jswrap_pb_blitScreen improvements - 41ms for h.flip() down to 35ms

Author: gfwilliams

libs/pipboy/jswrap_pipboy.c

@@ -2039,34 +2039,34 @@ void jswrap_pb_blitScreen(JsVar *image, JsVar *options) {
     getPaletteForLine4bpp(y, palette);
 
     uint32_t pixels = 0;
-    uint8_t pixelUp, pixelDn;
     for (int x=0;x<w4;x++) {
-      uint8_t p4up = *(row1++);
-      uint8_t p4 = *(row2++);
-      uint8_t p4dn = *(row3++);
+      uint16_t p4up = *(row1++); // line above for blurring
+      uint8_t p4 = *(row2++); // actual line
+      uint16_t p4dn = *(row3++); // line below for blurring
+      // Adding upper and lower lines. Spread 2bpp out so we can do 4 adds in one go.
+      p4up = (p4up | (p4up << 6)) & 0x3333; // 0bAABBCCDD -> 0b00AA00CC 00BB00DD
+      p4dn = (p4dn | (p4dn << 6)) & 0x3333;
+      uint16_t p4sum = p4up + p4dn; // Now 0b0AAA0CCC 0BBB0DDD
       uint32_t filterResult32;
+      uint8_t pixelSum;
       #define BLITPIXEL \
         filterResult32 = (pixels&mulx1) + ((pixels&mulx2)<<1); /* Apply 8x 2bit filter */ \
         filterResult32 = (filterResult32&0x0F0F0F0F) + ((filterResult32>>4)&0x0F0F0F0F) ; /* Merge each nibble, now we have 8 bits to play with */ \
         filterResult32 = filterResult32 + (filterResult32>>8); \
         filterResult32 = (filterResult32 + (filterResult32>>16)) & 255; \
-        filterResult32 = ((((pixelUp+pixelDn)<<1) + filterResult32)*3)>>3; /* blur upper and lower lines */ \
+        filterResult32 = ((pixelSum + filterResult32)*3)>>3; /* blur upper and lower lines */ \
         if (filterResult32>15) filterResult32=15; \
         lcdFSMC_blitPixel(palette[filterResult32]);
-      pixelUp = (p4up>>5);
-      pixelDn = (p4dn>>5);
+      pixelSum = (p4sum>>11)&14;
       pixels = (pixels<<4) | (p4>>6);
       BLITPIXEL
-      pixelUp = (p4up>>3)&6;
-      pixelDn = (p4dn>>3)&6;
+      pixelSum = (p4sum>>3)&14;
       pixels = (pixels<<4) | ((p4>>4)&3);
       BLITPIXEL
-      pixelUp = (p4up>>1)&6;
-      pixelDn = (p4dn>>1)&6;
+      pixelSum = (p4sum>>7)&14;
       pixels = (pixels<<4) | ((p4>>2)&3);
       BLITPIXEL
-      pixelUp = (p4up&3)<<1;
-      pixelDn = (p4dn&3)<<1;
+      pixelSum = (p4sum<<1)&14;
       pixels = (pixels<<4) | (p4&3);
       BLITPIXEL
       #undef BLITPIXEL
@@ -2422,12 +2422,12 @@ void jswrap_pb_init() {
 #endif
   bool hasBootCode = jsfFindFile(jsfNameFromString(".bootcde"), NULL) ;
   if (
-#ifndef LINUX 
-  upgradeInProgress || 
+#ifndef LINUX
+  upgradeInProgress ||
 #endif
   res || !hasBootCode) {
     JsVar *msg;
-#ifndef LINUX 
+#ifndef LINUX
     if (upgradeInProgress) msg = jsvNewFromString("UPGRADE IN PROGRESS");
     else
 #endif