a bit more speedup

Firmware/sdk/apps/gameboy/include/cardboy/apps/peanut_gb.h

@@ -568,7 +568,6 @@ union cart_rtc
 	} reg;
 	uint8_t bytes[5];
 };
-extern "C" uint8_t gb_rom_read(struct gb_s*, const uint_fast32_t);
 
 /**
  * Emulator context.
@@ -730,9 +729,14 @@ struct gb_s
 
 		/* Implementation defined data. Set to NULL if not required. */
 		void *priv;
+		const uint8_t* rom;
 	} direct;
 };
 
+__attribute__((always_inline)) inline uint8_t gb_rom_read(struct gb_s* s, const uint_fast32_t addr) {
+	return s->direct.rom[addr];
+}
+
 #ifndef PEANUT_GB_HEADER_ONLY
 
 #define IO_JOYP	0x00
@@ -3687,7 +3691,7 @@ enum gb_init_error_e gb_init(struct gb_s *gb,
 			     uint8_t (*gb_cart_ram_read)(struct gb_s*, const uint_fast32_t),
 			     void (*gb_cart_ram_write)(struct gb_s*, const uint_fast32_t, const uint8_t),
 			     void (*gb_error)(struct gb_s*, const enum gb_error_e, const uint16_t),
-			     void *priv)
+			     void *priv, const uint8_t* rom_data)
 {
 	const uint16_t mbc_location = 0x0147;
 	const uint16_t bank_count_location = 0x0148;
@@ -3726,6 +3730,7 @@ enum gb_init_error_e gb_init(struct gb_s *gb,
 	gb->gb_cart_ram_write = gb_cart_ram_write;
 	gb->gb_error = gb_error;
 	gb->direct.priv = priv;
+	gb->direct.rom = rom_data;
 
 	/* Initialise serial transfer function to NULL. If the front-end does
 	 * not provide serial support, Peanut-GB will emulate no cable connected
@@ -3879,7 +3884,7 @@ enum gb_init_error_e gb_init(struct gb_s *gb,
 			     uint8_t (*gb_cart_ram_read)(struct gb_s*, const uint_fast32_t),
 			     void (*gb_cart_ram_write)(struct gb_s*, const uint_fast32_t, const uint8_t),
 			     void (*gb_error)(struct gb_s*, const enum gb_error_e, const uint16_t),
-			     void *priv);
+			     void *priv, const uint8_t* rom_data);
 
 /**
  * Executes the emulator and runs for the duration of time equal to one frame.

Firmware/sdk/apps/gameboy/src/gameboy_app.cpp

@@ -76,6 +76,29 @@ static constexpr LUTFull buildNibbleLUT() {
 
 inline constexpr LUTFull kNibbleLUT = buildNibbleLUT();
 
+static constexpr LUTFull buildFullHeightWideByteLUT() {
+    LUTFull L{};
+    for (int yp = 0; yp < 2; ++yp)
+        for (int xp = 0; xp < 2; ++xp)
+            for (int p = 0; p < 256; ++p) {
+                const uint8_t p0 = static_cast<uint8_t>(p & 0x03);
+                const uint8_t p1 = static_cast<uint8_t>((p >> 2) & 0x03);
+                const uint8_t p2 = static_cast<uint8_t>((p >> 4) & 0x03);
+                const uint8_t p3 = static_cast<uint8_t>((p >> 6) & 0x03);
+
+                const uint8_t p4a = static_cast<uint8_t>(p0 | (p0 << 2) | (p1 << 4) | (p1 << 6));
+                const uint8_t p4b = static_cast<uint8_t>(p2 | (p2 << 2) | (p3 << 4) | (p3 << 6));
+
+                const uint8_t n0 = makeNibble(p4a, xp, yp);
+                const uint8_t n1 = makeNibble(p4b, xp, yp);
+
+                L[yp][xp][p] = static_cast<uint8_t>((n0 << 4) | (n1 & 0x0F));
+            }
+    return L;
+}
+
+inline constexpr LUTFull kFullHeightWideByteLUT = buildFullHeightWideByteLUT();
+
 
 namespace apps {
 namespace {
@@ -869,8 +892,8 @@ private:
         }
 
         std::memset(&gb, 0, sizeof(gb));
-        const auto initResult =
+        const auto initResult = gb_init(&gb, &GameboyApp::cartRamRead, &GameboyApp::cartRamWrite,
-                gb_init(&gb, &GameboyApp::cartRamRead, &GameboyApp::cartRamWrite, &GameboyApp::errorCallback, this);
+                                        &GameboyApp::errorCallback, this, romDataView);
         if (initResult != GB_INIT_NO_ERROR) {
             setStatus(initErrorToString(initResult));
             romData.clear();
@@ -1276,23 +1299,17 @@ private:
         CARDBOY_CHECK(yEnd > yStart);
         CARDBOY_CHECK((kFullHeightWideOffsetX % 8) == 0);
 
+        const int xParityBase = kFullHeightWideOffsetX & 1;
+
         for (int dstY = yStart; dstY < yEnd; ++dstY) {
             const int yParity = dstY & 1;
             int       dstX    = kFullHeightWideOffsetX;
 
             for (int srcX = 0; srcX < LCD_WIDTH; srcX += 4) {
-                const uint8_t p0 = pixels[srcX + 0];
+                const uint8_t packIndex = static_cast<uint8_t>((pixels[srcX + 0]) | ((pixels[srcX + 1]) << 2) |
-                const uint8_t p1 = pixels[srcX + 1];
+                                                               ((pixels[srcX + 2]) << 4) | ((pixels[srcX + 3]) << 6));
-                const uint8_t p2 = pixels[srcX + 2];
-                const uint8_t p3 = pixels[srcX + 3];
 
-                const uint8_t p4a = static_cast<uint8_t>(p0 | (p0 << 2) | (p1 << 4) | (p1 << 6));
+                const uint8_t pack = kFullHeightWideByteLUT[yParity][xParityBase][packIndex];
-                const uint8_t p4b = static_cast<uint8_t>(p2 | (p2 << 2) | (p3 << 4) | (p3 << 6));
-
-                const int     xParity = dstX & 1;
-                const uint8_t n0      = kNibbleLUT[yParity][xParity][p4a];
-                const uint8_t n1      = kNibbleLUT[yParity][xParity][p4b];
-                const uint8_t pack    = static_cast<uint8_t>((n0 << 4) | (n1 & 0x0F));
 
                 fb.drawBits8(dstX, dstY, pack);
                 dstX += 8;
@@ -1403,9 +1420,6 @@ private:
     }
 };
 
-extern "C" __attribute__((always_inline)) uint8_t gb_rom_read(struct gb_s* gb, const uint_fast32_t addr) {
-    return GameboyApp::romRead(gb, addr);
-}
 class GameboyAppFactory final : public cardboy::sdk::IAppFactory {
 public:
     const char*                         name() const override { return kGameboyAppName; }