
// channel to trigger when it completes.
void attachTriggerOnCompletion(DMAChannel &channel) {

}

// An interrupt routine can be run when the DMA channel completes
// the entire transfer.
void attachInterrupt(void (*isr)(void)) {
_VectorsRam[channel + IRQ_DMA_CH0 + 16] = isr;
NVIC_ENABLE_IRQ(IRQ_DMA_CH0 + channel);
}

// Use a single variable as the data source. Typically a register
// for receiving data from one of the hardware peripherals is used.
void source(const signed char &p) { src(&p, 1); }
void source(const unsigned char &p) { src(&p, 1); }
void source(const signed short &p) { src(&p, 2); }
void source(const unsigned short &p) { src(&p, 2); }
void source(const signed int &p) { src(&p, 4); }
void source(const unsigned int &p) { src(&p, 4); }
void source(const signed long &p) { src(&p, 4); }
void source(const unsigned long &p) { src(&p, 4); }

// Use a buffer (array of data) as the data source. Typically a
// buffer for transmitting data is used.
void sourceBuffer(const signed char p[], unsigned int len) { src(p, 1, len); }
void sourceBuffer(const unsigned char p[], unsigned int len) { src(p, 1, len); }
void sourceBuffer(const signed short p[], unsigned int len) { src(p, 2, len); }
void sourceBuffer(const unsigned short p[], unsigned int len) { src(p, 2, len); }
void sourceBuffer(const signed int p[], unsigned int len) { src(p, 4, len); }
void sourceBuffer(const unsigned int p[], unsigned int len) { src(p, 4, len); }
void sourceBuffer(const signed long p[], unsigned int len) { src(p, 4, len); }
void sourceBuffer(const unsigned long p[], unsigned int len) { src(p, 4, len); }

// Use a circular buffer as the data source
void sourceCircular(const signed char p[], unsigned int len) { srcc(p, 1, len); }
void sourceCircular(const unsigned char p[], unsigned int len) { srcc(p, 1, len); }
void sourceCircular(const signed short p[], unsigned int len) { srcc(p, 2, len); }
void sourceCircular(const unsigned short p[], unsigned int len) { srcc(p, 2, len); }
void sourceCircular(const signed int p[], unsigned int len) { srcc(p, 4, len); }
void sourceCircular(const unsigned int p[], unsigned int len) { srcc(p, 4, len); }
void sourceCircular(const signed long p[], unsigned int len) { srcc(p, 4, len); }
void sourceCircular(const unsigned long p[], unsigned int len) { srcc(p, 4, len); }


// Use a single variable as the data destination. Typically a register
// for transmitting data to one of the hardware peripherals is used.
void destination(signed char &p) { src(&p, 1); }
void destination(unsigned char &p) { src(&p, 1); }
void destination(signed short &p) { src(&p, 2); }
void destination(unsigned short &p) { src(&p, 2); }
void destination(signed int &p) { src(&p, 4); }
void destination(unsigned int &p) { src(&p, 4); }
void destination(signed long &p) { src(&p, 4); }
void destination(unsigned long &p) { src(&p, 4); }

// Use a buffer (array of data) as the data destination. Typically a
// buffer for receiving data is used.
void destinationBuffer(signed char p[], unsigned int len) { src(p, 1, len); }
void destinationBuffer(unsigned char p[], unsigned int len) { src(p, 1, len); }
void destinationBuffer(signed short p[], unsigned int len) { src(p, 2, len); }
void destinationBuffer(unsigned short p[], unsigned int len) { src(p, 2, len); }
void destinationBuffer(signed int p[], unsigned int len) { src(p, 4, len); }
void destinationBuffer(unsigned int p[], unsigned int len) { src(p, 4, len); }
void destinationBuffer(signed long p[], unsigned int len) { src(p, 4, len); }
void destinationBuffer(unsigned long p[], unsigned int len) { src(p, 4, len); }

// Use a circular buffer as the data destination
void destinationCircular(signed char p[], unsigned int len) { srcc(p, 1, len); }
void destinationCircular(unsigned char p[], unsigned int len) { srcc(p, 1, len); }
void destinationCircular(signed short p[], unsigned int len) { srcc(p, 2, len); }
void destinationCircular(unsigned short p[], unsigned int len) { srcc(p, 2, len); }
void destinationCircular(signed int p[], unsigned int len) { srcc(p, 4, len); }
void destinationCircular(unsigned int p[], unsigned int len) { srcc(p, 4, len); }
void destinationCircular(signed long p[], unsigned int len) { srcc(p, 4, len); }
void destinationCircular(unsigned long p[], unsigned int len) { srcc(p, 4, len); }

// TODO: explicit function for configuring transfer length....
// should we try to automatically pick it up from the array lengths?

// TODO: functions to configure major/minor loop
// option #1 - trigger moves 1 byte/word (minor=1, major=count)
// option #2 - trigger moves all data (minor=count, major=1)
// option ?? - more complex config, write TCD manually....

// TODO: functions to set other options, functions to enable
// manual start function, etc

// TODO: "get" functions, to read important stuff, like SADDR & DADDR...

// For complex and unusual configurations not possible with the above
// functions, the Transfer Control Descriptor (TCD) and channel number
// can be used directly. This leads to less portable and less readable
// code, but direct control of all parameters is possible.
TCD_t &TCD;
uint8_t channel;

protected:
void src(const void *p, uint32_t size) {
TCD.SADDR = p;
TCD.SOFF = 0;
if (size == 1) {
TCD.ATTR_SRC = 0; // 8 bits
} else if (size == 2) {
TCD.ATTR_SRC = 1; // 16 bits
} else {
TCD.ATTR_SRC = 2; // 32 bits
}
//TCD.NBYTES = size;
TCD.SLAST = 0;
}
void src(const void *p, uint32_t size, uint32_t len) {
TCD.SADDR = p;
TCD.SOFF = size;
if (size == 1) {
TCD.ATTR_SRC = 0; // 8 bits
} else if (size == 2) {
TCD.ATTR_SRC = 1; // 16 bits
} else {
TCD.ATTR_SRC = 2; // 32 bits
}
//TCD.NBYTES = size;
TCD.SLAST = -len;
}
void srcc(const void *p, uint32_t size, uint32_t len) {
TCD.SADDR = p;
TCD.SOFF = size;
if (size == 1) {
TCD.ATTR_SRC = 0 | ((31 - __builtin_clz(len)) << 3); // 8 bits
} else if (size == 2) {
TCD.ATTR_SRC = 1 | ((31 - __builtin_clz(len)) << 3); // 16 bits
} else {
TCD.ATTR_SRC = 2 | ((31 - __builtin_clz(len)) << 3); // 32 bits
}
//TCD.NBYTES = size;
TCD.SLAST = 0;
}

void dst(void *p, uint32_t size) {
TCD.DADDR = p;
TCD.DOFF = 0;
if (size == 1) {
TCD.ATTR_SRC = 0; // 8 bits
} else if (size == 2) {
TCD.ATTR_SRC = 1; // 16 bits
} else {
TCD.ATTR_SRC = 2; // 32 bits
}
//TCD.NBYTES = size;
TCD.DLASTSGA = 0;
}
void dst(void *p, uint32_t size, uint32_t len) {
TCD.DADDR = p;
TCD.DOFF = size;
if (size == 1) {
TCD.ATTR_DST = 0; // 8 bits
} else if (size == 2) {
TCD.ATTR_DST = 1; // 16 bits
} else {
TCD.ATTR_DST = 2; // 32 bits
}
//TCD.NBYTES = size;
TCD.DLASTSGA = -len;
}
void dstc(void *p, uint32_t size, uint32_t len) {
TCD.DADDR = p;
TCD.DOFF = size;
if (size == 1) {
TCD.ATTR_DST = 0 | ((31 - __builtin_clz(len)) << 3); // 8 bits
} else if (size == 2) {
TCD.ATTR_DST = 1 | ((31 - __builtin_clz(len)) << 3); // 16 bits
} else {
TCD.ATTR_DST = 2 | ((31 - __builtin_clz(len)) << 3); // 32 bits
}
//TCD.NBYTES = size;
TCD.DLASTSGA = 0;
}

};


extern "C" {
#endif
extern uint16_t dma_channel_allocated_mask;
#ifdef __cplusplus
}
#endif



#endif

Chargement…
Annuler
Enregistrer