tested

Marlin/src/lcd/dogm/ultralcd_st7920_u8glib_rrd_AVR.cpp

 // file u8g_dev_st7920_128x64_HAL.cpp for the HAL version.
 
 #include "../../inc/MarlinConfig.h"
+#include <U8glib.h>
 
 #if ENABLED(U8GLIB_ST7920)
 
 #define LCD_PIXEL_WIDTH 128
 #define LCD_PIXEL_HEIGHT 64
 
-#include <U8glib.h>
-
 //set optimization so ARDUINO optimizes this file
 #pragma GCC optimize (3)
 
   #define ST7920_DELAY_3 CPU_ST7920_DELAY_3
 #endif
 
+#define ST7920_SND_BIT \
+  WRITE(ST7920_CLK_PIN, LOW);        ST7920_DELAY_1; \
+  WRITE(ST7920_DAT_PIN, val & 0x80); ST7920_DELAY_2; \
+  WRITE(ST7920_CLK_PIN, HIGH);       ST7920_DELAY_3; \
+  val <<= 1
+
+static void ST7920_SWSPI_SND_8BIT(uint8_t val) {
+  ST7920_SND_BIT; // 1
+  ST7920_SND_BIT; // 2
+  ST7920_SND_BIT; // 3
+  ST7920_SND_BIT; // 4
+  ST7920_SND_BIT; // 5
+  ST7920_SND_BIT; // 6
+  ST7920_SND_BIT; // 7
+  ST7920_SND_BIT; // 8
+}
+
 #if defined(DOGM_SPI_DELAY_US) && DOGM_SPI_DELAY_US > 0
   #define U8G_DELAY() delayMicroseconds(DOGM_SPI_DELAY_US)
 #else
   #define U8G_DELAY() u8g_10MicroDelay()
 #endif
 
-static void ST7920_WRITE_BYTE(uint8_t val) {
-  for (uint8_t i = 0; i < 8; i++) {
-    WRITE(ST7920_DAT_PIN, val & 0x80);
-    WRITE(ST7920_CLK_PIN, HIGH);
-    WRITE(ST7920_CLK_PIN, LOW);
-    val <<= 1;
-  }
-}
-
-#define ST7920_SET_CMD()         { ST7920_WRITE_BYTE(0xF8); U8G_DELAY(); }
-#define ST7920_SET_DAT()         { ST7920_WRITE_BYTE(0xFA); U8G_DELAY(); }
-#define ST7920_WRITE_NIBBLES(a)  { ST7920_WRITE_BYTE((uint8_t)((a)&0xF0u)); ST7920_WRITE_BYTE((uint8_t)((a)<<4u)); U8G_DELAY(); }
-#define ST7920_WRITE_NIBBLES_P(p,l)  { for (uint8_t i = l + 1; --i;) { ST7920_WRITE_BYTE(*p&0xF0); ST7920_WRITE_BYTE(*p<<4); p++; } U8G_DELAY(); }
-
-
 #define ST7920_CS()              { WRITE(ST7920_CS_PIN,1); U8G_DELAY(); }
 #define ST7920_NCS()             { WRITE(ST7920_CS_PIN,0); }
+#define ST7920_SET_CMD()         { ST7920_SWSPI_SND_8BIT(0xF8); U8G_DELAY(); }
+#define ST7920_SET_DAT()         { ST7920_SWSPI_SND_8BIT(0xFA); U8G_DELAY(); }
+#define ST7920_WRITE_BYTE(a)     { ST7920_SWSPI_SND_8BIT((uint8_t)((a)&0xF0u)); ST7920_SWSPI_SND_8BIT((uint8_t)((a)<<4u)); U8G_DELAY(); }
+#define ST7920_WRITE_BYTES(p,l)  { for (uint8_t i = l + 1; --i;) { ST7920_SWSPI_SND_8BIT(*p&0xF0); ST7920_SWSPI_SND_8BIT(*p<<4); p++; } U8G_DELAY(); }
 
 uint8_t u8g_dev_rrd_st7920_128x64_fn(u8g_t *u8g, u8g_dev_t *dev, uint8_t msg, void *arg) {
   uint8_t i, y;
     case U8G_DEV_MSG_INIT: {
       OUT_WRITE(ST7920_CS_PIN, LOW);
       OUT_WRITE(ST7920_DAT_PIN, LOW);
-      OUT_WRITE(ST7920_CLK_PIN, LOW);
+      OUT_WRITE(ST7920_CLK_PIN, HIGH);
 
       ST7920_CS();
       u8g_Delay(120);                 //initial delay for boot up
       ST7920_SET_CMD();
-      ST7920_WRITE_NIBBLES(0x08);       //display off, cursor+blink off
-      ST7920_WRITE_NIBBLES(0x01);       //clear CGRAM ram
+      ST7920_WRITE_BYTE(0x08);       //display off, cursor+blink off
+      ST7920_WRITE_BYTE(0x01);       //clear CGRAM ram
       u8g_Delay(15);                 //delay for CGRAM clear
-      ST7920_WRITE_NIBBLES(0x3E);       //extended mode + GDRAM active
+      ST7920_WRITE_BYTE(0x3E);       //extended mode + GDRAM active
       for (y = 0; y < (LCD_PIXEL_HEIGHT) / 2; y++) { //clear GDRAM
-        ST7920_WRITE_NIBBLES(0x80 | y); //set y
-        ST7920_WRITE_NIBBLES(0x80);     //set x = 0
+        ST7920_WRITE_BYTE(0x80 | y); //set y
+        ST7920_WRITE_BYTE(0x80);     //set x = 0
         ST7920_SET_DAT();
         for (i = 0; i < 2 * (LCD_PIXEL_WIDTH) / 8; i++) //2x width clears both segments
-          ST7920_WRITE_NIBBLES(0);
+          ST7920_WRITE_BYTE(0);
         ST7920_SET_CMD();
       }
-
-      ST7920_WRITE_NIBBLES(0x0C); //display on, cursor+blink off
-
+      ST7920_WRITE_BYTE(0x0C); //display on, cursor+blink off
       ST7920_NCS();
     }
     break;
       for (i = 0; i < PAGE_HEIGHT; i ++) {
         ST7920_SET_CMD();
         if (y < 32) {
-          ST7920_WRITE_NIBBLES(0x80 | y);       //y
-          ST7920_WRITE_NIBBLES(0x80);           //x=0
+          ST7920_WRITE_BYTE(0x80 | y);       //y
+          ST7920_WRITE_BYTE(0x80);           //x=0
         }
         else {
-          ST7920_WRITE_NIBBLES(0x80 | (y - 32)); //y
-          ST7920_WRITE_NIBBLES(0x80 | 8);       //x=64
+          ST7920_WRITE_BYTE(0x80 | (y - 32)); //y
+          ST7920_WRITE_BYTE(0x80 | 8);       //x=64
         }
         ST7920_SET_DAT();
-        ST7920_WRITE_NIBBLES_P(ptr, (LCD_PIXEL_WIDTH) / 8); //ptr is incremented inside of macro
+        ST7920_WRITE_BYTES(ptr, (LCD_PIXEL_WIDTH) / 8); //ptr is incremented inside of macro
         y++;
       }
       ST7920_NCS();