improve SAMD audio DMA

This commit is contained in:
Dan Halbert 2021-08-20 20:30:13 -04:00
parent 1825fc86d3
commit b8b23c97d9
6 changed files with 126 additions and 80 deletions

View File

@ -24,6 +24,8 @@
* THE SOFTWARE.
*/
#include <string.h>
#include "audio_dma.h"
#include "samd/clocks.h"
#include "samd/events.h"
@ -31,6 +33,7 @@
#include "shared-bindings/audiocore/RawSample.h"
#include "shared-bindings/audiocore/WaveFile.h"
#include "shared-bindings/microcontroller/__init__.h"
#include "supervisor/background_callback.h"
#include "py/mpstate.h"
@ -38,6 +41,10 @@
#if CIRCUITPY_AUDIOIO || CIRCUITPY_AUDIOBUSIO
// Flag value for dma->buffer_to_load, indicating there is nothing to do.
// Otherwise dma->buffer_to_load is 0 or 1.
#define NO_BUFFER_TO_LOAD 0xff
static audio_dma_t *audio_dma_state[AUDIO_DMA_CHANNEL_COUNT];
// This cannot be in audio_dma_state because it's volatile.
@ -85,70 +92,79 @@ void audio_dma_enable_channel(uint8_t channel) {
dma_enable_channel(channel);
}
void audio_dma_convert_signed(audio_dma_t *dma, uint8_t *buffer, uint32_t buffer_length,
uint8_t **output_buffer, uint32_t *output_buffer_length,
static void audio_dma_convert_samples(
audio_dma_t *dma,
uint8_t *input, uint32_t input_length,
uint8_t *available_output_buffer, uint32_t available_output_buffer_length,
uint8_t **output, uint32_t *output_length,
uint8_t *output_spacing) {
if (dma->first_buffer_free) {
*output_buffer = dma->first_buffer;
} else {
*output_buffer = dma->second_buffer;
}
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wcast-align"
if (dma->signed_to_unsigned || dma->unsigned_to_signed) {
*output_buffer_length = buffer_length / dma->spacing;
// Must convert.
// Write the conversion into the passed-in output buffer
*output = available_output_buffer;
*output_length = input_length / dma->spacing;
*output_spacing = 1;
if (*output_length > available_output_buffer_length) {
mp_raise_RuntimeError(translate("Internal audio buffer too small"));
}
uint32_t out_i = 0;
if (dma->bytes_per_sample == 1) {
for (uint32_t i = 0; i < buffer_length; i += dma->spacing) {
for (uint32_t i = 0; i < input_length; i += dma->spacing) {
if (dma->signed_to_unsigned) {
((uint8_t *)*output_buffer)[out_i] = ((int8_t *)buffer)[i] + 0x80;
((uint8_t *)*output)[out_i] = ((int8_t *)input)[i] + 0x80;
} else {
((int8_t *)*output_buffer)[out_i] = ((uint8_t *)buffer)[i] - 0x80;
((int8_t *)*output)[out_i] = ((uint8_t *)input)[i] - 0x80;
}
out_i += 1;
}
} else if (dma->bytes_per_sample == 2) {
for (uint32_t i = 0; i < buffer_length / 2; i += dma->spacing) {
for (uint32_t i = 0; i < input_length / 2; i += dma->spacing) {
if (dma->signed_to_unsigned) {
((uint16_t *)*output_buffer)[out_i] = ((int16_t *)buffer)[i] + 0x8000;
((uint16_t *)*output)[out_i] = ((int16_t *)input)[i] + 0x8000;
} else {
((int16_t *)*output_buffer)[out_i] = ((uint16_t *)buffer)[i] - 0x8000;
((int16_t *)*output)[out_i] = ((uint16_t *)input)[i] - 0x8000;
}
out_i += 1;
}
}
} else {
*output_buffer = buffer;
*output_buffer_length = buffer_length;
*output = input;
*output_length = input_length;
*output_spacing = dma->spacing;
}
#pragma GCC diagnostic pop
dma->first_buffer_free = !dma->first_buffer_free;
}
void audio_dma_load_next_block(audio_dma_t *dma) {
uint8_t *buffer;
uint32_t buffer_length;
static void audio_dma_load_next_block(audio_dma_t *dma, size_t buffer_idx) {
uint8_t *sample_buffer;
uint32_t sample_buffer_length;
audioio_get_buffer_result_t get_buffer_result =
audiosample_get_buffer(dma->sample, dma->single_channel_output, dma->audio_channel,
&buffer, &buffer_length);
&sample_buffer, &sample_buffer_length);
DmacDescriptor *descriptor = dma->second_descriptor;
if (dma->first_descriptor_free) {
descriptor = dma_descriptor(dma->dma_channel);
}
dma->first_descriptor_free = !dma->first_descriptor_free;
DmacDescriptor *descriptor = dma->descriptor[buffer_idx];
if (get_buffer_result == GET_BUFFER_ERROR) {
audio_dma_stop(dma);
return;
}
// Use one of the allocated buffers for conversion. But if there's no conversion,
// this will be set to buffer in audio_dma_convert_samples() to avoid any copying.
uint8_t *output_buffer;
uint32_t output_buffer_length;
uint8_t output_spacing;
audio_dma_convert_signed(dma, buffer, buffer_length, &output_buffer, &output_buffer_length,
audio_dma_convert_samples(dma, sample_buffer, sample_buffer_length,
// Available output buffer: may be used or not.
dma->buffer[buffer_idx], dma->buffer_length[buffer_idx],
// Buffer where output was placed.
&output_buffer, &output_buffer_length,
&output_spacing);
descriptor->BTCNT.reg = output_buffer_length / dma->beat_size / output_spacing;
@ -157,9 +173,10 @@ void audio_dma_load_next_block(audio_dma_t *dma) {
if (dma->loop) {
audiosample_reset_buffer(dma->sample, dma->single_channel_output, dma->audio_channel);
} else {
if ((output_buffer_length == 0) && dma_transfer_status(SHARED_RX_CHANNEL) & 0x3) {
if (output_buffer_length == 0) {
// Nothing further to read and previous buffer is finished.
audio_dma_stop(dma);
return;
} else {
// Break descriptor chain.
descriptor->DESCADDR.reg = 0;
@ -206,15 +223,15 @@ audio_dma_result audio_dma_setup_playback(audio_dma_t *dma,
dma->dma_channel = dma_channel;
dma->signed_to_unsigned = false;
dma->unsigned_to_signed = false;
dma->second_descriptor = NULL;
dma->spacing = 1;
dma->first_descriptor_free = true;
audiosample_reset_buffer(sample, single_channel_output, audio_channel);
dma->buffer_to_load = NO_BUFFER_TO_LOAD;
dma->descriptor[0] = dma_descriptor(dma_channel);
dma->descriptor[1] = &dma->second_descriptor;
bool single_buffer;
bool samples_signed;
uint32_t max_buffer_length;
audiosample_get_buffer_structure(sample, single_channel_output, &single_buffer, &samples_signed,
audiosample_get_buffer_structure(sample, single_channel_output, &dma->single_buffer, &samples_signed,
&max_buffer_length, &dma->spacing);
uint8_t output_spacing = dma->spacing;
if (output_signed != samples_signed) {
@ -222,14 +239,17 @@ audio_dma_result audio_dma_setup_playback(audio_dma_t *dma,
max_buffer_length /= dma->spacing;
}
dma->first_buffer = (uint8_t *)m_realloc(dma->first_buffer, max_buffer_length);
if (dma->first_buffer == NULL) {
dma->buffer[0] = (uint8_t *)m_realloc(dma->buffer[0], max_buffer_length);
dma->buffer_length[0] = max_buffer_length;
if (dma->buffer[0] == NULL) {
return AUDIO_DMA_MEMORY_ERROR;
}
dma->first_buffer_free = true;
if (!single_buffer) {
dma->second_buffer = (uint8_t *)m_realloc(dma->second_buffer, max_buffer_length);
if (dma->second_buffer == NULL) {
if (!dma->single_buffer) {
dma->buffer[1] = (uint8_t *)m_realloc(dma->buffer[1], max_buffer_length);
dma->buffer_length[1] = max_buffer_length;
if (dma->buffer[1] == NULL) {
return AUDIO_DMA_MEMORY_ERROR;
}
}
@ -238,12 +258,7 @@ audio_dma_result audio_dma_setup_playback(audio_dma_t *dma,
dma->unsigned_to_signed = output_signed && !samples_signed;
dma->event_channel = 0xff;
if (!single_buffer) {
dma->second_descriptor = (DmacDescriptor *)m_malloc(sizeof(DmacDescriptor), false);
if (dma->second_descriptor == NULL) {
return AUDIO_DMA_MEMORY_ERROR;
}
if (!dma->single_buffer) {
// We're likely double buffering so set up the block interrupts.
turn_on_event_system();
dma->event_channel = find_sync_event_channel_raise();
@ -280,25 +295,27 @@ audio_dma_result audio_dma_setup_playback(audio_dma_t *dma,
int irq = EVSYS_IRQn;
#endif
DmacDescriptor *first_descriptor = dma_descriptor(dma_channel);
setup_audio_descriptor(first_descriptor, dma->beat_size, output_spacing, output_register_address);
if (single_buffer) {
first_descriptor->DESCADDR.reg = 0;
setup_audio_descriptor(dma->descriptor[0], dma->beat_size, output_spacing, output_register_address);
if (dma->single_buffer) {
dma->descriptor[0]->DESCADDR.reg = 0;
if (dma->loop) {
first_descriptor->DESCADDR.reg = (uint32_t)first_descriptor;
// The descriptor chains to itself.
dma->descriptor[0]->DESCADDR.reg = (uint32_t)dma->descriptor[0];
}
} else {
first_descriptor->DESCADDR.reg = (uint32_t)dma->second_descriptor;
setup_audio_descriptor(dma->second_descriptor, dma->beat_size, output_spacing, output_register_address);
dma->second_descriptor->DESCADDR.reg = (uint32_t)first_descriptor;
// Set up the two descriptors to chain to each other.
dma->descriptor[0]->DESCADDR.reg = (uint32_t)dma->descriptor[1];
setup_audio_descriptor(dma->descriptor[1], dma->beat_size, output_spacing, output_register_address);
dma->descriptor[1]->DESCADDR.reg = (uint32_t)dma->descriptor[0];
}
// Load the first two blocks up front.
audio_dma_load_next_block(dma);
if (!single_buffer) {
audio_dma_load_next_block(dma);
audio_dma_load_next_block(dma, 0);
if (!dma->single_buffer) {
audio_dma_load_next_block(dma, 1);
}
dma->playing_in_progress = true;
dma_configure(dma_channel, dma_trigger_source, true);
audio_dma_enable_channel(dma_channel);
@ -317,6 +334,7 @@ void audio_dma_stop(audio_dma_t *dma) {
dma_free_channel(dma->dma_channel);
}
dma->dma_channel = AUDIO_DMA_CHANNEL_COUNT;
dma->playing_in_progress = false;
}
void audio_dma_pause(audio_dma_t *dma) {
@ -355,12 +373,7 @@ bool audio_dma_get_playing(audio_dma_t *dma) {
if (dma->dma_channel >= AUDIO_DMA_CHANNEL_COUNT) {
return false;
}
uint32_t status = dma_transfer_status(dma->dma_channel);
if ((status & DMAC_CHINTFLAG_TCMPL) != 0 || (status & DMAC_CHINTFLAG_TERR) != 0) {
audio_dma_stop(dma);
}
return (status & DMAC_CHINTFLAG_TERR) == 0;
return dma->playing_in_progress;
}
// WARN(tannewt): DO NOT print from here, or anything it calls. Printing calls
@ -371,7 +384,16 @@ STATIC void dma_callback_fun(void *arg) {
return;
}
audio_dma_load_next_block(dma);
common_hal_mcu_disable_interrupts();
uint8_t buffer_to_load = dma->buffer_to_load;
dma->buffer_to_load = NO_BUFFER_TO_LOAD;
common_hal_mcu_enable_interrupts();
if (buffer_to_load == NO_BUFFER_TO_LOAD) {
audio_dma_stop(dma);
} else {
audio_dma_load_next_block(dma, buffer_to_load);
}
}
void audio_evsys_handler(void) {
@ -384,6 +406,28 @@ void audio_evsys_handler(void) {
if (!block_done) {
continue;
}
// By the time we get here, the write-back descriptor has been set to the
// current running descriptor. Fill the buffer that the next chained descriptor
// will play.
//
// The state of the write-back descriptor was determined empirically,
// The datasheet appears to imply that the descriptor that just finished would
// be in the write-back descriptor. But the VALID bit is set in the write-back descriptor,
// and reversing which buffer to fill produces crackly output. So the choice
// of which buffer to fill here appears correct.
DmacDescriptor *next_descriptor =
(DmacDescriptor *)dma_write_back_descriptor(dma->dma_channel)->DESCADDR.reg;
if (next_descriptor == dma->descriptor[0]) {
dma->buffer_to_load = 0;
} else if (next_descriptor == dma->descriptor[1]) {
dma->buffer_to_load = 1;
} else if (next_descriptor == NULL) {
dma->buffer_to_load = NO_BUFFER_TO_LOAD;
} else {
continue;
}
background_callback_add(&dma->callback, dma_callback_fun, (void *)dma);
}
}

View File

@ -35,22 +35,24 @@
typedef struct {
mp_obj_t sample;
uint8_t *buffer[2];
size_t buffer_length[2];
DmacDescriptor *descriptor[2];
DmacDescriptor second_descriptor;
background_callback_t callback;
uint8_t dma_channel;
uint8_t event_channel;
uint8_t audio_channel;
uint8_t bytes_per_sample;
uint8_t beat_size;
uint8_t spacing;
uint8_t buffer_to_load; // Index
bool loop;
bool single_buffer;
bool single_channel_output;
bool signed_to_unsigned;
bool unsigned_to_signed;
bool first_buffer_free;
uint8_t *first_buffer;
uint8_t *second_buffer;
bool first_descriptor_free;
DmacDescriptor *second_descriptor;
background_callback_t callback;
bool playing_in_progress;
} audio_dma_t;
typedef enum {

View File

@ -248,7 +248,7 @@ void common_hal_audioio_audioout_construct(audioio_audioout_obj_t *self,
}
self->tc_index = tc_index;
// Use the 48mhz clocks on both the SAMD21 and 51 because we will be going much slower.
// Use the 48MHz clocks on both the SAMD21 and 51 because we will be going much slower.
uint8_t tc_gclk = 0;
#ifdef SAM_D5X_E5X
tc_gclk = 1;
@ -469,8 +469,8 @@ bool common_hal_audioio_audioout_get_paused(audioio_audioout_obj_t *self) {
}
void common_hal_audioio_audioout_stop(audioio_audioout_obj_t *self) {
Tc *timer = tc_insts[self->tc_index];
timer->COUNT16.CTRLBSET.reg = TC_CTRLBSET_CMD_STOP;
// Do not stop the timer here. There are occasional audible artifacts if the DMA-triggering timer
// is stopped between audio plays. (Heard this only on PyPortal with one particular 32kHz sample.)
audio_dma_stop(&self->left_dma);
#ifdef SAM_D5X_E5X
audio_dma_stop(&self->right_dma);

@ -1 +1 @@
Subproject commit a7e39c4d01aa5916015beecb021777617e77b0ad
Subproject commit d3b20192cf94fdea68a412596e082108ba5ebbf0

View File

@ -137,7 +137,7 @@ STATIC void audio_dma_convert_samples(
#pragma GCC diagnostic pop
}
// channel_idx is 0 or 1.
// buffer_idx is 0 or 1.
STATIC void audio_dma_load_next_block(audio_dma_t *dma, size_t buffer_idx) {
size_t dma_channel = dma->channel[buffer_idx];

View File

@ -34,23 +34,23 @@
typedef struct {
mp_obj_t sample;
uint8_t *buffer[2];
size_t buffer_length[2];
uint32_t channels_to_load_mask;
uint32_t output_register_address;
background_callback_t callback;
uint8_t channel[2];
uint8_t audio_channel;
uint8_t output_size;
uint8_t sample_spacing;
uint8_t output_resolution; // in bits
uint8_t sample_resolution; // in bits
bool loop;
bool single_channel_output;
bool signed_to_unsigned;
bool unsigned_to_signed;
bool output_signed;
bool playing_in_progress;
uint8_t output_resolution; // in bits
uint8_t sample_resolution; // in bits
uint8_t *buffer[2];
size_t buffer_length[2];
uint32_t channels_to_load_mask;
uint32_t output_register_address;
background_callback_t callback;
} audio_dma_t;
typedef enum {