Merge branch 'OptimizeLine'

jtothebell · Sep 5, 2021 · dc8bacd · dc8bacd
2 parents 5f8b0c1 + f55b6c4
commit dc8bacd
Show file tree

Hide file tree

Showing 6 changed files with 459 additions and 156 deletions.
diff --git a/platform/SDL1_2/Makefile b/platform/SDL1_2/Makefile
@@ -16,13 +16,18 @@ INCLUDES	:=	${INCLUDES}
 #---------------------------------------------------------------------------------
 CC = $(CXX)
 
-CFLAGS	:=	-g -Wall -ffunction-sections -DVER_STR=\"$(APP_VERSION)\" \
+CFLAGS	:=	-g -Wall -Ofast -ffunction-sections -DVER_STR=\"$(APP_VERSION)\" \
 			$(DEFINES)
 
 CFLAGS	+=	$(INCLUDE)
 
 CXXFLAGS	:= $(CFLAGS) -fno-rtti -std=gnu++17
 
+#to profile, link to gperftools profiler
+#LIBS	:= -Wl,--no-as-needed -lprofiler -Wl,--as-needed -lSDL 
+#alternatively, use operf
+#sudo operf ./platform/SDL1_2/FAKE08 ~/p8carts/png-x-zero.p8.png 
+#opreport --demangle=smart --symbols > fake08-x-zero-master-preoptim2.txt
 LIBS	:= -lSDL 
 
 LDFLAGS	:= $(LIBS)

diff --git a/platform/bittboy/source/BittBoyHost.cpp b/platform/bittboy/source/BittBoyHost.cpp
@@ -41,6 +41,7 @@ int _maxNoStretchHeight = 128;
 
 const int PicoScreenWidth = 128;
 const int PicoScreenHeight = 128;
+const int pixelBlocksPerLine = PicoScreenWidth / 8;
 
 
 StretchOption stretch;
@@ -493,12 +494,50 @@ void Host::drawFrame(uint8_t* picoFb, uint8_t* screenPaletteMap, uint8_t drawMod
     }
     else { //default
         for (int y = 0; y < PicoScreenHeight; y ++){
-            for (int x = 0; x < PicoScreenWidth; x ++){
-                uint8_t c = getPixelNibble(x, y, picoFb);
-                uint16_t col = _mapped16BitColors[screenPaletteMap[c]];
-
-                base = ((uint16_t *)pixels) + ( y * PicoScreenHeight + x);
-                base[0] = col;
+            for (int x = 0; x < pixelBlocksPerLine; x ++){
+                int32_t eightPix = ((int32_t*)picoFb)[y * pixelBlocksPerLine + x];
+
+                int h = (eightPix >> 28) & 0x0f;
+                int g = (eightPix >> 24) & 0x0f;
+                int f = (eightPix >> 20) & 0x0f;
+                int e = (eightPix >> 16) & 0x0f;
+                int d = (eightPix >> 12) & 0x0f;
+                int c = (eightPix >>  8) & 0x0f;
+                int b = (eightPix >>  4) & 0x0f;
+                int a = (eightPix)       & 0x0f;
+
+                int32_t cola = _mapped16BitColors[screenPaletteMap[a]];
+                int32_t colb = _mapped16BitColors[screenPaletteMap[b]];
+                int32_t colc = _mapped16BitColors[screenPaletteMap[c]];
+                int32_t cold = _mapped16BitColors[screenPaletteMap[d]];
+                int32_t cole = _mapped16BitColors[screenPaletteMap[e]];
+                int32_t colf = _mapped16BitColors[screenPaletteMap[f]];
+                int32_t colg = _mapped16BitColors[screenPaletteMap[g]];
+                int32_t colh = _mapped16BitColors[screenPaletteMap[h]];
+
+
+                base = ((uint16_t *)pixels + (y * PicoScreenHeight + x * 8));
+                base[0] = cola;
+                base[1] = colb;
+                base[2] = colc;
+                base[3] = cold;
+                base[4] = cole;
+                base[5] = colf;
+                base[6] = colg;
+                base[7] = colh;
+
+
+                //----OR something like this for further optimization?
+                //not exactly this. its broken
+                /*
+                int32_t* base32 = ((int32_t *)pixels + (y * PicoScreenHeight + x * 8));
+                base32[0] = cola << 16 & colb;
+                base32[1] = colc << 16 & cold;
+                base32[2] = cole << 16 & colf;
+                base32[3] = colg << 16 & colh;
+                */
+
+
             }
         }
     }