diff --git a/Extensions/Smooth_font.cpp b/Extensions/Smooth_font.cpp
index 7d573be..fe77885 100644
--- a/Extensions/Smooth_font.cpp
+++ b/Extensions/Smooth_font.cpp
@@ -78,6 +78,12 @@ void TFT_eSPI::loadFont(String fontName)
     
   _gFontFilename = "/" + fontName + ".vlw";
 
+  // Avoid a crash on the ESP32 if the file does not exist
+  if (SPIFFS.exists(_gFontFilename) == false) {
+    Serial.println("Font file " + fontName + " not found!");
+    return;
+  }
+
   fontFile = SPIFFS.open( _gFontFilename, "r");
 
   if(!fontFile) return;
@@ -479,7 +485,6 @@ void TFT_eSPI::drawGlyph(uint16_t code)
 void TFT_eSPI::showFont(uint32_t td)
 {
   if(!fontLoaded) return;
-//  fontFile = SPIFFS.open( _gFontFilename, "r" );
 
   if(!fontFile)
   {
diff --git a/TFT_Drivers/ST7735_Defines.h b/TFT_Drivers/ST7735_Defines.h
index 82a3c30..494c673 100644
--- a/TFT_Drivers/ST7735_Defines.h
+++ b/TFT_Drivers/ST7735_Defines.h
@@ -9,15 +9,15 @@
 
 
 // Enumerate the different configurations
-#define INITR_GREENTAB  0x0
-#define INITR_REDTAB    0x1
-#define INITR_BLACKTAB  0x2
-#define INITR_GREENTAB2 0x3 // Use if you get random pixels on two edges of green tab display
-#define INITR_GREENTAB3 0x4 // Use if you get random pixels on edge(s) of 128x128 screen
-#define INITR_GREENTAB128 0x5 // Use if you only get part of 128x128 screen in rotation 0 & 1
+#define INITR_GREENTAB       0x0
+#define INITR_REDTAB         0x1
+#define INITR_BLACKTAB       0x2 // Display with no offsets
+#define INITR_GREENTAB2      0x3 // Use if you get random pixels on two edges of green tab display
+#define INITR_GREENTAB3      0x4 // Use if you get random pixels on edge(s) of 128x128 screen
+#define INITR_GREENTAB128    0x5 // Use if you only get part of 128x128 screen in rotation 0 & 1
 #define INITR_GREENTAB160x80 0x6 // Use if you only get part of 128x128 screen in rotation 0 & 1
-#define INITR_REDTAB160x80 0x7 // Added for https://www.aliexpress.com/item/ShengYang-1pcs-IPS-0-96-inch-7P-SPI-HD-65K-Full-Color-OLED-Module-ST7735-Drive/32918394604.html
-#define INITB           0xB
+#define INITR_REDTAB160x80   0x7 // Added for https://www.aliexpress.com/item/ShengYang-1pcs-IPS-0-96-inch-7P-SPI-HD-65K-Full-Color-OLED-Module-ST7735-Drive/32918394604.html
+#define INITB                0xB
 
 
 // Setup the tab color that will be used by the library setRotation() and setup command list
diff --git a/TFT_eSPI.cpp b/TFT_eSPI.cpp
index afaedd1..10f9bb3 100644
--- a/TFT_eSPI.cpp
+++ b/TFT_eSPI.cpp
@@ -185,8 +185,8 @@ TFT_eSPI::TFT_eSPI(int16_t w, int16_t h)
 
   _booted = true;
 
-  addr_row = 0xFFFF;
-  addr_col = 0xFFFF;
+  ys_row = ye_row = addr_row = 0xFFFF;
+  xs_col = xe_col = addr_col = 0xFFFF;
 
 #ifdef LOAD_GLCD
   fontsloaded  = 0x0002; // Bit 1 set
@@ -530,7 +530,7 @@ uint8_t TFT_eSPI::readcommand8(uint8_t cmd_function, uint8_t index)
   CS_L;
   tft_Write_8(cmd_function);
   DC_D;
-  reg = tft_Write_8(0);
+  reg = tft_Read_8(0);
   CS_H;
 
   spi_end_read();
@@ -628,15 +628,15 @@ uint16_t TFT_eSPI::readPixel(int32_t x0, int32_t y0)
     // Read the 3 RGB bytes, colour is actually only in the top 6 bits of each byte
     // as the TFT stores colours as 18 bits
   #if !defined (ILI9488_DRIVER)
-    uint8_t r = tft_Write_8(0);
-    uint8_t g = tft_Write_8(0);
-    uint8_t b = tft_Write_8(0);
+    uint8_t r = tft_Read_8(0);
+    uint8_t g = tft_Read_8(0);
+    uint8_t b = tft_Read_8(0);
   #else
     // The 6 colour bits are in MS 6 bits of each byte, but the ILI9488 needs an extra clock pulse
     // so bits appear shifted right 1 bit, so mask the middle 6 bits then shift 1 place left
-    uint8_t r = (tft_Write_8(0)&0x7E)<<1;
-    uint8_t g = (tft_Write_8(0)&0x7E)<<1;
-    uint8_t b = (tft_Write_8(0)&0x7E)<<1;
+    uint8_t r = (tft_Read_8(0)&0x7E)<<1;
+    uint8_t g = (tft_Read_8(0)&0x7E)<<1;
+    uint8_t b = (tft_Read_8(0)&0x7E)<<1;
   #endif
 
   CS_H;
@@ -768,15 +768,15 @@ void TFT_eSPI::readRect(uint32_t x, uint32_t y, uint32_t w, uint32_t h, uint16_t
     // Read the 3 RGB bytes, colour is actually only in the top 6 bits of each byte
     // as the TFT stores colours as 18 bits
   #if !defined (ILI9488_DRIVER)
-    uint8_t r = tft_Write_8(0);
-    uint8_t g = tft_Write_8(0);
-    uint8_t b = tft_Write_8(0);
+    uint8_t r = tft_Read_8(0);
+    uint8_t g = tft_Read_8(0);
+    uint8_t b = tft_Read_8(0);
   #else
     // The 6 colour bits are in LS 6 bits of each byte but we do not include the extra clock pulse
     // so we use a trick and mask the middle 6 bits of the byte, then only shift 1 place left
-    uint8_t r = (tft_Write_8(0)&0x7E)<<1;
-    uint8_t g = (tft_Write_8(0)&0x7E)<<1;
-    uint8_t b = (tft_Write_8(0)&0x7E)<<1;
+    uint8_t r = (tft_Read_8(0)&0x7E)<<1;
+    uint8_t g = (tft_Read_8(0)&0x7E)<<1;
+    uint8_t b = (tft_Read_8(0)&0x7E)<<1;
   #endif
 
     // Swapped colour byte order for compatibility with pushRect()
@@ -1354,15 +1354,15 @@ void  TFT_eSPI::readRectRGB(int32_t x0, int32_t y0, int32_t w, int32_t h, uint8_
     // Read the 3 RGB bytes, colour is actually only in the top 6 bits of each byte
     // as the TFT stores colours as 18 bits
   #if !defined (ILI9488_DRIVER)
-    *data++ = tft_Write_8(0);
-    *data++ = tft_Write_8(0);
-    *data++ = tft_Write_8(0);
+    *data++ = tft_Read_8(0);
+    *data++ = tft_Read_8(0);
+    *data++ = tft_Read_8(0);
   #else
     // The 6 colour bits are in MS 6 bits of each byte, but the ILI9488 needs an extra clock pulse
     // so bits appear shifted right 1 bit, so mask the middle 6 bits then shift 1 place left
-    *data++ = (tft_Write_8(0)&0x7E)<<1;
-    *data++ = (tft_Write_8(0)&0x7E)<<1;
-    *data++ = (tft_Write_8(0)&0x7E)<<1;
+    *data++ = (tft_Read_8(0)&0x7E)<<1;
+    *data++ = (tft_Read_8(0)&0x7E)<<1;
+    *data++ = (tft_Read_8(0)&0x7E)<<1;
   #endif
 
   }
@@ -2473,7 +2473,7 @@ void TFT_eSPI::setWindow(int16_t x0, int16_t y0, int16_t x1, int16_t y1)
 // Chip select stays low, use setWindow() from sketches
 
 #if defined (ESP8266) && !defined (RPI_WRITE_STROBE) && !defined (RPI_ILI9486_DRIVER)
-inline void TFT_eSPI::setAddrWindow(int32_t xs, int32_t ys, int32_t xe, int32_t ye)
+void TFT_eSPI::setAddrWindow(int32_t xs, int32_t ys, int32_t xe, int32_t ye)
 {
   //spi_begin();
 
@@ -2563,8 +2563,8 @@ void TFT_eSPI::setAddrWindow(int32_t xs, int32_t ys, int32_t xe, int32_t ye)
 
   DC_D;
 
-  uint8_t xBin[] = { 0, (uint8_t) (xs>>8), 0, (uint8_t) (xs>>0), 0, (uint8_t) (xe>>8), 0, (uint8_t) (xe>>0), };
-  SPI.writePattern(&xBin[0], 8, 1);
+  uint8_t xb[] = { 0, (uint8_t) (xs>>8), 0, (uint8_t) (xs>>0), 0, (uint8_t) (xe>>8), 0, (uint8_t) (xe>>0), };
+  SPI.writePattern(&xb[0], 8, 1);
 
   // Row addr set
   DC_C;
@@ -2577,8 +2577,8 @@ void TFT_eSPI::setAddrWindow(int32_t xs, int32_t ys, int32_t xe, int32_t ye)
 
   DC_D;
 
-  uint8_t yBin[] = { 0, (uint8_t) (ys>>8), 0, (uint8_t) (ys>>0), 0, (uint8_t) (ye>>8), 0, (uint8_t) (ye>>0), };
-  SPI.writePattern(&yBin[0], 8, 1);
+  uint8_t yb[] = { 0, (uint8_t) (ys>>8), 0, (uint8_t) (ys>>0), 0, (uint8_t) (ye>>8), 0, (uint8_t) (ye>>0), };
+  SPI.writePattern(&yb[0], 8, 1);
 
   // write to RAM
   DC_C;
@@ -2596,7 +2596,7 @@ void TFT_eSPI::setAddrWindow(int32_t xs, int32_t ys, int32_t xe, int32_t ye)
 #else
 
 #if defined (ESP8266) && defined (RPI_ILI9486_DRIVER) // This is for the RPi display that needs 16 bits
-inline void TFT_eSPI::setAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t y1)
+void TFT_eSPI::setAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t y1)
 {
   //spi_begin();
 
@@ -2673,7 +2673,7 @@ inline void TFT_eSPI::setAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t
 
 #else // This is for the ESP32
 
-inline void TFT_eSPI::setAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t y1)
+void TFT_eSPI::setAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t y1)
 {
   //spi_begin();
 
@@ -2687,42 +2687,48 @@ inline void TFT_eSPI::setAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t
   y1+=rowstart;
 #endif
 
-#if !defined (RPI_ILI9486_DRIVER)
-  uint32_t xaw = ((uint32_t)x0 << 16) | x1;
-  uint32_t yaw = ((uint32_t)y0 << 16) | y1;
-#endif
+  CS_L_DC_C;
 
-  // Column addr set
-  DC_C;
-  CS_L;
+  // No need to send x if it has not changed (small speed-up of complex transparent Sprites and bitmaps)
+  // Nb: caller must fill the set address area so that setting same area again will fill from xs, ys
+  if (xs_col != x0 || xe_col != x1) {
+    // Column addr set
 
-  tft_Write_8(TFT_CASET);
+    tft_Write_8(TFT_CASET);
 
-  DC_D;
+    DC_D;
 
 #if defined (RPI_ILI9486_DRIVER)
-  uint8_t xBin[] = { 0, (uint8_t) (x0>>8), 0, (uint8_t) (x0>>0), 0, (uint8_t) (x1>>8), 0, (uint8_t) (x1>>0), };
-  SPI.writePattern(&xBin[0], 8, 1);
+    uint8_t xb[] = { 0, (uint8_t) (x0>>8), 0, (uint8_t) (x0>>0), 0, (uint8_t) (x1>>8), 0, (uint8_t) (x1>>0), };
+    SPI.writePattern(&xb[0], 8, 1);
 #else
-  tft_Write_32(xaw);
+    tft_Write_32(SPI_32(x0, x1));
 #endif
 
-  // Row addr set
-  DC_C;
-  tft_Write_8(TFT_PASET);
+    DC_C;
+    xs_col = x0; xe_col = x1;
+  }
 
-  DC_D;
+  // No need to send y if it has not changed
+  if (ys_row != y0 || ye_row != y1) {
+
+    // Row addr set
+    tft_Write_8(TFT_PASET);
+
+    DC_D;
 
 #if defined (RPI_ILI9486_DRIVER)
-  uint8_t yBin[] = { 0, (uint8_t) (y0>>8), 0, (uint8_t) (y0>>0), 0, (uint8_t) (y1>>8), 0, (uint8_t) (y1>>0), };
-  SPI.writePattern(&yBin[0], 8, 1);
+    uint8_t yb[] = { 0, (uint8_t) (y0>>8), 0, (uint8_t) (y0>>0), 0, (uint8_t) (y1>>8), 0, (uint8_t) (y1>>0), };
+    SPI.writePattern(&yb[0], 8, 1);
 #else
-  tft_Write_32(yaw);
+    tft_Write_32(SPI_32(y0, y1));
 #endif
 
+    DC_C;
+    ys_row = y0; ye_row = y1;
+  }
+
   // write to RAM
-  DC_C;
-
   tft_Write_8(TFT_RAMWR);
 
   DC_D;
@@ -2819,19 +2825,15 @@ void TFT_eSPI::readAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t y1)
   y1+=rowstart;
 #endif
 
-  uint32_t xaw = ((uint32_t)x0 << 16) | x1;
-  uint32_t yaw = ((uint32_t)y0 << 16) | y1;
-  
   // Column addr set
-  DC_C;
-  CS_L;
+  CS_L_DC_C;
 
   tft_Write_8(TFT_CASET);
 
 
   DC_D;
 
-  tft_Write_32(xaw);
+  tft_Write_32(SPI_32(x0, x1));
 
   // Row addr set
   DC_C;
@@ -2840,7 +2842,7 @@ void TFT_eSPI::readAddrWindow(int32_t x0, int32_t y0, int32_t x1, int32_t y1)
 
   DC_D;
 
-  tft_Write_32(yaw);
+  tft_Write_32(SPI_32(y0, y1));
   
   DC_C;
   tft_Write_8(TFT_RAMRD); // Read CGRAM command
@@ -3053,52 +3055,49 @@ void TFT_eSPI::drawPixel(uint32_t x, uint32_t y, uint32_t color)
   y+=rowstart;
 #endif
 
-#if !defined (RPI_ILI9486_DRIVER)
-  uint32_t xaw = ((uint32_t)x << 16) | x;
-  uint32_t yaw = ((uint32_t)y << 16) | y;
-#endif
-
-  CS_L;
+  CS_L_DC_C;
 
   // No need to send x if it has not changed (speeds things up)
   if (addr_col != x) {
 
-    DC_C;
-
     tft_Write_8(TFT_CASET);
 
     DC_D;
 
 #if defined (RPI_ILI9486_DRIVER)
-    uint8_t xBin[] = { 0, (uint8_t) (x>>8), 0, (uint8_t) (x>>0), 0, (uint8_t) (x>>8), 0, (uint8_t) (x>>0), };
-    SPI.writePattern(&xBin[0], 8, 1);
+    uint8_t xb[] = { 0, (uint8_t) (x>>8), 0, (uint8_t) (x>>0), 0, (uint8_t) (x>>8), 0, (uint8_t) (x>>0), };
+    SPI.writePattern(&xb[0], 8, 1);
 #else
-    tft_Write_32(xaw);
+    tft_Write_32(SPI_32(x, x));
 #endif
-    
+
+    DC_C;
+
     addr_col = x;
+    xs_col = xe_col = x;
+    
   }
 
   // No need to send y if it has not changed (speeds things up)
   if (addr_row != y) {
 
-    DC_C;
-
     tft_Write_8(TFT_PASET);
 
     DC_D;
 
 #if defined (RPI_ILI9486_DRIVER)
-    uint8_t yBin[] = { 0, (uint8_t) (y>>8), 0, (uint8_t) (y>>0), 0, (uint8_t) (y>>8), 0, (uint8_t) (y>>0), };
-    SPI.writePattern(&yBin[0], 8, 1);
+    uint8_t yb[] = { 0, (uint8_t) (y>>8), 0, (uint8_t) (y>>0), 0, (uint8_t) (y>>8), 0, (uint8_t) (y>>0), };
+    SPI.writePattern(&yb[0], 8, 1);
 #else
-    tft_Write_32(yaw);
+    tft_Write_32(SPI_32(y, y));
 #endif
 
+    DC_C;
+
     addr_row = y;
+    ys_row = ye_row = y;
   }
 
-  DC_C;
 
   tft_Write_8(TFT_RAMWR);
 
@@ -3159,6 +3158,48 @@ void TFT_eSPI::pushColor(uint16_t color, uint32_t len)
   spi_end();
 }
 
+/***************************************************************************************
+** Function name:           startWrite
+** Description:             begin transaction with CS low, MUST later call endWrite
+***************************************************************************************/
+void TFT_eSPI::startWrite(void)
+{
+  spi_begin();
+
+  CS_L;
+}
+
+/***************************************************************************************
+** Function name:           endWrite
+** Description:             end transaction with CS high
+***************************************************************************************/
+void TFT_eSPI::endWrite(void)
+{
+
+  CS_H;
+
+  spi_end();
+}
+
+/***************************************************************************************
+** Function name:           writeColor (use startWrite() and endWrite() before & after)
+** Description:             raw write of "len" pixels avoiding transaction check
+***************************************************************************************/
+void TFT_eSPI::writeColor(uint16_t color, uint32_t len)
+{
+#ifdef RPI_WRITE_STROBE
+  uint8_t colorBin[] = { (uint8_t) (color >> 8), (uint8_t) color };
+  if(len) SPI.writePattern(&colorBin[0], 2, 1); len--;
+  while(len--) {WR_L; WR_H;}
+#else
+  #if defined (ESP32_PARALLEL)
+    while (len--) {tft_Write_16(color);}
+  #else
+    writeBlock(color, len);
+  #endif
+#endif
+}
+
 /***************************************************************************************
 ** Function name:           pushColors
 ** Description:             push an array of pixels for 16 bit raw image drawing
@@ -4655,7 +4696,7 @@ void writeBlock(uint16_t color, uint32_t repeat)
 #elif defined (ILI9488_DRIVER)
 
 #ifdef ESP8266
-void writeBlock(uint16_t color, uint32_t repeat)
+void TFT_eSPI::writeBlock(uint16_t color, uint32_t repeat)
 {
 
   uint32_t mask = ~(SPIMMOSI << SPILMOSI);
@@ -4722,9 +4763,6 @@ void writeBlock(uint16_t color, uint32_t repeat)
 }
 #else // Now the code for ESP32 and ILI9488
 
-#include "soc/spi_reg.h"
-#define SPI_NUM 0x3
-
 void writeBlock(uint16_t color, uint32_t repeat)
 {
   // Split out the colours
@@ -4795,38 +4833,46 @@ void writeBlock(uint16_t color, uint32_t repeat)
 
 #else // Low level register based ESP32 code for 16 bit colour SPI TFTs
 
-#include "soc/spi_reg.h"
-#define SPI_NUM 0x3
-
-
 void writeBlock(uint16_t color, uint32_t repeat)
 {
-  uint16_t color16 = (color >> 8) | (color << 8);
-  uint32_t color32 = color16 | color16 << 16;
+  uint32_t color32 = SPI_32(color, color);
 
-  if (repeat > 15)
+  if (repeat > 31) // Revert legacy toggle buffer change
   {
-    SET_PERI_REG_BITS(SPI_MOSI_DLEN_REG(SPI_NUM), SPI_USR_MOSI_DBITLEN, 255, SPI_USR_MOSI_DBITLEN_S);
-
-    while(repeat>15)
+    WRITE_PERI_REG(SPI_MOSI_DLEN_REG(SPI_NUM), 511);
+    while(repeat>31)
     {
       while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
-      for (uint32_t i=0; i<8; i++) WRITE_PERI_REG((SPI_W0_REG(SPI_NUM) + (i << 2)), color32);
+      WRITE_PERI_REG(SPI_W0_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W1_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W2_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W3_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W4_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W5_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W6_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W7_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W8_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W9_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W10_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W11_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W12_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W13_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W14_REG(SPI_NUM), color32);
+      WRITE_PERI_REG(SPI_W15_REG(SPI_NUM), color32);
       SET_PERI_REG_MASK(SPI_CMD_REG(SPI_NUM), SPI_USR);
-      repeat -= 16;
+      repeat -= 32;
     }
     while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
   }
 
   if (repeat)
   {
-    repeat = (repeat << 4) - 1;
-    SET_PERI_REG_BITS(SPI_MOSI_DLEN_REG(SPI_NUM), SPI_USR_MOSI_DBITLEN, repeat, SPI_USR_MOSI_DBITLEN_S);
-    for (uint32_t i=0; i<8; i++) WRITE_PERI_REG((SPI_W0_REG(SPI_NUM) + (i << 2)), color32);
+    // Revert toggle buffer change
+    WRITE_PERI_REG(SPI_MOSI_DLEN_REG(SPI_NUM), (repeat << 4) - 1);
+    for (uint32_t i=0; i <= (repeat>>1); i++) WRITE_PERI_REG((SPI_W0_REG(SPI_NUM) + (i << 2)), color32);
     SET_PERI_REG_MASK(SPI_CMD_REG(SPI_NUM), SPI_USR);
     while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
   }
-
 }
 #endif
 
@@ -4975,10 +5021,6 @@ void TFT_eSPI::getSetup(setup_t &tft_settings)
 
 #include "Extensions/Sprite.cpp"
 
-//  #ifdef ESP32
-//    #include "Extensions/pSprite.cpp"
-//  #endif
-
 #ifdef SMOOTH_FONT
   #include "Extensions/Smooth_font.cpp"
 #endif
diff --git a/TFT_eSPI.h b/TFT_eSPI.h
index 2c93392..456b066 100644
--- a/TFT_eSPI.h
+++ b/TFT_eSPI.h
@@ -103,6 +103,11 @@
 
 #include <SPI.h>
 
+#ifdef ESP32
+  #include "soc/spi_reg.h"
+  #define SPI_NUM 0x3
+#endif
+
 #ifdef SMOOTH_FONT
   // Call up the SPIFFS FLASH filing system for the anti-aliased fonts
   #define FS_NO_GLOBALS
@@ -127,16 +132,16 @@
 
     #else
       #if TFT_DC >= 32
-        #define DC_C GPIO.out1_w1ts.val = (1 << (TFT_DC - 32)); \
-                     GPIO.out1_w1tc.val = (1 << (TFT_DC - 32))
-        #define DC_D GPIO.out1_w1tc.val = (1 << (TFT_DC - 32)); \
-                     GPIO.out1_w1ts.val = (1 << (TFT_DC - 32))
+        #define DC_C GPIO.out1_w1tc.val = (1 << (TFT_DC - 32)); //\
+                     //GPIO.out1_w1tc.val = (1 << (TFT_DC - 32))
+        #define DC_D GPIO.out1_w1ts.val = (1 << (TFT_DC - 32)); //\
+                     //GPIO.out1_w1ts.val = (1 << (TFT_DC - 32))
       #else
         #if TFT_DC >= 0
-          #define DC_C GPIO.out_w1ts = (1 << TFT_DC); \
-                       GPIO.out_w1tc = (1 << TFT_DC)
-          #define DC_D GPIO.out_w1tc = (1 << TFT_DC); \
-                       GPIO.out_w1ts = (1 << TFT_DC)
+          #define DC_C GPIO.out_w1tc = (1 << TFT_DC); //\
+                       //GPIO.out_w1tc = (1 << TFT_DC)
+          #define DC_D GPIO.out_w1ts = (1 << TFT_DC); //\
+                       //GPIO.out_w1ts = (1 << TFT_DC)
         #else
           #define DC_C
           #define DC_D
@@ -166,14 +171,14 @@
       #define CS_H
     #else
       #if TFT_CS >= 32
-        #define CS_L GPIO.out1_w1ts.val = (1 << (TFT_CS - 32)); \
-                     GPIO.out1_w1tc.val = (1 << (TFT_CS - 32))
-        #define CS_H GPIO.out1_w1tc.val = (1 << (TFT_CS - 32)); \
-                     GPIO.out1_w1ts.val = (1 << (TFT_CS - 32))
+        #define CS_L GPIO.out1_w1tc.val = (1 << (TFT_CS - 32)); //\
+                     //GPIO.out1_w1tc.val = (1 << (TFT_CS - 32))
+        #define CS_H GPIO.out1_w1ts.val = (1 << (TFT_CS - 32)); //\
+                     //GPIO.out1_w1ts.val = (1 << (TFT_CS - 32))
       #else
         #if TFT_CS >= 0
-          #define CS_L GPIO.out_w1ts = (1 << TFT_CS);GPIO.out_w1tc = (1 << TFT_CS)
-          #define CS_H GPIO.out_w1tc = (1 << TFT_CS);GPIO.out_w1ts = (1 << TFT_CS)
+          #define CS_L GPIO.out_w1tc = (1 << TFT_CS); //GPIO.out_w1tc = (1 << TFT_CS)
+          #define CS_H GPIO.out_w1ts = (1 << TFT_CS); //GPIO.out_w1ts = (1 << TFT_CS)
         #else
           #define CS_L
           #define CS_H
@@ -186,6 +191,18 @@
   #endif
 #endif
 
+// Use single register write for CS_L and DC_C if pins are both in range 0-31
+#ifdef ESP32
+  #if (TFT_CS >= 0) && (TFT_CS < 32) && (TFT_DC >= 0) && (TFT_DC < 32)
+    #define CS_L_DC_C GPIO.out_w1tc = ((1 << TFT_CS) | (1 << TFT_DC)); //\
+                    //GPIO.out_w1tc = ((1 << TFT_CS) | (1 << TFT_DC))
+  #else
+    #define CS_L_DC_C CS_L; DC_C
+  #endif
+#else // ESP8266
+  #define CS_L_DC_C CS_L; DC_C
+#endif
+
 // chip select signal for touchscreen
 #ifndef TOUCH_CS
   #define T_CS_L // No macro allocated so it generates no code
@@ -206,6 +223,14 @@
   #endif
 #endif
 
+#ifdef ESP8266
+  // Concatenate two 16 bit values for the SPI 32 bit register write
+  #define SPI_32(H,L) ( (H)<<16 | (L) )
+#else
+  // Swap byte order for concatenated 16 bit window address or colors
+  // AB CD -> DCBA for SPI 32 bit register write
+  #define SPI_32(H,L) ( ((H)<<8 | (H)>>8) | (((L)<<8 | (L)>>8)<<16 ) )
+#endif
 
 #if defined (ESP32) && defined (ESP32_PARALLEL)
   // Mask for the 8 data bits to set pin directions
@@ -253,34 +278,75 @@
 #elif  defined (ILI9488_DRIVER) // 16 bit colour converted to 3 bytes for 18 bit RGB
 
   // Write 8 bits to TFT
-  #define tft_Write_8(C)  SPI.transfer(C)
+  #define tft_Write_8(C)   SPI.transfer(C)
 
   // Convert 16 bit colour to 18 bit and write in 3 bytes
-  #define tft_Write_16(C) SPI.transfer((C & 0xF800)>>8); \
-                          SPI.transfer((C & 0x07E0)>>3); \
-                          SPI.transfer((C & 0x001F)<<3)
+  #define tft_Write_16(C)  SPI.transfer((C & 0xF800)>>8); \
+                           SPI.transfer((C & 0x07E0)>>3); \
+                           SPI.transfer((C & 0x001F)<<3)
 
   // Convert swapped byte 16 bit colour to 18 bit and write in 3 bytes
   #define tft_Write_16S(C) SPI.transfer(C & 0xF8); \
                            SPI.transfer((C & 0xE0)>>11 | (C & 0x07)<<5); \
                            SPI.transfer((C & 0x1F00)>>5)
   // Write 32 bits to TFT
-  #define tft_Write_32(C) SPI.write32(C)
+  #define tft_Write_32(C)  SPI.write32(C)
 
 #elif  defined (RPI_ILI9486_DRIVER)
 
-  #define tft_Write_8(C)  SPI.transfer(0); SPI.transfer(C)
-  #define tft_Write_16(C) SPI.write16(C)
-  #define tft_Write_32(C) SPI.write32(C)
+  #define tft_Write_8(C)   SPI.transfer(0); SPI.transfer(C)
+  #define tft_Write_16(C)  SPI.write16(C)
+  #define tft_Write_16S(C) SPI.write16(C<<8 | C>>8)
+  #define tft_Write_32(C)  SPI.write32(C)
 
-#else
+#elif defined ESP8266
 
-  #define tft_Write_8(C)  SPI.transfer(C)
-  #define tft_Write_16(C) SPI.write16(C)
-  #define tft_Write_32(C) SPI.write32(C)
+  #define tft_Write_8(C)   SPI.write(C)
+  #define tft_Write_16(C)  SPI.write16(C)
+  #define tft_Write_32(C)  SPI.write32(C)
+
+#else // ESP32 using SPI with 16 bit color display
+
+  // ESP32 low level SPI writes for 8, 16 and 32 bit values
+  // to avoid the function call overhead
+  
+  // Write 8 bits
+  #define tft_Write_8(C) \
+  WRITE_PERI_REG(SPI_MOSI_DLEN_REG(SPI_NUM), 8-1); \
+  WRITE_PERI_REG(SPI_W0_REG(SPI_NUM), C); \
+  SET_PERI_REG_MASK(SPI_CMD_REG(SPI_NUM), SPI_USR); \
+  while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
+
+  // Write 16 bits with corrected endianess for 16 bit colours
+  #define tft_Write_16(C) \
+  WRITE_PERI_REG(SPI_MOSI_DLEN_REG(SPI_NUM), 16-1); \
+  WRITE_PERI_REG(SPI_W0_REG(SPI_NUM), C<<8 | C>>8); \
+  SET_PERI_REG_MASK(SPI_CMD_REG(SPI_NUM), SPI_USR); \
+  while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
+
+  // Write 16 bits (used for ESP32_PARALLEL or ILI9488_DRIVER)
+  #define tft_Write_16S(C) \
+  WRITE_PERI_REG(SPI_MOSI_DLEN_REG(SPI_NUM), 16-1); \
+  WRITE_PERI_REG(SPI_W0_REG(SPI_NUM), C); \
+  SET_PERI_REG_MASK(SPI_CMD_REG(SPI_NUM), SPI_USR); \
+  while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
+
+  // Write 32 bits
+  #define tft_Write_32(C) \
+  WRITE_PERI_REG(SPI_MOSI_DLEN_REG(SPI_NUM), 32-1); \
+  WRITE_PERI_REG(SPI_W0_REG(SPI_NUM), C); \
+  SET_PERI_REG_MASK(SPI_CMD_REG(SPI_NUM), SPI_USR); \
+  while (READ_PERI_REG(SPI_CMD_REG(SPI_NUM))&SPI_USR);
 
 #endif
 
+
+#ifndef ESP32_PARALLEL
+  // Support SPI TFT reads (not all displays support this)
+  #define tft_Read_8(C) SPI.transfer(C)
+#endif
+
+
 #ifdef LOAD_GFXFF
   // We can include all the free fonts and they will only be built into
   // the sketch if they are used
@@ -406,6 +472,7 @@ template <typename T> static inline void
 swap_coord(T& a, T& b) { T t = a; a = b; b = t; }
 
 #ifndef min
+  // Return minimum of two numbers, may already be defined
   #define min(a,b) (((a) < (b)) ? (a) : (b))
 #endif
 
@@ -667,6 +734,10 @@ class TFT_eSPI : public Print {
 
   void     setAddrWindow(int32_t xs, int32_t ys, int32_t xe, int32_t ye);
 
+           // These 3 functions are used together, for every startWrite() there must be an endWrite()
+  void     startWrite(void); // Begin SPI transaction
+  void     writeColor(uint16_t color, uint32_t len); // Write a colour without transaction overhead
+  void     endWrite(void);   // End SPI transaction
 
   size_t   write(uint8_t);
 
@@ -713,6 +784,8 @@ class TFT_eSPI : public Print {
   uint32_t _init_width, _init_height; // Display w/h as input, used by setRotation()
   uint32_t _width, _height;           // Display w/h as modified by current rotation
   uint32_t addr_row, addr_col;
+  uint32_t xs_col, xe_col;
+  uint32_t ys_row, ye_row;
 
   uint32_t fontsloaded;
 
@@ -750,9 +823,4 @@ class TFT_eSPI : public Print {
 // Load the Sprite Class
 #include "Extensions/Sprite.h"
 
-//  #ifdef ESP32
-//    // Load the Sprite Class
-//    #include "Extensions/pSprite.h"
-//  #endif
-
 #endif
diff --git a/User_Setup.h b/User_Setup.h
index 9b2359c..e8a5d79 100644
--- a/User_Setup.h
+++ b/User_Setup.h
@@ -113,14 +113,15 @@
 // but saves pins for other functions.
 // Use NodeMCU SD0=MISO, SD1=MOSI, CLK=SCLK to connect to TFT in overlap mode
 
+// In ESP8266 overlap mode the following must be defined
+//#define TFT_SPI_OVERLAP
+
 // In ESP8266 overlap mode the TFT chip select MUST connect to pin D3
 //#define TFT_CS   PIN_D3
 //#define TFT_DC   PIN_D5  // Data Command control pin
 //#define TFT_RST  PIN_D4  // Reset pin (could connect to NodeMCU RST, see next line)
 //#define TFT_RST  -1  // Set TFT_RST to -1 if the display RESET is connected to NodeMCU RST or 3.3V
 
-// In ESP8266 overlap mode the following must be defined
-//#define TFT_SPI_OVERLAP
 
 // ###### EDIT THE PIN NUMBERS IN THE LINES FOLLOWING TO SUIT YOUR ESP32 SETUP   ######
 
@@ -136,6 +137,9 @@
 //#define TFT_RST  -1  // Set TFT_RST to -1 if display RESET is connected to ESP32 board RST
 //#define TFT_BL   32  // LED back-light (only for ST7789 with backlight control pin)
 
+//#define TOUCH_CS 21     // Chip select pin (T_CS) of touch screen
+
+//#define TFT_WR 22    // Write strobe for modified Raspberry Pi TFT only
 
 // For the M5Stack module use these #define lines
 //#define TFT_MISO 19
@@ -146,10 +150,6 @@
 //#define TFT_RST  33  // Reset pin (could connect to Arduino RESET pin)
 //#define TFT_BL   32  // LED back-light (required for M5Stack)
 
-//#define TOUCH_CS 21     // Chip select pin (T_CS) of touch screen
-
-//#define TFT_WR 22    // Write strobe for modified Raspberry Pi TFT only
-
 // ######       EDIT THE PINs BELOW TO SUIT YOUR ESP32 PARALLEL TFT SETUP        ######
 
 // The library supports 8 bit parallel TFTs with the ESP32, the pin
diff --git a/User_Setups/SetupX_Template.h b/User_Setups/SetupX_Template.h
index 9b2359c..900da04 100644
--- a/User_Setups/SetupX_Template.h
+++ b/User_Setups/SetupX_Template.h
@@ -102,7 +102,7 @@
 
 //#define TFT_BL PIN_D1  // LED back-light (only for ST7789 with backlight control pin)
 
-//#define TOUCH_CS PIN_D2     // Chip select pin (T_CS) of touch screen
+//#define TOUCH_CS PIN_D1     // Chip select pin (T_CS) of touch screen
 
 //#define TFT_WR PIN_D2       // Write strobe for modified Raspberry Pi TFT only
 
@@ -113,14 +113,15 @@
 // but saves pins for other functions.
 // Use NodeMCU SD0=MISO, SD1=MOSI, CLK=SCLK to connect to TFT in overlap mode
 
+// In ESP8266 overlap mode the following must be defined
+//#define TFT_SPI_OVERLAP
+
 // In ESP8266 overlap mode the TFT chip select MUST connect to pin D3
 //#define TFT_CS   PIN_D3
 //#define TFT_DC   PIN_D5  // Data Command control pin
 //#define TFT_RST  PIN_D4  // Reset pin (could connect to NodeMCU RST, see next line)
 //#define TFT_RST  -1  // Set TFT_RST to -1 if the display RESET is connected to NodeMCU RST or 3.3V
 
-// In ESP8266 overlap mode the following must be defined
-//#define TFT_SPI_OVERLAP
 
 // ###### EDIT THE PIN NUMBERS IN THE LINES FOLLOWING TO SUIT YOUR ESP32 SETUP   ######
 
@@ -134,8 +135,10 @@
 //#define TFT_DC    2  // Data Command control pin
 //#define TFT_RST   4  // Reset pin (could connect to RST pin)
 //#define TFT_RST  -1  // Set TFT_RST to -1 if display RESET is connected to ESP32 board RST
-//#define TFT_BL   32  // LED back-light (only for ST7789 with backlight control pin)
 
+//#define TOUCH_CS 22     // Chip select pin (T_CS) of touch screen
+
+//#define TFT_WR 21    // Write strobe for modified Raspberry Pi TFT only
 
 // For the M5Stack module use these #define lines
 //#define TFT_MISO 19
@@ -144,11 +147,7 @@
 //#define TFT_CS   14  // Chip select control pin
 //#define TFT_DC   27  // Data Command control pin
 //#define TFT_RST  33  // Reset pin (could connect to Arduino RESET pin)
-//#define TFT_BL   32  // LED back-light (required for M5Stack)
-
-//#define TOUCH_CS 21     // Chip select pin (T_CS) of touch screen
-
-//#define TFT_WR 22    // Write strobe for modified Raspberry Pi TFT only
+//#define TFT_BL   32  // LED back-light (if needed)
 
 // ######       EDIT THE PINs BELOW TO SUIT YOUR ESP32 PARALLEL TFT SETUP        ######
 
@@ -209,9 +208,9 @@
 #define LOAD_FONT2  // Font 2. Small 16 pixel high font, needs ~3534 bytes in FLASH, 96 characters
 #define LOAD_FONT4  // Font 4. Medium 26 pixel high font, needs ~5848 bytes in FLASH, 96 characters
 #define LOAD_FONT6  // Font 6. Large 48 pixel font, needs ~2666 bytes in FLASH, only characters 1234567890:-.apm
-#define LOAD_FONT7  // Font 7. 7 segment 48 pixel font, needs ~2438 bytes in FLASH, only characters 1234567890:-.
-#define LOAD_FONT8  // Font 8. Large 75 pixel font needs ~3256 bytes in FLASH, only characters 1234567890:-.
+#define LOAD_FONT7  // Font 7. 7 segment 48 pixel font, needs ~2438 bytes in FLASH, only characters 1234567890:.
 //#define LOAD_FONT8N // Font 8. Alternative to Font 8 above, slightly narrower, so 3 digits fit a 160 pixel TFT
+#define LOAD_FONT8  // Font 8. Large 75 pixel font needs ~3256 bytes in FLASH, only characters 1234567890:-.
 #define LOAD_GFXFF  // FreeFonts. Include access to the 48 Adafruit_GFX free fonts FF1 to FF48 and custom fonts
 
 // Comment out the #define below to stop the SPIFFS filing system and smooth font code being loaded
@@ -263,4 +262,4 @@
 // Transactions are automatically enabled by the library for an ESP32 (to use HAL mutex)
 // so changing it here has no effect
 
-//#define SUPPORT_TRANSACTIONS
+// #define SUPPORT_TRANSACTIONS
diff --git a/keywords.txt b/keywords.txt
index b197c5a..9fb1758 100644
--- a/keywords.txt
+++ b/keywords.txt
@@ -6,6 +6,9 @@ drawChar	KEYWORD2
 setAddrWindow	KEYWORD2
 setWindow	KEYWORD2
 readAddrWindow	KEYWORD2
+startWrite	KEYWORD2
+writeColor	KEYWORD2
+endWrite	KEYWORD2
 pushColor	KEYWORD2
 pushColors	KEYWORD2
 fillScreen	KEYWORD2
diff --git a/library.json b/library.json
index 496bd2e..0867346 100644
--- a/library.json
+++ b/library.json
@@ -1,6 +1,6 @@
 {
   "name": "TFT_eSPI",
-  "version": "1.1.4",
+  "version": "1.2.0",
   "keywords": "tft, ePaper, display, ESP8266, NodeMCU, ESP32, M5Stack, ILI9341, ST7735, ILI9163, S6D02A1, ILI9486, ST7789",
   "description": "A TFT and ePaper SPI graphics library for ESP8266 and ESP32",
   "repository":
diff --git a/library.properties b/library.properties
index fd62f0e..eb7a239 100644
--- a/library.properties
+++ b/library.properties
@@ -1,5 +1,5 @@
 name=TFT_eSPI
-version=1.1.4
+version=1.2.0
 author=Bodmer
 maintainer=Bodmer
 sentence=A fast TFT graphics library for ESP8266 and ESP32 processors for the Arduino IDE