diff --git a/libgambatte/src/cpu.h b/libgambatte/src/cpu.h
index 2b75eef7a4..85df4edfe6 100644
--- a/libgambatte/src/cpu.h
+++ b/libgambatte/src/cpu.h
@@ -73,8 +73,8 @@ public:
 		mem_.setLinkCallback(callback);
 	}
 
-	LoadRes load(char const *romfiledata, unsigned romfilelength, bool forceDmg, bool multicartCompat) {
-		return mem_.loadROM(romfiledata, romfilelength, forceDmg, multicartCompat);
+	LoadRes load(char const *romfiledata, unsigned romfilelength, unsigned flags) {
+		return mem_.loadROM(romfiledata, romfilelength, flags);
 	}
 
 	bool loaded() const { return mem_.loaded(); }
@@ -94,7 +94,6 @@ public:
 	void setRtcDivisorOffset(long const rtcDivisorOffset) { mem_.setRtcDivisorOffset(rtcDivisorOffset); }
 
 	void setBios(char const *buffer, std::size_t size) { mem_.setBios(buffer, size); }
-	bool gbIsCgb() { return mem_.gbIsCgb(); }
 
 	unsigned char externalRead(unsigned short addr) {return mem_.peek(addr); }
 
diff --git a/libgambatte/src/gambatte.cpp b/libgambatte/src/gambatte.cpp
index 0bc0639a76..50a3447fce 100644
--- a/libgambatte/src/gambatte.cpp
+++ b/libgambatte/src/gambatte.cpp
@@ -88,7 +88,7 @@ void GB::reset() {
 
 		SaveState state;
 		p_->cpu.setStatePtrs(state);
-		setInitState(state, !(p_->loadflags & FORCE_DMG), p_->loadflags & GBA_CGB);
+		setInitState(state, !(p_->loadflags & FORCE_DMG));
 		p_->cpu.loadState(state);
 		if (length > 0)
 		{
@@ -139,13 +139,13 @@ void GB::setRtcDivisorOffset(long const rtcDivisorOffset) {
 }
 
 LoadRes GB::load(char const *romfiledata, unsigned romfilelength, unsigned const flags) {
-	LoadRes const loadres = p_->cpu.load(romfiledata, romfilelength, flags & FORCE_DMG, flags & MULTICART_COMPAT);
+	LoadRes const loadres = p_->cpu.load(romfiledata, romfilelength, flags);
 
 	if (loadres == LOADRES_OK) {
 		SaveState state;
 		p_->cpu.setStatePtrs(state);
 		p_->loadflags = flags;
-		setInitState(state, !(flags & FORCE_DMG), flags & GBA_CGB);
+		setInitState(state, !(flags & FORCE_DMG));
 		p_->cpu.loadState(state);
 	}
 
diff --git a/libgambatte/src/initstate.cpp b/libgambatte/src/initstate.cpp
index d71543e0ad..fb3588dc1b 100644
--- a/libgambatte/src/initstate.cpp
+++ b/libgambatte/src/initstate.cpp
@@ -1147,7 +1147,7 @@ static void setInitialDmgIoamhram(unsigned char ioamhram[]) {
 
 } // anon namespace
 
-void gambatte::setInitState(SaveState &state, const bool cgb, const bool gbaCgbMode) {
+void gambatte::setInitState(SaveState &state, const bool cgb) {
 	static unsigned char const cgbObjpDump[0x40] = {
 		0x00, 0x00, 0xF2, 0xAB,
 		0x61, 0xC2, 0xD9, 0xBA,
@@ -1182,8 +1182,6 @@ void gambatte::setInitState(SaveState &state, const bool cgb, const bool gbaCgbM
 	state.cpu.prefetched = false;
 	state.cpu.skip = false;
 	state.mem.biosMode = true;
-	state.mem.cgbSwitching = false;
-	state.mem.agbMode = gbaCgbMode;
 
 	std::memset(state.mem.sram.ptr, 0xFF, state.mem.sram.size());
 
@@ -1222,7 +1220,6 @@ void gambatte::setInitState(SaveState &state, const bool cgb, const bool gbaCgbM
 	state.mem.enableRam = false;
 	state.mem.rambankMode = false;
 	state.mem.hdmaTransfer = false;
-	state.mem.gbIsCgb = cgb;
 	state.mem.stopped = false;
 
 
@@ -1269,7 +1266,7 @@ void gambatte::setInitState(SaveState &state, const bool cgb, const bool gbaCgbM
 	state.ppu.nextM0Irq = 0;
 	state.ppu.oldWy = state.mem.ioamhram.get()[0x14A];
 	state.ppu.pendingLcdstatIrq = false;
-	state.ppu.isCgb = cgb;
+	state.ppu.notCgbDmg = true;
 
 	// spu.cycleCounter >> 12 & 7 represents the frame sequencer position.
 	state.spu.cycleCounter = state.cpu.cycleCounter >> 1;
diff --git a/libgambatte/src/initstate.h b/libgambatte/src/initstate.h
index 0e31dd35bb..13b0f28e16 100644
--- a/libgambatte/src/initstate.h
+++ b/libgambatte/src/initstate.h
@@ -23,7 +23,7 @@
 
 namespace gambatte {
 
-void setInitState(struct SaveState &state, bool cgb, bool gbaCgbMode);
+void setInitState(struct SaveState &state, bool cgb);
 }
 
 #endif
diff --git a/libgambatte/src/mem/time.cpp b/libgambatte/src/mem/time.cpp
index 87ffb7e10c..f85433bea9 100644
--- a/libgambatte/src/mem/time.cpp
+++ b/libgambatte/src/mem/time.cpp
@@ -43,7 +43,7 @@ void Time::loadState(SaveState const &state) {
 	lastTime_.tv_sec = state.time.lastTimeSec;
 	lastTime_.tv_usec = state.time.lastTimeUsec;
 	lastCycles_ = state.time.lastCycles;
-	ds_ = state.ppu.isCgb & state.mem.ioamhram.get()[0x14D] >> 7;
+	ds_ = state.mem.ioamhram.get()[0x14D] >> 7;
 }
 
 std::uint32_t Time::get(unsigned long const cc) {
diff --git a/libgambatte/src/memory.cpp b/libgambatte/src/memory.cpp
index 95a02ea4a4..3b255d7d62 100644
--- a/libgambatte/src/memory.cpp
+++ b/libgambatte/src/memory.cpp
@@ -17,6 +17,7 @@
 //
 
 #include "memory.h"
+#include "gambatte.h"
 #include "savestate.h"
 #include "sound.h"
 #include "video.h"
@@ -80,9 +81,6 @@ void Memory::setStatePtrs(SaveState &state) {
 
 void Memory::loadState(SaveState const &state) {
 	biosMode_ = state.mem.biosMode;
-	cgbSwitching_ = state.mem.cgbSwitching;
-	agbMode_ = state.mem.agbMode;
-	gbIsCgb_ = state.mem.gbIsCgb;
 	stopped_ = state.mem.stopped;
 	psg_.loadState(state);
 	lcd_.loadState(state, state.mem.oamDmaPos < oam_size ? cart_.rdisabledRam() : ioamhram_);
@@ -1014,17 +1012,17 @@ void Memory::nontrivial_ff_write(unsigned const p, unsigned data, unsigned long
 		oamDmaInitSetup();
 		return;
 	case 0x47:
-		if (!isCgb())
+		if (!isCgb() || isCgbDmg())
 			lcd_.dmgBgPaletteChange(data, cc);
 
 		break;
 	case 0x48:
-		if (!isCgb())
+		if (!isCgb() || isCgbDmg())
 			lcd_.dmgSpPalette1Change(data, cc);
 
 		break;
 	case 0x49:
-		if (!isCgb())
+		if (!isCgb() || isCgbDmg())
 			lcd_.dmgSpPalette2Change(data, cc);
 
 		break;
@@ -1033,25 +1031,34 @@ void Memory::nontrivial_ff_write(unsigned const p, unsigned data, unsigned long
 		break;
 	case 0x4B:
 		lcd_.wxChange(data, cc);
+		break;
+	case 0x4C:
+		if (!biosMode_)
+			return;
+
 		break;
 	case 0x4D:
-		if (isCgb())
+		if (isCgb() && !isCgbDmg())
 			ioamhram_[0x14D] = (ioamhram_[0x14D] & ~1u) | (data & 1);
 
 		return;
 	case 0x4F:
-		if (isCgb()) {
+		if (isCgb() && !isCgbDmg()) {
 			cart_.setVrambank(data & 1);
 			ioamhram_[0x14F] = 0xFE | data;
 		}
 
 		return;
 	case 0x50:
-		biosMode_ = false;
-		if(cgbSwitching_) {
+		if (!biosMode_)
+			return;
+
+		if (isCgb() && (ioamhram_[0x14C] & 0x04)) {
 			lcd_.copyCgbPalettesToDmg();
-			lcd_.setCgb(false);
+			lcd_.setCgbDmg(true);
 		}
+
+		biosMode_ = false;
 		return;
 	case 0x51:
 		dmaSource_ = data << 8 | (dmaSource_ & 0xFF);
@@ -1120,14 +1127,12 @@ void Memory::nontrivial_ff_write(unsigned const p, unsigned data, unsigned long
 
 		return;
 	case 0x6C:
-		if (isCgb()) {
+		if (isCgb())
 			ioamhram_[0x16C] = data | 0xFE;
-			cgbSwitching_ = true;
-		}
 
 		return;
 	case 0x70:
-		if (isCgb()) {
+		if (isCgb() && !isCgbDmg()) {
 			cart_.setWrambank(data & 0x07 ? data & 0x07 : 1);
 			ioamhram_[0x170] = data | 0xF8;
 		}
@@ -1211,13 +1216,18 @@ void Memory::nontrivial_write(unsigned const p, unsigned const data, unsigned lo
 		ioamhram_[p - mm_oam_begin] = data;
 }
 
-LoadRes Memory::loadROM(char const *romfiledata, unsigned romfilelength, const bool forceDmg, const bool multicartCompat) {
+LoadRes Memory::loadROM(char const *romfiledata, unsigned romfilelength, unsigned const flags) {
+	bool const forceDmg = flags & GB::LoadFlag::FORCE_DMG;
+	bool const multicartCompat = flags & GB::LoadFlag::MULTICART_COMPAT;
+
 	if (LoadRes const fail = cart_.loadROM(romfiledata, romfilelength, forceDmg, multicartCompat))
 		return fail;
 
 	psg_.init(cart_.isCgb());
 	lcd_.reset(ioamhram_, cart_.vramdata(), cart_.isCgb());
 
+	agbMode_ = flags & GB::LoadFlag::GBA_CGB;
+
 	return LOADRES_OK;
 }
 
@@ -1300,9 +1310,6 @@ SYNCFUNC(Memory)
 	NSS(serialCnt_);
 	NSS(blanklcd_);
 	NSS(biosMode_);
-	NSS(cgbSwitching_);
-	NSS(agbMode_);
-	NSS(gbIsCgb_);
 	NSS(stopped_);
 	NSS(LINKCABLE_);
 	NSS(linkClockTrigger_);
diff --git a/libgambatte/src/memory.h b/libgambatte/src/memory.h
index babf029853..601c0bf0c9 100644
--- a/libgambatte/src/memory.h
+++ b/libgambatte/src/memory.h
@@ -55,12 +55,12 @@ public:
 		memcpy(bios_, buffer, size);
 		biosSize_ = size;
 	}
-	bool gbIsCgb() { return gbIsCgb_; }
 
 	bool getMemoryArea(int which, unsigned char **data, int *length);
 
 	unsigned long stop(unsigned long cycleCounter, bool& skip);
 	bool isCgb() const { return lcd_.isCgb(); }
+	bool isCgbDmg() const { return lcd_.isCgbDmg(); }
 	bool ime() const { return intreq_.ime(); }
 	bool halted() const { return intreq_.halted(); }
 	unsigned long nextEventTime() const { return intreq_.minEventTime(); }
@@ -226,7 +226,7 @@ public:
 
 	unsigned long event(unsigned long cycleCounter);
 	unsigned long resetCounters(unsigned long cycleCounter);
-	LoadRes loadROM(char const *romfiledata, unsigned romfilelength, bool forceDmg, bool multicartCompat);
+	LoadRes loadROM(char const *romfiledata, unsigned romfilelength, unsigned flags);
 
 	void setInputGetter(unsigned (*getInput)()) {
 		getInput_ = getInput;
@@ -299,9 +299,7 @@ private:
 	unsigned char serialCnt_;
 	bool blanklcd_;
 	bool biosMode_;
-	bool cgbSwitching_;
 	bool agbMode_;
-	bool gbIsCgb_;
 	unsigned long basetime_;
 	bool stopped_;
 	enum HdmaState { hdma_low, hdma_high, hdma_requested } haltHdmaState_;
diff --git a/libgambatte/src/savestate.h b/libgambatte/src/savestate.h
index 5e5bf02349..e07908cacf 100644
--- a/libgambatte/src/savestate.h
+++ b/libgambatte/src/savestate.h
@@ -36,7 +36,7 @@ struct SaveState {
 		void set(T *p, std::size_t size) { ptr = p; size_ = size; }
 
 		friend class SaverList;
-		friend void setInitState(SaveState &, bool, bool);
+		friend void setInitState(SaveState &, bool);
 	private:
 		T *ptr;
 		std::size_t size_;
@@ -83,9 +83,6 @@ struct SaveState {
 		unsigned char /*bool*/ rambankMode;
 		unsigned char /*bool*/ hdmaTransfer;
 		unsigned char /*bool*/ biosMode;
-		unsigned char /*bool*/ cgbSwitching;
-		unsigned char /*bool*/ agbMode;
-		unsigned char /*bool*/ gbIsCgb;
 		unsigned char /*bool*/ stopped;
 	} mem;
 
@@ -122,7 +119,7 @@ struct SaveState {
 		unsigned char wscx;
 		unsigned char /*bool*/ weMaster;
 		unsigned char /*bool*/ pendingLcdstatIrq;
-		unsigned char /*bool*/ isCgb;
+		unsigned char /*bool*/ notCgbDmg;
 	} ppu;
 
 	struct SPU {
diff --git a/libgambatte/src/sound/channel4.cpp b/libgambatte/src/sound/channel4.cpp
index 2b3f99533c..17c7fde5a1 100644
--- a/libgambatte/src/sound/channel4.cpp
+++ b/libgambatte/src/sound/channel4.cpp
@@ -96,6 +96,7 @@ inline void Channel4::Lfsr::event() {
 void Channel4::Lfsr::nr3Change(unsigned newNr3, unsigned long cc) {
 	updateBackupCounter(cc);
 	nr3_ = newNr3;
+	counter_ = cc;
 }
 
 void Channel4::Lfsr::nr4Init(unsigned long cc) {
diff --git a/libgambatte/src/video.cpp b/libgambatte/src/video.cpp
index ce58c4a392..732d5e4572 100644
--- a/libgambatte/src/video.cpp
+++ b/libgambatte/src/video.cpp
@@ -24,8 +24,8 @@
 using namespace gambatte;
 
 unsigned long LCD::gbcToRgb32(const unsigned bgr15) {
-	unsigned long const r = bgr15       & 0x1F;
-	unsigned long const g = bgr15 >>  5 & 0x1F;
+	unsigned long const r = bgr15 & 0x1F;
+	unsigned long const g = bgr15 >> 5 & 0x1F;
 	unsigned long const b = bgr15 >> 10 & 0x1F;
 
 	return cgbColorsRgb32_[bgr15 & 0x7FFF];
@@ -69,18 +69,13 @@ namespace {
 			&& cc >= m0TimeOfCurrentLy;
 	}
 
-	void doCgbColorChange(unsigned char* pdata,
-		unsigned long* palette, unsigned index, unsigned data, bool trueColor) {
-		pdata[index] = data;
-		index /= 2;
-		palette[index] = gbcToRgb32(pdata[index * 2] | pdata[index * 2 + 1] * 0x100l);
-	}
-
 } // unnamed namespace.
 
 void LCD::setDmgPalette(unsigned long palette[], const unsigned long dmgColors[], unsigned data) {
-		for (int i = 0; i < num_palette_entries; ++i, data /= num_palette_entries)
-		palette[i] = gbcToRgb32(dmgColors[data % num_palette_entries]);
+	palette[0] = dmgColors[data & 3];
+	palette[1] = dmgColors[data >> 2 & 3];
+	palette[2] = dmgColors[data >> 4 & 3];
+	palette[3] = dmgColors[data >> 6 & 3];
 }
 
 void LCD::setCgbPalette(unsigned *lut) {
@@ -114,10 +109,6 @@ void LCD::reset(unsigned char const *oamram, unsigned char const *vram, bool cgb
 	refreshPalettes();
 }
 
-void LCD::setCgb(bool cgb) {
-	ppu_.setCgb(cgb);
-}
-
 void LCD::setStatePtrs(SaveState &state) {
 	state.ppu.bgpData.set(  bgpData_, sizeof  bgpData_);
 	state.ppu.objpData.set(objpData_, sizeof objpData_);
@@ -162,7 +153,7 @@ void LCD::loadState(SaveState const &state, unsigned char const *const oamram) {
 }
 
 void LCD::refreshPalettes() {
-	if (ppu_.cgb()) {
+	if (isCgb() && !isCgbDmg()) {
 		for (int i = 0; i < max_num_palettes * num_palette_entries; ++i) {
 			ppu_.bgPalette()[i] = gbcToRgb32(bgpData_[2 * i] | bgpData_[2 * i + 1] * 0x100l);
 			ppu_.spPalette()[i] = gbcToRgb32(objpData_[2 * i] | objpData_[2 * i + 1] * 0x100l);
@@ -186,7 +177,7 @@ void LCD::copyCgbPalettesToDmg() {
 namespace {
 
 template<typename T>
-static void clear(T *buf, unsigned long color, std::ptrdiff_t dpitch) {
+void clear(T *buf, unsigned long color, std::ptrdiff_t dpitch) {
 	unsigned lines = 144;
 
 	while (lines--) {
@@ -856,6 +847,4 @@ SYNCFUNC(LCD)
 	SSS(lycIrq_);
 	SSS(nextM0Time_);
 	NSS(statReg_);
-	NSS(m2IrqStatReg_);
-	NSS(m1IrqStatReg_);
 }
diff --git a/libgambatte/src/video.h b/libgambatte/src/video.h
index cedd2cf86a..ab8b25475c 100644
--- a/libgambatte/src/video.h
+++ b/libgambatte/src/video.h
@@ -50,13 +50,13 @@ public:
 	LCD(unsigned char const *oamram, unsigned char const *vram,
 	    VideoInterruptRequester memEventRequester);
 	void reset(unsigned char const *oamram, unsigned char const *vram, bool cgb);
+	void setCgbDmg(bool enabled) { ppu_.setCgbDmg(enabled); }
 	void setStatePtrs(SaveState &state);
 	void loadState(SaveState const &state, unsigned char const *oamram);
 	void setDmgPaletteColor(unsigned palNum, unsigned colorNum, unsigned long rgb32);
 	void setCgbPalette(unsigned *lut);
 	void setVideoBuffer(uint_least32_t *videoBuf, std::ptrdiff_t pitch);
 	void setLayers(unsigned mask) { ppu_.setLayers(mask); }
-	void setCgb(bool cgb);
 	void copyCgbPalettesToDmg();
 
 	int debugGetLY() const { return ppu_.lyCounter().ly(); }
@@ -147,6 +147,7 @@ public:
 	bool hdmaIsEnabled() const { return eventTimes_(memevent_hdma) != disabled_time; }
 	void update(unsigned long cycleCounter);
 	bool isCgb() const { return ppu_.cgb(); }
+	bool isCgbDmg() const { return ppu_.cgbDmg(); }
 	bool isDoubleSpeed() const { return ppu_.lyCounter().isDoubleSpeed(); }
 
 	unsigned long *bgPalette() { return ppu_.bgPalette(); }
diff --git a/libgambatte/src/video/ppu.cpp b/libgambatte/src/video/ppu.cpp
index 257a7906a3..61a8e470db 100644
--- a/libgambatte/src/video/ppu.cpp
+++ b/libgambatte/src/video/ppu.cpp
@@ -140,6 +140,13 @@ inline void nextCall(int const cycles, PPUState const &state, PPUPriv &p) {
 	p.nextCallPtr = &state;
 }
 
+inline unsigned long const* cgbSpPalette(PPUPriv const& p, unsigned const attrib) {
+	if (!p.cgbDmg)
+		return p.spPalette + (attrib & attr_cgbpalno) * num_palette_entries;
+	else
+		return p.spPalette + (attrib & attr_dmgpalno ? num_palette_entries : 0);
+}
+
 namespace M2_Ly0 {
 	void f0(PPUPriv& p) {
 		p.weMaster = lcdcWinEn(p) && 0 == p.wy;
@@ -444,6 +451,7 @@ namespace M3Loop {
 						+ expand_lut[(tileDataLine + ts * tno - 2 * ts * (tno & tileIndexSign))[1]] * 2;
 				} while (dst != dstend);
 
+
 				p.ntileword = ntileword;
 				continue;
 			}
@@ -455,6 +463,7 @@ namespace M3Loop {
 				p.cycles = cycles;
 			}
 
+
 			uint_least32_t* const dst = dbufline + (xpos - tile_len);
 			unsigned const tileword = -(p.lcdc & 1u * lcdc_bgen) & p.ntileword;
 
@@ -541,6 +550,7 @@ namespace M3Loop {
 				} while (i >= 0 && spx(p.spriteList[i]) > xpos - tile_len);
 			}
 
+
 			unsigned const tno = tileMapLine[tileMapXpos % tile_map_len];
 			int const ts = tile_size;
 			tileMapXpos = tileMapXpos % tile_map_len + 1;
@@ -766,6 +776,7 @@ namespace M3Loop {
 				} while (i >= 0 && spx(p.spriteList[i]) > xpos - tile_len);
 			}
 
+
 			{
 				unsigned const tno = tileMapLine[tileMapXpos % tile_map_len];
 				unsigned const nattrib = tileMapLine[tileMapXpos % tile_map_len + vram_bank_size];
@@ -786,7 +797,6 @@ namespace M3Loop {
 		p.xpos = xpos;
 	}
 
-
 	void doFullTilesUnrolled(PPUPriv& p) {
 		int xpos = p.xpos;
 		int const xend = p.wx < p.xpos || p.wx >= xpos_end
@@ -1768,6 +1778,7 @@ void PPU::loadState(SaveState const& ss, unsigned char const* const oamram) {
 void PPU::reset(unsigned char const *oamram, unsigned char const *vram, bool cgb) {
 	p_.vram = vram;
 	p_.cgb = cgb;
+	p_.cgbDmg = false;
 	p_.spriteMapper.reset(oamram, cgb);
 }
 
diff --git a/libgambatte/src/video/ppu.h b/libgambatte/src/video/ppu.h
index 927613b3f1..b711cea3dd 100644
--- a/libgambatte/src/video/ppu.h
+++ b/libgambatte/src/video/ppu.h
@@ -150,7 +150,6 @@ public:
 	unsigned long * spPalette() { return p_.spPalette; }
 	void update(unsigned long cc);
 	void setLayers(unsigned mask) { p_.layersMask = mask; }
-	void setCgb(bool cgb) { p_.cgb = cgb; }
 
 private:
 	PPUPriv p_;