diff --git a/README.md b/README.md
index 88c3c8a..6f32916 100644
--- a/README.md
+++ b/README.md
@@ -34,12 +34,16 @@ For the standalone you will at least need freeglut as well as openal-soft, it sh
 Also you can use this as a core for retroarch by using the Makefile in the libretro folder.    
 
 There are several defines you can add/remove for different options for both the standalone version and the retroarch core:
+AUDIO_LOWERFREQ  
+Reduces the output frequency of the emulator for more performance  
 COL_32BIT  
 Changes the output texture to be 32bit per pixel (RGBA8) instead of 16bit (RGB565)  
 COL_BGRA  
 Used in combination with COL_32BIT, this will change the output texture to BGRA8 instead of RGBA8  
 COL_TEX_BSWAP  
-In 32bit per pixel mode, this will change the byte order of the output texture, in 16bit mode this will output in BGR565    
+In 32bit per pixel mode, this will change the byte order of the output texture, in 16bit mode this will output in BGR565  
+DO_INLINE_ATTRIBS  
+Will use optimized function inline attributes for more overall performance    
 
 These defines are specific to the standalone version:  
 AUDIO_FLOAT  
diff --git a/apu.c b/apu.c
index 8feab1b..2e582d1 100644
--- a/apu.c
+++ b/apu.c
@@ -188,7 +188,11 @@ void apuInitBufs()
 	//apu.Frequency = nesPAL ? 831187 : 893415;
 	//effective frequencies for Original PPU Video out
 	//apu.Frequency = nesPAL ? 831303 : 894886;
+#if AUDIO_LOWERFREQ
+	apu.Frequency = nesPAL ? 51956 : 55930;
+#else
 	apu.Frequency = nesPAL ? 207825 : 223721;
+#endif
 	audioExpansion = 0;
 	double dt = 1.0/((double)apu.Frequency);
 	//LP at 22kHz
@@ -213,11 +217,11 @@ void apuInitBufs()
 #if AUDIO_FLOAT
 	apu.BufSizeBytes = apu.BufSize*sizeof(float);
 	apu.OutBuf = (float*)malloc(apu.BufSizeBytes);
-	printf("Audio: 32-bit Float Output\n"); 
+	printf("Audio: 32-bit Float Output at %iHz\n", apu.Frequency); 
 #else
 	apu.BufSizeBytes = apu.BufSize*sizeof(int16_t);
 	apu.OutBuf = (int16_t*)malloc(apu.BufSizeBytes);
-	printf("Audio: 16-bit Short Output\n");
+	printf("Audio: 16-bit Short Output at %iHz\n", apu.Frequency);
 #endif
 	/* https://wiki.nesdev.com/w/index.php/APU_Mixer#Lookup_Table */
 	uint8_t i;
@@ -314,7 +318,7 @@ void apuWriteDMCBuf(uint8_t val)
 
 extern bool cpu_odd_cycle;
 
-static void apuChangeMode()
+FIXNES_NOINLINE static void apuChangeMode()
 {
 	if(!cpu_odd_cycle)
 		return;
@@ -356,57 +360,49 @@ void doEnvelopeLogic(envelope_t *env)
 	//env->envelope = (env->constant ? env->vol : env->decay);
 }
 
-void sweepUpdateFreq(sweep_t *sw, uint16_t *freq)
+static void sweepUpdateFreq(sweep_t *sw, uint16_t *freq)
 {
-	uint16_t inFreq = *freq;
-	if(sw->shift > 0)
+	//any freq update causes target freq update
+	sw->targetFreq = *freq;
+	if(sw->targetFreq >= 8)
 	{
 		if(sw->negative)
 		{
-			inFreq -= (inFreq >> sw->shift);
-			if(sw->chan1 == true) inFreq--;
+			sw->targetFreq -= (sw->targetFreq >> sw->shift);
+			if(sw->chan1 == true) sw->targetFreq--;
 		}
 		else
-			inFreq += (inFreq >> sw->shift);
-	}
-	if(inFreq > 8 && (inFreq < 0x7FF))
-	{
-		sw->mute = false;
-		if(sw->enabled && sw->shift)
-			*freq = inFreq;
+			sw->targetFreq += (sw->targetFreq >> sw->shift);
+		if(sw->targetFreq <= 0x7FF)
+			sw->mute = false;
+		else //larger than freq register
+			sw->mute = true;
 	}
-	else
+	else //any input < 8 gets muted
 		sw->mute = true;
 }
 
-void doSweepLogic(sweep_t *sw, uint16_t *freq)
+static void doSweepLogic(sweep_t *sw, uint16_t *freq)
 {
-	if(sw->start)
-	{
-		uint8_t prevDiv = sw->divider;
-		sw->divider = sw->period;
-		sw->start = false;
-		if(prevDiv == 0)
-			sweepUpdateFreq(sw, freq);
-	}
-	else
+	if(sw->divider == 0)
 	{
-		if(sw->divider == 0)
+		if(sw->enabled && sw->shift && !sw->mute)
 		{
+			*freq = sw->targetFreq;
 			sweepUpdateFreq(sw, freq);
-			sw->divider = sw->period;
 		}
-		else
-			sw->divider--;
+		sw->divider = sw->period;
 	}
-	//gets clocked too little on its own?
-	/*if(inFreq < 8 || (inFreq >= 0x7FF))
-		sw->mute = true;
 	else
-		sw->mute = false;*/
+		sw->divider--;
+	if(sw->start)
+	{
+		sw->divider = sw->period;
+		sw->start = false;
+	}
 }
 
-void apuClockA()
+FIXNES_NOINLINE static void apuClockA()
 {
 	if(apu.p1LengthCtr)
 	{
@@ -426,7 +422,7 @@ void apuClockA()
 		apu.noiseLengthCtr--;
 }
 
-void apuClockB()
+FIXNES_NOINLINE static void apuClockB()
 {
 	if(apu.p1LengthCtr)
 		doEnvelopeLogic(&apu.p1Env);
@@ -442,19 +438,23 @@ void apuClockB()
 		apu.trireload = false;
 }
 
-void apuCycle()
+FIXNES_ALWAYSINLINE void apuCycle()
 {
 	uint8_t aExp = audioExpansion;
+#if AUDIO_LOWERFREQ
+	if(!(apu.apuClock&31))
+#else
 	if(!(apu.apuClock&7))
+#endif
 	{
 		if(apu.p1LengthCtr && (apu.reg[0x15] & P1_ENABLE))
 		{
-			if(!apu.p1Sweep.mute && apu.freq1 >= 8 && apu.freq1 < 0x7FF)
+			if(!apu.p1Sweep.mute)
 				apu.p1Out = apu.p1seq[apu.p1Cycle] ? (apu.p1Env.constant ? apu.p1Env.vol : apu.p1Env.decay) : 0;
 		}
 		if(apu.p2LengthCtr && (apu.reg[0x15] & P2_ENABLE))
 		{
-			if(!apu.p2Sweep.mute && apu.freq2 >= 8 && apu.freq2 < 0x7FF)
+			if(!apu.p2Sweep.mute)
 				apu.p2Out = apu.p2seq[apu.p2Cycle] ? (apu.p2Env.constant ? apu.p2Env.vol : apu.p2Env.decay) : 0;
 		}
 		if(apu.triLengthCtr && apu.triCurLinearCtr && (apu.reg[0x15] & TRI_ENABLE))
@@ -713,8 +713,6 @@ void apuSetReg00(uint16_t addr, uint8_t val)
 	apu.p1seq = pulseSeqs[val>>6];
 	apu.p1Env.constant = ((val&PULSE_CONST_V) != 0);
 	apu.p1Env.loop = apu.p1haltloop = ((val&PULSE_HALT_LOOP) != 0);
-	if(apu.freq1 > 8 && (apu.freq1 < 0x7FF))
-		apu.p1Sweep.mute = false; //to be safe
 }
 void apuSetReg01(uint16_t addr, uint8_t val)
 {
@@ -726,9 +724,8 @@ void apuSetReg01(uint16_t addr, uint8_t val)
 	apu.p1Sweep.period = (val>>4)&7;
 	apu.p1Sweep.negative = ((val&0x8) != 0);
 	apu.p1Sweep.start = true;
-	if(apu.freq1 > 8 && (apu.freq1 < 0x7FF))
-		apu.p1Sweep.mute = false; //to be safe
-	doSweepLogic(&apu.p1Sweep, &apu.freq1);
+	//adjust for new sweep regs
+	sweepUpdateFreq(&apu.p1Sweep, &apu.freq1);
 }
 void apuSetReg02(uint16_t addr, uint8_t val)
 {
@@ -736,8 +733,7 @@ void apuSetReg02(uint16_t addr, uint8_t val)
 	apu.reg[2] = val;
 	//printf("P1 time low %02x\n", val);
 	apu.freq1 = ((apu.freq1&~0xFF) | val);
-	if(apu.freq1 > 8 && (apu.freq1 < 0x7FF))
-		apu.p1Sweep.mute = false; //to be safe
+	sweepUpdateFreq(&apu.p1Sweep, &apu.freq1);
 }
 void apuSetReg03(uint16_t addr, uint8_t val)
 {
@@ -747,8 +743,7 @@ void apuSetReg03(uint16_t addr, uint8_t val)
 	if(apu.reg[0x15] & P1_ENABLE)
 		apu.p1LengthCtr = apu.lengthLookupTbl[val>>3];
 	apu.freq1 = (apu.freq1&0xFF) | ((val&7)<<8);
-	if(apu.freq1 > 8 && (apu.freq1 < 0x7FF))
-		apu.p1Sweep.mute = false; //to be safe
+	sweepUpdateFreq(&apu.p1Sweep, &apu.freq1);
 	//printf("P1 new freq %04x\n", apu.freq1);
 	apu.p1Env.start = true;
 }
@@ -760,8 +755,6 @@ void apuSetReg04(uint16_t addr, uint8_t val)
 	apu.p2seq = pulseSeqs[val>>6];
 	apu.p2Env.constant = ((val&PULSE_CONST_V) != 0);
 	apu.p2Env.loop = apu.p2haltloop = ((val&PULSE_HALT_LOOP) != 0);
-	if(apu.freq2 > 8 && (apu.freq2 < 0x7FF))
-		apu.p2Sweep.mute = false; //to be safe
 }
 void apuSetReg05(uint16_t addr, uint8_t val)
 {
@@ -773,9 +766,8 @@ void apuSetReg05(uint16_t addr, uint8_t val)
 	apu.p2Sweep.period = (val>>4)&7;
 	apu.p2Sweep.negative = ((val&0x8) != 0);
 	apu.p2Sweep.start = true;
-	if(apu.freq2 > 8 && (apu.freq2 < 0x7FF))
-		apu.p2Sweep.mute = false; //to be safe
-	doSweepLogic(&apu.p2Sweep, &apu.freq2);
+	//adjust for new sweep regs
+	sweepUpdateFreq(&apu.p2Sweep, &apu.freq2);
 }
 void apuSetReg06(uint16_t addr, uint8_t val)
 {
@@ -783,8 +775,7 @@ void apuSetReg06(uint16_t addr, uint8_t val)
 	apu.reg[6] = val;
 	//printf("P2 time low %02x\n", val);
 	apu.freq2 = ((apu.freq2&~0xFF) | val);
-	if(apu.freq2 > 8 && (apu.freq2 < 0x7FF))
-		apu.p2Sweep.mute = false; //to be safe
+	sweepUpdateFreq(&apu.p2Sweep, &apu.freq2);
 }
 void apuSetReg07(uint16_t addr, uint8_t val)
 {
@@ -794,8 +785,7 @@ void apuSetReg07(uint16_t addr, uint8_t val)
 	if(apu.reg[0x15] & P2_ENABLE)
 		apu.p2LengthCtr = apu.lengthLookupTbl[val>>3];
 	apu.freq2 = (apu.freq2&0xFF) | ((val&7)<<8);
-	if(apu.freq2 > 8 && (apu.freq2 < 0x7FF))
-		apu.p2Sweep.mute = false; //to be safe
+	sweepUpdateFreq(&apu.p2Sweep, &apu.freq2);
 	//printf("P2 new freq %04x\n", apu.freq2);
 	apu.p2Env.start = true;
 }
diff --git a/apu.h b/apu.h
index c5cc455..0953c23 100644
--- a/apu.h
+++ b/apu.h
@@ -8,6 +8,8 @@
 #ifndef _apu_h_
 #define _apu_h_
 
+#include "common.h"
+
 #define NUM_BUFFERS 4
 
 #define EXP_VRC6 (1<<0)
@@ -20,7 +22,7 @@
 void apuInitBufs();
 void apuDeinitBufs();
 void apuInit();
-void apuCycle();
+FIXNES_ALWAYSINLINE void apuCycle();
 void apuWriteDMCBuf(uint8_t val);
 uint8_t *apuGetBuf();
 uint32_t apuGetBufSize();
@@ -69,6 +71,9 @@ typedef struct _sweep_t {
 	uint8_t period;
 	uint8_t divider;
 	uint8_t shift;
+	//specifically signed for pulse 1 as a shift value of 0 and negative
+	//flag enabled would cause an overflow and a mute otherwise
+	int16_t targetFreq;
 } sweep_t;
 
 void doEnvelopeLogic(envelope_t *env);
diff --git a/build_linux.sh b/build_linux.sh
index 41299e3..bc8d9d3 100755
--- a/build_linux.sh
+++ b/build_linux.sh
@@ -2,6 +2,6 @@
 
 #Need to replace this with a makefile
 
-gcc -DZIPSUPPORT main.c apu.c audio.c audio_fds.c audio_mmc5.c audio_vrc6.c audio_vrc7.c audio_n163.c audio_s5b.c alhelpers.c cpu.c ppu.c mem.c input.c mapper.c mapperList.c fm2play.c vrc_irq.c mapper/*.c unzip/*.c -DFREEGLUT_STATIC -lglut -lopenal -lGL -lGLU -lm -Wall -lz -Wextra -O3 -flto -s -o fixNES
+gcc -DZIPSUPPORT main.c apu.c audio.c audio_fds.c audio_mmc5.c audio_vrc6.c audio_vrc7.c audio_n163.c audio_s5b.c alhelpers.c cpu.c ppu.c mem.c input.c mapper.c mapperList.c fm2play.c vrc_irq.c mapper/*.c unzip/*.c -DFREEGLUT_STATIC -lglut -lopenal -lGL -lGLU -lm -Wall -lz -Wextra -DDO_INLINE_ATTRIBS=1 -Wno-attributes -O3 -flto -s -o fixNES
 echo "Succesfully built fixNES"
 
diff --git a/build_msys32.sh b/build_msys32.sh
index f4c70e5..62e36d2 100644
--- a/build_msys32.sh
+++ b/build_msys32.sh
@@ -1,2 +1,2 @@
 #!/bin/sh
-gcc -DWINDOWS_BUILD -DZIPSUPPORT main.c apu.c audio.c audio_fds.c audio_mmc5.c audio_vrc6.c audio_vrc7.c audio_n163.c audio_s5b.c alhelpers.c cpu.c ppu.c mem.c input.c mapper.c mapperList.c fm2play.c vrc_irq.c mapper/*.c unzip/*.c -DFREEGLUT_STATIC -lfreeglut_static -lopenal32 -lopengl32 -lglu32 -lgdi32 -lwinmm -lz -Wall -Wextra -O3 -flto -s -o fixNES
\ No newline at end of file
+gcc -DWINDOWS_BUILD -DZIPSUPPORT main.c apu.c audio.c audio_fds.c audio_mmc5.c audio_vrc6.c audio_vrc7.c audio_n163.c audio_s5b.c alhelpers.c cpu.c ppu.c mem.c input.c mapper.c mapperList.c fm2play.c vrc_irq.c mapper/*.c unzip/*.c -DFREEGLUT_STATIC -lfreeglut_static -lopenal32 -lopengl32 -lglu32 -lgdi32 -lwinmm -lz -Wall -Wextra -DDO_INLINE_ATTRIBS=1 -Wno-attributes -O3 -flto -s -o fixNES
\ No newline at end of file
diff --git a/build_windows_console.bat b/build_windows_console.bat
deleted file mode 100644
index 25de828..0000000
--- a/build_windows_console.bat
+++ /dev/null
@@ -1,2 +0,0 @@
-gcc -DWINDOWS_BUILD -DZIPSUPPORT main.c apu.c audio.c audio_fds.c audio_mmc5.c audio_vrc6.c audio_vrc7.c audio_n163.c audio_s5b.c alhelpers.c cpu.c ppu.c mem.c input.c mapper.c mapperList.c fm2play.c vrc_irq.c mapper/*.c unzip/*.c -DFREEGLUT_STATIC -lfreeglut_static -lopenal32 -lopengl32 -lglu32 -lgdi32 -lwinmm -lz -Wall -Wextra -O3 -flto -s -o fixNES
-pause
\ No newline at end of file
diff --git a/build_windows_noconsole.bat b/build_windows_noconsole.bat
deleted file mode 100644
index e81209c..0000000
--- a/build_windows_noconsole.bat
+++ /dev/null
@@ -1,2 +0,0 @@
-gcc -DWINDOWS_BUILD -DZIPSUPPORT main.c apu.c audio.c audio_fds.c audio_mmc5.c audio_vrc6.c audio_vrc7.c audio_n163.c audio_s5b.c alhelpers.c cpu.c ppu.c mem.c input.c mapper.c mapperList.c fm2play.c vrc_irq.c mapper/*.c unzip/*.c -DFREEGLUT_STATIC -lfreeglut_static -lopenal32 -lopengl32 -lglu32 -lgdi32 -lwinmm -lz -Wall -Wextra -O3 -flto -s -o fixNES -Wl,--subsystem,windows
-pause 
diff --git a/common.h b/common.h
index fd08cf6..c2548e7 100644
--- a/common.h
+++ b/common.h
@@ -1,10 +1,17 @@
 #ifndef _common_h_
 #define _common_h_
 
+#if DO_INLINE_ATTRIBS
 #ifdef _MSC_VER
 #define FIXNES_NOINLINE __declspec(noinline)
+#define FIXNES_ALWAYSINLINE __forceinline
 #else
 #define FIXNES_NOINLINE __attribute__((noinline))
+#define FIXNES_ALWAYSINLINE __attribute__((always_inline))
+#endif
+#else
+#define FIXNES_NOINLINE
+#define FIXNES_ALWAYSINLINE
 #endif
 
 #endif
diff --git a/cpu.c b/cpu.c
index 84407c7..86df8f8 100644
--- a/cpu.c
+++ b/cpu.c
@@ -25,6 +25,9 @@
 #define P_FLAG_OVERFLOW (1<<6)
 #define P_FLAG_NEGATIVE (1<<7)
 
+#define CPU_OAM_DMA (1<<0)
+#define CPU_DMC_DMA (1<<1)
+
 extern bool nesPause;
 
 //used externally
@@ -41,12 +44,11 @@ static struct {
 	uint8_t instr;
 	uint8_t irq;
 	uint8_t irqMask;
+	uint8_t dma;
 	bool boot;
 	bool reset;
 	bool needsIndFix;
 	bool allow_update_irq;
-	bool oam_dma;
-	bool dmc_dma;
 	bool currently_dma;
 	bool oam_ready;
 	bool dmc_dma_dummyread;
@@ -84,8 +86,7 @@ void cpuInit()
 	cpu.instr = 0;
 	cpu.irq = 0;
 	cpu.irqMask = 0;
-	cpu.oam_dma = false;
-	cpu.dmc_dma = false;
+	cpu.dma = 0;
 	cpu.currently_dma = false;
 	cpu.oam_ready = false;
 	cpu.dmc_dma_dummyread = false;
@@ -108,7 +109,7 @@ void cpuInit()
 	nsf_endPlayback = false;
 }
 
-static void setRegStats(uint8_t reg)
+FIXNES_ALWAYSINLINE inline static void setRegStats(uint8_t reg)
 {
 	if(reg == 0)
 	{
@@ -125,7 +126,7 @@ static void setRegStats(uint8_t reg)
 	}
 }
 
-static void cpuSetARRRegs()
+FIXNES_ALWAYSINLINE inline static void cpuSetARRRegs()
 {
 	if((cpu.a & ((1<<5) | (1<<6))) == ((1<<5) | (1<<6)))
 	{
@@ -151,25 +152,25 @@ static void cpuSetARRRegs()
 
 /* Helper functions for updating reg sets */
 
-static inline void cpuSetA(uint8_t val)
+FIXNES_ALWAYSINLINE inline static void cpuSetA(uint8_t val)
 {
 	cpu.a = val;
 	setRegStats(cpu.a);
 }
 
-static inline void cpuSetX(uint8_t val)
+FIXNES_ALWAYSINLINE inline static void cpuSetX(uint8_t val)
 {
 	cpu.x = val;
 	setRegStats(cpu.x);
 }
 
-static inline void cpuSetY(uint8_t val)
+FIXNES_ALWAYSINLINE inline static void cpuSetY(uint8_t val)
 {
 	cpu.y = val;
 	setRegStats(cpu.y);
 }
 
-static inline uint8_t cpuSetTMP(uint8_t val)
+FIXNES_ALWAYSINLINE inline static uint8_t cpuSetTMP(uint8_t val)
 {
 	setRegStats(val);
 	return val;
@@ -177,25 +178,25 @@ static inline uint8_t cpuSetTMP(uint8_t val)
 
 /* Various Instructions used multiple times */
 
-static inline void cpuAND()
+FIXNES_ALWAYSINLINE inline static void cpuAND()
 {
 	cpu.a &= cpu.tmp;
 	setRegStats(cpu.a);
 }
 
-static inline void cpuORA()
+FIXNES_ALWAYSINLINE inline static void cpuORA()
 {
 	cpu.a |= cpu.tmp;
 	setRegStats(cpu.a);
 }
 
-static inline void cpuEOR()
+FIXNES_ALWAYSINLINE inline static void cpuEOR()
 {
 	cpu.a ^= cpu.tmp;
 	setRegStats(cpu.a);
 }
 
-static uint8_t cpuASL(uint8_t val)
+FIXNES_ALWAYSINLINE inline static uint8_t cpuASL(uint8_t val)
 {
 	if(val & (1<<7))
 		cpu.p |= P_FLAG_CARRY;
@@ -206,10 +207,10 @@ static uint8_t cpuASL(uint8_t val)
 	return val;
 }
 
-static void cpuASLa() { cpuSetA(cpuASL(cpu.a)); };
-static void cpuASLt() { cpu.tmp = cpuASL(cpu.tmp); };
+FIXNES_ALWAYSINLINE inline static void cpuASLa() { cpuSetA(cpuASL(cpu.a)); };
+FIXNES_ALWAYSINLINE inline static void cpuASLt() { cpu.tmp = cpuASL(cpu.tmp); };
 
-static uint8_t cpuLSR(uint8_t val)
+FIXNES_ALWAYSINLINE inline static uint8_t cpuLSR(uint8_t val)
 {
 	if(val & (1<<0))
 		cpu.p |= P_FLAG_CARRY;
@@ -220,10 +221,10 @@ static uint8_t cpuLSR(uint8_t val)
 	return val;
 }
 
-static void cpuLSRa() { cpuSetA(cpuLSR(cpu.a)); };
-static void cpuLSRt() { cpu.tmp = cpuLSR(cpu.tmp); };
+FIXNES_ALWAYSINLINE inline static void cpuLSRa() { cpuSetA(cpuLSR(cpu.a)); };
+FIXNES_ALWAYSINLINE inline static void cpuLSRt() { cpu.tmp = cpuLSR(cpu.tmp); };
 
-static uint8_t cpuROL(uint8_t val)
+FIXNES_ALWAYSINLINE inline static uint8_t cpuROL(uint8_t val)
 {
 	uint8_t oldP = cpu.p;
 	if(val & (1<<7))
@@ -237,10 +238,10 @@ static uint8_t cpuROL(uint8_t val)
 	return val;
 }
 
-static void cpuROLa() { cpuSetA(cpuROL(cpu.a)); };
-static void cpuROLt() { cpu.tmp = cpuROL(cpu.tmp); };
+FIXNES_ALWAYSINLINE inline static void cpuROLa() { cpuSetA(cpuROL(cpu.a)); };
+FIXNES_ALWAYSINLINE inline static void cpuROLt() { cpu.tmp = cpuROL(cpu.tmp); };
 
-static uint8_t cpuROR(uint8_t val)
+FIXNES_ALWAYSINLINE inline static uint8_t cpuROR(uint8_t val)
 {
 	uint8_t oldP = cpu.p;
 	if(val & (1<<0))
@@ -254,16 +255,16 @@ static uint8_t cpuROR(uint8_t val)
 	return val;
 }
 
-static void cpuRORa() { cpuSetA(cpuROR(cpu.a)); };
-static void cpuRORt() { cpu.tmp = cpuROR(cpu.tmp); };
+FIXNES_ALWAYSINLINE inline static void cpuRORa() { cpuSetA(cpuROR(cpu.a)); };
+FIXNES_ALWAYSINLINE inline static void cpuRORt() { cpu.tmp = cpuROR(cpu.tmp); };
 
-static void cpuKIL()
+FIXNES_ALWAYSINLINE inline static void cpuKIL()
 {
 	printf("Processor Requested Lock-Up at %04x\n", cpu.pc-1);
 	nesPause = true;
 }
 
-static void cpuADCv(uint8_t in)
+FIXNES_ALWAYSINLINE inline static void cpuADCv(uint8_t in)
 {
 	//use uint16_t here to easly detect carry
 	uint16_t res = cpu.a + in;
@@ -286,14 +287,14 @@ static void cpuADCv(uint8_t in)
 	cpuSetA(res);
 }
 
-static void cpuADC() { cpuADCv(cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuADC() { cpuADCv(cpu.tmp); }
 
-static void cpuSBC() { cpuADCv(~cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuSBC() { cpuADCv(~cpu.tmp); }
 
-static void cpuISC() { cpu.tmp = cpuSetTMP(cpu.tmp+1); cpuSBC(); }
+FIXNES_ALWAYSINLINE inline static void cpuISC() { cpu.tmp = cpuSetTMP(cpu.tmp+1); cpuSBC(); }
 
 
-static uint8_t cpuCMP(uint8_t reg)
+FIXNES_ALWAYSINLINE inline static uint8_t cpuCMP(uint8_t reg)
 {
 	if(reg >= cpu.tmp)
 		cpu.p |= P_FLAG_CARRY;
@@ -305,13 +306,13 @@ static uint8_t cpuCMP(uint8_t reg)
 	return cmpVal;
 }
 
-static void cpuCMPa() { cpuCMP(cpu.a); }
-static void cpuCMPx() { cpuCMP(cpu.x); }
-static void cpuCMPy() { cpuCMP(cpu.y); }
-static void cpuCMPax() { cpu.x = cpuCMP(cpu.a&cpu.x); }
-static void cpuDCP() { cpu.tmp = cpuSetTMP(cpu.tmp-1); cpuCMPa(); }
+FIXNES_ALWAYSINLINE inline static void cpuCMPa() { cpuCMP(cpu.a); }
+FIXNES_ALWAYSINLINE inline static void cpuCMPx() { cpuCMP(cpu.x); }
+FIXNES_ALWAYSINLINE inline static void cpuCMPy() { cpuCMP(cpu.y); }
+FIXNES_ALWAYSINLINE inline static void cpuCMPax() { cpu.x = cpuCMP(cpu.a&cpu.x); }
+FIXNES_ALWAYSINLINE inline static void cpuDCP() { cpu.tmp = cpuSetTMP(cpu.tmp-1); cpuCMPa(); }
 
-static void cpuBIT()
+FIXNES_ALWAYSINLINE inline static void cpuBIT()
 {
 	if((cpu.a & cpu.tmp) == 0)
 		cpu.p |= P_FLAG_ZERO;
@@ -329,20 +330,20 @@ static void cpuBIT()
 		cpu.p &= ~P_FLAG_NEGATIVE;
 }
 
-static void cpuSLO() {	cpuASLt(); cpuORA(); }
-static void cpuRLA() {	cpuROLt(); cpuAND(); }
-static void cpuSRE() { cpuLSRt(); cpuEOR(); }
-static void cpuRRA() { cpuRORt(); cpuADC(); }
-static void cpuASR() {	cpuAND(); cpuLSRa(); }
+FIXNES_ALWAYSINLINE inline static void cpuSLO() {	cpuASLt(); cpuORA(); }
+FIXNES_ALWAYSINLINE inline static void cpuRLA() {	cpuROLt(); cpuAND(); }
+FIXNES_ALWAYSINLINE inline static void cpuSRE() { cpuLSRt(); cpuEOR(); }
+FIXNES_ALWAYSINLINE inline static void cpuRRA() { cpuRORt(); cpuADC(); }
+FIXNES_ALWAYSINLINE inline static void cpuASR() {	cpuAND(); cpuLSRa(); }
 
-static void cpuARR()
+FIXNES_ALWAYSINLINE inline static void cpuARR()
 {
 	cpuAND();
 	cpuRORa();
 	cpuSetARRRegs();
 }
 
-static void cpuAAC()
+FIXNES_ALWAYSINLINE inline static void cpuAAC()
 {
 	cpuAND();
 	if(cpu.p & P_FLAG_NEGATIVE)
@@ -351,40 +352,40 @@ static void cpuAAC()
 		cpu.p &= ~P_FLAG_CARRY;
 }
 
-static void cpuCLC() { cpu.p &= ~P_FLAG_CARRY; }
-static void cpuSEC() { cpu.p |= P_FLAG_CARRY; }
+FIXNES_ALWAYSINLINE inline static void cpuCLC() { cpu.p &= ~P_FLAG_CARRY; }
+FIXNES_ALWAYSINLINE inline static void cpuSEC() { cpu.p |= P_FLAG_CARRY; }
 
-static void cpuCLI() { cpu.p_irq_req = 2; }
-static void cpuSEI() { cpu.p_irq_req = 1; }
+FIXNES_ALWAYSINLINE inline static void cpuCLI() { cpu.p_irq_req = 2; }
+FIXNES_ALWAYSINLINE inline static void cpuSEI() { cpu.p_irq_req = 1; }
 
-static void cpuCLV() { cpu.p &= ~P_FLAG_OVERFLOW; }
+FIXNES_ALWAYSINLINE inline static void cpuCLV() { cpu.p &= ~P_FLAG_OVERFLOW; }
 
-static void cpuCLD() { cpu.p &= ~P_FLAG_DECIMAL; }
-static void cpuSED() { cpu.p |= P_FLAG_DECIMAL; }
+FIXNES_ALWAYSINLINE inline static void cpuCLD() { cpu.p &= ~P_FLAG_DECIMAL; }
+FIXNES_ALWAYSINLINE inline static void cpuSED() { cpu.p |= P_FLAG_DECIMAL; }
 
-static void cpuINX() { cpuSetX(cpu.x+1); }
-static void cpuINY() { cpuSetY(cpu.y+1); }
-static void cpuINCt() { cpu.tmp = cpuSetTMP(cpu.tmp+1); }
+FIXNES_ALWAYSINLINE inline static void cpuINX() { cpuSetX(cpu.x+1); }
+FIXNES_ALWAYSINLINE inline static void cpuINY() { cpuSetY(cpu.y+1); }
+FIXNES_ALWAYSINLINE inline static void cpuINCt() { cpu.tmp = cpuSetTMP(cpu.tmp+1); }
 
-static void cpuDEX() { cpuSetX(cpu.x-1); }
-static void cpuDEY() { cpuSetY(cpu.y-1); }
-static void cpuDECt() { cpu.tmp = cpuSetTMP(cpu.tmp-1); }
+FIXNES_ALWAYSINLINE inline static void cpuDEX() { cpuSetX(cpu.x-1); }
+FIXNES_ALWAYSINLINE inline static void cpuDEY() { cpuSetY(cpu.y-1); }
+FIXNES_ALWAYSINLINE inline static void cpuDECt() { cpu.tmp = cpuSetTMP(cpu.tmp-1); }
 
-static void cpuTXA() { cpuSetA(cpu.x); }
-static void cpuTYA() { cpuSetA(cpu.y); }
-static void cpuTSX() { cpuSetX(cpu.s); }
-static void cpuTXS() { cpu.s = cpu.x; }
+FIXNES_ALWAYSINLINE inline static void cpuTXA() { cpuSetA(cpu.x); }
+FIXNES_ALWAYSINLINE inline static void cpuTYA() { cpuSetA(cpu.y); }
+FIXNES_ALWAYSINLINE inline static void cpuTSX() { cpuSetX(cpu.s); }
+FIXNES_ALWAYSINLINE inline static void cpuTXS() { cpu.s = cpu.x; }
 
-static void cpuTAY() { cpuSetY(cpu.a); }
-static void cpuTAX() { cpuSetX(cpu.a); }
+FIXNES_ALWAYSINLINE inline static void cpuTAY() { cpuSetY(cpu.a); }
+FIXNES_ALWAYSINLINE inline static void cpuTAX() { cpuSetX(cpu.a); }
 
-static void cpuLDA() { cpuSetA(cpu.tmp); }
-static void cpuLDX() { cpuSetX(cpu.tmp); }
-static void cpuLDY() { cpuSetY(cpu.tmp); }
-static void cpuXAA() { cpuSetA(cpu.x&cpu.tmp); }
-static void cpuAXT() { cpuSetA(cpu.tmp); cpuSetX(cpu.a); }
-static void cpuLAX() { cpuSetA(cpu.tmp); cpuSetX(cpu.tmp); }
-static void cpuLAR() { cpuSetA(cpu.tmp); cpuAND(cpu.s); cpuSetX(cpu.a); cpu.s = cpu.a; }
+FIXNES_ALWAYSINLINE inline static void cpuLDA() { cpuSetA(cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuLDX() { cpuSetX(cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuLDY() { cpuSetY(cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuXAA() { cpuSetA(cpu.x&cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuAXT() { cpuSetA(cpu.tmp); cpuSetX(cpu.a); }
+FIXNES_ALWAYSINLINE inline static void cpuLAX() { cpuSetA(cpu.tmp); cpuSetX(cpu.tmp); }
+FIXNES_ALWAYSINLINE inline static void cpuLAR() { cpu.a = cpu.s; cpuAND(); cpuSetX(cpu.a); cpu.s = cpu.a; }
 
 /* For Interrupt Handling */
 
@@ -541,7 +542,7 @@ static const uint8_t cpu_mem_type[CPU_STATE_END] = {
 
 static const uint8_t cpu_start_arr[1] = { CPU_GET_INSTRUCTION };
 
-static void cpuSetStartArray()
+FIXNES_ALWAYSINLINE inline static void cpuSetStartArray()
 {
 	cpu.action_arr = cpu_start_arr;
 	cpu.arr_pos = 0;
@@ -604,20 +605,20 @@ static const uint8_t cpu_bmi_arr[4] = { CPU_TMP_READ8_PC_INC_CHECK_BMI, CPU_BRAN
 static const uint8_t cpu_bvc_arr[4] = { CPU_TMP_READ8_PC_INC_CHECK_BVC, CPU_BRANCH_SETUP, CPU_NULL_READ8_PC_CHK, CPU_GET_INSTRUCTION };
 static const uint8_t cpu_bvs_arr[4] = { CPU_TMP_READ8_PC_INC_CHECK_BVS, CPU_BRANCH_SETUP, CPU_NULL_READ8_PC_CHK, CPU_GET_INSTRUCTION };
 
-static void cpuCheckIrq()
+FIXNES_ALWAYSINLINE inline static void cpuCheckIrq()
 {
 	cpu.irq |= (interrupt & cpu.irqMask) | ppuNMI();
 }
 
 /* useful for the branch checks */
-static void cpuBranchCheck(bool takeBranch)
+FIXNES_ALWAYSINLINE inline static void cpuBranchCheck(bool takeBranch)
 {
 	cpuCheckIrq();
 	if(!takeBranch) //get next instruction
 		cpuSetStartArray();
 }
 
-static void cpuBranchSetup()
+FIXNES_ALWAYSINLINE inline static void cpuBranchSetup()
 {
 	cpu.indVal = cpu.pc + (int8_t)cpu.tmp;
 	//only need extra cycle if it needs fixup
@@ -812,7 +813,7 @@ static bool cpuHandleIrqUpdates()
 	return false;
 }
 
-static void cpuSetAddrIndFix()
+FIXNES_ALWAYSINLINE inline static void cpuSetAddrIndFix()
 {
 	//first read will be at wrong pos
 	//so let cpu know it needs fixup
@@ -823,7 +824,7 @@ static void cpuSetAddrIndFix()
 	}
 }
 
-static bool cpuDoAddrIndFix()
+FIXNES_ALWAYSINLINE inline static bool cpuDoAddrIndFix()
 {
 	if(cpu.needsIndFix)
 	{
@@ -897,16 +898,16 @@ static bool cpuTryTakeover()
 	return false;
 }
 
-static void cpuDoDMA()
+FIXNES_NOINLINE static void cpuDoDMA()
 {
-	if(cpu.dmc_dma)
+	if(cpu.dma & CPU_DMC_DMA)
 	{
 		if(cpu.currently_dma && cpu.dmc_halted)
 		{
 			if(cpu.dmc_dma_dummyread) //1st read always dummy read
 			{
 				cpu.dmc_dma_dummyread = false;
-				if(!cpu.oam_dma || cpu.oam_dma_pause)
+				if(!(cpu.dma&CPU_OAM_DMA) || cpu.oam_dma_pause)
 				{
 					cpuDMATryHalt();
 					return;
@@ -916,7 +917,7 @@ static void cpuDoDMA()
 			{
 				uint8_t dmc_dma_val = memGet8(cpu.dmc_dma_addr);
 				apuWriteDMCBuf(dmc_dma_val);
-				cpu.dmc_dma = false;
+				cpu.dma &= ~CPU_DMC_DMA;
 				if(!cpu.oam_halted) //done with DMA
 					cpu.currently_dma = false;
 				//for next dmc dma
@@ -926,7 +927,7 @@ static void cpuDoDMA()
 			}
 			else //WRITE on odd, possibly another dummy read
 			{
-				if(!cpu.oam_dma || cpu.oam_dma_pause)
+				if(!(cpu.dma&CPU_OAM_DMA) || cpu.oam_dma_pause)
 				{
 					cpuDMATryHalt();
 					return;
@@ -943,7 +944,7 @@ static void cpuDoDMA()
 				{
 					cpu.dmc_halted = true;
 					//fully ignore OAM DMA if this is out 2nd/3rd attempt
-					if(cpu.oam_dma && cpu.dmc_halt_attempt)
+					if((cpu.dma&CPU_OAM_DMA) && cpu.dmc_halt_attempt)
 						cpu.oam_dma_pause = true;
 					else //was first halt attempt, possibly allows next OAM DMA cycle
 						cpu.oam_dma_pause = false;
@@ -953,7 +954,7 @@ static void cpuDoDMA()
 			}
 		}
 	}
-	if(cpu.oam_dma)
+	if(cpu.dma&CPU_OAM_DMA)
 	{
 		if(cpu.currently_dma && cpu.oam_halted)
 		{
@@ -975,7 +976,7 @@ static void cpuDoDMA()
 					if(cpu.oam_dma_ptr == 0)
 					{
 						//printf("OAM DMA Done\n");
-						cpu.oam_dma = false;
+						cpu.dma &= ~CPU_OAM_DMA;
 						if(!cpu.dmc_halted) //done with DMA
 							cpu.currently_dma = false;
 						cpu.oam_halted = false;
@@ -991,14 +992,17 @@ static void cpuDoDMA()
 }
 
 /* Main CPU Interpreter */
-bool cpuCycle()
+FIXNES_ALWAYSINLINE bool cpuCycle()
 {
 	cpu_odd_cycle^=true;
 	//printf("CPU Cycle\n");
 	//do DMC and OAM DMA first
-	cpuDoDMA();
-	if(cpu.currently_dma)
-		return true;
+	if(cpu.dma)
+	{
+		cpuDoDMA();
+		if(cpu.currently_dma)
+			return true;
+	}
 	uint8_t instr, cpu_action;
 	cpu_action = cpu.action_arr[cpu.arr_pos];
 	cpu.arr_pos++;
@@ -1602,17 +1606,17 @@ void cpuSoftReset()
 void cpuDoOAM_DMA(uint16_t addr, uint8_t val)
 {
 	(void)addr;
-	cpu.oam_dma = true;
+	cpu.dma |= CPU_OAM_DMA;
 	cpu.oam_dma_addr = (val<<8);
 }
 
 void cpuDoDMC_DMA(uint16_t addr)
 {
-	cpu.dmc_dma = true;
+	cpu.dma |= CPU_DMC_DMA;
 	cpu.dmc_dma_addr = addr;
 }
 
 bool cpuInDMC_DMA()
 {
-	return cpu.dmc_dma;
+	return !!(cpu.dma&CPU_DMC_DMA);
 }
diff --git a/cpu.h b/cpu.h
index 6f0cacf..4eb75dc 100644
--- a/cpu.h
+++ b/cpu.h
@@ -8,12 +8,14 @@
 #ifndef _cpu_h_
 #define _cpu_h_
 
+#include "common.h"
+
 void cpuInit();
 void cpuInitNSF(uint16_t addr, uint8_t newA, uint8_t newX);
 void cpuStartPlayNSF();
 void cpuEndPlayNSF();
 void cpuSoftReset();
-bool cpuCycle();
+FIXNES_ALWAYSINLINE bool cpuCycle();
 void cpuDoOAM_DMA(uint16_t addr, uint8_t val);
 void cpuDoDMC_DMA(uint16_t addr);
 bool cpuInDMC_DMA();
diff --git a/libretro/Makefile b/libretro/Makefile
index 3730369..4c2007a 100644
--- a/libretro/Makefile
+++ b/libretro/Makefile
@@ -1,8 +1,9 @@
 
 TARGET_NAME := fixnes
 DEBUG = 0
-AUDIO_FLOAT = 0
+AUDIO_LOWERFREQ = 0
 DO_LTO = 0
+DO_INLINE_ATTRIBS = 0
 
 ifeq ($(platform),)
 	platform = unix
@@ -221,7 +222,9 @@ else ifeq ($(platform), ngc)
 	PLATFORM_DEFINES += -DGEKKO -DHW_DOL -mrvl -mcpu=750 -meabi -mhard-float
 	PLATFORM_DEFINES += -U__INT32_TYPE__ -U __UINT32_TYPE__ -D__INT32_TYPE__=int
 	STATIC_LINKING = 1
+	AUDIO_LOWERFREQ = 1
 	DO_LTO = 1
+	DO_INLINE_ATTRIBS = 1
 
 # Nintendo Wii
 else ifeq ($(platform), wii)
@@ -233,6 +236,7 @@ else ifeq ($(platform), wii)
 	PLATFORM_DEFINES += -U__INT32_TYPE__ -U __UINT32_TYPE__ -D__INT32_TYPE__=int
 	STATIC_LINKING = 1
 	DO_LTO = 1
+	DO_INLINE_ATTRIBS = 1
 
 # Nintendo WiiU
 else ifeq ($(platform), wiiu)
@@ -243,6 +247,7 @@ else ifeq ($(platform), wiiu)
 	PLATFORM_DEFINES += -DGEKKO -DWIIU -DHW_RVL -mwup -mcpu=750 -meabi -mhard-float
 	PLATFORM_DEFINES += -U__INT32_TYPE__ -U __UINT32_TYPE__ -D__INT32_TYPE__=int
 	STATIC_LINKING = 1
+	DO_INLINE_ATTRIBS = 1
 
 # ARM
 else ifneq (,$(findstring armv,$(platform)))
@@ -294,8 +299,8 @@ else
 
 endif
 
-ifeq ($(AUDIO_FLOAT),1)
-CFLAGS += -DAUDIO_FLOAT=1
+ifeq ($(AUDIO_LOWERFREQ),1)
+CFLAGS += -DAUDIO_LOWERFREQ=1
 endif
 
 ifeq ($(DO_LTO),1)
@@ -303,6 +308,10 @@ CFLAGS += -flto
 CXXFLAGS += -flto
 endif
 
+ifeq ($(DO_INLINE_ATTRIBS),1)
+CFLAGS += -DDO_INLINE_ATTRIBS=1 -Wno-attributes
+endif
+
 ifeq ($(DEBUG), 1)
 	CFLAGS += -O0 -g
 	CXXFLAGS += -O0 -g
diff --git a/main.c b/main.c
index 1c7edca..760249b 100644
--- a/main.c
+++ b/main.c
@@ -38,7 +38,7 @@
 #define DEBUG_KEY 0
 #define DEBUG_LOAD_INFO 1
 
-const char *VERSION_STRING = "fixNES Alpha v1.2.1";
+const char *VERSION_STRING = "fixNES Alpha v1.2.2";
 static char window_title[256];
 static char window_title_pause[256];
 
diff --git a/ppu.c b/ppu.c
index f67591b..b797ea7 100644
--- a/ppu.c
+++ b/ppu.c
@@ -468,22 +468,22 @@ static void saveSprite(uint8_t p0, uint8_t p1)
 	ppu.OAM2Ptr += 4;
 }
 
-static void updateBGTileAddress()
+FIXNES_ALWAYSINLINE inline static void updateBGTileAddress()
 {
 	if((ppu.VramAddr & 0x1F) == 0x1F)
 		ppu.VramAddr ^= 0x41F;
 	else
 		ppu.VramAddr++;
 }
-static void updateBGHoriAddress()
+FIXNES_ALWAYSINLINE inline static void updateBGHoriAddress()
 {
 	ppu.VramAddr = (ppu.VramAddr & (~PPU_VRAM_HORIZONTAL_MASK)) | (ppu.TmpVramAddr & PPU_VRAM_HORIZONTAL_MASK);
 }
-static void updateBGVertAddress()
+FIXNES_ALWAYSINLINE inline static void updateBGVertAddress()
 {
 	ppu.VramAddr = (ppu.VramAddr & (~PPU_VRAM_VERTICAL_MASK)) | (ppu.TmpVramAddr & PPU_VRAM_VERTICAL_MASK);
 }
-static void updateBGYAddress()
+FIXNES_ALWAYSINLINE inline static void updateBGYAddress()
 {
 	/* update Y position for writes */
 	if((ppu.VramAddr & 0x7000) != (7<<12))
@@ -496,13 +496,13 @@ static void updateBGYAddress()
 	}
 }
 
-static void setNTAddr()
+FIXNES_ALWAYSINLINE inline static void setNTAddr()
 {
 	ppu.NextAddr = (ppu.VramAddr & 0xFFF) | 0x2000;
 }
 
 extern void mmc5setTile(uint16_t dot);
-static void getNTAddr(uint16_t dot)
+FIXNES_ALWAYSINLINE inline static void getNTAddr(uint16_t dot)
 {
 	/* MMC5 Scanline/Scroll Related */
 	if(ppuMapper5) mmc5setTile(dot);
@@ -512,40 +512,40 @@ static void getNTAddr(uint16_t dot)
 	ppu.NextTile = chrROMBG+(ntByte<<4)+curTileY;
 }
 
-static void setATAddr()
+FIXNES_ALWAYSINLINE inline static void setATAddr()
 {
 	ppu.NextAddr = (ppu.VramAddr & 0xC00) | 0x23C0 | ((ppu.VramAddr>>4)&0x38) | ((ppu.VramAddr>>2)&7);
 }
 
-static void getATAddr()
+FIXNES_ALWAYSINLINE inline static void getATAddr()
 {
 	/* Select new BG Background Attribute */
 	ppu.BGAttribReg = memPPUGet8(ppu.NextAddr) >> ((ppu.VramAddr & 0x2) | (ppu.VramAddr >> 4 & 0x4));
 }
 
-static void setTileAddr(uint8_t add)
+FIXNES_ALWAYSINLINE inline static void setTileAddr(uint8_t add)
 {
 	ppu.NextAddr = ppu.NextTile+add;
 }
-static void getBGTileAddrLow()
+FIXNES_ALWAYSINLINE inline static void getBGTileAddrLow()
 {
 	mapperChrMode = 0;
 	uint8_t tmp = memPPUGet8(ppu.NextAddr);
 	ppu.BGRegB = tmp >> 0 & 0x55; ppu.BGRegA = tmp >> 1 & 0x55;
 }
-static void getBGTileAddrHigh()
+FIXNES_ALWAYSINLINE inline static void getBGTileAddrHigh()
 {
 	mapperChrMode = 0;
 	uint8_t tmp = memPPUGet8(ppu.NextAddr);
 	ppu.BGRegA |= tmp << 0 & 0xAA; ppu.BGRegB |= tmp << 1 & 0xAA;
 }
-static uint8_t getSpriteTileAddr()
+FIXNES_ALWAYSINLINE inline static uint8_t getSpriteTileAddr()
 {
 	mapperChrMode = 1;
 	return memPPUGet8(ppu.NextAddr);
 }
 
-static void spriteEvalInit()
+FIXNES_ALWAYSINLINE inline static void spriteEvalInit()
 {
 	ppu.OAMzSpritePos = ppu.OAMpos;
 	ppu.OAM2pos = 0;
@@ -553,7 +553,7 @@ static void spriteEvalInit()
 	ppu.spriteEvalB = spriteEvalB_P1;
 }
 
-static void spriteEvalA()
+FIXNES_ALWAYSINLINE inline static void spriteEvalA()
 {
 	ppu.TmpOAMVal = ppu.OAM[(ppu.OAMpos+ppu.OAMcpPos)&0xFF];
 	//printf("%i %i %i %02x\n", ppu.OAMpos,  ppu.OAMcpPos, (ppu.OAMpos+ppu.OAMcpPos)&0xFF, ppu.TmpOAMVal);
@@ -702,17 +702,17 @@ static void spriteEvalB_P9(uint16_t line)
 	ppu.OAMpos += 4;
 }
 
-static void initSpriteTileFetch()
+FIXNES_ALWAYSINLINE inline static void initSpriteTileFetch()
 {
 	ppu.OAM2Ptr = ppu.OAM2;
 }
 
-static void resetOAMPos()
+FIXNES_ALWAYSINLINE inline static void resetOAMPos()
 {
 	ppu.OAMpos = 0;
 }
 
-static void ClearOAM2Byte()
+FIXNES_ALWAYSINLINE inline static void ClearOAM2Byte()
 {
 	ppu.OAM2[ppu.OAM2pos++] = 0xFF;
 }
@@ -744,7 +744,7 @@ static uint16_t ppuLastDot(uint16_t line)
 	return line;
 }
 
-void ppuCycle()
+FIXNES_ALWAYSINLINE void ppuCycle()
 {
 	uint16_t dot = ppu.curDot, line = ppu.curLine, drawPos;
 	uint8_t curCol;
@@ -1635,7 +1635,7 @@ static uint8_t ppuDoSprites(uint8_t color, uint16_t dot)
 	return color;
 }
 
-bool ppuDrawDone()
+FIXNES_ALWAYSINLINE bool ppuDrawDone()
 {
 	if(ppu.FrameDone)
 	{
diff --git a/ppu.h b/ppu.h
index 928aae9..b66a631 100644
--- a/ppu.h
+++ b/ppu.h
@@ -8,10 +8,12 @@
 #ifndef _ppu_h_
 #define _ppu_h_
 
+#include "common.h"
+
 void ppuInit();
 void ppuAddCycles();
-void ppuCycle();
-bool ppuDrawDone();
+FIXNES_ALWAYSINLINE void ppuCycle();
+FIXNES_ALWAYSINLINE bool ppuDrawDone();
 uint8_t ppuGet8(uint16_t addr);
 void ppuSet8(uint16_t addr, uint8_t val);
 uint8_t ppuNMI();