optimize leds stage

firmware/src/gfx_decoder.cpp

@@ -40,6 +40,38 @@ int32_t gfx_decoder_loadNextFrame() {
     return false;
   }
 
+  // Convert framebuffer into raw shift register data for fast PIO pixel pushing
+  // Data will be held in buffers, one per pixel's depth bit (aka brightness stage),
+  // with each row split into 32-bit chunks, one per module
+  // (20 pixels, 24 shift register stages, 8 unused bits)
+  // TODO: Move this to leds.cpp
+  // TODO: Use a separate buffer, then copy to ledsBuffer to avoid tearing
+  for (int bi = 0; bi < 8; bi++) {
+    uint8_t bitPosition = 1 << bi;
+    for (int y = 0; y < ROW_COUNT; y++) {
+      auto yOffset = y * COL_COUNT;
+      for (int xModule = 0; xModule < COL_MODULES; xModule++) {
+        auto bufferXOffset = yOffset + xModule * 20;
+        uint32_t sample = 0;
+
+        for (int x = 0; x < 20; x++) {
+          // insert placeholders for unused stages
+          // (before pixels 0, 6, 13)
+          if (x == 0 || x == 6 || x == 13) {
+            sample >>= 1;
+          }
+          uint8_t px = buffer[bufferXOffset + x];
+          bool bit = px & bitPosition;
+          sample = (sample >> 1) | (bit ? 0x80000000 : 0);
+        }
+        // insert placeholder for unused last stage (after pixel 19)
+        sample >>=1;
+
+        ledBuffer[bi][y * COL_MODULES + xModule] = sample;
+      }
+    }
+  }
+
   // copy to framebuffer
   // TODO: mutex? double buffer? or something...
   memcpy(framebuffer, buffer, ROW_COUNT * COL_COUNT);

firmware/src/leds.cpp

@@ -38,6 +38,7 @@ uint8_t brightnessPhaseDelays[COLOR_BITS] = {0, 1, 6, 20, 60};
 
 // NOTE: Alignment required to allow 4-byte reads
 uint8_t framebuffer[ROW_COUNT * COL_COUNT]  __attribute__((aligned(32))) = {0};
+uint32_t ledBuffer[8][ROW_COUNT * COL_MODULES] = {0};
 
 void leds_init() {
   memset(framebuffer, 0, sizeof(framebuffer));
@@ -104,7 +105,7 @@ void leds_render() {
 
   for (int yCount = 0; yCount < ROW_COUNT; yCount++) {
     int y = ROW_COUNT - 1 - yCount;
-    // brigthness - pushing data takes 40us, so to maximize brightness (at high brightness phases)
+    // brigthness - pushing data takes time, so to maximize brightness (at high brightness phases)
     // we want to keep the matrix on during update (except during latch). At low brightness phases,
     // we want it off to actually be dim
     bool brightPhase = brightnessPhase >= 2;
@@ -132,46 +133,23 @@ void leds_render() {
     // silly shit
     // TODO: Some ideas for future optimization:
     // - see if we can disable px pusher delays on improved electric interface
-    // - use a profiler to see how the inner loop can be improved
-    // - do the shift register bullshit once per frame, so that data can be loaded into
-    //   registers with aligned access, DMA, etc.
     // - improve outer loop which adds 2us of processing on each loop
     // - change busy wait into some kind of interrupt-based thing so that processing can continue
     // - latch row and clock simultaneously, avoid disabling output
-    uint8_t *buffer = framebuffer + (y * COL_COUNT);
+    // - DMA?
     for (int xModule = 0; xModule < COL_MODULES; xModule++) {
-      uint32_t pxValues;
-
-      // placeholder at 0; pixels 0, 1, 2
-      pxValues = *(reinterpret_cast<uint32_t *>(buffer));
-      pxValues = pxValues << 8;
-      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues >> brightnessPhase);
-
-      // pixels 3, 4, 5, placeholder at 6
-      pxValues = *(reinterpret_cast<uint32_t *>(buffer + 3));
-      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues >> brightnessPhase);
-
-      // pixels 6, 7, 8, 9
-      pxValues = *(reinterpret_cast<uint32_t *>(buffer + 6));
-      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues >> brightnessPhase);
-
-      // pixels 10, 11, 12, placeholder at 13
-      pxValues = *(reinterpret_cast<uint32_t *>(buffer + 10));
-      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues >> brightnessPhase);
-
-      // pixels 13, 14, 15, 16
-      pxValues = *(reinterpret_cast<uint32_t *>(buffer + 13));
-      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues >> brightnessPhase);
-
-      // pixels 17, 18, 19, placeholder
-      pxValues = *(reinterpret_cast<uint32_t *>(buffer + 17));
-      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues >> brightnessPhase);
-
-      buffer += 20;
+      uint32_t pxValues = ledBuffer[brightnessPhase + 3][y * COL_MODULES + xModule];
+      pio_sm_put_blocking(pusher_pio, pusher_sm, pxValues);
     }
 
     // wait for all data to be shifted out
-    pio_sm_drain_tx_fifo(pusher_pio, pusher_sm);
+    while (!pio_sm_is_tx_fifo_empty(pusher_pio, pusher_sm)) {
+      tight_loop_contents();
+    }
+    // TODO: Is there an API to wait for PIO to actually become idle?
+    // pio_sm_drain_tx_fifo doesn't seem to do the trick
+    // if not, we might need to use irqs or something
+    busy_wait_us(4);
 
     // disable columns before latch
     outputEnable(ROW_OE, false);
@@ -201,7 +179,7 @@ void leds_initPusher() {
   uint latchPin = COL_SRCLK;
 
   pio_sm_config config = leds_px_pusher_program_get_default_config(offset);
-  sm_config_set_clkdiv_int_frac(&config, 2, 0);
+  sm_config_set_clkdiv_int_frac(&config, 1, 0);
 
   // Shift OSR to the right, autopull
   sm_config_set_out_shift(&config, true, true, 32);

firmware/src/leds.h

@@ -31,5 +31,6 @@ void leds_loop();
 void leds_render();
 
 extern uint8_t framebuffer[ROW_COUNT * COL_COUNT];
+extern uint32_t ledBuffer[8][ROW_COUNT * COL_MODULES];
 
 #endif

firmware/src/leds.pio

@@ -1,9 +1,21 @@
 .program leds_px_pusher
 .side_set 1 opt
-.wrap_target
 public entry_point:
-  out null, 3   side 0 [0]    ; ignore least significant digits
-  out pins, 1                 ; set bit (shifted for brightness phase by C code)
-  out null, 4   side 1 [1]    ; ignore remaining bits, latch data, allow time for latching
-  nop           side 0        ; return to 0 (weird glitches happen otherwise)
+.wrap_target
+  ; get 32 bits from fifo (not required with autopull, useful for debug)
+  ; pull
+  ; push 24 bits to the shift registers
+  ; also, return latch bit to 0
+  set x, 23   side 0
+  ; ignore the 8 least significant bits
+  out null, 8
+loop:
+  ; lower clock edge
+  nop side 0
+  ; set bit
+  out pins, 1
+  ; loop; latch bit (rising edge)
+  ; TODO: check if this delay can be lowered with a PCB
+  jmp x-- loop   side 1 [2]
+end:
 .wrap

firmware/src/leds.pio.h

@@ -13,23 +13,24 @@
 // -------------- //
 
 #define leds_px_pusher_wrap_target 0
-#define leds_px_pusher_wrap 3
+#define leds_px_pusher_wrap 4
 
 #define leds_px_pusher_offset_entry_point 0u
 
 static const uint16_t leds_px_pusher_program_instructions[] = {
             //     .wrap_target
-    0x7063, //  0: out    null, 3         side 0     
-    0x6001, //  1: out    pins, 1                    
-    0x7964, //  2: out    null, 4         side 1 [1] 
-    0xb042, //  3: nop                    side 0     
+    0xf037, //  0: set    x, 23           side 0     
+    0x6068, //  1: out    null, 8                    
+    0xb042, //  2: nop                    side 0     
+    0x6001, //  3: out    pins, 1                    
+    0x1a42, //  4: jmp    x--, 2          side 1 [2] 
             //     .wrap
 };
 
 #if !PICO_NO_HARDWARE
 static const struct pio_program leds_px_pusher_program = {
     .instructions = leds_px_pusher_program_instructions,
-    .length = 4,
+    .length = 5,
     .origin = -1,
 };