fix rp2040_pio_led flicker and proper multi-strip support (#6194)

Co-authored-by: Jesse Hills <3060199+jesserockz@users.noreply.github.com>

esphome/components/rp2040_pio_led_strip/led_strip.cpp

@@ -6,6 +6,7 @@
 #include "esphome/core/log.h"
 
 #include <hardware/clocks.h>
+#include <hardware/dma.h>
 #include <hardware/pio.h>
 #include <pico/stdlib.h>
 
@@ -14,6 +15,15 @@ namespace rp2040_pio_led_strip {
 
 static const char *TAG = "rp2040_pio_led_strip";
 
+static uint8_t num_instance_[2] = {0, 0};
+static std::map<Chipset, uint> chipset_offsets_ = {
+    {CHIPSET_WS2812, 0}, {CHIPSET_WS2812B, 0}, {CHIPSET_SK6812, 0}, {CHIPSET_SM16703, 0}, {CHIPSET_CUSTOM, 0},
+};
+static std::map<Chipset, bool> conf_count_ = {
+    {CHIPSET_WS2812, false},  {CHIPSET_WS2812B, false}, {CHIPSET_SK6812, false},
+    {CHIPSET_SM16703, false}, {CHIPSET_CUSTOM, false},
+};
+
 void RP2040PIOLEDStripLightOutput::setup() {
   ESP_LOGCONFIG(TAG, "Setting up RP2040 LED Strip...");
 
@@ -34,24 +44,71 @@ void RP2040PIOLEDStripLightOutput::setup() {
     return;
   }
 
+  // Initialize the PIO program
+
   // Select PIO instance to use (0 or 1)
-  this->pio_ = pio0;
   if (this->pio_ == nullptr) {
     ESP_LOGE(TAG, "Failed to claim PIO instance");
     this->mark_failed();
     return;
   }
 
-  // Load the assembled program into the PIO and get its location in the PIO's instruction memory
+  // if there are multiple strips, we can reuse the same PIO program and save space
-  uint offset = pio_add_program(this->pio_, this->program_);
+  // but there are only 4 state machines on each PIO so we can only have 4 strips per PIO
+  uint offset = 0;
+
+  if (num_instance_[this->pio_ == pio0 ? 0 : 1] > 4) {
+    ESP_LOGE(TAG, "Too many instances of PIO program");
+    this->mark_failed();
+    return;
+  }
+  // keep track of how many instances of the PIO program are running on each PIO
+  num_instance_[this->pio_ == pio0 ? 0 : 1]++;
+
+  // if there are multiple strips of the same chipset, we can reuse the same PIO program and save space
+  if (this->conf_count_[this->chipset_]) {
+    offset = chipset_offsets_[this->chipset_];
+  } else {
+    // Load the assembled program into the PIO and get its location in the PIO's instruction memory and save it
+    offset = pio_add_program(this->pio_, this->program_);
+    chipset_offsets_[this->chipset_] = offset;
+    conf_count_[this->chipset_] = true;
+  }
 
   // Configure the state machine's PIO, and start it
   this->sm_ = pio_claim_unused_sm(this->pio_, true);
   if (this->sm_ < 0) {
+    // in theory this code should never be reached
     ESP_LOGE(TAG, "Failed to claim PIO state machine");
     this->mark_failed();
     return;
   }
+
+  // Initalize the DMA channel (Note: There are 12 DMA channels and 8 state machines so we won't run out)
+
+  this->dma_chan_ = dma_claim_unused_channel(true);
+  if (this->dma_chan_ < 0) {
+    ESP_LOGE(TAG, "Failed to claim DMA channel");
+    this->mark_failed();
+    return;
+  }
+
+  this->dma_config_ = dma_channel_get_default_config(this->dma_chan_);
+  channel_config_set_transfer_data_size(
+      &this->dma_config_,
+      DMA_SIZE_8);  // 8 bit transfers (could be 32 but the pio program would need to be changed to handle junk data)
+  channel_config_set_read_increment(&this->dma_config_, true);    // increment the read address
+  channel_config_set_write_increment(&this->dma_config_, false);  // don't increment the write address
+  channel_config_set_dreq(&this->dma_config_,
+                          pio_get_dreq(this->pio_, this->sm_, true));  // set the DREQ to the state machine's TX FIFO
+
+  dma_channel_configure(this->dma_chan_, &this->dma_config_,
+                        &this->pio_->txf[this->sm_],                     // write to the state machine's TX FIFO
+                        this->buf_,                                      // read from memory
+                        this->is_rgbw_ ? num_leds_ * 4 : num_leds_ * 3,  // number of bytes to transfer
+                        false                                            // don't start yet
+  );
+
   this->init_(this->pio_, this->sm_, offset, this->pin_, this->max_refresh_rate_);
 }
 
@@ -68,16 +125,8 @@ void RP2040PIOLEDStripLightOutput::write_state(light::LightState *state) {
     return;
   }
 
-  // assemble bits in buffer to 32 bit words with ex for GBR: 0bGGGGGGGGRRRRRRRRBBBBBBBB00000000
+  // the bits are already in the correct order for the pio program so we can just copy the buffer using DMA
-  for (int i = 0; i < this->num_leds_; i++) {
+  dma_channel_transfer_from_buffer_now(this->dma_chan_, this->buf_, this->get_buffer_size_());
-    uint8_t multiplier = this->is_rgbw_ ? 4 : 3;
-    uint8_t c1 = this->buf_[(i * multiplier) + 0];
-    uint8_t c2 = this->buf_[(i * multiplier) + 1];
-    uint8_t c3 = this->buf_[(i * multiplier) + 2];
-    uint8_t w = this->is_rgbw_ ? this->buf_[(i * 4) + 3] : 0;
-    uint32_t color = encode_uint32(c1, c2, c3, w);
-    pio_sm_put_blocking(this->pio_, this->sm_, color);
-  }
 }
 
 light::ESPColorView RP2040PIOLEDStripLightOutput::get_view_internal(int32_t index) const {

esphome/components/rp2040_pio_led_strip/led_strip.h

@@ -9,9 +9,11 @@
 #include "esphome/components/light/addressable_light.h"
 #include "esphome/components/light/light_output.h"
 
+#include <hardware/dma.h>
 #include <hardware/pio.h>
 #include <hardware/structs/pio.h>
 #include <pico/stdio.h>
+#include <map>
 
 namespace esphome {
 namespace rp2040_pio_led_strip {
@@ -25,6 +27,15 @@ enum RGBOrder : uint8_t {
   ORDER_BRG,
 };
 
+enum Chipset : uint8_t {
+  CHIPSET_WS2812,
+  CHIPSET_WS2812B,
+  CHIPSET_SK6812,
+  CHIPSET_SM16703,
+  CHIPSET_APA102,
+  CHIPSET_CUSTOM = 0xFF,
+};
+
 inline const char *rgb_order_to_string(RGBOrder order) {
   switch (order) {
     case ORDER_RGB:
@@ -69,6 +80,7 @@ class RP2040PIOLEDStripLightOutput : public light::AddressableLight {
   void set_program(const pio_program_t *program) { this->program_ = program; }
   void set_init_function(init_fn init) { this->init_ = init; }
 
+  void set_chipset(Chipset chipset) { this->chipset_ = chipset; };
   void set_rgb_order(RGBOrder rgb_order) { this->rgb_order_ = rgb_order; }
   void clear_effect_data() override {
     for (int i = 0; i < this->size(); i++) {
@@ -92,14 +104,22 @@ class RP2040PIOLEDStripLightOutput : public light::AddressableLight {
 
   pio_hw_t *pio_;
   uint sm_;
+  uint dma_chan_;
+  dma_channel_config dma_config_;
 
   RGBOrder rgb_order_{ORDER_RGB};
+  Chipset chipset_{CHIPSET_CUSTOM};
 
   uint32_t last_refresh_{0};
   float max_refresh_rate_;
 
   const pio_program_t *program_;
   init_fn init_;
+
+ private:
+  inline static int num_instance_[2];
+  inline static std::map<Chipset, bool> conf_count_;
+  inline static std::map<Chipset, int> chipset_offsets_;
 };
 
 }  // namespace rp2040_pio_led_strip

esphome/components/rp2040_pio_led_strip/light.py

@@ -68,12 +68,15 @@ static inline void rp2040_pio_led_strip_driver_{id}_init(PIO pio, uint sm, uint
 
     pio_sm_config c = rp2040_pio_led_strip_{id}_program_get_default_config(offset);
     sm_config_set_set_pins(&c, pin, 1);
-    sm_config_set_out_shift(&c, false, true, {32 if rgbw else 24});
+    sm_config_set_out_shift(&c, false, true, 8);
     sm_config_set_fifo_join(&c, PIO_FIFO_JOIN_TX);
 
-    int cycles_per_bit = 69;
+    // target frequency is 57.5MHz
-    float div = 2.409;
+    long clk = clock_get_hz(clk_sys);
-    sm_config_set_clkdiv(&c, div);
+    long target_freq = 57500000;
+    int n = 2;
+    int f = round(((clk / target_freq) - n ) * 256);
+    sm_config_set_clkdiv_int_frac(&c, n, f);
 
 
     pio_sm_init(pio, sm, offset, &c);
@@ -86,8 +89,9 @@ static inline void rp2040_pio_led_strip_driver_{id}_init(PIO pio, uint sm, uint
 .wrap_target
 awaiting_data:
     ; Wait for data in FIFO queue
+    ; out null, 24 ; discard the byte lane replication of the FIFO since we only need 8 bits (not needed????)
     pull block ; this will block until there is data in the FIFO queue and then it will pull it into the shift register
-    set y, {31 if rgbw else 23} ; set y to the number of bits to write counting 0, (23 if RGB, 31 if RGBW)
+    set y, 7 ; set y to the number of bits to write counting 0, (always 7 because we are doing one word at a time)
 
 mainloop:
     ; go through each bit in the shift register and jump to the appropriate label
@@ -95,7 +99,15 @@ mainloop:
 
     out x, 1
     jmp !x, writezero
-    jmp writeone
+
+writeone:
+    ; Write T1H and T1L bits to the output pin
+    set pins, 1 [{t1h}]
+{nops_t1h}
+    set pins, 0 [{t1l}]
+{nops_t1l}
+    jmp y--, mainloop
+    jmp awaiting_data
 
 writezero:
     ; Write T0H and T0L bits to the output pin
@@ -106,14 +118,7 @@ writezero:
     jmp y--, mainloop
     jmp awaiting_data
 
-writeone:
+
-    ; Write T1H and T1L bits to the output pin
-    set pins, 1 [{t1h}]
-{nops_t1h}
-    set pins, 0 [{t1l}]
-{nops_t1l}
-    jmp y--, mainloop
-    jmp awaiting_data
 
 .wrap"""
 
@@ -139,7 +144,15 @@ RP2040PIOLEDStripLightOutput = rp2040_pio_led_strip_ns.class_(
 
 RGBOrder = rp2040_pio_led_strip_ns.enum("RGBOrder")
 
-Chipsets = rp2040_pio_led_strip_ns.enum("Chipset")
+Chipset = rp2040_pio_led_strip_ns.enum("Chipset")
+
+CHIPSETS = {
+    "WS2812": Chipset.CHIPSET_WS2812,
+    "WS2812B": Chipset.CHIPSET_WS2812B,
+    "SK6812": Chipset.CHIPSET_SK6812,
+    "SM16703": Chipset.CHIPSET_SM16703,
+    "CUSTOM": Chipset.CHIPSET_CUSTOM,
+}
 
 
 @dataclass
@@ -159,10 +172,10 @@ RGB_ORDERS = {
     "BRG": RGBOrder.ORDER_BRG,
 }
 
-CHIPSETS = {
+CHIPSET_TIMINGS = {
-    "WS2812": LEDStripTimings(20, 43, 41, 31),
+    "WS2812": LEDStripTimings(20, 40, 46, 34),
-    "WS2812B": LEDStripTimings(23, 46, 46, 23),
+    "WS2812B": LEDStripTimings(23, 49, 46, 26),
-    "SK6812": LEDStripTimings(17, 52, 31, 31),
+    "SK6812": LEDStripTimings(17, 52, 34, 34),
     "SM16703": LEDStripTimings(17, 52, 52, 17),
 }
 
@@ -192,7 +205,7 @@ CONFIG_SCHEMA = cv.All(
             cv.Required(CONF_NUM_LEDS): cv.positive_not_null_int,
             cv.Required(CONF_RGB_ORDER): cv.enum(RGB_ORDERS, upper=True),
             cv.Required(CONF_PIO): cv.one_of(0, 1, int=True),
-            cv.Optional(CONF_CHIPSET): cv.one_of(*CHIPSETS, upper=True),
+            cv.Optional(CONF_CHIPSET): cv.enum(CHIPSETS, upper=True),
             cv.Optional(CONF_IS_RGBW, default=False): cv.boolean,
             cv.Inclusive(
                 CONF_BIT0_HIGH,
@@ -238,7 +251,8 @@ async def to_code(config):
 
     key = f"led_strip_{id}"
 
-    if CONF_CHIPSET in config:
+    if chipset := config.get(CONF_CHIPSET):
+        cg.add(var.set_chipset(chipset))
         _LOGGER.info("Generating PIO assembly code")
         rp2040.add_pio_file(
             __name__,
@@ -246,13 +260,14 @@ async def to_code(config):
             generate_assembly_code(
                 id,
                 config[CONF_IS_RGBW],
-                CHIPSETS[config[CONF_CHIPSET]].T0H,
+                CHIPSET_TIMINGS[chipset].T0H,
-                CHIPSETS[config[CONF_CHIPSET]].T0L,
+                CHIPSET_TIMINGS[chipset].T0L,
-                CHIPSETS[config[CONF_CHIPSET]].T1H,
+                CHIPSET_TIMINGS[chipset].T1H,
-                CHIPSETS[config[CONF_CHIPSET]].T1L,
+                CHIPSET_TIMINGS[chipset].T1L,
             ),
         )
     else:
+        cg.add(var.set_chipset(Chipset.CHIPSET_CUSTOM))
         _LOGGER.info("Generating custom PIO assembly code")
         rp2040.add_pio_file(
             __name__,