Fix clocking & add framebuffer mirroring examples

examples/fbmirror.py

@@ -0,0 +1,43 @@
+#!/usr/bin/python3
+"""
+Mirror a scaled copy of the framebuffer to a 64x32 matrix
+
+The upper left corner of the framebuffer is displayed until the user hits ctrl-c.
+
+The `/dev/fb0` special file will exist if a monitor is plugged in at boot time,
+or if `/boot/firmware/cmdline.txt` specifies a resolution such as
+`...  video=HDMI-A-1:640x480M@60D`.
+"""
+
+
+import adafruit_raspberry_pi5_piomatter
+import numpy as np
+
+yoffset = 0
+xoffset = 0
+
+with open("/sys/class/graphics/fb0/virtual_size") as f:
+    screenx, screeny = [int(word) for word in f.read().split(",")]
+
+with open("/sys/class/graphics/fb0/bits_per_pixel") as f:
+    bits_per_pixel = int(f.read())
+
+assert bits_per_pixel == 16
+
+bytes_per_pixel = bits_per_pixel // 8
+dtype = {2: np.uint16, 4: np.uint32}[bytes_per_pixel]
+
+with open("/sys/class/graphics/fb0/stride") as f:
+    stride = int(f.read())
+
+linux_framebuffer = np.memmap('/dev/fb0',mode='r', shape=(screeny, stride // bytes_per_pixel), dtype=dtype)
+
+width = 64
+height = 32
+geometry = adafruit_raspberry_pi5_piomatter.Geometry(width=width, height=height, n_addr_lines=4, rotation=adafruit_raspberry_pi5_piomatter.Orientation.Normal)
+matrix_framebuffer = np.zeros(shape=(geometry.height, geometry.width), dtype=dtype)
+matrix = adafruit_raspberry_pi5_piomatter.AdafruitMatrixBonnetRGB565(matrix_framebuffer, geometry)
+
+while True:
+    matrix_framebuffer[:,:] = linux_framebuffer[yoffset:yoffset+height, xoffset:xoffset+width]
+    matrix.show()

examples/fbmirror_scaled.py

@@ -0,0 +1,58 @@
+#!/usr/bin/python3
+"""
+Mirror a scaled copy of the framebuffer to a 64x32 matrix
+
+The upper left corner of the framebuffer is displayed until the user hits ctrl-c.
+
+The `/dev/fb0` special file will exist if a monitor is plugged in at boot time,
+or if `/boot/firmware/cmdline.txt` specifies a resolution such as
+`...  video=HDMI-A-1:640x480M@60D`.
+"""
+
+
+import adafruit_raspberry_pi5_piomatter
+import numpy as np
+import PIL.Image as Image
+
+with open("/sys/class/graphics/fb0/virtual_size") as f:
+    screenx, screeny = [int(word) for word in f.read().split(",")]
+
+with open("/sys/class/graphics/fb0/bits_per_pixel") as f:
+    bits_per_pixel = int(f.read())
+
+assert bits_per_pixel == 16
+
+bytes_per_pixel = bits_per_pixel // 8
+dtype = {2: np.uint16, 4: np.uint32}[bytes_per_pixel]
+
+with open("/sys/class/graphics/fb0/stride") as f:
+    stride = int(f.read())
+
+linux_framebuffer = np.memmap('/dev/fb0',mode='r', shape=(screeny, stride // bytes_per_pixel), dtype=dtype)
+
+xoffset = 0
+yoffset = 0
+width = 64
+height = 32
+scale = 3
+
+geometry = adafruit_raspberry_pi5_piomatter.Geometry(width=width, height=height, n_addr_lines=4, rotation=adafruit_raspberry_pi5_piomatter.Orientation.Normal)
+matrix_framebuffer = np.zeros(shape=(geometry.height, geometry.width, 3), dtype=np.uint8)
+matrix = adafruit_raspberry_pi5_piomatter.AdafruitMatrixBonnetRGB888Packed(matrix_framebuffer, geometry)
+
+while True:
+    tmp = linux_framebuffer[yoffset:yoffset+height*scale, xoffset:xoffset+width*scale]
+    # Convert the RGB565 framebuffer into RGB888Packed (so that we can use PIL image operations to rescale it)
+    r = (tmp & 0xf800) >> 8
+    r = r | (r >> 5)
+    r = r.astype(np.uint8)
+    g = (tmp & 0x07e0) >> 3
+    g = g | (g >> 6)
+    g = g.astype(np.uint8)
+    b = (tmp & 0x001f) << 3
+    b = b | (b >> 5)
+    b = b.astype(np.uint8)
+    img = Image.fromarray(np.stack([r, g, b], -1))
+    img = img.resize((width, height))
+    matrix_framebuffer[:,:] = np.asarray(img)
+    matrix.show()

examples/playframes.py

@@ -32,4 +32,4 @@
     t1 = time.monotonic()
     dt = t1 - t0
     fps = nimages/dt
-    print(f"{nimages} frames in {dt}s, {fps}fps")
+    print(f"{nimages} frames in {dt}s, {fps}fps [{matrix.fps}]")

src/assemble.py

@@ -1,29 +1,26 @@
-import sys
-from contextlib import contextmanager
+import io
+import pathlib
+from contextlib import redirect_stdout
 
 import adafruit_pioasm
 import click
 
 
-@contextmanager
-def temporary_stdout(filename):
-    old_stdout = sys.stdout
-    try:
-        with open(filename, "w", encoding="utf-8") as sys.stdout:
-            yield sys.stdout
-    finally:
-        sys.stdout = old_stdout
-
 @click.command
 @click.argument("infile")
 @click.argument("outfile")
 def main(infile, outfile):
-    program_name = infile.rpartition("/")[2].partition(".")[0]
-    print(program_name)
+    program_name = pathlib.Path(infile).stem
     program = adafruit_pioasm.Program.from_file(infile, build_debuginfo=True)
 
-    with temporary_stdout(outfile):
+    c_program = io.StringIO()
+    with redirect_stdout(c_program):
         program.print_c_program(program_name)
 
+    with open(outfile, "w", encoding="utf-8") as out:
+        print("#pragma once", file=out)
+        print("", file=out)
+        print(c_program.getvalue().rstrip().replace("True", "true"), file=out)
+
 if __name__ == '__main__':
     main()

src/include/piomatter/piomatter.h

@@ -12,6 +12,12 @@
 
 namespace piomatter {
 
+static uint64_t monotonicns64() {
+    struct timespec tp;
+    clock_gettime(CLOCK_MONOTONIC, &tp);
+    return tp.tv_sec * UINT64_C(1000000000) + tp.tv_nsec;
+}
+
 constexpr size_t MAX_XFER = 65532;
 
 void pio_sm_xfer_data_large(PIO pio, int sm, int direction, size_t size,
@@ -35,6 +41,8 @@ struct piomatter_base {
 
     virtual ~piomatter_base() {}
     virtual void show() = 0;
+
+    double fps;
 };
 
 template <class pinout = adafruit_matrix_bonnet_pinout,
@@ -111,11 +119,21 @@ struct piomatter : piomatter_base {
         pio_sm_config c = pio_get_default_sm_config();
         sm_config_set_wrap(&c, offset + protomatter_wrap_target,
                            offset + protomatter_wrap);
+        // 1 side-set pin
+        sm_config_set_sideset(&c, 2, true, false);
         sm_config_set_out_shift(&c, /* shift_right= */ false,
                                 /* auto_pull = */ true, 32);
         sm_config_set_fifo_join(&c, PIO_FIFO_JOIN_TX);
-        sm_config_set_clkdiv(&c, 1.0);
+        // Due to https://github.com/raspberrypi/utils/issues/116 it's not
+        // possible to keep the RP1 state machine fed at high rates. This target
+        // frequency is approximately the best sustainable clock with current
+        // FW & kernel.
+        constexpr double target_freq =
+            2700000 * 2; // 2.7MHz pixel clock, 2 PIO cycles per pixel
+        double div = clock_get_hz(clk_sys) / target_freq;
+        sm_config_set_clkdiv(&c, div);
         sm_config_set_out_pins(&c, 0, 28);
+        sm_config_set_sideset_pins(&c, pinout::PIN_CLK);
         pio_sm_init(pio, sm, offset, &c);
         pio_sm_set_enabled(pio, sm, true);
 
@@ -146,6 +164,8 @@ struct piomatter : piomatter_base {
         size_t datasize = 0;
         int old_buffer_idx = buffer_manager::no_buffer;
         int buffer_idx;
+        uint64_t t0, t1;
+        t0 = monotonicns64();
         while ((buffer_idx = manager.get_filled_buffer()) !=
                buffer_manager::exit_request) {
             if (buffer_idx != buffer_manager::no_buffer) {
@@ -160,6 +180,11 @@ struct piomatter : piomatter_base {
             if (datasize) {
                 pio_sm_xfer_data_large(pio, sm, PIO_DIR_TO_SM, datasize,
                                        (uint32_t *)databuf);
+                t1 = monotonicns64();
+                if (t0 != t1) {
+                    fps = 1e9 / (t1 - t0);
+                }
+                t0 = t1;
             } else {
                 std::this_thread::sleep_for(std::chrono::milliseconds(1));
             }

src/include/piomatter/protomatter.pio.h

@@ -2,24 +2,27 @@
 
 const int protomatter_wrap = 4;
 const int protomatter_wrap_target = 0;
-const int protomatter_sideset_pin_count = 0;
-const bool protomatter_sideset_enable = 0;
+const int protomatter_sideset_pin_count = 1;
+const bool protomatter_sideset_enable = true;
 const uint16_t protomatter[] = {
     // ; data format (out-shift-right):
     // ; MSB ... LSB
     // ; 0 ddd......ddd: 31-bit delay
     // ; 1 ccc......ccc: 31 bit data count
+    // .side_set 1 opt
     // .wrap_target
     // top:
     0x6021, //     out x, 1
     0x605f, //     out y, 31
-    0x0025, //     jmp !x delay_loop
+    0x0025, //     jmp !x do_delay
             // data_loop:
     0x6000, //     out pins, 32
-    0x0083, //     jmp y--, data_loop
+    0x1883, //     jmp y--, data_loop  side 1 ; assert clk bit
             // .wrap
+            // do_delay:
+    0x6000, //     out pins, 32
             // delay_loop:
-    0x0085, //     jmp y--, delay_loop
+    0x0086, //     jmp y--, delay_loop
     0x0000, //     jmp top
     //     ;; fill program out to 32 instructions so nothing else can load
     0xa042, //     nop

src/include/piomatter/render.h

@@ -7,12 +7,10 @@
 
 namespace piomatter {
 
-constexpr unsigned DATA_OVERHEAD = 3;
-// this is ... flatly wrong!? but it's the number that makes the ramp intensity
-// correct to my eye
-constexpr unsigned CLOCKS_PER_DATA = 128;
-constexpr unsigned DELAY_OVERHEAD = 5;
-constexpr unsigned CLOCKS_PER_DELAY = 1;
+constexpr int DATA_OVERHEAD = 3;
+constexpr int CLOCKS_PER_DATA = 2;
+constexpr int DELAY_OVERHEAD = 5;
+constexpr int CLOCKS_PER_DELAY = 1;
 
 constexpr uint32_t command_data = 1u << 31;
 constexpr uint32_t command_delay = 0;
@@ -139,12 +137,12 @@ void protomatter_render_rgb10(std::vector<uint32_t> &result,
 
     int data_count = 0;
 
-    auto do_delay = [&](uint32_t delay) {
-        if (delay == 0)
-            return;
+    auto do_data_delay = [&](uint32_t data, int32_t delay) {
+        delay = std::max((delay / CLOCKS_PER_DELAY) - DELAY_OVERHEAD, 1);
         assert(delay < 1000000);
         assert(!data_count);
         result.push_back(command_delay | (delay ? delay - 1 : 0));
+        result.push_back(data);
     };
 
     auto prep_data = [&data_count, &result](uint32_t n) {
@@ -155,27 +153,15 @@ void protomatter_render_rgb10(std::vector<uint32_t> &result,
         data_count = n;
     };
 
-    auto do_data = [&](uint32_t d) {
-        assert(data_count);
-        data_count--;
-        result.push_back(d);
-    };
-
     int32_t active_time;
 
-    auto do_data_active = [&active_time, &do_data](uint32_t d) {
+    auto do_data_clk_active = [&active_time, &data_count, &result](uint32_t d) {
         bool active = active_time > 0;
         active_time--;
         d |= active ? pinout::oe_active : pinout::oe_inactive;
-        do_data(d);
-    };
-
-    auto do_data_delay = [&prep_data, &do_data, &do_delay](uint32_t d,
-                                                           int32_t delay) {
-        prep_data(1);
-        do_data(d);
-        if (delay > 0)
-            do_delay(delay);
+        assert(data_count);
+        data_count--;
+        result.push_back(d);
     };
 
     auto calc_addr_bits = [](int addr) {
@@ -197,9 +183,9 @@ void protomatter_render_rgb10(std::vector<uint32_t> &result,
         return data;
     };
 
-    auto add_pixels = [&do_data_active, &result](uint32_t addr_bits, bool r0,
-                                                 bool g0, bool b0, bool r1,
-                                                 bool g1, bool b1) {
+    auto add_pixels = [&do_data_clk_active,
+                       &result](uint32_t addr_bits, bool r0, bool g0, bool b0,
+                                bool r1, bool g1, bool b1) {
         uint32_t data = addr_bits;
         if (r0)
             data |= (1 << pinout::PIN_RGB[0]);
@@ -214,8 +200,7 @@ void protomatter_render_rgb10(std::vector<uint32_t> &result,
         if (b1)
             data |= (1 << pinout::PIN_RGB[5]);
 
-        do_data_active(data);
-        do_data_active(data | pinout::clk_bit);
+        do_data_clk_active(data);
     };
 
     int last_bit = 0;
@@ -245,7 +230,7 @@ void protomatter_render_rgb10(std::vector<uint32_t> &result,
             active_time = 1 << last_bit;
             last_bit = bit;
 
-            prep_data(2 * pixels_across);
+            prep_data(pixels_across);
             auto mapiter = matrixmap.map.begin() + 2 * addr * pixels_across;
             for (size_t x = 0; x < pixels_across; x++) {
                 assert(mapiter != matrixmap.map.end());

src/protodemo.c

@@ -90,7 +90,6 @@ void test_pattern(int offs) {
 }
 
 static uint64_t monotonicns64() {
-
     struct timespec tp;
     clock_gettime(CLOCK_MONOTONIC, &tp);
     return tp.tv_sec * UINT64_C(1000000000) + tp.tv_nsec;

src/protomatter.pio

@@ -3,17 +3,20 @@
 ; 0 ddd......ddd: 31-bit delay
 ; 1 ccc......ccc: 31 bit data count
 
+.side_set 1 opt
 .wrap_target
 top:
     out x, 1
     out y, 31
-    jmp !x delay_loop
+    jmp !x do_delay
 
 data_loop:
     out pins, 32
-    jmp y--, data_loop
+    jmp y--, data_loop  side 1 ; assert clk bit
 .wrap
 
+do_delay:
+    out pins, 32
 delay_loop:
     jmp y--, delay_loop
     jmp top