Merge pull request #325 from MichaelBell/improve-performance

Badger2040: system_speed call, plus performance improvements

drivers/uc8151/uc8151.cpp

@@ -461,6 +461,10 @@ namespace pimoroni {
     *p |= b; // set bit value
   }
 
+  uint8_t* UC8151::get_frame_buffer() {
+	  return frame_buffer;
+  }
+
   void UC8151::invert(bool inv) {
     inverted = inv;
     command(CDI, {(uint8_t)(inverted ? 0b01'01'1100 : 0b01'00'1100)}); // vcom and data interval
@@ -477,15 +481,15 @@ namespace pimoroni {
   uint32_t UC8151::update_time() {
     switch(_update_speed) {
       case 0:
-        return 5500;
+        return 4500;
       case 1:
-        return 2600;
+        return 2000;
       case 2:
-        return 1000;
+        return 800;
       case 3:
-        return 300;
+        return 250;
       default:
-        return 5500;
+        return 4500;
     }
   }
 

drivers/uc8151/uc8151.hpp

@@ -208,6 +208,7 @@ namespace pimoroni {
     void off();
 
     void pixel(int x, int y, int v);
+    uint8_t* get_frame_buffer();
   };
 
 }

libraries/badger2040/badger2040.cpp

@@ -69,14 +69,6 @@ namespace pimoroni {
     // TODO: set default image?
   }
 
-  void Badger2040::clear() {
-    for(uint32_t y = 0; y < 128; y++) {
-      for(uint32_t x = 0; x < 296; x++) {
-        pixel(x, y);
-      }
-    }
-  }
-
   void Badger2040::halt() {
     gpio_put(ENABLE_3V3, 0);
 
@@ -100,15 +92,54 @@ namespace pimoroni {
       15,  7, 13,  5
     };
 
+    if (p == 0) {
+      return 1;
+    }
+    if (p == 15) {
+      return 0;
+    }
+
     // calculate dither matrix offset
     uint32_t dmo = (x & 0b11) | ((y & 0b11) << 2);
 
-    if(p == 0) {
-      return 1;
-    }else if(p == 15) {
+    return p <= _odm[dmo] ? 1 : 0;
+  }
+
+  // Return dither values for an entire byte in the column
+  uint8_t _dither_column_value(int32_t x, uint8_t p) {
+    if (p == 0) {
+      return 0xff;
+    }
+    if (p == 15) {
       return 0;
-    }else{
-      return p <= _odm[dmo] ? 1 : 0;
+    }
+
+    uint8_t val = 0;
+    for (int32_t y = 0; y < 4; ++y) {
+      val |= _dither_value(x, y, p) << (7 - y);
+    }
+    val |= val >> 4;
+    return val;
+  }
+
+
+  void Badger2040::clear() {
+    const uint32_t column_len = 128 / 8;
+    const uint32_t buf_len = column_len * 296;
+    uint8_t* buf = uc8151.get_frame_buffer();
+    
+    if (_pen == 0) {
+      memset(buf, 0xff, buf_len);
+    }
+    else if (_pen == 15) {
+      memset(buf, 0, buf_len);
+    }
+    else {
+      for(uint32_t x = 0; x < 296; x++) {
+        uint8_t val = _dither_column_value(x, _pen);
+        memset(buf, val, column_len);
+        buf += column_len;
+      }
     }
   }
 
@@ -130,14 +161,38 @@ namespace pimoroni {
     image(data, sheet_width, icon_size * index, 0, icon_size, icon_size, dx, dy);
   }
 
-  // Display an image that fills the screen (286*128)
+  // Display an image that fills the screen (296*128)
   void Badger2040::image(const uint8_t* data) {
-    image(data, 296, 0, 0, 296, 128, 0, 0);
+    uint8_t* ptr = uc8151.get_frame_buffer();
+    
+    for (uint32_t x = 0; x < 296; ++x) {
+      // extract bitmask for this pixel
+      uint32_t bm = 0b10000000 >> (x & 0b111);
+      
+      for (uint32_t y = 0; y < 128; y += 8) {
+        uint8_t val = 0;
+        for (uint32_t cy = 0; cy < 8; ++cy) {
+          // work out byte offset in source data
+          uint32_t o = ((y + cy) * (296 >> 3)) + (x >> 3);
+          
+          // Set bit in val if set in source data
+          if (data[o] & bm) {
+            val |= 0b10000000 >> cy;
+          }
+        }
+        *ptr++ = val;
+      }
+    }
   }
 
   // Display an image smaller than the screen (sw*sh) at dx, dy
   void Badger2040::image(const uint8_t *data, int w, int h, int x, int y) {
-    image(data, w, 0, 0, w, h, x, y);
+    if (x == 0 && y == 0 && w == 296 && h == 128) {
+      image(data);
+    }
+    else {
+      image(data, w, 0, 0, w, h, x, y);
+    }
   }
 
   void Badger2040::image(const uint8_t *data, int stride, int sx, int sy, int dw, int dh, int dx, int dy) {
@@ -156,9 +211,44 @@ namespace pimoroni {
   }
 
   void Badger2040::rectangle(int32_t x, int32_t y, int32_t w, int32_t h) {
-    for(int cy = y; cy < y + h; cy++) {
+    // Adjust for thickness
+    uint32_t ht = _thickness / 2;
+    x -= ht;
+    if (x < 0) {
+      w += x;
+      x = 0;
+    }
+    y -= ht;
+    if (y < 0) {
+      h += y;
+      y = 0;
+    }
+    w += _thickness - 1;
+    h += _thickness - 1;
+
+    if (h >= 8) {
+      // Directly write to the frame buffer when clearing a large area
+      uint8_t* buf = uc8151.get_frame_buffer();
+      
       for(int cx = x; cx < x + w; cx++) {
-        pixel(cx, cy);
+        uint8_t* buf_ptr = &buf[cx * 16 + y / 8];
+        uint8_t first_mask = 0xff >> (y & 7);
+        uint8_t last_mask = 0xff >> ((y + h) & 7);
+        uint32_t val = _dither_column_value(cx, _pen);
+        *buf_ptr &= ~first_mask;
+        *buf_ptr++ |= (val & first_mask);
+        for (int32_t c = h - (8 - (y & 7)); c >= 8; c -= 8) {
+          *buf_ptr++ = val;
+        }
+        *buf_ptr &= last_mask;
+        *buf_ptr |= (val & (~last_mask));
+      }
+    }
+    else {
+      for(int cx = x; cx < x + w; cx++) {
+        for(int cy = y; cy < y + h; cy++) {
+          uc8151.pixel(cx, cy, _dither_value(cx, cy, _pen));
+        }
       }
     }
   }

micropython/examples/badger2040/clock.py

@@ -2,6 +2,8 @@ import time
 import machine
 import badger2040
 
+# We're going to keep the badger on, so slow down the system clock if on battery
+badger2040.system_speed(badger2040.SYSTEM_SLOW)
 
 rtc = machine.RTC()
 display = badger2040.Badger2040()

micropython/examples/badger2040/conway.py

@@ -6,6 +6,9 @@ import machine
 
 import badger2040
 
+# Overclock the RP2040 to run the sim faster
+badger2040.system_speed(badger2040.SYSTEM_TURBO)
+
 # ------------------------------
 #        Program setup
 # ------------------------------
@@ -16,6 +19,7 @@ INITIAL_DENSITY = 0.3  # Density of cells at start
 
 # Create a new Badger and set it to update TURBO
 screen = badger2040.Badger2040()
+screen.led(128)
 screen.update_speed(badger2040.UPDATE_TURBO)
 
 restart = False  # should sim be restarted

micropython/examples/badger2040/launcher.py

@@ -1,14 +1,13 @@
 import gc
 import time
 import math
-import machine
 import badger2040
 from badger2040 import WIDTH
 import launchericons
 import badger_os
 
-# Reduce clock speed to 48MHz, that's fast enough!
-machine.freq(48000000)
+# Reduce clock speed to 48MHz
+badger2040.system_speed(badger2040.SYSTEM_NORMAL)
 
 changed = False
 exited_to_launcher = False

micropython/modules/badger2040/README.md

@@ -67,7 +67,19 @@ Badger 2040 features five buttons on its front, labelled A, B, C, ↑ (up), ↓
 * `BUTTON_DOWN` = `11`
 * `BUTTON_USER` = `23`
 
-Note, due to the BOOT/USR button performing both BOOT and USER functions, the output of reading it with `pressed()` will be inverted from the others.
+## System speed
+
+The system clock speed of the RP2040 can be controlled, allowing power to be saved if on battery, or faster computations to be performed.  Use `badger2040.system_speed(speed)` where `speed` is one of the following constants:
+
+* `SYSTEM_VERY_SLOW` = `0`  _4 MHz if on battery, 48 MHz if connected to USB_
+* `SYSTEM_SLOW` = `1`  _12 MHz if on battery, 48 MHz if connected to USB_
+* `SYSTEM_NORMAL` = `2`  _48 MHz_
+* `SYSTEM_FAST` = `3`  _133 MHz_
+* `SYSTEM_TURBO` = `4`  _250 MHz_
+
+On USB, the system will not run slower than 48MHz, as that is the minimum clock speed required to keep the USB connection stable.
+
+Note that `SYSTEM_TURBO` overclocks the RP2040 to 250MHz, and applies a small over voltage to ensure this is stable.  We've found that every RP2040 we've tested is happy to run at this speed without any issues.
 
 ## Other Functions
 
@@ -95,7 +107,6 @@ icon(data, icon_index, sheet_size, icon_size)
 clear()
 update()
 partial_update(x, y, w, h)
-update_speed(speed)
 invert(inverted)
 ```
 

micropython/modules/badger2040/badger2040.c

@@ -35,6 +35,7 @@ MP_DEFINE_CONST_FUN_OBJ_3(Badger2040_command_obj, Badger2040_command);
 MP_DEFINE_CONST_FUN_OBJ_1(Badger2040_pressed_to_wake_obj, Badger2040_pressed_to_wake);
 MP_DEFINE_CONST_FUN_OBJ_1(Badger2040_halt_obj, Badger2040_halt);
 MP_DEFINE_CONST_FUN_OBJ_0(Badger2040_woken_by_button_obj, Badger2040_woken_by_button);
+MP_DEFINE_CONST_FUN_OBJ_1(Badger2040_system_speed_obj, Badger2040_system_speed);
 
 /***** Binding of Methods *****/
 STATIC const mp_rom_map_elem_t Badger2040_locals_dict_table[] = {
@@ -90,6 +91,7 @@ STATIC const mp_rom_map_elem_t badger2040_globals_table[] = {
 
     { MP_ROM_QSTR(MP_QSTR_pressed_to_wake), MP_ROM_PTR(&Badger2040_pressed_to_wake_obj) },
     { MP_ROM_QSTR(MP_QSTR_woken_by_button), MP_ROM_PTR(&Badger2040_woken_by_button_obj) },
+    { MP_ROM_QSTR(MP_QSTR_system_speed), MP_ROM_PTR(&Badger2040_system_speed_obj) },
 
     { MP_ROM_QSTR(MP_QSTR_UPDATE_NORMAL), MP_ROM_INT(0) },
     { MP_ROM_QSTR(MP_QSTR_UPDATE_MEDIUM), MP_ROM_INT(1) },
@@ -97,6 +99,12 @@ STATIC const mp_rom_map_elem_t badger2040_globals_table[] = {
     { MP_ROM_QSTR(MP_QSTR_UPDATE_TURBO), MP_ROM_INT(3) },
     { MP_ROM_QSTR(MP_QSTR_UPDATE_SUPER_EXTRA_TURBO), MP_ROM_INT(3) }, // ho ho placebo!
 
+    { MP_ROM_QSTR(MP_QSTR_SYSTEM_VERY_SLOW), MP_ROM_INT(0) },
+    { MP_ROM_QSTR(MP_QSTR_SYSTEM_SLOW), MP_ROM_INT(1) },
+    { MP_ROM_QSTR(MP_QSTR_SYSTEM_NORMAL), MP_ROM_INT(2) },
+    { MP_ROM_QSTR(MP_QSTR_SYSTEM_FAST), MP_ROM_INT(3) },
+    { MP_ROM_QSTR(MP_QSTR_SYSTEM_TURBO), MP_ROM_INT(4) },
+
     { MP_ROM_QSTR(MP_QSTR_WIDTH), MP_ROM_INT(296) },
     { MP_ROM_QSTR(MP_QSTR_HEIGHT), MP_ROM_INT(128) },
 

micropython/modules/badger2040/badger2040.cpp

@@ -139,7 +139,7 @@ MICROPY_EVENT_POLL_HOOK
 
     // Ensure blocking for the minimum amount of time
     // in cases where "is_busy" is unreliable.
-    while(self->badger2040->is_busy() || absolute_time_diff_us(t_end, get_absolute_time()) > 0) {
+    while(self->badger2040->is_busy() || absolute_time_diff_us(get_absolute_time(), t_end) > 0) {
 #ifdef MICROPY_EVENT_POLL_HOOK
 MICROPY_EVENT_POLL_HOOK
 #endif
@@ -517,4 +517,75 @@ mp_obj_t Badger2040_measure_glyph(size_t n_args, const mp_obj_t *pos_args, mp_ma
     return mp_obj_new_int(self->badger2040->measure_glyph(c, scale));
 }
 
+#include "hardware/vreg.h"
+#include "hardware/clocks.h"
+#include "hardware/pll.h"
+
+mp_obj_t Badger2040_system_speed(mp_obj_t speed) {
+    uint32_t sys_freq;
+    uint32_t selected_speed = mp_obj_get_int(speed);
+
+    if (gpio_get(pimoroni::Badger2040::VBUS_DETECT) && selected_speed < 2) {
+      // If on USB never go slower than normal speed.
+      selected_speed = 2;
+    }
+
+    switch (selected_speed)
+    {
+    case 4: // TURBO: 250 MHZ, 1.2V
+        vreg_set_voltage(VREG_VOLTAGE_1_20);
+        set_sys_clock_khz(250000, true);
+        return mp_const_none;
+    case 3: // FAST: 133 MHZ
+        vreg_set_voltage(VREG_VOLTAGE_1_10);
+        set_sys_clock_khz(133000, true);
+        return mp_const_none;
+
+    default:
+    case 2: // NORMAL: 48 MHZ
+        vreg_set_voltage(VREG_VOLTAGE_1_10);
+        set_sys_clock_48mhz();
+        return mp_const_none;
+
+    case 1: // SLOW: 12 MHZ, 1.0V
+        sys_freq = 12 * MHZ;
+        break;
+
+    case 0: // VERY_SLOW: 4 MHZ, 1.0V
+        sys_freq = 4 * MHZ;
+        break;
+    }
+
+
+    // Set the configured clock speed, by dividing the USB PLL
+    clock_configure(clk_sys,
+                    CLOCKS_CLK_SYS_CTRL_SRC_VALUE_CLKSRC_CLK_SYS_AUX,
+                    CLOCKS_CLK_SYS_CTRL_AUXSRC_VALUE_CLKSRC_PLL_USB,
+                    48 * MHZ,
+                    sys_freq);
+
+    clock_configure(clk_peri,
+                    0,
+                    CLOCKS_CLK_PERI_CTRL_AUXSRC_VALUE_CLK_SYS,
+                    sys_freq,
+                    sys_freq);
+
+    clock_configure(clk_adc,
+                    0,
+                    CLOCKS_CLK_ADC_CTRL_AUXSRC_VALUE_CLKSRC_PLL_USB,
+                    48 * MHZ,
+                    sys_freq);
+
+    // No longer using the SYS PLL so disable it
+    pll_deinit(pll_sys);
+
+    // Not using USB so stop the clock
+    clock_stop(clk_usb);
+
+    // Drop the core voltage
+    vreg_set_voltage(VREG_VOLTAGE_1_00);
+
+    return mp_const_none;
+}
+
 }

micropython/modules/badger2040/badger2040.h

@@ -43,3 +43,5 @@ extern mp_obj_t Badger2040_command(mp_obj_t self_in, mp_obj_t reg, mp_obj_t data
 
 extern mp_obj_t Badger2040_pressed_to_wake(mp_obj_t button);
 extern mp_obj_t Badger2040_woken_by_button();
+
+extern mp_obj_t Badger2040_system_speed(mp_obj_t speed);

micropython/modules/badger2040/micropython.cmake

@@ -20,4 +20,7 @@ target_compile_definitions(usermod_${MOD_NAME} INTERFACE
     MODULE_BADGER2040_ENABLED=1
 )
 
-target_link_libraries(usermod INTERFACE usermod_${MOD_NAME})
+target_link_libraries(usermod INTERFACE usermod_${MOD_NAME}
+  hardware_vreg
+  hardware_pll
+  hardware_resets)