Merge pull request #663 from tannewt/dma3

Use DMA for long SPI transactions including those to the SPI Flash.

10 files changed, 417 insertions, 222 deletions

diff --git a/ports/atmel-samd/Makefile b/ports/atmel-samd/Makefile
index 4660d575e..73799e5f2 100644
--- a/ports/atmel-samd/Makefile
+++ b/ports/atmel-samd/Makefile
@@ -217,9 +217,6 @@ endif
 
 SRC_ASF := $(addprefix asf4/$(CHIP_FAMILY)/, $(SRC_ASF))
 
-# Skip this source for now.
-#	shared_dma.c \
-
 SRC_C = \
 	background.c \
 	fatfs_port.c \
@@ -229,6 +226,7 @@ SRC_C = \
 	$(CHIP_FAMILY)_peripherals.c \
 	peripherals.c \
 	$(CHIP_FAMILY)_pins.c \
+	shared_dma.c \
 	tick.c \
 	timers.c \
 	usb.c \
diff --git a/ports/atmel-samd/boards/itsybitsy_m0_express/mpconfigboard.h b/ports/atmel-samd/boards/itsybitsy_m0_express/mpconfigboard.h
index ca39792a9..675b9a30d 100644
--- a/ports/atmel-samd/boards/itsybitsy_m0_express/mpconfigboard.h
+++ b/ports/atmel-samd/boards/itsybitsy_m0_express/mpconfigboard.h
@@ -5,7 +5,7 @@
 #define MICROPY_HW_APA102_MOSI   (&pin_PA01)
 #define MICROPY_HW_APA102_SCK    (&pin_PA00)
 
-// Salae reads 12mhz which is the limit even though we set it to the safer 8mhz.
+// Saleae reads 12mhz which is the limit even though we set it to the safer 8mhz.
 #define SPI_FLASH_BAUDRATE  (8000000)
 
 #define SPI_FLASH_MOSI_PIN          PIN_PB22
diff --git a/ports/atmel-samd/boards/metro_m4_express_revb/mpconfigboard.h b/ports/atmel-samd/boards/metro_m4_express_revb/mpconfigboard.h
index 70c92ea44..7af9767ca 100644
--- a/ports/atmel-samd/boards/metro_m4_express_revb/mpconfigboard.h
+++ b/ports/atmel-samd/boards/metro_m4_express_revb/mpconfigboard.h
@@ -11,7 +11,7 @@
 
 #define MICROPY_HW_NEOPIXEL (&pin_PB17)
 
-#define SPI_FLASH_BAUDRATE  (8000000)
+#define SPI_FLASH_BAUDRATE  (60000000)
 
 // Rev B: single channel SPI
 // Rev C will be QSPI
diff --git a/ports/atmel-samd/common-hal/busio/SPI.c b/ports/atmel-samd/common-hal/busio/SPI.c
index 43288cc8b..eca73be78 100644
--- a/ports/atmel-samd/common-hal/busio/SPI.c
+++ b/ports/atmel-samd/common-hal/busio/SPI.c
@@ -38,7 +38,7 @@
 
 #include "peripherals.h"
 #include "pins.h"
-//#include "shared_dma.h"
+#include "shared_dma.h"
 
 void common_hal_busio_spi_construct(busio_spi_obj_t *self,
         const mcu_pin_obj_t * clock, const mcu_pin_obj_t * mosi,
@@ -113,7 +113,7 @@ void common_hal_busio_spi_construct(busio_spi_obj_t *self,
 
     // Set up SPI clocks on SERCOM.
     samd_peripherals_sercom_clock_init(sercom, sercom_index);
-    
+
     #if defined(MICROPY_HW_APA102_SCK) && defined(MICROPY_HW_APA102_MOSI) && !defined(CIRCUITPY_BITBANG_APA102)
     // if we're re-using the dotstar sercom, make sure it is disabled or the init will fail out
     hri_sercomspi_clear_CTRLA_ENABLE_bit(sercom);
@@ -121,7 +121,7 @@ void common_hal_busio_spi_construct(busio_spi_obj_t *self,
     if (spi_m_sync_init(&self->spi_desc, sercom) != ERR_NONE) {
         mp_raise_OSError(MP_EIO);
     }
-    
+
     // Pads must be set after spi_m_sync_init(), which uses default values from
     // the prototypical SERCOM.
     hri_sercomspi_write_CTRLA_DOPO_bf(sercom, dopo);
@@ -135,7 +135,7 @@ void common_hal_busio_spi_construct(busio_spi_obj_t *self,
         // busy or not
         mp_raise_OSError(MP_EIO);
     }
-        
+
     gpio_set_pin_direction(clock->pin, GPIO_DIRECTION_OUT);
     gpio_set_pin_pull_mode(clock->pin, GPIO_PULL_OFF);
     gpio_set_pin_function(clock->pin, clock_pinmux);
@@ -194,7 +194,7 @@ bool common_hal_busio_spi_configure(busio_spi_obj_t *self,
         return true;
     }
 
-    // Disable, set values (most or all are enable-protected), and re-enable. 
+    // Disable, set values (most or all are enable-protected), and re-enable.
     spi_m_sync_disable(&self->spi_desc);
     hri_sercomspi_wait_for_sync(hw, SERCOM_SPI_SYNCBUSY_MASK);
 
@@ -235,14 +235,14 @@ bool common_hal_busio_spi_write(busio_spi_obj_t *self,
         return true;
     }
     int32_t status;
-//    if (len >= 16) {
-//        status = shared_dma_write(self->spi_desc.dev.prvt, data, len);
-//    } else {
+    if (len >= 16) {
+        status = sercom_dma_write(self->spi_desc.dev.prvt, data, len);
+    } else {
         struct io_descriptor *spi_io;
         spi_m_sync_get_io_descriptor(&self->spi_desc, &spi_io);
         status = spi_io->write(spi_io, data, len);
-//    }
-        return status >= 0; // Status is number of chars read or an error code < 0.
+    }
+    return status >= 0; // Status is number of chars read or an error code < 0.
 }
 
 bool common_hal_busio_spi_read(busio_spi_obj_t *self,
@@ -251,17 +251,17 @@ bool common_hal_busio_spi_read(busio_spi_obj_t *self,
         return true;
     }
     int32_t status;
-//    if (len >= 16) {
-//        status = shared_dma_read(self->spi_desc.dev.prvt, data, len, write_value);
-//    } else {
+    if (len >= 16) {
+        status = sercom_dma_read(self->spi_desc.dev.prvt, data, len, write_value);
+    } else {
         self->spi_desc.dev.dummy_byte = write_value;
 
         struct io_descriptor *spi_io;
         spi_m_sync_get_io_descriptor(&self->spi_desc, &spi_io);
 
         status = spi_io->read(spi_io, data, len);
-//    }
-        return status >= 0; // Status is number of chars read or an error code < 0.
+    }
+    return status >= 0; // Status is number of chars read or an error code < 0.
 }
 
 bool common_hal_busio_spi_transfer(busio_spi_obj_t *self, uint8_t *data_out, uint8_t *data_in, size_t len) {
@@ -269,16 +269,16 @@ bool common_hal_busio_spi_transfer(busio_spi_obj_t *self, uint8_t *data_out, uin
         return true;
     }
     int32_t status;
-//    if (len >= 16) {
-//        status = shared_dma_transfer(self->spi_master_instance.hw, data_out, data_in, len, 0 /*ignored*/);
-//    } else {
+    if (len >= 16) {
+        status = sercom_dma_transfer(self->spi_desc.dev.prvt, data_out, data_in, len);
+    } else {
         struct spi_xfer xfer;
         xfer.txbuf = data_out;
         xfer.rxbuf = data_in;
         xfer.size = len;
         status = spi_m_sync_transfer(&self->spi_desc, &xfer);
-//    }
-        return status >= 0; // Status is number of chars read or an error code < 0.
+    }
+    return status >= 0; // Status is number of chars read or an error code < 0.
 }
 
 uint32_t common_hal_busio_spi_get_frequency(busio_spi_obj_t* self) {
diff --git a/ports/atmel-samd/external_flash/external_flash.c b/ports/atmel-samd/external_flash/external_flash.c
index ca8121bc9..304bf6462 100644
--- a/ports/atmel-samd/external_flash/external_flash.c
+++ b/ports/atmel-samd/external_flash/external_flash.c
@@ -199,7 +199,6 @@ void external_flash_init(void) {
 
     spi_flash_init();
 
-
     for (uint8_t i = 0; i < num_possible_devices; i++) {
         const external_flash_device* possible_device = &possible_devices[i];
         uint8_t jedec_id_response[3] = {0x00, 0x00, 0x00};
@@ -213,7 +212,6 @@ void external_flash_init(void) {
     }
 
     if (flash_device == NULL) {
-        asm("bkpt");
         return;
     }
 
diff --git a/ports/atmel-samd/external_flash/qspi_flash.c b/ports/atmel-samd/external_flash/qspi_flash.c
index 54f1af0dc..46e3cb92a 100644
--- a/ports/atmel-samd/external_flash/qspi_flash.c
+++ b/ports/atmel-samd/external_flash/qspi_flash.c
@@ -30,6 +30,7 @@
 #include <string.h>
 
 #include "external_flash/common_commands.h"
+#include "shared_dma.h"
 
 #include "atmel_start_pins.h"
 #include "hal_gpio.h"
@@ -125,6 +126,8 @@ bool spi_flash_write_data(uint32_t address, uint8_t* data, uint32_t length) {
                            QSPI_INSTRFRAME_DATAEN;
 
     memcpy(((uint8_t *) QSPI_AHB) + address, data, length);
+    // TODO(tannewt): Fix DMA and enable it.
+    // qspi_dma_write(address, data, length);
 
     QSPI->CTRLA.reg = QSPI_CTRLA_ENABLE | QSPI_CTRLA_LASTXFER;
 
@@ -148,6 +151,8 @@ bool spi_flash_read_data(uint32_t address, uint8_t* data, uint32_t length) {
                            QSPI_INSTRFRAME_DUMMYLEN(8);
 
     memcpy(data, ((uint8_t *) QSPI_AHB) + address, length);
+    // TODO(tannewt): Fix DMA and enable it.
+    // qspi_dma_read(address, data, length);
 
     QSPI->CTRLA.reg = QSPI_CTRLA_ENABLE | QSPI_CTRLA_LASTXFER;
 
@@ -167,12 +172,15 @@ void spi_flash_init(void) {
     QSPI->CTRLA.reg = QSPI_CTRLA_SWRST;
     // We don't need to wait because we're running as fast as the CPU.
 
-    QSPI->BAUD.bit.BAUD = 1;
+    // Slow, good for debugging with Saleae
+    // QSPI->BAUD.bit.BAUD = 32;
+    // Super fast
+    QSPI->BAUD.bit.BAUD = 2;
     QSPI->CTRLB.reg = QSPI_CTRLB_MODE_MEMORY |
                       QSPI_CTRLB_DATALEN_8BITS |
                       QSPI_CTRLB_CSMODE_LASTXFER;
 
-    QSPI->CTRLA.bit.ENABLE = 1;
+    QSPI->CTRLA.reg = QSPI_CTRLA_ENABLE;
 
     // The QSPI is only connected to one set of pins in the SAMD51 so we can hard code it.
     uint32_t pins[6] = {PIN_PA08, PIN_PA09, PIN_PA10, PIN_PA11, PIN_PB10, PIN_PB11};
diff --git a/ports/atmel-samd/external_flash/spi_flash.c b/ports/atmel-samd/external_flash/spi_flash.c
index 4b63132ee..fcaedcaac 100644
--- a/ports/atmel-samd/external_flash/spi_flash.c
+++ b/ports/atmel-samd/external_flash/spi_flash.c
@@ -30,6 +30,7 @@
 
 #include "external_flash/common_commands.h"
 #include "peripherals.h"
+#include "shared_dma.h"
 
 #include "hal_gpio.h"
 #include "hal_spi_m_sync.h"
@@ -91,14 +92,28 @@ bool spi_flash_write_data(uint32_t address, uint8_t* data, uint32_t data_length)
     uint8_t request[4] = {CMD_PAGE_PROGRAM, 0x00, 0x00, 0x00};
     // Write the SPI flash write address into the bytes following the command byte.
     address_to_bytes(address, request + 1);
-    return transfer(request, 4, data, NULL, data_length);
+    struct spi_xfer xfer = { request, NULL, 4 };
+    flash_enable();
+    int32_t status = spi_m_sync_transfer(&spi_flash_desc, &xfer);
+    if (status >= 0) {
+        status = sercom_dma_write(spi_flash_desc.dev.prvt, data, data_length);
+    }
+    flash_disable();
+    return status >= 0;
 }
 
 bool spi_flash_read_data(uint32_t address, uint8_t* data, uint32_t data_length) {
     uint8_t request[4] = {CMD_READ_DATA, 0x00, 0x00, 0x00};
     // Write the SPI flash write address into the bytes following the command byte.
     address_to_bytes(address, request + 1);
-    return transfer(request, 4, NULL, data, data_length);
+    struct spi_xfer xfer = { request, NULL, 4 };
+    flash_enable();
+    int32_t status = spi_m_sync_transfer(&spi_flash_desc, &xfer);
+    if (status >= 0) {
+        status = sercom_dma_read(spi_flash_desc.dev.prvt, data, data_length, 0xff);
+    }
+    flash_disable();
+    return status >= 0;
 }
 
 void spi_flash_init(void) {
diff --git a/ports/atmel-samd/shared_dma.c b/ports/atmel-samd/shared_dma.c
index 59fca30c8..8a6a606bb 100644
--- a/ports/atmel-samd/shared_dma.c
+++ b/ports/atmel-samd/shared_dma.c
@@ -23,245 +23,414 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-
-#include <stdbool.h>
 #include "shared_dma.h"
 
+#include <string.h>
+
 #include "py/gc.h"
 #include "py/mpstate.h"
 
-#undef ENABLE
+#include "hal/utils/include/utils.h"
+
+#include "shared-bindings/microcontroller/__init__.h"
 
-// We allocate two DMA resources for the entire lifecycle of the board (not the
+// We allocate three DMA resources for the entire lifecycle of the board (not the
 // vm) because the general_dma resource will be shared between the REPL and SPI
 // flash. Both uses must block each other in order to prevent conflict.
-struct dma_resource audio_dma;
-struct dma_resource general_dma_tx;
-struct dma_resource general_dma_rx;
+COMPILER_ALIGNED(16) static DmacDescriptor dma_descriptors[3];
+
+// Don't use these directly. They are used by the DMA engine itself.
+COMPILER_ALIGNED(16) static DmacDescriptor write_back_descriptors[3];
+
+#define AUDIO_DMA_CHANNEL 0
+#define SHARED_TX_CHANNEL 1
+#define SHARED_RX_CHANNEL 2
+
+
+#ifdef SAMD21
+#define FIRST_SERCOM_RX_TRIGSRC 0x01
+#define FIRST_SERCOM_TX_TRIGSRC 0x02
+#endif
+#ifdef SAMD51
+#define FIRST_SERCOM_RX_TRIGSRC 0x04
+#define FIRST_SERCOM_TX_TRIGSRC 0x05
+#endif
+
+// static void dma_configure_audio(uint8_t channel) {
+//     system_interrupt_enter_critical_section();
+//     /** Select the DMA channel and clear software trigger */
+//     DMAC->CHID.reg = DMAC_CHID_ID(channel);
+//     DMAC->CHCTRLA.reg &= ~DMAC_CHCTRLA_ENABLE;
+//     DMAC->CHCTRLA.reg = DMAC_CHCTRLA_SWRST;
+//     DMAC->SWTRIGCTRL.reg &= (uint32_t)(~(1 << channel));
+//     uint32_t event_output_enable = 0;
+//     if (output_event) {
+//         event_output_enable = DMAC_CHCTRLB_EVOE;
+//     }
+//     DMAC->CHCTRLB.reg = DMAC_CHCTRLB_LVL(DMA_PRIORITY_LEVEL_0) |
+//             DMAC_CHCTRLB_TRIGSRC(trigsrc) |
+//             DMAC_CHCTRLB_TRIGACT(DMA_TRIGGER_ACTION_BEAT) |
+//             event_output_enable;
+//     // config.peripheral_trigger = DAC_DMAC_ID_EMPTY;
+//     // config.trigger_action = DMA_TRIGGER_ACTION_BEAT;
+//     // config.event_config.input_action = DMA_EVENT_INPUT_TRIG;
+//     // config.event_config.event_output_enable = true;
+//     system_interrupt_leave_critical_section();
+// }
 
 void init_shared_dma(void) {
-    struct dma_resource_config config;
-    dma_get_config_defaults(&config);
-
-    // See asf4_conf/hpl_dmac_config.h for initial settings for DMA channels
-    // DMA Channel 0: audio, highest priority,
-    // normal transfer on input, DAC 0 empty is trigger source, trigger on each beat, beat is one byte
-    // output enable true.
-    // asf3 settings:
-    //config.peripheral_trigger = DAC_DMAC_ID_EMPTY;
-    //config.trigger_action = DMA_TRIGGER_ACTION_BEAT;
-    //config.event_config.input_action = DMA_EVENT_INPUT_TRIG;
-    //config.event_config.event_output_enable = true;
-
-    // Turn on the transfer complete interrupt so that the job_status changes to done.
-    g_chan_interrupt_flag[audio_dma.channel_id] |= (1UL << DMA_CALLBACK_TRANSFER_DONE);
-
-    // Prioritize the RX channel over the TX channel because TX can cause an RX
-    // overflow.
-    
-    // DMA Channel 1: rx channel,
-    // normal transfer on input, trigger on each beat, beat is one byte
-    //config.trigger_action = DMA_TRIGGER_ACTION_BEAT;
-    //config.event_config.input_action = DMA_EVENT_INPUT_TRIG;
-    dma_allocate(&general_dma_rx, &config);
-    g_chan_interrupt_flag[general_dma_rx.channel_id] |= (1UL << DMA_CALLBACK_TRANSFER_DONE);
-
-    // DMA Channel 1: rx channel,
-    // normal transfer on input, trigger on each beat, beat is one byte
-    //config.trigger_action = DMA_TRIGGER_ACTION_BEAT;
-    //config.event_config.input_action = DMA_EVENT_INPUT_TRIG;
-    g_chan_interrupt_flag[general_dma_tx.channel_id] |= (1UL << DMA_CALLBACK_TRANSFER_DONE);
-
-    // Be sneaky and reuse the active descriptor memory.
-    audio_dma.descriptor = &descriptor_section[audio_dma.channel_id];
-    general_dma_rx.descriptor = &descriptor_section[general_dma_rx.channel_id];
-    general_dma_tx.descriptor = &descriptor_section[general_dma_tx.channel_id];
+    // Turn on the clocks
+    #ifdef SAMD51
+    MCLK->AHBMASK.reg |= MCLK_AHBMASK_DMAC;
+    #endif
+
+    #ifdef SAMD21
+    PM->AHBMASK.reg |= PM_AHBMASK_DMAC;
+    PM->APBBMASK.reg |= PM_APBBMASK_DMAC;
+    #endif
+
+    DMAC->CTRL.reg = DMAC_CTRL_SWRST;
+
+    DMAC->BASEADDR.reg = (uint32_t) dma_descriptors;
+    DMAC->WRBADDR.reg = (uint32_t) write_back_descriptors;
+
+    DMAC->CTRL.reg = DMAC_CTRL_DMAENABLE | DMAC_CTRL_LVLEN0;
+
+    // This allocates the lowest channel first so make sure the audio is first
+    // so it gets the highest priority.
+    // dma_configure_audio(0);
 }
 
 static uint8_t sercom_index(Sercom* sercom) {
+    #ifdef SAMD21
     return ((uint32_t) sercom - (uint32_t) SERCOM0) / 0x400;
+    #else
+    const Sercom* sercoms[SERCOM_INST_NUM] = SERCOM_INSTS;
+    for (uint8_t i = 0; i < SERCOM_INST_NUM; i++) {
+        if (sercoms[i] == sercom) {
+            return i;
+        }
+    }
+    return 0;
+    #endif
 }
 
-static void dma_configure(uint8_t channel, uint8_t trigsrc, bool output_event) {
-    system_interrupt_enter_critical_section();
+static void dma_configure(uint8_t channel_number, uint8_t trigsrc, bool output_event) {
+    #ifdef SAMD21
+    common_hal_mcu_disable_interrupts();
     /** Select the DMA channel and clear software trigger */
-    DMAC->CHID.reg = DMAC_CHID_ID(channel);
+    DMAC->CHID.reg = DMAC_CHID_ID(channel_number);
     DMAC->CHCTRLA.reg &= ~DMAC_CHCTRLA_ENABLE;
     DMAC->CHCTRLA.reg = DMAC_CHCTRLA_SWRST;
-    DMAC->SWTRIGCTRL.reg &= (uint32_t)(~(1 << channel));
+    DMAC->SWTRIGCTRL.reg &= (uint32_t)(~(1 << channel_number));
     uint32_t event_output_enable = 0;
     if (output_event) {
         event_output_enable = DMAC_CHCTRLB_EVOE;
     }
-    DMAC->CHCTRLB.reg = DMAC_CHCTRLB_LVL(DMA_PRIORITY_LEVEL_0) |
+    DMAC->CHCTRLB.reg = DMAC_CHCTRLB_LVL_LVL0 |
             DMAC_CHCTRLB_TRIGSRC(trigsrc) |
-            DMAC_CHCTRLB_TRIGACT(DMA_TRIGGER_ACTION_BEAT) |
+            DMAC_CHCTRLB_TRIGACT_BEAT |
             event_output_enable;
-    system_interrupt_leave_critical_section();
-}
+    common_hal_mcu_enable_interrupts();
+    #endif
 
-int32_t shared_dma_write(Sercom* sercom, const uint8_t* buffer, uint32_t length) {
-    if (general_dma_tx.job_status != STATUS_OK) {
-        return general_dma_tx.job_status;
-    }
-    dma_configure(general_dma_tx.channel_id, sercom_index(sercom) * 2 + 2, false);
-
-    // Set up TX. There is no RX job.
-    struct dma_descriptor_config descriptor_config;
-    dma_descriptor_get_config_defaults(&descriptor_config);
-    descriptor_config.beat_size = DMA_BEAT_SIZE_BYTE;
-    descriptor_config.dst_increment_enable = false;
-    descriptor_config.block_transfer_count = length;
-    descriptor_config.source_address = ((uint32_t)buffer + length);
-    // DATA register is consistently addressed across all SERCOM modes.
-    descriptor_config.destination_address = ((uint32_t)&sercom->SPI.DATA.reg);
-
-    dma_descriptor_create(general_dma_tx.descriptor, &descriptor_config);
-    enum status_code status = dma_start_transfer_job(&general_dma_tx);
-    if (status != ERR_NONE) {
-        return status;
+    #ifdef SAMD51
+    DmacChannel* channel = &DMAC->Channel[channel_number];
+    channel->CHCTRLA.reg &= ~DMAC_CHCTRLA_ENABLE;
+    channel->CHCTRLA.reg = DMAC_CHCTRLA_SWRST;
+    if (output_event) {
+        channel->CHEVCTRL.reg = DMAC_CHEVCTRL_EVOE;
     }
+    channel->CHCTRLA.reg = DMAC_CHCTRLA_TRIGSRC(trigsrc) |
+                           DMAC_CHCTRLA_TRIGACT_BURST |
+                           DMAC_CHCTRLA_BURSTLEN_SINGLE;
+    #endif
+}
 
-    // Wait for the dma transfer to finish.
-    while (general_dma_tx.job_status == STATUS_BUSY) {}
-
-    // Wait for the SPI transfer to complete.
-    while (sercom->SPI.INTFLAG.bit.TXC == 0) {}
-
-    // This transmit will cause the RX buffer overflow but we're OK with that.
-    // So, read the garbage and clear the overflow flag.
-    while (sercom->SPI.INTFLAG.bit.RXC == 1) {
-        sercom->SPI.DATA.reg;
-    }
-    sercom->SPI.STATUS.bit.BUFOVF = 1;
-    sercom->SPI.INTFLAG.reg = SERCOM_SPI_INTFLAG_ERROR;
+static void enable_channel(uint8_t channel_number) {
+    #ifdef SAMD21
+    common_hal_mcu_disable_interrupts();
+    /** Select the DMA channel and clear software trigger */
+    DMAC->CHID.reg = DMAC_CHID_ID(channel_number);
+    DMAC->CHCTRLA.bit.ENABLE = true;
+    common_hal_mcu_enable_interrupts();
+    #endif
+
+    #ifdef SAMD51
+    DmacChannel* channel = &DMAC->Channel[channel_number];
+    channel->CHCTRLA.bit.ENABLE = true;
+    #endif
+}
 
-    return general_dma_tx.job_status;
+static uint8_t transfer_status(uint8_t channel_number) {
+    #ifdef SAMD21
+    common_hal_mcu_disable_interrupts();
+    /** Select the DMA channel and clear software trigger */
+    DMAC->CHID.reg = DMAC_CHID_ID(channel_number);
+    uint8_t status = DMAC->CHINTFLAG.reg;
+    common_hal_mcu_enable_interrupts();
+    return status;
+    #endif
+
+    #ifdef SAMD51
+    DmacChannel* channel = &DMAC->Channel[channel_number];
+    return channel->CHINTFLAG.reg;
+    #endif
 }
 
-int32_t shared_dma_read(Sercom* sercom, uint8_t* buffer, uint32_t length, uint8_t tx) {
-    if (general_dma_tx.job_status != ERR_NONE) {
+static bool channel_free(uint8_t channel_number) {
+    #ifdef SAMD21
+    common_hal_mcu_disable_interrupts();
+    /** Select the DMA channel and clear software trigger */
+    DMAC->CHID.reg = DMAC_CHID_ID(channel_number);
+    bool channel_free = DMAC->CHSTATUS.reg == 0;
+    common_hal_mcu_enable_interrupts();
+    return channel_free;
+    #endif
+
+    #ifdef SAMD51
+    DmacChannel* channel = &DMAC->Channel[channel_number];
+    return channel->CHSTATUS.reg == 0;
+    #endif
 }
 
 // Do write and read simultaneously. If buffer_out is NULL, write the tx byte over and over.
 // If buffer_out is a real buffer, ignore tx.
-enum status_code shared_dma_transfer(Sercom* sercom, uint8_t* buffer_out, uint8_t* buffer_in, uint32_t length, uint8_t tx) {
-        return general_dma_tx.job_status;
+// DMAs buffer_out -> dest
+// DMAs src -> buffer_in
+static int32_t shared_dma_transfer(void* peripheral,
+                                   const uint8_t* buffer_out, volatile uint32_t* dest,
+                                   volatile uint32_t* src, uint8_t* buffer_in,
+                                   uint32_t length, uint8_t tx) {
+    if (!channel_free(SHARED_TX_CHANNEL) ||
+        (buffer_in != NULL && !channel_free(SHARED_RX_CHANNEL))) {
+        return -1;
     }
 
-    dma_configure(general_dma_tx.channel_id, sercom_index(sercom) * 2 + 2, false);
-    dma_configure(general_dma_rx.channel_id, sercom_index(sercom) * 2 + 1, false);
+    uint32_t beat_size = DMAC_BTCTRL_BEATSIZE_BYTE;
+    bool sercom = true;
+    bool tx_active = false;
+    bool rx_active = false;
+    uint16_t beat_length = length;
+    #ifdef SAMD51
+    if (peripheral == QSPI) {
+        // Check input alignment on word boundaries.
+        if ((((uint32_t) buffer_in) & 0x3) != 0 ||
+            (((uint32_t) buffer_out) & 0x3) != 0) {
+            return -3;
+        }
+        beat_size = DMAC_BTCTRL_BEATSIZE_WORD | DMAC_BTCTRL_SRCINC | DMAC_BTCTRL_DSTINC;
+        beat_length /= 4;
+        sercom = false;
+        if (buffer_out != NULL) {
+            dma_configure(SHARED_TX_CHANNEL, QSPI_DMAC_ID_TX, false);
+            tx_active = true;
+        } else {
+            dma_configure(SHARED_RX_CHANNEL, QSPI_DMAC_ID_RX, false);
+            rx_active = true;
+        }
+
+    } else {
+    #endif
+
+        // sercom index is incorrect for SAMD51
+        dma_configure(SHARED_TX_CHANNEL, sercom_index(peripheral) * 2 + FIRST_SERCOM_TX_TRIGSRC, false);
+        tx_active = true;
+        if (buffer_in != NULL) {
+            dma_configure(SHARED_RX_CHANNEL, sercom_index(peripheral) * 2 + FIRST_SERCOM_RX_TRIGSRC, false);
+            rx_active = true;
+        }
+
+    #ifdef SAMD51
+    }
+    #endif
 
     // Set up RX first.
-    struct dma_descriptor_config descriptor_config;
-    dma_descriptor_get_config_defaults(&descriptor_config);
-    descriptor_config.beat_size = DMA_BEAT_SIZE_BYTE;
-    descriptor_config.src_increment_enable = false;
-    descriptor_config.dst_increment_enable = true;
-    descriptor_config.block_transfer_count = length;
-    // DATA register is consistently addressed across all SERCOM modes.
-    descriptor_config.source_address = ((uint32_t)&sercom->SPI.DATA.reg);
-    descriptor_config.destination_address = ((uint32_t)buffer_in + length);
-
-    dma_descriptor_create(general_dma_rx.descriptor, &descriptor_config);
+    if (rx_active) {
+        DmacDescriptor* rx_descriptor = &dma_descriptors[SHARED_RX_CHANNEL];
+        rx_descriptor->BTCTRL.reg = beat_size | DMAC_BTCTRL_DSTINC;
+        rx_descriptor->BTCNT.reg = beat_length;
+        rx_descriptor->SRCADDR.reg = ((uint32_t) src);
+        #ifdef SAMD51
+        if (peripheral == QSPI) {
+            rx_descriptor->SRCADDR.reg = ((uint32_t) src + length);
+        }
+        #endif
+        rx_descriptor->DSTADDR.reg = ((uint32_t)buffer_in + length);
+        rx_descriptor->BTCTRL.bit.VALID = true;
+    }
 
     // Set up TX second.
-    dma_descriptor_get_config_defaults(&descriptor_config);
-    descriptor_config.beat_size = DMA_BEAT_SIZE_BYTE;
-    // Increment write address only if we have a real buffer.
-    descriptor_config.src_increment_enable = buffer_out != NULL;
-    descriptor_config.dst_increment_enable = false;
-    descriptor_config.block_transfer_count = length;
-    // 
-    descriptor_config.source_address = ((uint32_t) (buffer_out != NULL ? buffer_out + length : &tx));
-    // DATA register is consistently addressed across all SERCOM modes.
-    descriptor_config.destination_address = ((uint32_t)&sercom->SPI.DATA.reg);
-
-    dma_descriptor_create(general_dma_tx.descriptor, &descriptor_config);
-
+    if (tx_active) {
+        DmacDescriptor* tx_descriptor = &dma_descriptors[SHARED_TX_CHANNEL];
+        tx_descriptor->BTCTRL.reg = beat_size;
+        tx_descriptor->BTCNT.reg = beat_length;
+
+        if (buffer_out != NULL) {
+            tx_descriptor->SRCADDR.reg = ((uint32_t)buffer_out + length);
+            tx_descriptor->BTCTRL.reg |= DMAC_BTCTRL_SRCINC;
+        } else {
+            tx_descriptor->SRCADDR.reg = ((uint32_t) &tx);
+        }
+        tx_descriptor->DSTADDR.reg = ((uint32_t) dest);
+        tx_descriptor->BTCTRL.bit.VALID = true;
+    }
+    if (sercom) {
+        SercomSpi *s = &((Sercom*) peripheral)->SPI;
+        s->INTFLAG.reg = SERCOM_SPI_INTFLAG_RXC | SERCOM_SPI_INTFLAG_DRE;
+    } else {
+        //QSPI->INTFLAG.reg = QSPI_INTFLAG_RXC | QSPI_INTFLAG_DRE;
+    }
     // Start the RX job first so we don't miss the first byte. The TX job clocks
     // the output.
-    general_dma_rx.transfered_size = 0;
-    dma_start_transfer_job(&general_dma_rx);
-    general_dma_tx.transfered_size = 0;
-    dma_start_transfer_job(&general_dma_tx);
-
-    // Wait for the transfer to finish.
-    while (general_dma_rx.job_status == STATUS_BUSY) {}
+    if (rx_active) {
+        enable_channel(SHARED_RX_CHANNEL);
+    }
+    if (tx_active) {
+        enable_channel(SHARED_TX_CHANNEL);
+    }
 
-    while (sercom->SPI.INTFLAG.bit.RXC == 1) {}
-    return general_dma_rx.job_status;
-}
 
-bool allocate_block_counter() {
-    // Find a timer to count DMA block completions.
-    Tc *t = NULL;
-    Tc *tcs[TC_INST_NUM] = TC_INSTS;
-    for (uint8_t i = TC_INST_NUM; i > 0; i--) {
-        if (tcs[i - 1]->COUNT16.CTRLA.bit.ENABLE == 0) {
-            t = tcs[i - 1];
-            break;
+    if (sercom) {
+        //DMAC->SWTRIGCTRL.reg |= (1 << SHARED_TX_CHANNEL);
+    } else {
+        // Do a manual copy to trigger then DMA. We do 32-bit accesses to match the DMA.
+        #pragma GCC diagnostic push
+        #pragma GCC diagnostic ignored "-Wcast-align"
+        if (rx_active) {
+            //buffer_in[0] = *src;
+            DMAC->SWTRIGCTRL.reg |= (1 << SHARED_RX_CHANNEL);
+        } else {
+            //*(uint32_t*)dest = ((uint32_t*) buffer_out)[0];
         }
+        #pragma GCC diagnostic pop
     }
-    if (t == NULL) {
-        return false;
+
+    // Channels cycle between Suspend -> Pending -> Busy and back while transfering. So, we check
+    // the channels transfer status for an error or completion.
+    if (rx_active) {
+        while ((transfer_status(SHARED_RX_CHANNEL) & 0x3) == 0) {}
     }
-    MP_STATE_VM(audiodma_block_counter) = gc_alloc(sizeof(struct tc_module), false);
-    if (MP_STATE_VM(audiodma_block_counter) == NULL) {
-        return false;
+    if (tx_active) {
+        while ((transfer_status(SHARED_TX_CHANNEL) & 0x3) == 0) {}
     }
 
-    // Don't bother setting the period. We set it before you playback anything.
-    struct tc_config config_tc;
-    tc_get_config_defaults(&config_tc);
-    config_tc.counter_size    = TC_COUNTER_SIZE_16BIT;
-    config_tc.clock_prescaler = TC_CLOCK_PRESCALER_DIV1;
-    if (tc_init(MP_STATE_VM(audiodma_block_counter), t, &config_tc) != STATUS_OK) {
-        return false;
-    };
-
-    struct tc_events events_tc;
-    events_tc.generate_event_on_overflow = false;
-    events_tc.on_event_perform_action = true;
-    events_tc.event_action = TC_EVENT_ACTION_INCREMENT_COUNTER;
-    tc_enable_events(MP_STATE_VM(audiodma_block_counter), &events_tc);
-
-    // Connect the timer overflow event, which happens at the target frequency,
-    // to the DAC conversion trigger.
-    MP_STATE_VM(audiodma_block_event) = gc_alloc(sizeof(struct events_resource), false);
-    if (MP_STATE_VM(audiodma_block_event) == NULL) {
-        return false;
-    }
-    struct events_config config;
-    events_get_config_defaults(&config);
-
-    uint8_t user = EVSYS_ID_USER_TC3_EVU;
-    if (t == TC4) {
-        user = EVSYS_ID_USER_TC4_EVU;
-    } else if (t == TC5) {
-        user = EVSYS_ID_USER_TC5_EVU;
-#ifdef TC6
-    } else if (t == TC6) {
-        user = EVSYS_ID_USER_TC6_EVU;
-#endif
-#ifdef TC7
-    } else if (t == TC7) {
-        user = EVSYS_ID_USER_TC7_EVU;
-#endif
+    if (sercom) {
+        Sercom* s = (Sercom*) peripheral;
+        // Wait for the SPI transfer to complete.
+        while (s->SPI.INTFLAG.bit.TXC == 0) {}
+
+        // This transmit will cause the RX buffer overflow but we're OK with that.
+        // So, read the garbage and clear the overflow flag.
+        if (!rx_active) {
+            while (s->SPI.INTFLAG.bit.RXC == 1) {
+                s->SPI.DATA.reg;
+            }
+            s->SPI.STATUS.bit.BUFOVF = 1;
+            s->SPI.INTFLAG.reg = SERCOM_SPI_INTFLAG_ERROR;
+        }
     }
 
-    config.generator    = EVSYS_ID_GEN_DMAC_CH_0;
-    config.path         = EVENTS_PATH_ASYNCHRONOUS;
-    if (events_allocate(MP_STATE_VM(audiodma_block_event), &config) != STATUS_OK ||
-        events_attach_user(MP_STATE_VM(audiodma_block_event), user) != STATUS_OK) {
-        return false;
+    if ((!rx_active || transfer_status(SHARED_RX_CHANNEL) == DMAC_CHINTFLAG_TCMPL) &&
+        (!tx_active || transfer_status(SHARED_TX_CHANNEL) == DMAC_CHINTFLAG_TCMPL)) {
+        return length;
     }
+    return -2;
+}
+
+
+int32_t sercom_dma_transfer(Sercom* sercom, const uint8_t* buffer_out, uint8_t* buffer_in,
+                            uint32_t length) {
+    return shared_dma_transfer(sercom, buffer_out, &sercom->SPI.DATA.reg, &sercom->SPI.DATA.reg, buffer_in, length, 0);
+}
+
+int32_t sercom_dma_write(Sercom* sercom, const uint8_t* buffer, uint32_t length) {
+    return shared_dma_transfer(sercom, buffer, &sercom->SPI.DATA.reg, NULL, NULL, length, 0);
+}
 
-    tc_enable(MP_STATE_VM(audiodma_block_counter));
-    tc_stop_counter(MP_STATE_VM(audiodma_block_counter));
+int32_t sercom_dma_read(Sercom* sercom, uint8_t* buffer, uint32_t length, uint8_t tx) {
+    return shared_dma_transfer(sercom, NULL, &sercom->SPI.DATA.reg, &sercom->SPI.DATA.reg, buffer, length, tx);
+}
+
+#ifdef SAMD51
+int32_t qspi_dma_write(uint32_t address, const uint8_t* buffer, uint32_t length) {
+    return shared_dma_transfer(QSPI, buffer, (uint32_t*) (QSPI_AHB + address), NULL, NULL, length, 0);
+}
+
+int32_t qspi_dma_read(uint32_t address, uint8_t* buffer, uint32_t length) {
+    return shared_dma_transfer(QSPI, NULL, NULL, (uint32_t*) (QSPI_AHB + address), buffer, length, 0);
+}
+#endif
+
+bool allocate_block_counter() {
+//     // Find a timer to count DMA block completions.
+//     Tc *t = NULL;
+//     Tc *tcs[TC_INST_NUM] = TC_INSTS;
+//     for (uint8_t i = TC_INST_NUM; i > 0; i--) {
+//         if (tcs[i - 1]->COUNT16.CTRLA.bit.ENABLE == 0) {
+//             t = tcs[i - 1];
+//             break;
+//         }
+//     }
+//     if (t == NULL) {
+//         return false;
+//     }
+//     MP_STATE_VM(audiodma_block_counter) = gc_alloc(sizeof(struct tc_module), false);
+//     if (MP_STATE_VM(audiodma_block_counter) == NULL) {
+//         return false;
+//     }
+//
+//     // Don't bother setting the period. We set it before you playback anything.
+//     struct tc_config config_tc;
+//     tc_get_config_defaults(&config_tc);
+//     config_tc.counter_size    = TC_COUNTER_SIZE_16BIT;
+//     config_tc.clock_prescaler = TC_CLOCK_PRESCALER_DIV1;
+//     if (tc_init(MP_STATE_VM(audiodma_block_counter), t, &config_tc) != STATUS_OK) {
+//         return false;
+//     };
+//
+//     struct tc_events events_tc;
+//     events_tc.generate_event_on_overflow = false;
+//     events_tc.on_event_perform_action = true;
+//     events_tc.event_action = TC_EVENT_ACTION_INCREMENT_COUNTER;
+//     tc_enable_events(MP_STATE_VM(audiodma_block_counter), &events_tc);
+//
+//     // Connect the timer overflow event, which happens at the target frequency,
+//     // to the DAC conversion trigger.
+//     MP_STATE_VM(audiodma_block_event) = gc_alloc(sizeof(struct events_resource), false);
+//     if (MP_STATE_VM(audiodma_block_event) == NULL) {
+//         return false;
+//     }
+//     struct events_config config;
+//     events_get_config_defaults(&config);
+//
+//     uint8_t user = EVSYS_ID_USER_TC3_EVU;
+//     if (t == TC4) {
+//         user = EVSYS_ID_USER_TC4_EVU;
+//     } else if (t == TC5) {
+//         user = EVSYS_ID_USER_TC5_EVU;
+// #ifdef TC6
+//     } else if (t == TC6) {
+//         user = EVSYS_ID_USER_TC6_EVU;
+// #endif
+// #ifdef TC7
+//     } else if (t == TC7) {
+//         user = EVSYS_ID_USER_TC7_EVU;
+// #endif
+//     }
+//
+//     config.generator    = EVSYS_ID_GEN_DMAC_CH_0;
+//     config.path         = EVENTS_PATH_ASYNCHRONOUS;
+//     if (events_allocate(MP_STATE_VM(audiodma_block_event), &config) != STATUS_OK ||
+//         events_attach_user(MP_STATE_VM(audiodma_block_event), user) != STATUS_OK) {
+//         return false;
+//     }
+//
+//     tc_enable(MP_STATE_VM(audiodma_block_counter));
+//     tc_stop_counter(MP_STATE_VM(audiodma_block_counter));
     return true;
 }
 
 void switch_audiodma_trigger(uint8_t trigger_dmac_id) {
-    dma_configure(audio_dma.channel_id, trigger_dmac_id, true);
+    //dma_configure(audio_dma.channel_id, trigger_dmac_id, true);
 }
diff --git a/ports/atmel-samd/shared_dma.h b/ports/atmel-samd/shared_dma.h
index 2ddd81c4e..536a95af3 100644
--- a/ports/atmel-samd/shared_dma.h
+++ b/ports/atmel-samd/shared_dma.h
@@ -27,17 +27,23 @@
 #ifndef MICROPY_INCLUDED_ATMEL_SAMD_SHARED_DMA_H
 #define MICROPY_INCLUDED_ATMEL_SAMD_SHARED_DMA_H
 
-extern struct dma_resource audio_dma;
-extern struct dma_resource general_dma_tx;
-extern struct dma_resource general_dma_rx;
+#include <stdbool.h>
+#include <stdint.h>
+
+#include "include/sam.h"
 
 volatile bool audio_dma_in_use;
 
 void init_shared_dma(void);
 
-enum status_code shared_dma_write(Sercom* sercom, const uint8_t* buffer, uint32_t length);
-enum status_code shared_dma_read(Sercom* sercom, uint8_t* buffer, uint32_t length, uint8_t tx);
-enum status_code shared_dma_transfer(Sercom* sercom, uint8_t* buffer_out, uint8_t* buffer_in, uint32_t length, uint8_t tx);
+#ifdef SAMD51
+int32_t qspi_dma_write(uint32_t address, const uint8_t* buffer, uint32_t length);
+int32_t qspi_dma_read(uint32_t address, uint8_t* buffer, uint32_t length);
+#endif
+
+int32_t sercom_dma_write(Sercom* sercom, const uint8_t* buffer, uint32_t length);
+int32_t sercom_dma_read(Sercom* sercom, uint8_t* buffer, uint32_t length, uint8_t tx);
+int32_t sercom_dma_transfer(Sercom* sercom, const uint8_t* buffer_out, uint8_t* buffer_in, uint32_t length);
 
 // Allocate a counter to track how far along we are in a DMA double buffer.
 bool allocate_block_counter(void);
diff --git a/ports/atmel-samd/supervisor/port.c b/ports/atmel-samd/supervisor/port.c
index fe6591510..7935c87d3 100644
--- a/ports/atmel-samd/supervisor/port.c
+++ b/ports/atmel-samd/supervisor/port.c
@@ -49,6 +49,7 @@
 #include "common-hal/pulseio/PulseIn.h"
 #include "common-hal/pulseio/PulseOut.h"
 #include "common-hal/pulseio/PWMOut.h"
+#include "shared_dma.h"
 #include "tick.h"
 
 extern volatile bool mp_msc_enabled;
@@ -120,7 +121,7 @@ safe_mode_t port_init(void) {
     // config_nvm.manual_page_write = false;
     // nvm_set_config(&config_nvm);
 
-    // init_shared_dma();
+    init_shared_dma();
     #ifdef CIRCUITPY_CANARY_WORD
     // Run in safe mode if the canary is corrupt.
     if (_ezero != CIRCUITPY_CANARY_WORD) {
@@ -206,7 +207,7 @@ void reset_port(void) {
     reset_all_pins();
 
     // Set up debugging pins after reset_all_pins().
-    
+
     // Uncomment to init PIN_PA17 for debugging.
     // struct port_config pin_conf;
     // port_get_config_defaults(&pin_conf);