diff --git a/stepmania/src/InputFilter.cpp b/stepmania/src/InputFilter.cpp
index 303f3dedb3..3e2a9f674d 100644
--- a/stepmania/src/InputFilter.cpp
+++ b/stepmania/src/InputFilter.cpp
@@ -27,7 +27,7 @@ InputFilter::InputFilter()
 	memset( m_BeingForced, 0, sizeof(m_BeingForced) );
 	memset( m_fSecsHeld, 0, sizeof(m_fSecsHeld) );
 	for( int d=0; d<NUM_INPUT_DEVICES; d++ )	// foreach InputDevice
-		for( int b=0; b < NUM_DEVICE_BUTTONS; b++ )	// foreach button
+		for( int b=0; b < NUM_DEVICE_BUTTONS[d]; b++ )	// foreach button
 			m_fSecsToForce[d][b] = -1;	
 
 	Reset();
@@ -142,7 +142,7 @@ void InputFilter::StopForcingKey( DeviceInput di )
 /* Release all buttons on the given device. */
 void InputFilter::ResetDevice( InputDevice dev )
 {
-	for( int button = 0; button < NUM_DEVICE_BUTTONS; ++button )
+	for( int button = 0; button < NUM_DEVICE_BUTTONS[dev]; ++button )
 		ButtonPressed( DeviceInput(dev, button), false );
 }
 
@@ -174,11 +174,17 @@ void InputFilter::Update(float fDeltaTime)
 	 * things like "key pressed, key release, key repeat". */
 	LockMut(*queuemutex);
 
+	// Don't reconstruct "di" inside the loop.  This line alone is 
+	// taking 4% of the CPU on a P3-666.
+	DeviceInput di( (InputDevice)0,0,now);
+
 	for( int d=0; d<NUM_INPUT_DEVICES; d++ )	// foreach InputDevice
 	{
-		for( int b=0; b < NUM_DEVICE_BUTTONS; b++ )	// foreach button
+		di.device = (InputDevice)d;
+
+		for( int b=0; b < NUM_DEVICE_BUTTONS[d]; b++ )	// foreach button
 		{
-			DeviceInput di( (InputDevice)d,b,now);
+			di.button = b;
 
 			if( m_fSecsToForce[d][b] > 0 )
 			{
diff --git a/stepmania/src/InputFilter.h b/stepmania/src/InputFilter.h
index c5f9f07c59..863273ad0b 100644
--- a/stepmania/src/InputFilter.h
+++ b/stepmania/src/InputFilter.h
@@ -38,12 +38,12 @@ typedef vector<InputEvent> InputEventArray;
 class RageMutex;
 class InputFilter
 {
-	bool m_BeingHeld[NUM_INPUT_DEVICES][NUM_DEVICE_BUTTONS];
-	bool m_BeingForced[NUM_INPUT_DEVICES][NUM_DEVICE_BUTTONS];
-	float m_fSecsHeld[NUM_INPUT_DEVICES][NUM_DEVICE_BUTTONS];
+	bool m_BeingHeld[NUM_INPUT_DEVICES][MAX_DEVICE_BUTTONS];
+	bool m_BeingForced[NUM_INPUT_DEVICES][MAX_DEVICE_BUTTONS];
+	float m_fSecsHeld[NUM_INPUT_DEVICES][MAX_DEVICE_BUTTONS];
 
 	/* If > 0, then when it reaches 0, stop forcing. */
-	float m_fSecsToForce[NUM_INPUT_DEVICES][NUM_DEVICE_BUTTONS];
+	float m_fSecsToForce[NUM_INPUT_DEVICES][MAX_DEVICE_BUTTONS];
 	InputEventArray queue;
 	RageMutex *queuemutex;
 
diff --git a/stepmania/src/InputMapper.cpp b/stepmania/src/InputMapper.cpp
index db69097154..56db61e3eb 100644
--- a/stepmania/src/InputMapper.cpp
+++ b/stepmania/src/InputMapper.cpp
@@ -448,7 +448,7 @@ void InputMapper::UpdateTempDItoGI()
 	// clear out m_tempDItoGI
 	for( int d=0; d<NUM_INPUT_DEVICES; d++ )
 	{
-		for( int b=0; b<NUM_DEVICE_BUTTONS; b++ )
+		for( int b=0; b<NUM_DEVICE_BUTTONS[d]; b++ )
 		{
 			m_tempDItoGI[d][b].MakeInvalid();
 		}
diff --git a/stepmania/src/InputMapper.h b/stepmania/src/InputMapper.h
index bd2da7dcf7..94ad6a8dd1 100644
--- a/stepmania/src/InputMapper.h
+++ b/stepmania/src/InputMapper.h
@@ -68,7 +68,7 @@ protected:
 
 	// lookup for efficiency from a DeviceInput to a GameInput
 	// This is repopulated every time m_PItoDI changes by calling UpdateTempDItoPI().
-	GameInput m_tempDItoGI[NUM_INPUT_DEVICES][NUM_DEVICE_BUTTONS];
+	GameInput m_tempDItoGI[NUM_INPUT_DEVICES][MAX_DEVICE_BUTTONS];
 	void UpdateTempDItoGI();
 };
 
diff --git a/stepmania/src/RageInputDevice.h b/stepmania/src/RageInputDevice.h
index 1570cda901..5fe18ab495 100644
--- a/stepmania/src/RageInputDevice.h
+++ b/stepmania/src/RageInputDevice.h
@@ -9,8 +9,6 @@ const int NUM_JOYSTICKS = 4;
 const int NUM_JOYSTICK_HATS = 1;
 const int NUM_PUMPS = 2;
 
-const int NUM_DEVICE_BUTTONS = NUM_KEYBOARD_BUTTONS;
-
 enum InputDevice {
 	DEVICE_KEYBOARD = 0,
 	DEVICE_JOY1,
@@ -56,6 +54,20 @@ enum PumpButton {
 	NUM_PUMP_PAD_BUTTONS	// leave this at the end
 };
 
+
+const int NUM_DEVICE_BUTTONS[NUM_INPUT_DEVICES] =
+{
+	NUM_KEYBOARD_BUTTONS, // DEVICE_KEYBOARD
+	NUM_JOYSTICK_BUTTONS, // DEVICE_JOY1
+	NUM_JOYSTICK_BUTTONS, // DEVICE_JOY2
+	NUM_JOYSTICK_BUTTONS, // DEVICE_JOY3
+	NUM_JOYSTICK_BUTTONS, // DEVICE_JOY4
+	NUM_PUMP_PAD_BUTTONS, // DEVICE_PUMP1
+	NUM_PUMP_PAD_BUTTONS, // DEVICE_PUMP2
+};
+
+const int MAX_DEVICE_BUTTONS = NUM_KEYBOARD_BUTTONS;
+
 struct DeviceInput
 {
 public: