From 5c3bcc8f5ff4fbf7d1b10fd669e83296a98dab28 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 07:08:15 +0100
Subject: [PATCH 01/28] Revert "Readd the old AXWii (rename the new version to
 NewAXWii) and set it as default"

Put back the new AX Wii as the default in this development branch.
---
 Source/Core/Core/CMakeLists.txt               |   3 +-
 Source/Core/Core/Core.vcxproj                 |   6 +-
 Source/Core/Core/Core.vcxproj.filters         |  14 +-
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp    |   4 -
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h |   2 +-
 .../{UCode_AX_Structs.h => UCode_AXStructs.h} |   0
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 478 +++++++++++-------
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  75 ++-
 .../Src/HW/DSPHLE/UCodes/UCode_AXWii_ADPCM.h  |  92 ----
 .../HW/DSPHLE/UCodes/UCode_AXWii_Structs.h    | 365 -------------
 .../Src/HW/DSPHLE/UCodes/UCode_AXWii_Voice.h  | 271 ----------
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     |   2 +-
 .../Src/HW/DSPHLE/UCodes/UCode_NewAXWii.cpp   | 383 --------------
 .../Src/HW/DSPHLE/UCodes/UCode_NewAXWii.h     |  80 ---
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCodes.cpp |  11 +-
 15 files changed, 354 insertions(+), 1432 deletions(-)
 rename Source/Core/Core/Src/HW/DSPHLE/UCodes/{UCode_AX_Structs.h => UCode_AXStructs.h} (100%)
 delete mode 100644 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_ADPCM.h
 delete mode 100644 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Structs.h
 delete mode 100644 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Voice.h
 delete mode 100644 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.cpp
 delete mode 100644 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.h

diff --git a/Source/Core/Core/CMakeLists.txt b/Source/Core/Core/CMakeLists.txt
index 6ffeeaec3a..79ad6b60cb 100644
--- a/Source/Core/Core/CMakeLists.txt
+++ b/Source/Core/Core/CMakeLists.txt
@@ -70,8 +70,7 @@ set(SRCS	Src/ActionReplay.cpp
 			Src/HW/CPU.cpp
 			Src/HW/DSP.cpp
 			Src/HW/DSPHLE/UCodes/UCode_AX.cpp
-            Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
-            Src/HW/DSPHLE/UCodes/UCode_NewAXWii.cpp
+			Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
 			Src/HW/DSPHLE/UCodes/UCode_CARD.cpp
 			Src/HW/DSPHLE/UCodes/UCode_InitAudioSystem.cpp
 			Src/HW/DSPHLE/UCodes/UCode_ROM.cpp
diff --git a/Source/Core/Core/Core.vcxproj b/Source/Core/Core/Core.vcxproj
index bd3d97f11b..ddd8b0dfd8 100644
--- a/Source/Core/Core/Core.vcxproj
+++ b/Source/Core/Core/Core.vcxproj
@@ -261,7 +261,6 @@
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCodes.cpp" />
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_AX.cpp" />
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_AXWii.cpp" />
-    <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_NewAXWii.cpp" />
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_CARD.cpp" />
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_GBA.cpp" />
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_InitAudioSystem.cpp" />
@@ -464,11 +463,8 @@
     <ClInclude Include="Src\HW\DSPHLE\MailHandler.h" />
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCodes.h" />
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AX.h" />
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AX_Structs.h" />
+    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXStructs.h" />
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXWii.h" />
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXWii_Structs.h" />
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXWii_Voice.h" />
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_NewAXWii.h" />
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AX_Voice.h" />
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_CARD.h" />
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_GBA.h" />
diff --git a/Source/Core/Core/Core.vcxproj.filters b/Source/Core/Core/Core.vcxproj.filters
index 246dafd972..a9de67dbb5 100644
--- a/Source/Core/Core/Core.vcxproj.filters
+++ b/Source/Core/Core/Core.vcxproj.filters
@@ -197,9 +197,6 @@
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_AXWii.cpp">
       <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
     </ClCompile>
-    <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_NewAXWii.cpp">
-      <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
-    </ClCompile>
     <ClCompile Include="Src\HW\DSPHLE\UCodes\UCode_CARD.cpp">
       <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
     </ClCompile>
@@ -736,21 +733,12 @@
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AX_Voice.h">
       <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
     </ClInclude>
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AX_Structs.h">
+    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXStructs.h">
       <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
     </ClInclude>
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXWii.h">
       <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
     </ClInclude>
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXWii_Structs.h">
-      <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
-    </ClInclude>
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_AXWii_Voice.h">
-      <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
-    </ClInclude>
-    <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_NewAXWii.h">
-      <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
-    </ClInclude>
     <ClInclude Include="Src\HW\DSPHLE\UCodes\UCode_CARD.h">
       <Filter>HW %28Flipper/Hollywood%29\DSP Interface + HLE\HLE\uCodes</Filter>
     </ClInclude>
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
index 81ef7c6482..57032a774e 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
@@ -200,11 +200,7 @@ void CUCode_AX::HandleCommandList()
 				u16 idx = m_cmdlist[curr_idx++];
 				addr_hi = m_cmdlist[curr_idx++];
 				addr_lo = m_cmdlist[curr_idx++];
-
 				// TODO
-				(void)samp_val;
-				(void)idx;
-
 				break;
 			}
 
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
index 8fd2a2af5f..7158c197e8 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
@@ -27,7 +27,7 @@
 #define _UCODE_AX_H
 
 #include "UCodes.h"
-#include "UCode_AX_Structs.h"
+#include "UCode_AXStructs.h"
 
 // We can't directly use the mixer_control field from the PB because it does
 // not mean the same in all AX versions. The AX UCode converts the
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Structs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
similarity index 100%
rename from Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Structs.h
rename to Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index f4effbba6d..8c74c7a1e7 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -21,247 +21,363 @@
 #include "Mixer.h"
 
 #include "UCodes.h"
-#include "UCode_AXWii_Structs.h"
-#include "UCode_AX.h" // for some functions in CUCode_AX
+#include "UCode_AXStructs.h"
 #include "UCode_AXWii.h"
-#include "UCode_AXWii_Voice.h"
+
+#define AX_WII
+#include "UCode_AX_Voice.h"
 
 
 CUCode_AXWii::CUCode_AXWii(DSPHLE *dsp_hle, u32 l_CRC)
-	: IUCode(dsp_hle, l_CRC)
-	, m_addressPBs(0xFFFFFFFF)
+	: CUCode_AX(dsp_hle, l_CRC)
 {
-	// we got loaded
-	m_rMailHandler.PushMail(DSP_INIT);
-
-	templbuffer = new int[1024 * 1024];
-	temprbuffer = new int[1024 * 1024];
-
-	wiisportsHack = m_CRC == 0xfa450138;
+	WARN_LOG(DSPHLE, "Instantiating CUCode_AXWii");
 }
 
 CUCode_AXWii::~CUCode_AXWii()
 {
-	m_rMailHandler.Clear();
-	delete [] templbuffer;
-	delete [] temprbuffer;
 }
 
-void CUCode_AXWii::HandleMail(u32 _uMail)
+void CUCode_AXWii::HandleCommandList()
 {
-	if (m_UploadSetupInProgress) 
+	// Temp variables for addresses computation
+	u16 addr_hi, addr_lo;
+	u16 addr2_hi, addr2_lo;
+	u16 volume;
+
+//	WARN_LOG(DSPHLE, "Command list:");
+//	for (u32 i = 0; m_cmdlist[i] != CMD_END; ++i)
+//		WARN_LOG(DSPHLE, "%04x", m_cmdlist[i]);
+//	WARN_LOG(DSPHLE, "-------------");
+
+	u32 curr_idx = 0;
+	bool end = false;
+	while (!end)
 	{
-		PrepareBootUCode(_uMail);
-		return;
-	}
-	else if ((_uMail & 0xFFFF0000) == MAIL_AX_ALIST)
-	{
-		// We are expected to get a new CmdBlock
-		DEBUG_LOG(DSPHLE, "GetNextCmdBlock (%ibytes)", (u16)_uMail);
-	}
-	else switch(_uMail)
-	{
-		case 0xCDD10000: // Action 0 - AX_ResumeTask()
-			m_rMailHandler.PushMail(DSP_RESUME);
-			break;
+		u16 cmd = m_cmdlist[curr_idx++];
 
-		case 0xCDD10001: // Action 1 - new ucode upload
-			DEBUG_LOG(DSPHLE,"DSP IROM - New Ucode!");
-			// TODO find a better way to protect from HLEMixer?
-			soundStream->GetMixer()->SetHLEReady(false);
-			m_UploadSetupInProgress = true;
-			break;
+		switch (cmd)
+		{
+			// Some of these commands are unknown, or unused in this AX HLE.
+			// We still need to skip their arguments using "curr_idx += N".
 
-		case 0xCDD10002: // Action 2 - IROM_Reset(); ( WII: De Blob, Cursed Mountain,...)
-			DEBUG_LOG(DSPHLE,"DSP IROM - Reset!");
-			m_DSPHLE->SetUCode(UCODE_ROM);
-			return;
+			case CMD_SETUP:
+				addr_hi = m_cmdlist[curr_idx++];
+				addr_lo = m_cmdlist[curr_idx++];
+				SetupProcessing(HILO_TO_32(addr));
+				break;
 
-		case 0xCDD10003: // Action 3 - AX_GetNextCmdBlock()
-			break;
+			case CMD_UNK_01: curr_idx += 2; break;
+			case CMD_UNK_02: curr_idx += 2; break;
+			case CMD_UNK_03: curr_idx += 2; break;
 
-		default:
-			DEBUG_LOG(DSPHLE, " >>>> u32 MAIL : AXTask Mail (%08x)", _uMail);
-			AXTask(_uMail);
-			break;
+			case CMD_PROCESS:
+				addr_hi = m_cmdlist[curr_idx++];
+				addr_lo = m_cmdlist[curr_idx++];
+				ProcessPBList(HILO_TO_32(addr));
+				break;
+
+			case CMD_MIX_AUXA:
+			case CMD_MIX_AUXB:
+			case CMD_MIX_AUXC:
+				volume = m_cmdlist[curr_idx++];
+				addr_hi = m_cmdlist[curr_idx++];
+				addr_lo = m_cmdlist[curr_idx++];
+				addr2_hi = m_cmdlist[curr_idx++];
+				addr2_lo = m_cmdlist[curr_idx++];
+				MixAUXSamples(cmd - CMD_MIX_AUXA, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
+				break;
+
+			// These two go together and manipulate some AUX buffers.
+			case CMD_UNK_08: curr_idx += 13; break;
+			case CMD_UNK_09: curr_idx += 13; break;
+
+			case CMD_UNK_0A: curr_idx += 4; break;
+
+			case CMD_OUTPUT:
+				volume = m_cmdlist[curr_idx++];
+				addr_hi = m_cmdlist[curr_idx++];
+				addr_lo = m_cmdlist[curr_idx++];
+				addr2_hi = m_cmdlist[curr_idx++];
+				addr2_lo = m_cmdlist[curr_idx++];
+				OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), volume);
+				break;
+
+			case CMD_UNK_0C: curr_idx += 5; break;
+
+			case CMD_WM_OUTPUT:
+			{
+				u32 addresses[4] = {
+					(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
+					(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
+					(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
+					(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
+				};
+				curr_idx += 8;
+				OutputWMSamples(addresses);
+				break;
+			}
+
+			case CMD_END:
+				end = true;
+				break;
+		}
 	}
 }
 
-void CUCode_AXWii::MixAdd(short* _pBuffer, int _iSize)
+void CUCode_AXWii::SetupProcessing(u32 init_addr)
 {
-	AXPBWii PB;
+	// TODO: should be easily factorizable with AX
+	s16 init_data[60];
 
-	if (_iSize > 1024 * 1024)
-		_iSize = 1024 * 1024;
+	for (u32 i = 0; i < 60; ++i)
+		init_data[i] = HLEMemory_Read_U16(init_addr + 2 * i);
 
-	memset(templbuffer, 0, _iSize * sizeof(int));
-	memset(temprbuffer, 0, _iSize * sizeof(int));
+	// List of all buffers we have to initialize
+	struct {
+		int* ptr;
+		u32 samples;
+	} buffers[] = {
+		{ m_samples_left, 32 },
+		{ m_samples_right, 32 },
+		{ m_samples_surround, 32 },
+		{ m_samples_auxA_left, 32 },
+		{ m_samples_auxA_right, 32 },
+		{ m_samples_auxA_surround, 32 },
+		{ m_samples_auxB_left, 32 },
+		{ m_samples_auxB_right, 32 },
+		{ m_samples_auxB_surround, 32 },
+		{ m_samples_auxC_left, 32 },
+		{ m_samples_auxC_right, 32 },
+		{ m_samples_auxC_surround, 32 },
 
-	u32 blockAddr = m_addressPBs;
-	if (!blockAddr)
-		return;
+		{ m_samples_wm0, 6 },
+		{ m_samples_aux0, 6 },
+		{ m_samples_wm1, 6 },
+		{ m_samples_aux1, 6 },
+		{ m_samples_wm2, 6 },
+		{ m_samples_aux2, 6 },
+		{ m_samples_wm3, 6 },
+		{ m_samples_aux3, 6 }
+	};
 
-	for (int i = 0; i < NUMBER_OF_PBS; i++)
+	u32 init_idx = 0;
+	for (u32 i = 0; i < sizeof (buffers) / sizeof (buffers[0]); ++i)
 	{
-		if (!ReadPB(blockAddr, PB))
-			break;
+		s32 init_val = (s32)((init_data[init_idx] << 16) | init_data[init_idx + 1]);
+		s16 delta = (s16)init_data[init_idx + 2];
 
-		if (wiisportsHack)
-			MixAddVoice(*(AXPBWiiSports*)&PB, templbuffer, temprbuffer, _iSize);
+		init_idx += 3;
+
+		if (!init_val)
+			memset(buffers[i].ptr, 0, 3 * buffers[i].samples * sizeof (int));
 		else
-			MixAddVoice(PB, templbuffer, temprbuffer, _iSize);
-
-		if (!WritePB(blockAddr, PB))
-			break;
-		
-		// next PB, or done
-		blockAddr = (PB.next_pb_hi << 16) | PB.next_pb_lo;
-		if (!blockAddr)
-			break;
-	}		
-
-	// We write the sound to _pBuffer
-	if (_pBuffer)
-	{
-		for (int i = 0; i < _iSize; i++)
 		{
-			// Clamp into 16-bit. Maybe we should add a volume compressor here.
-			int left  = templbuffer[i] + _pBuffer[0];
-			int right = temprbuffer[i] + _pBuffer[1];
-			if (left  < -32767)  left = -32767;
-			else if (left  >  32767)  left =  32767;
-			if (right < -32767) right = -32767;
-			else if (right >  32767) right =  32767;
-			*_pBuffer++ = left;
-			*_pBuffer++ = right;
+			for (u32 j = 0; j < 3 * buffers[i].samples; ++j)
+			{
+				buffers[i].ptr[j] = init_val;
+				init_val += delta;
+			}
 		}
 	}
 }
 
-
-void CUCode_AXWii::Update(int cycles)
+AXMixControl CUCode_AXWii::ConvertMixerControl(u32 mixer_control)
 {
-	if (NeedsResumeMail())
+	u32 ret = 0;
+
+	if (mixer_control & 0x00000001) ret |= MIX_L;
+	if (mixer_control & 0x00000002) ret |= MIX_R;
+	if (mixer_control & 0x00000004) ret |= MIX_L_RAMP | MIX_R_RAMP;
+	if (mixer_control & 0x00000008) ret |= MIX_S;
+	if (mixer_control & 0x00000010) ret |= MIX_S_RAMP;
+	if (mixer_control & 0x00010000) ret |= MIX_AUXA_L;
+	if (mixer_control & 0x00020000) ret |= MIX_AUXA_R;
+	if (mixer_control & 0x00040000) ret |= MIX_AUXA_L_RAMP | MIX_AUXA_R_RAMP;
+	if (mixer_control & 0x00080000) ret |= MIX_AUXA_S;
+	if (mixer_control & 0x00100000) ret |= MIX_AUXA_S_RAMP;
+	if (mixer_control & 0x00200000) ret |= MIX_AUXB_L;
+	if (mixer_control & 0x00400000) ret |= MIX_AUXB_R;
+	if (mixer_control & 0x00800000) ret |= MIX_AUXB_L_RAMP | MIX_AUXB_R_RAMP;
+	if (mixer_control & 0x01000000) ret |= MIX_AUXB_S;
+	if (mixer_control & 0x02000000) ret |= MIX_AUXB_S_RAMP;
+	if (mixer_control & 0x04000000) ret |= MIX_AUXC_L;
+	if (mixer_control & 0x08000000) ret |= MIX_AUXC_R;
+	if (mixer_control & 0x10000000) ret |= MIX_AUXC_L_RAMP | MIX_AUXC_R_RAMP;
+	if (mixer_control & 0x20000000) ret |= MIX_AUXC_S;
+	if (mixer_control & 0x40000000) ret |= MIX_AUXC_S_RAMP;
+
+	return (AXMixControl)ret;
+}
+
+void CUCode_AXWii::ProcessPBList(u32 pb_addr)
+{
+	const u32 spms = 32;
+
+	AXPBWii pb;
+
+	while (pb_addr)
 	{
-		m_rMailHandler.PushMail(DSP_RESUME);
-		DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
-	}
-	// check if we have to send something
-	else if (!m_rMailHandler.IsEmpty())
-	{
-		DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
+		AXBuffers buffers = {{
+			m_samples_left,
+			m_samples_right,
+			m_samples_surround,
+			m_samples_auxA_left,
+			m_samples_auxA_right,
+			m_samples_auxA_surround,
+			m_samples_auxB_left,
+			m_samples_auxB_right,
+			m_samples_auxB_surround,
+			m_samples_auxC_left,
+			m_samples_auxC_right,
+			m_samples_auxC_surround
+		}};
+
+		if (!ReadPB(pb_addr, pb))
+			break;
+
+		for (int curr_ms = 0; curr_ms < 3; ++curr_ms)
+		{
+			Process1ms(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)));
+
+			// Forward the buffers
+			for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
+				buffers.ptrs[i] += spms;
+		}
+
+		WritePB(pb_addr, pb);
+		pb_addr = HILO_TO_32(pb.next_pb);
 	}
 }
 
-// AX seems to bootup one task only and waits for resume-callbacks
-// everytime the DSP has "spare time" it sends a resume-mail to the CPU
-// and the __DSPHandler calls a AX-Callback which generates a new AXFrame
-bool CUCode_AXWii::AXTask(u32& _uMail)
+void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume)
 {
-	u32 uAddress = _uMail;
-	//u32 Addr__AXStudio;
-	//u32 Addr__AXOutSBuffer;
-	bool bExecuteList = true;
+	int* buffers[3] = { 0 };
+	int* main_buffers[3] = {
+		m_samples_left,
+		m_samples_right,
+		m_samples_surround
+	};
 
-/*
-	for (int i=0;i<64;i++) {
-		NOTICE_LOG(DSPHLE,"%x - %08x",uAddress+(i*4),HLEMemory_Read_U32(uAddress+(i*4)));
-	}
-*/
-
-	while (bExecuteList)
+	switch (aux_id)
 	{
-		u16 iCommand = HLEMemory_Read_U16(uAddress);
-		uAddress += 2;
-		//NOTICE_LOG(DSPHLE,"AXWII - AXLIST CMD %X",iCommand);
+	case 0:
+		buffers[0] = m_samples_auxA_left;
+		buffers[1] = m_samples_auxA_right;
+		buffers[2] = m_samples_auxA_surround;
+		break;
 
-		switch (iCommand)
+	case 1:
+		buffers[0] = m_samples_auxB_left;
+		buffers[1] = m_samples_auxB_right;
+		buffers[2] = m_samples_auxB_surround;
+		break;
+
+	case 2:
+		buffers[0] = m_samples_auxC_left;
+		buffers[1] = m_samples_auxC_right;
+		buffers[2] = m_samples_auxC_surround;
+		break;
+	}
+
+	// Send the content of AUX buffers to the CPU
+	if (write_addr)
+	{
+		int* ptr = (int*)HLEMemory_Get_Pointer(write_addr);
+		for (u32 i = 0; i < 3; ++i)
+			for (u32 j = 0; j < 3 * 32; ++j)
+				*ptr++ = Common::swap32(buffers[i][j]);
+	}
+
+	// Then read the buffers from the CPU and add to our main buffers.
+	int* ptr = (int*)HLEMemory_Get_Pointer(read_addr);
+	for (u32 i = 0; i < 3; ++i)
+		for (u32 j = 0; j < 3 * 32; ++j)
 		{
-	    case 0x0000: 
-		    //Addr__AXStudio = HLEMemory_Read_U32(uAddress);
-		    uAddress += 4;
-		    break;
+			s64 new_val = main_buffers[i][j] + Common::swap32(*ptr++);
+			main_buffers[i][j] = (new_val * volume) >> 15;
+		}
+}
 
-	    case 0x0001:
-		    uAddress += 4;
-		    break;
+void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
+{
+	int surround_buffer[3 * 32] = { 0 };
 
-		case 0x0003:
-		    uAddress += 4;
-		    break;
+	for (u32 i = 0; i < 3 * 32; ++i)
+		surround_buffer[i] = Common::swap32(m_samples_surround[i]);
+	memcpy(HLEMemory_Get_Pointer(surround_addr), surround_buffer, sizeof (surround_buffer));
 
-	    case 0x0004: 
-			// PBs are here now
-			m_addressPBs = HLEMemory_Read_U32(uAddress);
-			if (soundStream)
-				soundStream->GetMixer()->SetHLEReady(true);
-//			soundStream->Update();
-			uAddress += 4;
-		    break;
+	short buffer[3 * 32 * 2];
 
-	    case 0x0005:
-			if (!wiisportsHack)
-				uAddress += 10;
-		    break;
+	// Clamp internal buffers to 16 bits.
+	for (u32 i = 0; i < 3 * 32; ++i)
+	{
+		int left  = m_samples_left[i];
+		int right = m_samples_right[i];
 
-	    case 0x0006:
-		    uAddress += 10;
-		    break; 
+		// Apply global volume. Cast to s64 to avoid overflow.
+		left = ((s64)left * volume) >> 15;
+		right = ((s64)right * volume) >> 15;
 
-		case 0x0007:   // AXLIST_SBUFFER
-		    //Addr__AXOutSBuffer = HLEMemory_Read_U32(uAddress);
-		    uAddress += 10;
-		    break;
+		if (left < -32767)  left = -32767;
+		if (left > 32767)   left = 32767;
+		if (right < -32767) right = -32767;
+		if (right >  32767) right = 32767;
 
-		case 0x0008:
-		    uAddress += 26;
-			break;
+		m_samples_left[i] = left;
+		m_samples_right[i] = right;
+	}
 
-		case 0x000a:
-			uAddress += wiisportsHack ? 4 : 8; // AXLIST_COMPRESSORTABLE
-		    break;
+	for (u32 i = 0; i < 3 * 32; ++i)
+	{
+		buffer[2 * i] = Common::swap16(m_samples_left[i]);
+		buffer[2 * i + 1] = Common::swap16(m_samples_right[i]);
+	}
 
-		case 0x000b:
-			uAddress += wiisportsHack ? 2 : 10;
-			break;
+	memcpy(HLEMemory_Get_Pointer(lr_addr), buffer, sizeof (buffer));
+}
 
-		case 0x000c:
-			uAddress += wiisportsHack ? 8 : 10;
-			break;
+void CUCode_AXWii::OutputWMSamples(u32* addresses)
+{
+	int* buffers[] = {
+		m_samples_wm0,
+		m_samples_wm1,
+		m_samples_wm2,
+		m_samples_wm3
+	};
 
-		case 0x000d:
-			uAddress += 16;
-			break;
-
-	    case 0x000e:
-			if (wiisportsHack) 
-				uAddress += 16;
-			else
-				bExecuteList = false;
-			break;
-
-	    case 0x000f: // only for Wii Sports uCode
-			bExecuteList = false;
-			break;
-
-	    default:
-			INFO_LOG(DSPHLE,"DSPHLE - AXwii - AXLIST - Unknown CMD: %x",iCommand);
-		    // unknown command so stop the execution of this TaskList
-		    bExecuteList = false;
-			break;
+	for (u32 i = 0; i < 4; ++i)
+	{
+		int* in = buffers[i];
+		u16* out = (u16*)HLEMemory_Get_Pointer(addresses[i]);
+		for (u32 j = 0; j < 3 * 6; ++j)
+		{
+			int sample = in[j];
+			if (sample < -32767) sample = -32767;
+			if (sample > 32767) sample = 32767;
+			out[j] = Common::swap16((u16)sample);
 		}
 	}
-
-	m_rMailHandler.PushMail(DSP_YIELD); //its here in case there is a CMD fuckup
-	return true;
 }
 
 void CUCode_AXWii::DoState(PointerWrap &p)
 {
-	std::lock_guard<std::mutex> lk(m_csMix);
-
-	p.Do(m_addressPBs);
-	p.Do(wiisportsHack);
+	std::lock_guard<std::mutex> lk(m_processing);
 
 	DoStateShared(p);
+	DoAXState(p);
+
+	p.Do(m_samples_auxC_left);
+	p.Do(m_samples_auxC_right);
+	p.Do(m_samples_auxC_surround);
+
+	p.Do(m_samples_wm0);
+	p.Do(m_samples_wm1);
+	p.Do(m_samples_wm2);
+	p.Do(m_samples_wm3);
+
+	p.Do(m_samples_aux0);
+	p.Do(m_samples_aux1);
+	p.Do(m_samples_aux2);
+	p.Do(m_samples_aux3);
 }
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index dc07e71a63..3a66c30868 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -12,44 +12,69 @@
 // A copy of the GPL 2.0 should have been included with the program.
 // If not, see http://www.gnu.org/licenses/
 
-// Official SVN repository and contact information can be found at
+// Official Git repository and contact information can be found at
 // http://code.google.com/p/dolphin-emu/
 
-#ifndef _UCODE_AXWII
-#define _UCODE_AXWII
+#ifndef _UCODE_AXWII_H
+#define _UCODE_AXWII_H
 
-#include "UCode_AXWii_Structs.h"
+#include "UCode_AX.h"
 
-#define NUMBER_OF_PBS 128
-
-class CUCode_AXWii : public IUCode
+class CUCode_AXWii : public CUCode_AX
 {
 public:
 	CUCode_AXWii(DSPHLE *dsp_hle, u32 _CRC);
 	virtual ~CUCode_AXWii();
 
-	void HandleMail(u32 _uMail);
-	void MixAdd(short* _pBuffer, int _iSize);
-	void Update(int cycles);
-	void DoState(PointerWrap &p);
+	virtual void DoState(PointerWrap &p);
+
+protected:
+	int m_samples_auxC_left[32 * 3];
+	int m_samples_auxC_right[32 * 3];
+	int m_samples_auxC_surround[32 * 3];
+
+	// Wiimote buffers
+	int m_samples_wm0[6 * 3];
+	int m_samples_aux0[6 * 3];
+	int m_samples_wm1[6 * 3];
+	int m_samples_aux1[6 * 3];
+	int m_samples_wm2[6 * 3];
+	int m_samples_aux2[6 * 3];
+	int m_samples_wm3[6 * 3];
+	int m_samples_aux3[6 * 3];
+
+	// Convert a mixer_control bitfield to our internal representation for that
+	// value. Required because that bitfield has a different meaning in some
+	// versions of AX.
+	AXMixControl ConvertMixerControl(u32 mixer_control);
+
+	virtual void HandleCommandList();
+
+	void SetupProcessing(u32 init_addr);
+	void ProcessPBList(u32 pb_addr);
+	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
+	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume);
+	void OutputWMSamples(u32* addresses);	// 4 addresses
 
 private:
-	enum
+	enum CmdType
 	{
-		MAIL_AX_ALIST = 0xBABE0000,
+		CMD_SETUP = 0x00,
+		CMD_UNK_01 = 0x01,
+		CMD_UNK_02 = 0x02,
+		CMD_UNK_03 = 0x03,
+		CMD_PROCESS = 0x04,
+		CMD_MIX_AUXA = 0x05,
+		CMD_MIX_AUXB = 0x06,
+		CMD_MIX_AUXC = 0x07,
+		CMD_UNK_08 = 0x08,
+		CMD_UNK_09 = 0x09,
+		CMD_UNK_0A = 0x0A,
+		CMD_OUTPUT = 0x0B,
+		CMD_UNK_0C = 0x0C,
+		CMD_WM_OUTPUT = 0x0D,
+		CMD_END = 0x0E
 	};
-
-	// PBs
-	u32 m_addressPBs;
-
-	bool wiisportsHack;
-
-	int *templbuffer;
-	int *temprbuffer;
-
-	// ax task message handler
-	bool AXTask(u32& _uMail);
-	void SendMail(u32 _uMail);
 };
 
 #endif  // _UCODE_AXWII
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_ADPCM.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_ADPCM.h
deleted file mode 100644
index 9130bb9da4..0000000000
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_ADPCM.h
+++ /dev/null
@@ -1,92 +0,0 @@
-// Copyright (C) 2003 Dolphin Project.
-
-// This program is free software: you can redistribute it and/or modify
-// it under the terms of the GNU General Public License as published by
-// the Free Software Foundation, version 2.0.
-
-// This program is distributed in the hope that it will be useful,
-// but WITHOUT ANY WARRANTY; without even the implied warranty of
-// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-// GNU General Public License 2.0 for more details.
-
-// A copy of the GPL 2.0 should have been included with the program.
-// If not, see http://www.gnu.org/licenses/
-
-// Official SVN repository and contact information can be found at
-// http://code.google.com/p/dolphin-emu/
-
-#ifndef _UCODE_AX_ADPCM_H
-#define _UCODE_AX_ADPCM_H
-
-#include "../../DSP.h"
-
-static inline s16 ADPCM_Step(PBADPCMInfo &adpcm, u32& samplePos, u32 newSamplePos, u16 frac)
-{
-	while (samplePos < newSamplePos)
-	{
-		if ((samplePos & 15) == 0)
-		{
-			adpcm.pred_scale = DSP::ReadARAM((samplePos & ~15) >> 1);
-			samplePos    += 2;
-			newSamplePos += 2;
-		}
-
-		int scale = 1 << (adpcm.pred_scale & 0xF);
-		int coef_idx = (adpcm.pred_scale >> 4) & 7;
-
-		s32 coef1 = adpcm.coefs[coef_idx * 2 + 0];
-		s32 coef2 = adpcm.coefs[coef_idx * 2 + 1];
-
-		int temp = (samplePos & 1) ?
-			   (DSP::ReadARAM(samplePos >> 1) & 0xF) :
-			   (DSP::ReadARAM(samplePos >> 1) >> 4);
-
-		if (temp >= 8)
-			temp -= 16;
-
-		// 0x400 = 0.5 in 11-bit fixed point
-		int val = (scale * temp) + ((0x400 + coef1 * adpcm.yn1 + coef2 * adpcm.yn2) >> 11);
-
-		if (val > 0x7FFF)
-			val = 0x7FFF;
-		else if (val < -0x7FFF)
-			val = -0x7FFF;
-		
-		adpcm.yn2 = adpcm.yn1;
-		adpcm.yn1 = val;
-
-		samplePos++;
-	}
-
-	return adpcm.yn1;
-}
-
-// TODO: WTF is going on here?!?
-// Volume control (ramping)
-static inline u16 ADPCM_Vol(u16 vol, u16 delta)
-{
-	int x = vol;
-	if (delta && delta < 0x5000)
-		x += delta * 20 * 8; // unsure what the right step is
-		//x += 1 * 20 * 8;
-	else if (delta && delta > 0x5000)
-		//x -= (0x10000 - delta); // this is to small, it's often 1
-		x -= (0x10000 - delta) * 20 * 16; // if this was 20 * 8 the sounds in Fire Emblem and Paper Mario
-			// did not have time to go to zero before the were closed
-		//x -= 1 * 20 * 16;
-
-	 // make lower limits
-	if (x < 0) x = 0;	
-	//if (pb.mixer_control < 1000 && x < pb.mixer_control) x = pb.mixer_control; // does this make
-		// any sense?
-
-	// make upper limits
-	//if (mixer_control > 1000 && x > mixer_control) x = mixer_control; // maybe mixer_control also
-		// has a volume target?
-	//if (x >= 0x7fff) x = 0x7fff; // this seems a little high
-	//if (x >= 0x4e20) x = 0x4e20; // add a definitive limit at 20 000
-	if (x >= 0x8000) x = 0x8000; // clamp to 32768;
-	return x; // update volume
-}
-
-#endif  // _UCODE_AX_ADPCM_H
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Structs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Structs.h
deleted file mode 100644
index 7f082740de..0000000000
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Structs.h
+++ /dev/null
@@ -1,365 +0,0 @@
-// Copyright (C) 2003 Dolphin Project.
-
-// This program is free software: you can redistribute it and/or modify
-// it under the terms of the GNU General Public License as published by
-// the Free Software Foundation, version 2.0.
-
-// This program is distributed in the hope that it will be useful,
-// but WITHOUT ANY WARRANTY; without even the implied warranty of
-// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-// GNU General Public License 2.0 for more details.
-
-// A copy of the GPL 2.0 should have been included with the program.
-// If not, see http://www.gnu.org/licenses/
-
-// Official SVN repository and contact information can be found at
-// http://code.google.com/p/dolphin-emu/
-
-#ifndef _UCODE_AX_STRUCTS_H
-#define _UCODE_AX_STRUCTS_H
-
-struct PBMixer
-{
-	u16 left;
-	u16 left_delta;
-	u16 right;
-	u16 right_delta;
-
-	u16 unknown3[8];
-	u16 unknown4[6];
-};
-
-struct PBMixerWii
-{
-	// volume mixing values in .15, 0x8000 = ca. 1.0
-	u16 left;
-	u16 left_delta;
-	u16 right;
-	u16 right_delta;
-
-	u16 auxA_left;
-	u16 auxA_left_delta;
-	u16 auxA_right;
-	u16 auxA_right_delta;
-
-	u16 auxB_left;
-	u16 auxB_left_delta;
-	u16 auxB_right;
-	u16 auxB_right_delta;
-
-	// Note: the following elements usage changes a little in DPL2 mode
-	// TODO: implement and comment it in the mixer
-	u16 auxC_left;
-	u16 auxC_left_delta;
-	u16 auxC_right;
-	u16 auxC_right_delta;
-
-	u16 surround;
-	u16 surround_delta;
-	u16 auxA_surround;
-	u16 auxA_surround_delta;
-	u16 auxB_surround;
-	u16 auxB_surround_delta;
-	u16 auxC_surround;
-	u16 auxC_surround_delta;
-};
-
-struct PBMixerWM
-{
-	u16 main0;
-	u16 main0_delta;
-	u16 aux0;
-	u16 aux0_delta;
-
-	u16 main1;
-	u16 main1_delta;
-	u16 aux1;
-	u16 aux1_delta;
-
-	u16 main2;
-	u16 main2_delta;
-	u16 aux2;
-	u16 aux2_delta;
-
-	u16 main3;
-	u16 main3_delta;
-	u16 aux3;
-	u16 aux3_delta;
-};
-
-struct PBInitialTimeDelay
-{
-	u16 on;
-	u16 addrMemHigh;
-	u16 addrMemLow;
-	u16 offsetLeft;
-	u16 offsetRight;
-	u16 targetLeft;
-	u16 targetRight;
-};
-
-// Update data - read these each 1ms subframe and use them!
-// It seems that to provide higher time precisions for MIDI events, some games
-// use this thing to update the parameter blocks per 1ms sub-block (a block is 5ms).
-// Using this data should fix games that are missing MIDI notes.
-struct PBUpdates
-{
-	u16 num_updates[5];
-	u16 data_hi;  // These point to main RAM. Not sure about the structure of the data.
-	u16 data_lo;
-};
-
-// The DSP stores the final sample values for each voice after every frame of processing.
-// The values are then accumulated for all dropped voices, added to the next frame of audio,
-// and ramped down on a per-sample basis to provide a gentle "roll off."
-struct PBDpop
-{
-	s16 unknown[9];
-};
-
-struct PBDpopWii
-{
-	s16 left;
-	s16 auxA_left;
-	s16 auxB_left;
-	s16 auxC_left;
-
-	s16 right;
-	s16 auxA_right;
-	s16 auxB_right;
-	s16 auxC_right;
-
-	s16 surround;
-	s16 auxA_surround;
-	s16 auxB_surround;
-	s16 auxC_surround;
-};
-
-struct PBDpopWM
-{
-	s16 aMain0;
-	s16 aMain1;
-	s16 aMain2;
-	s16 aMain3;
-
-	s16 aAux0;
-	s16 aAux1;
-	s16 aAux2;
-	s16 aAux3;
-};
-
-struct PBVolumeEnvelope
-{
-	u16 cur_volume;			// volume at start of frame
-	s16 cur_volume_delta;	// signed per sample delta (96 samples per frame)
-};
-
-struct PBUnknown2
-{
-	u16 unknown_reserved[3];
-};
-
-struct PBAudioAddr
-{
-	u16 looping;
-	u16 sample_format;
-	u16 loop_addr_hi;  // Start of loop (this will point to a shared "zero" buffer if one-shot mode is active)
-	u16 loop_addr_lo;
-	u16 end_addr_hi;   // End of sample (and loop), inclusive
-	u16 end_addr_lo;
-	u16 cur_addr_hi;
-	u16 cur_addr_lo;
-};
-
-struct PBADPCMInfo
-{
-	s16 coefs[16];
-	u16 gain;
-	u16 pred_scale;
-	s16 yn1;
-	s16 yn2;
-};
-
-struct PBSampleRateConverter
-{
-	// ratio = (f32)ratio * 0x10000;
-	// valid range is 1/512 to 4.0000
-	u16 ratio_hi; // integer part of sampling ratio
-	u16 ratio_lo; // fraction part of sampling ratio
-	u16 cur_addr_frac;
-	u16 last_samples[4];
-};
-
-struct PBSampleRateConverterWM
-{
-	u16 currentAddressFrac;
-	u16 last_samples[4];
-};
-
-struct PBADPCMLoopInfo
-{
-	u16 pred_scale;
-	u16 yn1;
-	u16 yn2;
-};
-
-struct AXPB
-{
-	u16 next_pb_hi;
-	u16 next_pb_lo;
-	u16 this_pb_hi;
-	u16 this_pb_lo;
-
-	u16 src_type;     // Type of sample rate converter (none, ?, linear)
-	u16 coef_select;
-	u16 mixer_control;
-
-	u16 running;       // 1=RUN 0=STOP
-	u16 is_stream;     // 1 = stream, 0 = one shot
-
-	PBMixer mixer;
-	PBInitialTimeDelay initial_time_delay;  
-	PBUpdates updates;
-	PBDpop dpop;
-	PBVolumeEnvelope vol_env;
-	PBUnknown2 unknown3;
-	PBAudioAddr audio_addr;
-	PBADPCMInfo adpcm;
-	PBSampleRateConverter src;
-	PBADPCMLoopInfo adpcm_loop_info;
-	u16 unknown_maybe_padding[3];
-};
-
-struct PBLowPassFilter
-{
-	u16 enabled;
-	u16 yn1;
-	u16 a0;
-	u16 b0;
-};
-
-struct PBBiquadFilter
-{
-
-	u16 on;		// on = 2, off = 0
-	u16 xn1;	// History data
-	u16 xn2;
-	u16 yn1;
-	u16 yn2;
-	u16 b0;		// Filter coefficients
-	u16 b1;
-	u16 b2;
-	u16 a1;
-	u16 a2;
-
-};
-
-union PBInfImpulseResponseWM
-{
-	PBLowPassFilter lpf;
-	PBBiquadFilter biquad;
-};
-
-struct AXPBWii
-{
-	u16 next_pb_hi;
-	u16 next_pb_lo;
-	u16 this_pb_hi;
-	u16 this_pb_lo;
-
-	u16 src_type;		// Type of sample rate converter (none, 4-tap, linear)
-	u16 coef_select;	// coef for the 4-tap src
-	u32 mixer_control;
-
-	u16 running;		// 1=RUN   0=STOP
-	u16 is_stream;		// 1 = stream, 0 = one shot
-
-	PBMixerWii mixer;
-	PBInitialTimeDelay initial_time_delay;
-	PBDpopWii dpop;
-	PBVolumeEnvelope vol_env;
-	PBAudioAddr audio_addr;
-	PBADPCMInfo adpcm;
-	PBSampleRateConverter src;
-	PBADPCMLoopInfo adpcm_loop_info;
-	PBLowPassFilter lpf;
-	PBBiquadFilter biquad;
-
-	// WIIMOTE :D
-	u16 remote;
-	u16 remote_mixer_control;
-
-	PBMixerWM remote_mixer;
-	PBDpopWM remote_dpop;
-	PBSampleRateConverterWM remote_src;
-	PBInfImpulseResponseWM remote_iir;
-
-	u16 pad[12]; // align us, captain! (32B)
-};
-
-// Seems like nintendo used an early version of AXWii and forgot to remove the update functionality ;p
-struct PBUpdatesWiiSports
-{
-	u16 num_updates[3];
-	u16 data_hi;
-	u16 data_lo;
-};
-
-struct AXPBWiiSports
-{
-	u16 next_pb_hi;
-	u16 next_pb_lo;
-	u16 this_pb_hi;
-	u16 this_pb_lo;
-
-	u16 src_type;		// Type of sample rate converter (none, 4-tap, linear)
-	u16 coef_select;	// coef for the 4-tap src
-	u32 mixer_control;
-
-	u16 running;		// 1=RUN   0=STOP
-	u16 is_stream;		// 1 = stream, 0 = one shot
-
-	PBMixerWii mixer;
-	PBInitialTimeDelay initial_time_delay;
-	PBUpdatesWiiSports updates;
-	PBDpopWii dpop;
-	PBVolumeEnvelope vol_env;
-	PBAudioAddr audio_addr;
-	PBADPCMInfo adpcm;
-	PBSampleRateConverter src;
-	PBADPCMLoopInfo adpcm_loop_info;
-	PBLowPassFilter lpf;
-	PBBiquadFilter biquad;
-
-	// WIIMOTE :D
-	u16 remote;
-	u16 remote_mixer_control;
-
-	PBMixerWM remote_mixer;
-	PBDpopWM remote_dpop;
-	PBSampleRateConverterWM remote_src;
-	PBInfImpulseResponseWM remote_iir;
-
-	u16 pad[7]; // align us, captain! (32B)
-};
-
-// TODO: All these enums have changed a lot for wii
-enum {
-    AUDIOFORMAT_ADPCM = 0,
-    AUDIOFORMAT_PCM8  = 0x19,
-	AUDIOFORMAT_PCM16 = 0xA,
-};
-
-enum {
-	SRCTYPE_LINEAR  = 1,
-	SRCTYPE_NEAREST = 2,
-	MIXCONTROL_RAMPING = 8,
-};
-
-// Both may be used at once
-enum {
-	FILTER_LOWPASS = 1,
-	FILTER_BIQUAD = 2,
-};
-
-#endif  // _UCODE_AX_STRUCTS_H
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Voice.h
deleted file mode 100644
index 55f7face27..0000000000
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii_Voice.h
+++ /dev/null
@@ -1,271 +0,0 @@
-// Copyright (C) 2003 Dolphin Project.
-
-// This program is free software: you can redistribute it and/or modify
-// it under the terms of the GNU General Public License as published by
-// the Free Software Foundation, version 2.0.
-
-// This program is distributed in the hope that it will be useful,
-// but WITHOUT ANY WARRANTY; without even the implied warranty of
-// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-// GNU General Public License 2.0 for more details.
-
-// A copy of the GPL 2.0 should have been included with the program.
-// If not, see http://www.gnu.org/licenses/
-
-// Official SVN repository and contact information can be found at
-// http://code.google.com/p/dolphin-emu/
-
-#ifndef _UCODE_AXWII_VOICE_H
-#define _UCODE_AXWII_VOICE_H
-
-#include "UCodes.h"
-#include "UCode_AXWii_ADPCM.h"
-#include "UCode_AX.h"
-#include "Mixer.h"
-#include "../../AudioInterface.h"
-
-// MRAM -> ARAM for GC
-inline bool ReadPB(u32 addr, AXPB &PB)
-{
-	const u16* PB_in_mram = (const u16*)Memory::GetPointer(addr);
-	if (PB_in_mram == NULL)
-		return false;
-	u16* PB_in_aram = (u16*)&PB;
-
-	for (size_t p = 0; p < (sizeof(AXPB) >> 1); p++)
-	{
-		PB_in_aram[p] = Common::swap16(PB_in_mram[p]);
-	}
-
-	return true;
-}
-
-// MRAM -> ARAM for Wii
-inline bool ReadPB(u32 addr, AXPBWii &PB)
-{
-	const u16* PB_in_mram = (const u16*)Memory::GetPointer(addr);
-	if (PB_in_mram == NULL)
-		return false;
-	u16* PB_in_aram = (u16*)&PB;
-
-	// preswap the mixer_control
-	PB.mixer_control = ((u32)PB_in_mram[7] << 16) | ((u32)PB_in_mram[6] >> 16);
-
-	for (size_t p = 0; p < (sizeof(AXPBWii) >> 1); p++)
-	{
-		PB_in_aram[p] = Common::swap16(PB_in_mram[p]);
-	}
-
-	return true;
-}
-
-// ARAM -> MRAM for GC
-inline bool WritePB(u32 addr, AXPB &PB)
-{
-	const u16* PB_in_aram  = (const u16*)&PB;
-	u16* PB_in_mram = (u16*)Memory::GetPointer(addr);
-	if (PB_in_mram == NULL)
-		return false;
-
-	for (size_t p = 0; p < (sizeof(AXPB) >> 1); p++)
-	{
-		PB_in_mram[p] = Common::swap16(PB_in_aram[p]);
-	}
-
-	return true;
-}
-
-// ARAM -> MRAM for Wii
-inline bool WritePB(u32 addr, AXPBWii &PB)
-{
-	const u16* PB_in_aram  = (const u16*)&PB;
-	u16* PB_in_mram = (u16*)Memory::GetPointer(addr);
-	if (PB_in_mram == NULL)
-		return false;
-
-	// preswap the mixer_control
-	*(u32*)&PB_in_mram[6] = (PB.mixer_control << 16) | (PB.mixer_control >> 16);
-
-	for (size_t p = 0; p < (sizeof(AXPBWii) >> 1); p++)
-	{
-		PB_in_mram[p] = Common::swap16(PB_in_aram[p]);
-	}
-
-	return true;
-}
-
-//////////////////////////////////////////////////////////////////////////
-// TODO: fix handling of gc/wii PB differences
-// TODO: generally fix up the mess - looks crazy and kinda wrong
-template<class ParamBlockType>
-inline void MixAddVoice(ParamBlockType &pb,
-						int *templbuffer, int *temprbuffer,
-						int _iSize)
-{
-	if (pb.running)
-	{
-		const u32 ratio = (u32)(((pb.src.ratio_hi << 16) + pb.src.ratio_lo)
-			* /*ratioFactor:*/((float)AudioInterface::GetAIDSampleRate() / (float)soundStream->GetMixer()->GetSampleRate()));
-		u32 sampleEnd = (pb.audio_addr.end_addr_hi << 16) | pb.audio_addr.end_addr_lo;
-		u32 loopPos   = (pb.audio_addr.loop_addr_hi << 16) | pb.audio_addr.loop_addr_lo;
-
-		u32 samplePos = (pb.audio_addr.cur_addr_hi << 16) | pb.audio_addr.cur_addr_lo;
-		u32 frac = pb.src.cur_addr_frac;
-
-		// =======================================================================================
-		// Handle No-SRC streams - No src streams have pb.src_type == 2 and have pb.src.ratio_hi = 0
-		// and pb.src.ratio_lo = 0. We handle that by setting the sampling ratio integer to 1. This
-		// makes samplePos update in the correct way. I'm unsure how we are actually supposed to
-		// detect that this setting. Updates did not fix this automatically.
-		// ---------------------------------------------------------------------------------------
-		// Stream settings
-		// src_type = 2 (most other games have src_type = 0)
-		// Affected games:
-		// Baten Kaitos - Eternal Wings (2003)
-		// Baten Kaitos - Origins (2006)?
-		// Soul Calibur 2: The movie music use src_type 2 but it needs no adjustment, perhaps
-		// the sound format plays in to, Baten use ADPCM, SC2 use PCM16
-		//if (pb.src_type == 2 && (pb.src.ratio_hi == 0 && pb.src.ratio_lo == 0))
-		if (pb.running && (pb.src.ratio_hi == 0 && pb.src.ratio_lo == 0))
-		{
-			pb.src.ratio_hi = 1;
-		}
-
-		// =======================================================================================
-		// Games that use looping to play non-looping music streams - SSBM has info in all
-		// pb.adpcm_loop_info parameters but has pb.audio_addr.looping = 0. If we treat these streams
-		// like any other looping streams the music works. I'm unsure how we are actually supposed to
-		// detect that these kinds of blocks should be looping. It seems like pb.mixer_control == 0 may
-		// identify these types of blocks. Updates did not write any looping values.
-		if (
-			(pb.adpcm_loop_info.pred_scale || pb.adpcm_loop_info.yn1 || pb.adpcm_loop_info.yn2)
-			&& pb.mixer_control == 0 && pb.adpcm_loop_info.pred_scale <= 0x7F
-		   )
-		{
-			pb.audio_addr.looping = 1;
-		}
-
-
-
-		// Top Spin 3 Wii
-		if (pb.audio_addr.sample_format > 25)
-			pb.audio_addr.sample_format = 0;
-
-		// =======================================================================================
-		// Walk through _iSize. _iSize = numSamples. If the game goes slow _iSize will be higher to
-		// compensate for that. _iSize can be as low as 100 or as high as 2000 some cases.
-		for (int s = 0; s < _iSize; s++)
-		{
-			int sample = 0;
-			u32 oldFrac = frac;
-			frac += ratio;
-			u32 newSamplePos = samplePos + (frac >> 16); //whole number of frac
-
-			// =======================================================================================
-			// Process sample format
-			switch (pb.audio_addr.sample_format)
-			{
-			    case AUDIOFORMAT_PCM8:
-				    pb.adpcm.yn2 = ((s8)DSP::ReadARAM(samplePos)) << 8; //current sample
-				    pb.adpcm.yn1 = ((s8)DSP::ReadARAM(samplePos + 1)) << 8; //next sample
-
-				    if (pb.src_type == SRCTYPE_NEAREST)
-					    sample = pb.adpcm.yn2;
-				    else // linear interpolation
-						sample = (pb.adpcm.yn1 * (u16)oldFrac + pb.adpcm.yn2 * (u16)(0xFFFF - oldFrac) + pb.adpcm.yn2) >> 16;
-
-				    samplePos = newSamplePos;
-				    break;
-
-			    case AUDIOFORMAT_PCM16:
-				    pb.adpcm.yn2 = (s16)(u16)((DSP::ReadARAM(samplePos * 2) << 8) | (DSP::ReadARAM((samplePos * 2 + 1)))); //current sample
-				    pb.adpcm.yn1 = (s16)(u16)((DSP::ReadARAM((samplePos + 1) * 2) << 8) | (DSP::ReadARAM(((samplePos + 1) * 2 + 1)))); //next sample
-
-					if (pb.src_type == SRCTYPE_NEAREST)
-					    sample = pb.adpcm.yn2;
-				    else // linear interpolation
-					    sample = (pb.adpcm.yn1 * (u16)oldFrac + pb.adpcm.yn2 * (u16)(0xFFFF - oldFrac) + pb.adpcm.yn2) >> 16;
-
-				    samplePos = newSamplePos;
-				    break;
-
-			    case AUDIOFORMAT_ADPCM:
-					ADPCM_Step(pb.adpcm, samplePos, newSamplePos, frac);
-					
-					if (pb.src_type == SRCTYPE_NEAREST)
-					    sample = pb.adpcm.yn2;
-				    else // linear interpolation
-					    sample = (pb.adpcm.yn1 * (u16)frac + pb.adpcm.yn2 * (u16)(0xFFFF - frac) + pb.adpcm.yn2) >> 16; //adpcm moves on frac
-
-				    break;
-
-			    default:
-				    break;
-			}
-
-			// ===================================================================
-			// Overall volume control. In addition to this there is also separate volume settings to
-			// different channels (left, right etc).
-			frac &= 0xffff;
-
-			int vol = pb.vol_env.cur_volume >> 9;
-			sample = sample * vol >> 8;
-
-			if (pb.mixer_control & MIXCONTROL_RAMPING)
-			{
-				int x = pb.vol_env.cur_volume;
-				x += pb.vol_env.cur_volume_delta;	// I'm not sure about this, can anybody find a game
-													// that use this? Or how does it work?
-				if (x < 0)
-					x = 0;
-				if (x >= 0x7fff)
-					x = 0x7fff;
-				pb.vol_env.cur_volume = x; // maybe not per sample?? :P
-			}
-
-			int leftmix		= pb.mixer.left >> 5;
-			int rightmix	= pb.mixer.right >> 5;
-			int left		= sample * leftmix >> 8;
-			int right		= sample * rightmix >> 8;
-			// adpcm has to walk from oldSamplePos to samplePos here
-			templbuffer[s] += left;
-			temprbuffer[s] += right;
-
-			// Control the behavior when we reach the end of the sample
-			if (samplePos >= sampleEnd)
-			{
-				if (pb.audio_addr.looping == 1)
-				{
-					if ((samplePos & ~0x1f) == (sampleEnd & ~0x1f) || (pb.audio_addr.sample_format != AUDIOFORMAT_ADPCM))
-						samplePos = loopPos;
-					if ((!pb.is_stream) && (pb.audio_addr.sample_format == AUDIOFORMAT_ADPCM))
-					{
-						pb.adpcm.yn1 = pb.adpcm_loop_info.yn1;
-						pb.adpcm.yn2 = pb.adpcm_loop_info.yn2;
-						pb.adpcm.pred_scale = pb.adpcm_loop_info.pred_scale;
-					}
-				}
-				else
-				{
-					pb.running = 0;
-					samplePos = loopPos;
-					//samplePos = samplePos - sampleEnd + loopPos;
-					memset(&pb.dpop, 0, sizeof(pb.dpop));
-					memset(pb.src.last_samples, 0, 8);
-					break;
-				}
-			}
-		} // end of the _iSize loop
-
-		// Update volume
-		pb.mixer.left = ADPCM_Vol(pb.mixer.left, pb.mixer.left_delta);
-		pb.mixer.right = ADPCM_Vol(pb.mixer.right, pb.mixer.right_delta);
-
-		pb.src.cur_addr_frac = (u16)frac;
-		pb.audio_addr.cur_addr_hi = samplePos >> 16;
-		pb.audio_addr.cur_addr_lo = (u16)samplePos;
-		
-	} // if (pb.running)
-}
-
-#endif
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index e39b38b100..15261b6184 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -27,7 +27,7 @@
 #endif
 
 #include "Common.h"
-#include "UCode_AX_Structs.h"
+#include "UCode_AXStructs.h"
 #include "../../DSP.h"
 
 #ifdef AX_GC
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.cpp
deleted file mode 100644
index 40ad6e1947..0000000000
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.cpp
+++ /dev/null
@@ -1,383 +0,0 @@
-// Copyright (C) 2003 Dolphin Project.
-
-// This program is free software: you can redistribute it and/or modify
-// it under the terms of the GNU General Public License as published by
-// the Free Software Foundation, version 2.0.
-
-// This program is distributed in the hope that it will be useful,
-// but WITHOUT ANY WARRANTY; without even the implied warranty of
-// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-// GNU General Public License 2.0 for more details.
-
-// A copy of the GPL 2.0 should have been included with the program.
-// If not, see http://www.gnu.org/licenses/
-
-// Official SVN repository and contact information can be found at
-// http://code.google.com/p/dolphin-emu/
-
-#include "StringUtil.h"
-
-#include "../MailHandler.h"
-#include "Mixer.h"
-
-#include "UCodes.h"
-#include "UCode_AX_Structs.h"
-#include "UCode_NewAXWii.h"
-
-#define AX_WII
-#include "UCode_AX_Voice.h"
-
-
-CUCode_NewAXWii::CUCode_NewAXWii(DSPHLE *dsp_hle, u32 l_CRC)
-	: CUCode_AX(dsp_hle, l_CRC)
-{
-	WARN_LOG(DSPHLE, "Instantiating CUCode_NewAXWii");
-}
-
-CUCode_NewAXWii::~CUCode_NewAXWii()
-{
-}
-
-void CUCode_NewAXWii::HandleCommandList()
-{
-	// Temp variables for addresses computation
-	u16 addr_hi, addr_lo;
-	u16 addr2_hi, addr2_lo;
-	u16 volume;
-
-//	WARN_LOG(DSPHLE, "Command list:");
-//	for (u32 i = 0; m_cmdlist[i] != CMD_END; ++i)
-//		WARN_LOG(DSPHLE, "%04x", m_cmdlist[i]);
-//	WARN_LOG(DSPHLE, "-------------");
-
-	u32 curr_idx = 0;
-	bool end = false;
-	while (!end)
-	{
-		u16 cmd = m_cmdlist[curr_idx++];
-
-		switch (cmd)
-		{
-			// Some of these commands are unknown, or unused in this AX HLE.
-			// We still need to skip their arguments using "curr_idx += N".
-
-			case CMD_SETUP:
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				SetupProcessing(HILO_TO_32(addr));
-				break;
-
-			case CMD_UNK_01: curr_idx += 2; break;
-			case CMD_UNK_02: curr_idx += 2; break;
-			case CMD_UNK_03: curr_idx += 2; break;
-
-			case CMD_PROCESS:
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				ProcessPBList(HILO_TO_32(addr));
-				break;
-
-			case CMD_MIX_AUXA:
-			case CMD_MIX_AUXB:
-			case CMD_MIX_AUXC:
-				volume = m_cmdlist[curr_idx++];
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				addr2_hi = m_cmdlist[curr_idx++];
-				addr2_lo = m_cmdlist[curr_idx++];
-				MixAUXSamples(cmd - CMD_MIX_AUXA, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
-				break;
-
-			// These two go together and manipulate some AUX buffers.
-			case CMD_UNK_08: curr_idx += 13; break;
-			case CMD_UNK_09: curr_idx += 13; break;
-
-			case CMD_UNK_0A: curr_idx += 4; break;
-
-			case CMD_OUTPUT:
-				volume = m_cmdlist[curr_idx++];
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				addr2_hi = m_cmdlist[curr_idx++];
-				addr2_lo = m_cmdlist[curr_idx++];
-				OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), volume);
-				break;
-
-			case CMD_UNK_0C: curr_idx += 5; break;
-
-			case CMD_WM_OUTPUT:
-			{
-				u32 addresses[4] = {
-					(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
-					(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
-					(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
-					(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
-				};
-				curr_idx += 8;
-				OutputWMSamples(addresses);
-				break;
-			}
-
-			case CMD_END:
-				end = true;
-				break;
-		}
-	}
-}
-
-void CUCode_NewAXWii::SetupProcessing(u32 init_addr)
-{
-	// TODO: should be easily factorizable with AX
-	s16 init_data[60];
-
-	for (u32 i = 0; i < 60; ++i)
-		init_data[i] = HLEMemory_Read_U16(init_addr + 2 * i);
-
-	// List of all buffers we have to initialize
-	struct {
-		int* ptr;
-		u32 samples;
-	} buffers[] = {
-		{ m_samples_left, 32 },
-		{ m_samples_right, 32 },
-		{ m_samples_surround, 32 },
-		{ m_samples_auxA_left, 32 },
-		{ m_samples_auxA_right, 32 },
-		{ m_samples_auxA_surround, 32 },
-		{ m_samples_auxB_left, 32 },
-		{ m_samples_auxB_right, 32 },
-		{ m_samples_auxB_surround, 32 },
-		{ m_samples_auxC_left, 32 },
-		{ m_samples_auxC_right, 32 },
-		{ m_samples_auxC_surround, 32 },
-
-		{ m_samples_wm0, 6 },
-		{ m_samples_aux0, 6 },
-		{ m_samples_wm1, 6 },
-		{ m_samples_aux1, 6 },
-		{ m_samples_wm2, 6 },
-		{ m_samples_aux2, 6 },
-		{ m_samples_wm3, 6 },
-		{ m_samples_aux3, 6 }
-	};
-
-	u32 init_idx = 0;
-	for (u32 i = 0; i < sizeof (buffers) / sizeof (buffers[0]); ++i)
-	{
-		s32 init_val = (s32)((init_data[init_idx] << 16) | init_data[init_idx + 1]);
-		s16 delta = (s16)init_data[init_idx + 2];
-
-		init_idx += 3;
-
-		if (!init_val)
-			memset(buffers[i].ptr, 0, 3 * buffers[i].samples * sizeof (int));
-		else
-		{
-			for (u32 j = 0; j < 3 * buffers[i].samples; ++j)
-			{
-				buffers[i].ptr[j] = init_val;
-				init_val += delta;
-			}
-		}
-	}
-}
-
-AXMixControl CUCode_NewAXWii::ConvertMixerControl(u32 mixer_control)
-{
-	u32 ret = 0;
-
-	if (mixer_control & 0x00000001) ret |= MIX_L;
-	if (mixer_control & 0x00000002) ret |= MIX_R;
-	if (mixer_control & 0x00000004) ret |= MIX_L_RAMP | MIX_R_RAMP;
-	if (mixer_control & 0x00000008) ret |= MIX_S;
-	if (mixer_control & 0x00000010) ret |= MIX_S_RAMP;
-	if (mixer_control & 0x00010000) ret |= MIX_AUXA_L;
-	if (mixer_control & 0x00020000) ret |= MIX_AUXA_R;
-	if (mixer_control & 0x00040000) ret |= MIX_AUXA_L_RAMP | MIX_AUXA_R_RAMP;
-	if (mixer_control & 0x00080000) ret |= MIX_AUXA_S;
-	if (mixer_control & 0x00100000) ret |= MIX_AUXA_S_RAMP;
-	if (mixer_control & 0x00200000) ret |= MIX_AUXB_L;
-	if (mixer_control & 0x00400000) ret |= MIX_AUXB_R;
-	if (mixer_control & 0x00800000) ret |= MIX_AUXB_L_RAMP | MIX_AUXB_R_RAMP;
-	if (mixer_control & 0x01000000) ret |= MIX_AUXB_S;
-	if (mixer_control & 0x02000000) ret |= MIX_AUXB_S_RAMP;
-	if (mixer_control & 0x04000000) ret |= MIX_AUXC_L;
-	if (mixer_control & 0x08000000) ret |= MIX_AUXC_R;
-	if (mixer_control & 0x10000000) ret |= MIX_AUXC_L_RAMP | MIX_AUXC_R_RAMP;
-	if (mixer_control & 0x20000000) ret |= MIX_AUXC_S;
-	if (mixer_control & 0x40000000) ret |= MIX_AUXC_S_RAMP;
-
-	return (AXMixControl)ret;
-}
-
-void CUCode_NewAXWii::ProcessPBList(u32 pb_addr)
-{
-	const u32 spms = 32;
-
-	AXPBWii pb;
-
-	while (pb_addr)
-	{
-		AXBuffers buffers = {{
-			m_samples_left,
-			m_samples_right,
-			m_samples_surround,
-			m_samples_auxA_left,
-			m_samples_auxA_right,
-			m_samples_auxA_surround,
-			m_samples_auxB_left,
-			m_samples_auxB_right,
-			m_samples_auxB_surround,
-			m_samples_auxC_left,
-			m_samples_auxC_right,
-			m_samples_auxC_surround
-		}};
-
-		if (!ReadPB(pb_addr, pb))
-			break;
-
-		for (int curr_ms = 0; curr_ms < 3; ++curr_ms)
-		{
-			Process1ms(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)));
-
-			// Forward the buffers
-			for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
-				buffers.ptrs[i] += spms;
-		}
-
-		WritePB(pb_addr, pb);
-		pb_addr = HILO_TO_32(pb.next_pb);
-	}
-}
-
-void CUCode_NewAXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume)
-{
-	int* buffers[3] = { 0 };
-	int* main_buffers[3] = {
-		m_samples_left,
-		m_samples_right,
-		m_samples_surround
-	};
-
-	switch (aux_id)
-	{
-	case 0:
-		buffers[0] = m_samples_auxA_left;
-		buffers[1] = m_samples_auxA_right;
-		buffers[2] = m_samples_auxA_surround;
-		break;
-
-	case 1:
-		buffers[0] = m_samples_auxB_left;
-		buffers[1] = m_samples_auxB_right;
-		buffers[2] = m_samples_auxB_surround;
-		break;
-
-	case 2:
-		buffers[0] = m_samples_auxC_left;
-		buffers[1] = m_samples_auxC_right;
-		buffers[2] = m_samples_auxC_surround;
-		break;
-	}
-
-	// Send the content of AUX buffers to the CPU
-	if (write_addr)
-	{
-		int* ptr = (int*)HLEMemory_Get_Pointer(write_addr);
-		for (u32 i = 0; i < 3; ++i)
-			for (u32 j = 0; j < 3 * 32; ++j)
-				*ptr++ = Common::swap32(buffers[i][j]);
-	}
-
-	// Then read the buffers from the CPU and add to our main buffers.
-	int* ptr = (int*)HLEMemory_Get_Pointer(read_addr);
-	for (u32 i = 0; i < 3; ++i)
-		for (u32 j = 0; j < 3 * 32; ++j)
-		{
-			s64 new_val = main_buffers[i][j] + Common::swap32(*ptr++);
-			main_buffers[i][j] = (new_val * volume) >> 15;
-		}
-}
-
-void CUCode_NewAXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
-{
-	int surround_buffer[3 * 32] = { 0 };
-
-	for (u32 i = 0; i < 3 * 32; ++i)
-		surround_buffer[i] = Common::swap32(m_samples_surround[i]);
-	memcpy(HLEMemory_Get_Pointer(surround_addr), surround_buffer, sizeof (surround_buffer));
-
-	short buffer[3 * 32 * 2];
-
-	// Clamp internal buffers to 16 bits.
-	for (u32 i = 0; i < 3 * 32; ++i)
-	{
-		int left  = m_samples_left[i];
-		int right = m_samples_right[i];
-
-		// Apply global volume. Cast to s64 to avoid overflow.
-		left = ((s64)left * volume) >> 15;
-		right = ((s64)right * volume) >> 15;
-
-		if (left < -32767)  left = -32767;
-		if (left > 32767)   left = 32767;
-		if (right < -32767) right = -32767;
-		if (right >  32767) right = 32767;
-
-		m_samples_left[i] = left;
-		m_samples_right[i] = right;
-	}
-
-	for (u32 i = 0; i < 3 * 32; ++i)
-	{
-		buffer[2 * i] = Common::swap16(m_samples_left[i]);
-		buffer[2 * i + 1] = Common::swap16(m_samples_right[i]);
-	}
-
-	memcpy(HLEMemory_Get_Pointer(lr_addr), buffer, sizeof (buffer));
-}
-
-void CUCode_NewAXWii::OutputWMSamples(u32* addresses)
-{
-	int* buffers[] = {
-		m_samples_wm0,
-		m_samples_wm1,
-		m_samples_wm2,
-		m_samples_wm3
-	};
-
-	for (u32 i = 0; i < 4; ++i)
-	{
-		int* in = buffers[i];
-		u16* out = (u16*)HLEMemory_Get_Pointer(addresses[i]);
-		for (u32 j = 0; j < 3 * 6; ++j)
-		{
-			int sample = in[j];
-			if (sample < -32767) sample = -32767;
-			if (sample > 32767) sample = 32767;
-			out[j] = Common::swap16((u16)sample);
-		}
-	}
-}
-
-void CUCode_NewAXWii::DoState(PointerWrap &p)
-{
-	std::lock_guard<std::mutex> lk(m_processing);
-
-	DoStateShared(p);
-	DoAXState(p);
-
-	p.Do(m_samples_auxC_left);
-	p.Do(m_samples_auxC_right);
-	p.Do(m_samples_auxC_surround);
-
-	p.Do(m_samples_wm0);
-	p.Do(m_samples_wm1);
-	p.Do(m_samples_wm2);
-	p.Do(m_samples_wm3);
-
-	p.Do(m_samples_aux0);
-	p.Do(m_samples_aux1);
-	p.Do(m_samples_aux2);
-	p.Do(m_samples_aux3);
-}
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.h
deleted file mode 100644
index 4c9bc5757c..0000000000
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_NewAXWii.h
+++ /dev/null
@@ -1,80 +0,0 @@
-// Copyright (C) 2003 Dolphin Project.
-
-// This program is free software: you can redistribute it and/or modify
-// it under the terms of the GNU General Public License as published by
-// the Free Software Foundation, version 2.0.
-
-// This program is distributed in the hope that it will be useful,
-// but WITHOUT ANY WARRANTY; without even the implied warranty of
-// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-// GNU General Public License 2.0 for more details.
-
-// A copy of the GPL 2.0 should have been included with the program.
-// If not, see http://www.gnu.org/licenses/
-
-// Official Git repository and contact information can be found at
-// http://code.google.com/p/dolphin-emu/
-
-#ifndef _UCODE_NEWAXWII_H
-#define _UCODE_NEWAXWII_H
-
-#include "UCode_AX.h"
-
-class CUCode_NewAXWii : public CUCode_AX
-{
-public:
-	CUCode_NewAXWii(DSPHLE *dsp_hle, u32 _CRC);
-	virtual ~CUCode_NewAXWii();
-
-	virtual void DoState(PointerWrap &p);
-
-protected:
-	int m_samples_auxC_left[32 * 3];
-	int m_samples_auxC_right[32 * 3];
-	int m_samples_auxC_surround[32 * 3];
-
-	// Wiimote buffers
-	int m_samples_wm0[6 * 3];
-	int m_samples_aux0[6 * 3];
-	int m_samples_wm1[6 * 3];
-	int m_samples_aux1[6 * 3];
-	int m_samples_wm2[6 * 3];
-	int m_samples_aux2[6 * 3];
-	int m_samples_wm3[6 * 3];
-	int m_samples_aux3[6 * 3];
-
-	// Convert a mixer_control bitfield to our internal representation for that
-	// value. Required because that bitfield has a different meaning in some
-	// versions of AX.
-	AXMixControl ConvertMixerControl(u32 mixer_control);
-
-	virtual void HandleCommandList();
-
-	void SetupProcessing(u32 init_addr);
-	void ProcessPBList(u32 pb_addr);
-	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
-	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume);
-	void OutputWMSamples(u32* addresses);	// 4 addresses
-
-private:
-	enum CmdType
-	{
-		CMD_SETUP = 0x00,
-		CMD_UNK_01 = 0x01,
-		CMD_UNK_02 = 0x02,
-		CMD_UNK_03 = 0x03,
-		CMD_PROCESS = 0x04,
-		CMD_MIX_AUXA = 0x05,
-		CMD_MIX_AUXB = 0x06,
-		CMD_MIX_AUXC = 0x07,
-		CMD_UNK_08 = 0x08,
-		CMD_UNK_09 = 0x09,
-		CMD_UNK_0A = 0x0A,
-		CMD_OUTPUT = 0x0B,
-		CMD_UNK_0C = 0x0C,
-		CMD_WM_OUTPUT = 0x0D,
-		CMD_END = 0x0E
-	};
-};
-
-#endif  // _UCODE_AXWII
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCodes.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCodes.cpp
index 86773ad020..c04bf41403 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCodes.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCodes.cpp
@@ -19,7 +19,6 @@
 
 #include "UCode_AX.h"
 #include "UCode_AXWii.h"
-#include "UCode_NewAXWii.h"
 #include "UCode_Zelda.h"
 #include "UCode_ROM.h"
 #include "UCode_CARD.h"
@@ -27,12 +26,6 @@
 #include "UCode_GBA.h"
 #include "Hash.h"
 
-#if 0
-# define AXWII CUCode_NewAXWii
-#else
-# define AXWII CUCode_AXWii
-#endif
-
 IUCode* UCodeFactory(u32 _CRC, DSPHLE *dsp_hle, bool bWii)
 {
 	switch (_CRC)
@@ -97,13 +90,13 @@ IUCode* UCodeFactory(u32 _CRC, DSPHLE *dsp_hle, bool bWii)
 	case 0x4cc52064: // Bleach: Versus Crusade    
     case 0xd9c4bf34: // WiiMenu
 		INFO_LOG(DSPHLE, "CRC %08x: Wii - AXWii chosen", _CRC);
-		return new AXWII(dsp_hle, _CRC);
+		return new CUCode_AXWii(dsp_hle, _CRC);
 
 	default:
 		if (bWii)
 		{
 			PanicAlert("DSPHLE: Unknown ucode (CRC = %08x) - forcing AXWii.\n\nTry LLE emulator if this is homebrew.", _CRC);
-			return new AXWII(dsp_hle, _CRC);
+			return new CUCode_AXWii(dsp_hle, _CRC);
 		}
 		else
 		{

From fb64cfd18ad855abb03896a14d52bb832db082b5 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 07:36:48 +0100
Subject: [PATCH 02/28] AXWii: fix L/R channels being swapped when outputting
 audio samples

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 8c74c7a1e7..8aec99d700 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -330,8 +330,8 @@ void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
 
 	for (u32 i = 0; i < 3 * 32; ++i)
 	{
-		buffer[2 * i] = Common::swap16(m_samples_left[i]);
-		buffer[2 * i + 1] = Common::swap16(m_samples_right[i]);
+		buffer[2 * i] = Common::swap16(m_samples_right[i]);
+		buffer[2 * i + 1] = Common::swap16(m_samples_left[i]);
 	}
 
 	memcpy(HLEMemory_Get_Pointer(lr_addr), buffer, sizeof (buffer));

From 35249a8576c604ad7528c254c7b27e356f1a127b Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 14:42:37 +0100
Subject: [PATCH 03/28] Handle 3ms at a time in AXWii instead of 3x1ms

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp    |  2 +-
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 11 +-------
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 26 ++++++++++++-------
 3 files changed, 18 insertions(+), 21 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
index 57032a774e..cb34c2532d 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
@@ -405,7 +405,7 @@ void CUCode_AX::ProcessPBList(u32 pb_addr)
 		{
 			ApplyUpdatesForMs(pb, curr_ms);
 
-			Process1ms(pb, buffers, ConvertMixerControl(pb.mixer_control));
+			ProcessVoice(pb, buffers, ConvertMixerControl(pb.mixer_control));
 
 			// Forward the buffers
 			for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 8aec99d700..1b1097c7fe 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -212,8 +212,6 @@ AXMixControl CUCode_AXWii::ConvertMixerControl(u32 mixer_control)
 
 void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 {
-	const u32 spms = 32;
-
 	AXPBWii pb;
 
 	while (pb_addr)
@@ -236,14 +234,7 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 		if (!ReadPB(pb_addr, pb))
 			break;
 
-		for (int curr_ms = 0; curr_ms < 3; ++curr_ms)
-		{
-			Process1ms(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)));
-
-			// Forward the buffers
-			for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
-				buffers.ptrs[i] += spms;
-		}
+		ProcessVoice(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)));
 
 		WritePB(pb_addr, pb);
 		pb_addr = HILO_TO_32(pb.next_pb);
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 15261b6184..578330ba34 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -32,8 +32,12 @@
 
 #ifdef AX_GC
 # define PB_TYPE AXPB
+# define MS_PER_FRAME 1
+# define SAMPLES_PER_FRAME 32
 #else
 # define PB_TYPE AXPBWii
+# define MS_PER_FRAME 3
+# define SAMPLES_PER_FRAME 96
 #endif
 
 // Put all of that in an anonymous namespace to avoid stupid compilers merging
@@ -232,7 +236,8 @@ u16 AcceleratorGetSample()
 	return ret;
 }
 
-// Read 32 input samples from ARAM, decoding and converting rate if required.
+// Read SAMPLES_PER_FRAME input samples from ARAM, decoding and converting rate
+// if required.
 void GetInputSamples(PB_TYPE& pb, s16* samples)
 {
 	u32 cur_addr = HILO_TO_32(pb.audio_addr.cur_addr);
@@ -260,7 +265,7 @@ void GetInputSamples(PB_TYPE& pb, s16* samples)
 		s16 curr0 = pb.src.last_samples[2];
 		s16 curr1 = pb.src.last_samples[3];
 
-		for (u32 i = 0; i < 32; ++i)
+		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
 		{
 			// Get our current fractional position, used to know how much of
 			// curr0 and how much of curr1 the output sample should be.
@@ -290,12 +295,12 @@ void GetInputSamples(PB_TYPE& pb, s16* samples)
 	}
 	else // SRCTYPE_NEAREST
 	{
-		// No sample rate conversion here: simply read 32 samples from the
+		// No sample rate conversion here: simply read samples from the
 		// accelerator to the output buffer.
-		for (u32 i = 0; i < 32; ++i)
+		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
 			samples[i] = AcceleratorGetSample();
 
-		memcpy(pb.src.last_samples, samples + 28, 4 * sizeof (u16));
+		memcpy(pb.src.last_samples, samples + SAMPLES_PER_FRAME - 4, 4 * sizeof (u16));
 	}
 
 	// Update current position in the PB.
@@ -315,7 +320,7 @@ void MixAdd(int* out, const s16* input, u16* pvol, s16* dpop, bool ramp)
 	if (!ramp)
 		volume_delta = 0;
 
-	for (u32 i = 0; i < 32; ++i)
+	for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
 	{
 		s64 sample = input[i];
 		sample *= volume;
@@ -328,19 +333,20 @@ void MixAdd(int* out, const s16* input, u16* pvol, s16* dpop, bool ramp)
 	}
 }
 
-// Process 1ms of audio (32 samples) from a PB and mix it to the buffers.
-void Process1ms(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
+// Process 1ms of audio (for AX GC) or 3ms of audio (for AX Wii) from a PB and
+// mix it to the output buffers.
+void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
 {
 	// If the voice is not running, nothing to do.
 	if (!pb.running)
 		return;
 
 	// Read input samples, performing sample rate conversion if needed.
-	s16 samples[32];
+	s16 samples[SAMPLES_PER_FRAME];
 	GetInputSamples(pb, samples);
 
 	// Apply a global volume ramp using the volume envelope parameters.
-	for (u32 i = 0; i < 32; ++i)
+	for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
 	{
 		s64 sample = 2 * (s16)samples[i] * (s16)pb.vol_env.cur_volume;
 		samples[i] = (s16)(sample >> 16);

From 25419ed827b48027bfb2cc66bef8c7d9b142b2ae Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 15:32:55 +0100
Subject: [PATCH 04/28] Implement Wiimote sound mixing in AXWii

---
 .../Src/HW/DSPHLE/UCodes/UCode_AXStructs.h    | 18 ++--
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 10 +-
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 93 +++++++++++++++----
 3 files changed, 94 insertions(+), 27 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
index c92196dd49..d82088b34f 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
@@ -161,15 +161,15 @@ struct PBDpopWii
 
 struct PBDpopWM
 {
-	s16 aMain0;
-	s16 aMain1;
-	s16 aMain2;
-	s16 aMain3;
+	s16 main0;
+	s16 main1;
+	s16 main2;
+	s16 main3;
 
-	s16 aAux0;
-	s16 aAux1;
-	s16 aAux2;
-	s16 aAux3;
+	s16 aux0;
+	s16 aux1;
+	s16 aux2;
+	s16 aux3;
 };
 
 struct PBVolumeEnvelope
@@ -216,7 +216,7 @@ struct PBSampleRateConverter
 
 struct PBSampleRateConverterWM
 {
-	u16 currentAddressFrac;
+	u16 cur_addr_frac;
 	u16 last_samples[4];
 };
 
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 1b1097c7fe..84fab9108b 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -228,7 +228,15 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 			m_samples_auxB_surround,
 			m_samples_auxC_left,
 			m_samples_auxC_right,
-			m_samples_auxC_surround
+			m_samples_auxC_surround,
+			m_samples_wm0,
+			m_samples_aux0,
+			m_samples_wm1,
+			m_samples_aux1,
+			m_samples_wm2,
+			m_samples_aux2,
+			m_samples_wm3,
+			m_samples_aux3
 		}};
 
 		if (!ReadPB(pb_addr, pb))
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 578330ba34..bcf69ddb4b 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -32,11 +32,9 @@
 
 #ifdef AX_GC
 # define PB_TYPE AXPB
-# define MS_PER_FRAME 1
 # define SAMPLES_PER_FRAME 32
 #else
 # define PB_TYPE AXPBWii
-# define MS_PER_FRAME 3
 # define SAMPLES_PER_FRAME 96
 #endif
 
@@ -69,13 +67,22 @@ union AXBuffers
 		int* auxC_left;
 		int* auxC_right;
 		int* auxC_surround;
+
+		int* wm_main0;
+		int* wm_aux0;
+		int* wm_main1;
+		int* wm_aux1;
+		int* wm_main2;
+		int* wm_aux2;
+		int* wm_main3;
+		int* wm_aux3;
 #endif
 	};
 
 #ifdef AX_GC
 	int* ptrs[9];
 #else
-	int* ptrs[12];
+	int* ptrs[20];
 #endif
 };
 
@@ -309,7 +316,7 @@ void GetInputSamples(PB_TYPE& pb, s16* samples)
 }
 
 // Add samples to an output buffer, with optional volume ramping.
-void MixAdd(int* out, const s16* input, u16* pvol, s16* dpop, bool ramp)
+void MixAdd(int* out, const s16* input, u32 count, u16* pvol, s16* dpop, bool ramp)
 {
 	u16& volume = pvol[0];
 	u16 volume_delta = pvol[1];
@@ -320,7 +327,7 @@ void MixAdd(int* out, const s16* input, u16* pvol, s16* dpop, bool ramp)
 	if (!ramp)
 		volume_delta = 0;
 
-	for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
+	for (u32 i = 0; i < count; ++i)
 	{
 		s64 sample = input[i];
 		sample *= volume;
@@ -363,33 +370,33 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
 	// TODO: Handle DPL2 on AUXB.
 
 	if (mctrl & MIX_L)
-		MixAdd(buffers.left, samples, &pb.mixer.left, &pb.dpop.left, mctrl & MIX_L_RAMP);
+		MixAdd(buffers.left, samples, SAMPLES_PER_FRAME, &pb.mixer.left, &pb.dpop.left, mctrl & MIX_L_RAMP);
 	if (mctrl & MIX_R)
-		MixAdd(buffers.right, samples, &pb.mixer.right, &pb.dpop.right, mctrl & MIX_R_RAMP);
+		MixAdd(buffers.right, samples, SAMPLES_PER_FRAME, &pb.mixer.right, &pb.dpop.right, mctrl & MIX_R_RAMP);
 	if (mctrl & MIX_S)
-		MixAdd(buffers.surround, samples, &pb.mixer.surround, &pb.dpop.surround, mctrl & MIX_S_RAMP);
+		MixAdd(buffers.surround, samples, SAMPLES_PER_FRAME, &pb.mixer.surround, &pb.dpop.surround, mctrl & MIX_S_RAMP);
 
 	if (mctrl & MIX_AUXA_L)
-		MixAdd(buffers.auxA_left, samples, &pb.mixer.auxA_left, &pb.dpop.auxA_left, mctrl & MIX_AUXA_L_RAMP);
+		MixAdd(buffers.auxA_left, samples, SAMPLES_PER_FRAME, &pb.mixer.auxA_left, &pb.dpop.auxA_left, mctrl & MIX_AUXA_L_RAMP);
 	if (mctrl & MIX_AUXA_R)
-		MixAdd(buffers.auxA_right, samples, &pb.mixer.auxA_right, &pb.dpop.auxA_right, mctrl & MIX_AUXA_R_RAMP);
+		MixAdd(buffers.auxA_right, samples, SAMPLES_PER_FRAME, &pb.mixer.auxA_right, &pb.dpop.auxA_right, mctrl & MIX_AUXA_R_RAMP);
 	if (mctrl & MIX_AUXA_S)
-		MixAdd(buffers.auxA_surround, samples, &pb.mixer.auxA_surround, &pb.dpop.auxA_surround, mctrl & MIX_AUXA_S_RAMP);
+		MixAdd(buffers.auxA_surround, samples, SAMPLES_PER_FRAME, &pb.mixer.auxA_surround, &pb.dpop.auxA_surround, mctrl & MIX_AUXA_S_RAMP);
 
 	if (mctrl & MIX_AUXB_L)
-		MixAdd(buffers.auxB_left, samples, &pb.mixer.auxB_left, &pb.dpop.auxB_left, mctrl & MIX_AUXB_L_RAMP);
+		MixAdd(buffers.auxB_left, samples, SAMPLES_PER_FRAME, &pb.mixer.auxB_left, &pb.dpop.auxB_left, mctrl & MIX_AUXB_L_RAMP);
 	if (mctrl & MIX_AUXB_R)
-		MixAdd(buffers.auxB_right, samples, &pb.mixer.auxB_right, &pb.dpop.auxB_right, mctrl & MIX_AUXB_R_RAMP);
+		MixAdd(buffers.auxB_right, samples, SAMPLES_PER_FRAME, &pb.mixer.auxB_right, &pb.dpop.auxB_right, mctrl & MIX_AUXB_R_RAMP);
 	if (mctrl & MIX_AUXB_S)
-		MixAdd(buffers.auxB_surround, samples, &pb.mixer.auxB_surround, &pb.dpop.auxB_surround, mctrl & MIX_AUXB_S_RAMP);
+		MixAdd(buffers.auxB_surround, samples, SAMPLES_PER_FRAME, &pb.mixer.auxB_surround, &pb.dpop.auxB_surround, mctrl & MIX_AUXB_S_RAMP);
 
 #ifdef AX_WII
 	if (mctrl & MIX_AUXC_L)
-		MixAdd(buffers.auxC_left, samples, &pb.mixer.auxC_left, &pb.dpop.auxC_left, mctrl & MIX_AUXC_L_RAMP);
+		MixAdd(buffers.auxC_left, samples, SAMPLES_PER_FRAME, &pb.mixer.auxC_left, &pb.dpop.auxC_left, mctrl & MIX_AUXC_L_RAMP);
 	if (mctrl & MIX_AUXC_R)
-		MixAdd(buffers.auxC_right, samples, &pb.mixer.auxC_right, &pb.dpop.auxC_right, mctrl & MIX_AUXC_R_RAMP);
+		MixAdd(buffers.auxC_right, samples, SAMPLES_PER_FRAME, &pb.mixer.auxC_right, &pb.dpop.auxC_right, mctrl & MIX_AUXC_R_RAMP);
 	if (mctrl & MIX_AUXC_S)
-		MixAdd(buffers.auxC_surround, samples, &pb.mixer.auxC_surround, &pb.dpop.auxC_surround, mctrl & MIX_AUXC_S_RAMP);
+		MixAdd(buffers.auxC_surround, samples, SAMPLES_PER_FRAME, &pb.mixer.auxC_surround, &pb.dpop.auxC_surround, mctrl & MIX_AUXC_S_RAMP);
 #endif
 
 	// Optionally, phase shift left or right channel to simulate 3D sound.
@@ -397,6 +404,58 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
 	{
 		// TODO
 	}
+
+#ifdef AX_WII
+	// Wiimote mixing.
+	if (pb.remote)
+	{
+		// Interpolate 18 samples from the 96 samples we read before. The real
+		// DSP code does it using a polyphase interpolation, we just use a
+		// linear interpolation here.
+		s16 wm_samples[18];
+
+		s16 curr0 = pb.remote_src.last_samples[2];
+		s16 curr1 = pb.remote_src.last_samples[3];
+
+		u32 ratio = 0x55555; // about 96/18 = 5.33333
+		u32 curr_pos = pb.remote_src.cur_addr_frac;
+		for (u32 i = 0; i < 18; ++i)
+		{
+			s32 curr_frac_pos = curr_pos & 0xFFFF;
+			s16 sample = curr0 + (s16)(((curr1 - curr0) * (s32)curr_frac_pos) >> 16);
+			wm_samples[i] = sample;
+
+			curr_pos += ratio;
+			curr0 = curr1;
+			curr1 = samples[curr_pos >> 16];
+		}
+		pb.remote_src.last_samples[2] = curr0;
+		pb.remote_src.last_samples[3] = curr1;
+		pb.src.cur_addr_frac = curr_pos & 0xFFFF;
+
+		// Mix to main[0-3] and aux[0-3]
+#define WMCHAN_MIX_ON(n) ((pb.remote_mixer_control >> (2 * n)) & 3)
+#define WMCHAN_MIX_RAMP(n) ((pb.remote_mixer_control >> (2 * n)) & 2)
+
+		if (WMCHAN_MIX_ON(0))
+			MixAdd(buffers.wm_main0, wm_samples, 18, &pb.remote_mixer.main0, &pb.remote_dpop.main0, WMCHAN_MIX_RAMP(0));
+		if (WMCHAN_MIX_ON(1))
+			MixAdd(buffers.wm_aux0, wm_samples, 18, &pb.remote_mixer.aux0, &pb.remote_dpop.aux0, WMCHAN_MIX_RAMP(1));
+		if (WMCHAN_MIX_ON(2))
+			MixAdd(buffers.wm_main1, wm_samples, 18, &pb.remote_mixer.main1, &pb.remote_dpop.main1, WMCHAN_MIX_RAMP(2));
+		if (WMCHAN_MIX_ON(3))
+			MixAdd(buffers.wm_aux1, wm_samples, 18, &pb.remote_mixer.aux1, &pb.remote_dpop.aux1, WMCHAN_MIX_RAMP(3));
+		if (WMCHAN_MIX_ON(4))
+			MixAdd(buffers.wm_main2, wm_samples, 18, &pb.remote_mixer.main2, &pb.remote_dpop.main2, WMCHAN_MIX_RAMP(4));
+		if (WMCHAN_MIX_ON(5))
+			MixAdd(buffers.wm_aux2, wm_samples, 18, &pb.remote_mixer.aux2, &pb.remote_dpop.aux2, WMCHAN_MIX_RAMP(5));
+		if (WMCHAN_MIX_ON(6))
+			MixAdd(buffers.wm_main3, wm_samples, 18, &pb.remote_mixer.main3, &pb.remote_dpop.main3, WMCHAN_MIX_RAMP(6));
+		if (WMCHAN_MIX_ON(7))
+			MixAdd(buffers.wm_aux3, wm_samples, 18, &pb.remote_mixer.aux3, &pb.remote_dpop.aux3, WMCHAN_MIX_RAMP(7));
+	}
+
+#endif
 }
 
 } // namespace

From c6629015048bc9d4aa18d297c0c41d3e26293843 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 16:15:42 +0100
Subject: [PATCH 05/28] Implement low-pass voice filtering, untested because I
 couldn't find any game using it

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h     |  6 +++---
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h | 13 ++++++++++---
 2 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
index d82088b34f..cc24bd78b7 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
@@ -230,9 +230,9 @@ struct PBADPCMLoopInfo
 struct PBLowPassFilter
 {
 	u16 enabled;
-	u16 yn1;
-	u16 a0;
-	u16 b0;
+	s16 yn1;
+	s16 a0;
+	s16 b0;
 };
 
 struct AXPB
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index bcf69ddb4b..fcc1bb7999 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -340,6 +340,15 @@ void MixAdd(int* out, const s16* input, u32 count, u16* pvol, s16* dpop, bool ra
 	}
 }
 
+// Execute a low pass filter on the samples using one history value. Returns
+// the new history value.
+s16 LowPassFilter(s16* samples, u32 count, s16 yn1, s16 a0, s16 b0)
+{
+	for (u32 i = 0; i < count; ++i)
+		yn1 = samples[i] = (s32)a0 * samples[i] - (s32)b0 * yn1;
+	return yn1;
+}
+
 // Process 1ms of audio (for AX GC) or 3ms of audio (for AX Wii) from a PB and
 // mix it to the output buffers.
 void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
@@ -362,9 +371,7 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
 
 	// Optionally, execute a low pass filter
 	if (pb.lpf.enabled)
-	{
-		// TODO
-	}
+		pb.lpf.yn1 = LowPassFilter(samples, SAMPLES_PER_FRAME, pb.lpf.yn1, pb.lpf.a0, pb.lpf.b0);
 
 	// Mix LRS, AUXA and AUXB depending on mixer_control
 	// TODO: Handle DPL2 on AUXB.

From 7252becf24583e6bf34f89889944611ccf343bb4 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 17:20:30 +0100
Subject: [PATCH 06/28] Fix the voice LPF implementation

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h | 4 ++--
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h  | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
index cc24bd78b7..2ac9ddd2f0 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
@@ -231,8 +231,8 @@ struct PBLowPassFilter
 {
 	u16 enabled;
 	s16 yn1;
-	s16 a0;
-	s16 b0;
+	u16 a0;
+	u16 b0;
 };
 
 struct AXPB
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index fcc1bb7999..f41c4c47ff 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -342,10 +342,10 @@ void MixAdd(int* out, const s16* input, u32 count, u16* pvol, s16* dpop, bool ra
 
 // Execute a low pass filter on the samples using one history value. Returns
 // the new history value.
-s16 LowPassFilter(s16* samples, u32 count, s16 yn1, s16 a0, s16 b0)
+s16 LowPassFilter(s16* samples, u32 count, s16 yn1, u16 a0, u16 b0)
 {
 	for (u32 i = 0; i < count; ++i)
-		yn1 = samples[i] = (s32)a0 * samples[i] - (s32)b0 * yn1;
+		yn1 = samples[i] = (a0 * (s32)samples[i] + b0 * (s32)yn1) >> 15;
 	return yn1;
 }
 

From 4ef021e975293f7530c1f2115408aebca0a721e2 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 20 Jan 2013 17:36:57 +0100
Subject: [PATCH 07/28] Better global volume adjustment on a voice. Fixes one
 of the 2 issues with the HOME button menu sound.

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index f41c4c47ff..f731fc5360 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -364,8 +364,7 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
 	// Apply a global volume ramp using the volume envelope parameters.
 	for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
 	{
-		s64 sample = 2 * (s16)samples[i] * (s16)pb.vol_env.cur_volume;
-		samples[i] = (s16)(sample >> 16);
+		samples[i] = ((s32)samples[i] * pb.vol_env.cur_volume) >> 15;
 		pb.vol_env.cur_volume += pb.vol_env.cur_volume_delta;
 	}
 

From 1612225e79fc17b748cb5736999d883d8300e541 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Mon, 21 Jan 2013 11:51:55 +0100
Subject: [PATCH 08/28] Implement AXWii command 01: ADD_TO_LR

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 19 ++++++++++++++++++-
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  3 ++-
 2 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 84fab9108b..2377839e86 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -67,7 +67,12 @@ void CUCode_AXWii::HandleCommandList()
 				SetupProcessing(HILO_TO_32(addr));
 				break;
 
-			case CMD_UNK_01: curr_idx += 2; break;
+			case CMD_ADD_TO_LR:
+				addr_hi = m_cmdlist[curr_idx++];
+				addr_lo = m_cmdlist[curr_idx++];
+				AddToLR(HILO_TO_32(addr));
+				break;
+
 			case CMD_UNK_02: curr_idx += 2; break;
 			case CMD_UNK_03: curr_idx += 2; break;
 
@@ -182,6 +187,18 @@ void CUCode_AXWii::SetupProcessing(u32 init_addr)
 	}
 }
 
+void CUCode_AXWii::AddToLR(u32 val_addr)
+{
+	int* ptr = (int*)HLEMemory_Get_Pointer(val_addr);
+	for (int i = 0; i < 32 * 3; ++i)
+	{
+		int val = (int)Common::swap32(*ptr++);
+
+		m_samples_left[i] += val;
+		m_samples_right[i] += val;
+	}
+}
+
 AXMixControl CUCode_AXWii::ConvertMixerControl(u32 mixer_control)
 {
 	u32 ret = 0;
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index 3a66c30868..6ace48abc1 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -51,6 +51,7 @@ protected:
 	virtual void HandleCommandList();
 
 	void SetupProcessing(u32 init_addr);
+	void AddToLR(u32 val_addr);
 	void ProcessPBList(u32 pb_addr);
 	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
 	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume);
@@ -60,7 +61,7 @@ private:
 	enum CmdType
 	{
 		CMD_SETUP = 0x00,
-		CMD_UNK_01 = 0x01,
+		CMD_ADD_TO_LR = 0x01,
 		CMD_UNK_02 = 0x02,
 		CMD_UNK_03 = 0x03,
 		CMD_PROCESS = 0x04,

From 9776f135e231d7da2cb31573c4cef65ae1674c33 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 25 Jan 2013 08:25:28 +0100
Subject: [PATCH 09/28] Add polyphase resampling support in AX

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp    | 27 +++++++++-
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h |  8 +++
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp |  3 +-
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 51 +++++++++++++++++--
 4 files changed, 82 insertions(+), 7 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
index cb34c2532d..3675970917 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
@@ -17,6 +17,7 @@
 
 #include "UCode_AX.h"
 #include "../../DSP.h"
+#include "FileUtil.h"
 
 #define AX_GC
 #include "UCode_AX_Voice.h"
@@ -29,6 +30,8 @@ CUCode_AX::CUCode_AX(DSPHLE* dsp_hle, u32 crc)
 	WARN_LOG(DSPHLE, "Instantiating CUCode_AX: crc=%08x", crc);
 	m_rMailHandler.PushMail(DSP_INIT);
 	DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
+
+	LoadResamplingCoefficients();
 }
 
 CUCode_AX::~CUCode_AX()
@@ -40,6 +43,27 @@ CUCode_AX::~CUCode_AX()
 	m_rMailHandler.Clear();
 }
 
+void CUCode_AX::LoadResamplingCoefficients()
+{
+	m_coeffs_available = false;
+
+	std::string filename = File::GetUserPath(D_GCUSER_IDX) + "dsp_coef.bin";
+	if (!File::Exists(filename))
+		return;
+
+	if (File::GetSize(filename) != 0x1000)
+		return;
+
+	FILE* fp = fopen(filename.c_str(), "rb");
+	fread(m_coeffs, 1, 0x1000, fp);
+	fclose(fp);
+
+	for (u32 i = 0; i < 0x800; ++i)
+		m_coeffs[i] = Common::swap16(m_coeffs[i]);
+
+	m_coeffs_available = true;
+}
+
 void CUCode_AX::SpawnAXThread(CUCode_AX* self)
 {
 	self->AXThread();
@@ -405,7 +429,8 @@ void CUCode_AX::ProcessPBList(u32 pb_addr)
 		{
 			ApplyUpdatesForMs(pb, curr_ms);
 
-			ProcessVoice(pb, buffers, ConvertMixerControl(pb.mixer_control));
+			ProcessVoice(pb, buffers, ConvertMixerControl(pb.mixer_control),
+			             m_coeffs_available ? m_coeffs : NULL);
 
 			// Forward the buffers
 			for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
index 7158c197e8..68248fef60 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
@@ -114,6 +114,14 @@ protected:
 	std::condition_variable m_cmdlist_cv;
 	std::mutex m_cmdlist_mutex;
 
+	// Table of coefficients for polyphase sample rate conversion.
+	// The coefficients aren't always available (they are part of the DSP DROM)
+	// so we also need to know if they are valid or not.
+	bool m_coeffs_available;
+	s16 m_coeffs[0x800];
+
+	void LoadResamplingCoefficients();
+
 	// Copy a command list from memory to our temp buffer
 	void CopyCmdList(u32 addr, u16 size);
 
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 2377839e86..1fde3f74b4 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -259,7 +259,8 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 		if (!ReadPB(pb_addr, pb))
 			break;
 
-		ProcessVoice(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)));
+		ProcessVoice(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)),
+		             m_coeffs_available ? m_coeffs : NULL);
 
 		WritePB(pb_addr, pb);
 		pb_addr = HILO_TO_32(pb.next_pb);
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index f731fc5360..649ad3c37f 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -245,13 +245,54 @@ u16 AcceleratorGetSample()
 
 // Read SAMPLES_PER_FRAME input samples from ARAM, decoding and converting rate
 // if required.
-void GetInputSamples(PB_TYPE& pb, s16* samples)
+void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
 {
 	u32 cur_addr = HILO_TO_32(pb.audio_addr.cur_addr);
 	AcceleratorSetup(&pb, &cur_addr);
 
-	// TODO: support polyphase interpolation if coefficients are available.
-	if (pb.src_type == SRCTYPE_POLYPHASE || pb.src_type == SRCTYPE_LINEAR)
+	// If DSP DROM coefficients are available, support polyphase resampling.
+	if (coeffs && pb.src_type == SRCTYPE_POLYPHASE)
+	{
+		s16 temp[4];
+		u32 idx = 0;
+
+		u32 ratio = HILO_TO_32(pb.src.ratio);
+		u32 curr_pos = pb.src.cur_addr_frac;
+
+		temp[idx++ & 3] = pb.src.last_samples[0];
+		temp[idx++ & 3] = pb.src.last_samples[1];
+		temp[idx++ & 3] = pb.src.last_samples[2];
+		temp[idx++ & 3] = pb.src.last_samples[3];
+
+		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
+		{
+			curr_pos += ratio;
+			while (curr_pos >= 0x10000)
+			{
+				temp[idx++ & 3] = AcceleratorGetSample();
+				curr_pos -= 0x10000;
+			}
+
+			u16 curr_pos_frac = ((curr_pos & 0xFFFF) >> 9) << 2;
+			const s16* c = &coeffs[pb.coef_select * 0x200 + curr_pos_frac];
+
+			s64 t0 = temp[idx++ & 3];
+			s64 t1 = temp[idx++ & 3];
+			s64 t2 = temp[idx++ & 3];
+			s64 t3 = temp[idx++ & 3];
+
+			s64 samp = (t0 * c[0] + t1 * c[1] + t2 * c[2] + t3 * c[3]) >> 15;
+
+			samples[i] = (s16)samp;
+		}
+
+		pb.src.last_samples[3] = temp[--idx & 3];
+		pb.src.last_samples[2] = temp[--idx & 3];
+		pb.src.last_samples[1] = temp[--idx & 3];
+		pb.src.last_samples[0] = temp[--idx & 3];
+		pb.src.cur_addr_frac = curr_pos & 0xFFFF;
+	}
+	else if (pb.src_type == SRCTYPE_LINEAR || (!coeffs && pb.src_type == SRCTYPE_POLYPHASE))
 	{
 		// Convert the input to a higher or lower sample rate using a linear
 		// interpolation algorithm. The input to output ratio is set in
@@ -351,7 +392,7 @@ s16 LowPassFilter(s16* samples, u32 count, s16 yn1, u16 a0, u16 b0)
 
 // Process 1ms of audio (for AX GC) or 3ms of audio (for AX Wii) from a PB and
 // mix it to the output buffers.
-void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
+void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, const s16* coeffs)
 {
 	// If the voice is not running, nothing to do.
 	if (!pb.running)
@@ -359,7 +400,7 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl)
 
 	// Read input samples, performing sample rate conversion if needed.
 	s16 samples[SAMPLES_PER_FRAME];
-	GetInputSamples(pb, samples);
+	GetInputSamples(pb, samples, coeffs);
 
 	// Apply a global volume ramp using the volume envelope parameters.
 	for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)

From bad4f7f790b6038450488d1c23d8b65329cea4b4 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 25 Jan 2013 19:16:07 +0100
Subject: [PATCH 10/28] Rewrite the linear interpolation SRC to give the exact
 same results as the one in AXWii

---
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 64 +++++++++++++------
 1 file changed, 43 insertions(+), 21 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 649ad3c37f..92c797ad14 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -307,38 +307,60 @@ void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
 		// samples.
 		u32 curr_pos = pb.src.cur_addr_frac;
 
-		// These are the two samples between which we interpolate. The initial
-		// values are stored in the PB, and we update them when resampling the
-		// input data.
-		s16 curr0 = pb.src.last_samples[2];
-		s16 curr1 = pb.src.last_samples[3];
+		// This is the circular buffer containing samples to use for the
+		// interpolation. It is initialized with the values from the PB, and it
+		// will be stored back to the PB at the end.
+		s16 temp[4];
+		u32 idx = 0;
+
+		temp[idx++ & 3] = pb.src.last_samples[0];
+		temp[idx++ & 3] = pb.src.last_samples[1];
+		temp[idx++ & 3] = pb.src.last_samples[2];
+		temp[idx++ & 3] = pb.src.last_samples[3];
 
 		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
 		{
-			// Get our current fractional position, used to know how much of
-			// curr0 and how much of curr1 the output sample should be.
-			s32 curr_frac_pos = curr_pos & 0xFFFF;
-
-			// Linear interpolation: s1 + (s2 - s1) * pos
-			s16 sample = curr0 + (s16)(((curr1 - curr0) * (s32)curr_frac_pos) >> 16);
-			samples[i] = sample;
-
 			curr_pos += ratio;
 
-			// While our current position is >= 1.0, shift to the next 2
-			// samples for interpolation.
-			while ((curr_pos >> 16) != 0)
+			// While our current position is >= 1.0, push new samples to the
+			// circular buffer.
+			while (curr_pos >= 0x10000)
 			{
-				curr0 = curr1;
-				curr1 = AcceleratorGetSample();
+				temp[idx++ & 3] = AcceleratorGetSample();
 				curr_pos -= 0x10000;
 			}
+
+			// Get our current fractional position, used to know how much of
+			// curr0 and how much of curr1 the output sample should be.
+			u16 curr_frac = curr_pos & 0xFFFF;
+			u16 inv_curr_frac = -curr_frac;
+
+			// Interpolate! If curr_frac is 0, we can simply take the last
+			// sample without any multiplying.
+			s16 sample;
+			if (curr_frac)
+			{
+				s32 s0 = temp[idx++ & 3];
+				s32 s1 = temp[idx++ & 3];
+
+				sample = ((s0 * inv_curr_frac) + (s1 * curr_frac)) >> 16;
+				idx += 2;
+			}
+			else
+			{
+				sample = temp[idx++ & 3];
+				idx += 3;
+			}
+
+			samples[i] = sample;
 		}
 
-		// Update the two last_samples values in the PB as well as the current
+		// Update the four last_samples values in the PB as well as the current
 		// position.
-		pb.src.last_samples[2] = curr0;
-		pb.src.last_samples[3] = curr1;
+		pb.src.last_samples[3] = temp[--idx & 3];
+		pb.src.last_samples[2] = temp[--idx & 3];
+		pb.src.last_samples[1] = temp[--idx & 3];
+		pb.src.last_samples[0] = temp[--idx & 3];
 		pb.src.cur_addr_frac = curr_pos & 0xFFFF;
 	}
 	else // SRCTYPE_NEAREST

From 85b498ba97e829e561fe5b4cb48b30348e83d323 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 29 Mar 2013 13:18:30 +0100
Subject: [PATCH 11/28] Update the right cur_addr_frac after wiimote audio
 resampling

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 0f3dcccb88..51c1c1e728 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -500,7 +500,7 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, con
 		}
 		pb.remote_src.last_samples[2] = curr0;
 		pb.remote_src.last_samples[3] = curr1;
-		pb.src.cur_addr_frac = curr_pos & 0xFFFF;
+		pb.remote_src.cur_addr_frac = curr_pos & 0xFFFF;
 
 		// Mix to main[0-3] and aux[0-3]
 #define WMCHAN_MIX_ON(n) ((pb.remote_mixer_control >> (2 * n)) & 3)
@@ -523,7 +523,6 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, con
 		if (WMCHAN_MIX_ON(7))
 			MixAdd(buffers.wm_aux3, wm_samples, 18, &pb.remote_mixer.aux3, &pb.remote_dpop.aux3, WMCHAN_MIX_RAMP(7));
 	}
-
 #endif
 }
 

From 4dc1ffbb207b5085cb490bec2b1da0f76d2832eb Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 29 Mar 2013 13:49:36 +0100
Subject: [PATCH 12/28] Refactor the resampling code to avoid having two
 polyphase resampling implementations (normal/wm)

---
 .../Src/HW/DSPHLE/UCodes/UCode_AXStructs.h    |   4 +-
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 151 +++++++++---------
 2 files changed, 79 insertions(+), 76 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
index 2ac9ddd2f0..5fc4806e26 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
@@ -211,13 +211,13 @@ struct PBSampleRateConverter
 	u16 ratio_hi; // integer part of sampling ratio
 	u16 ratio_lo; // fraction part of sampling ratio
 	u16 cur_addr_frac;
-	u16 last_samples[4];
+	s16 last_samples[4];
 };
 
 struct PBSampleRateConverterWM
 {
 	u16 cur_addr_frac;
-	u16 last_samples[4];
+	s16 last_samples[4];
 };
 
 struct PBADPCMLoopInfo
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 51c1c1e728..7bf060b0d4 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -243,38 +243,52 @@ u16 AcceleratorGetSample()
 	return ret;
 }
 
-// Read SAMPLES_PER_FRAME input samples from ARAM, decoding and converting rate
-// if required.
-void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
+// Reads samples from the input callback, resamples them to <count> samples at
+// the wanted sample rate (computed from the ratio, see below).
+//
+// If srctype is SRCTYPE_POLYPHASE, coefficients need to be provided as well
+// (or the srctype will automatically be changed to LINEAR).
+//
+// Returns the current position after resampling (including fractional part).
+//
+// The input to output ratio is set in <ratio>, which is a floating point num
+// stored as a 32b integer:
+//  * Upper 16 bits of the ratio are the integer part
+//  * Lower 16 bits are the decimal part
+//
+// <curr_pos> is a 32b integer structured in the same way as the ratio: the
+// upper 16 bits are the integer part of the current position in the input
+// stream, and the lower 16 bits are the decimal part.
+//
+// We start getting samples not from sample 0, but 0.<curr_pos_frac>. This
+// avoids discontinuties in the audio stream, especially with very low ratios
+// which interpolate a lot of values between two "real" samples.
+u32 ResampleAudio(std::function<s16(u32)> input_callback, s16* output, u32 count,
+                  s16* last_samples, u32 curr_pos, u32 ratio, int srctype,
+                  const s16* coeffs)
 {
-	u32 cur_addr = HILO_TO_32(pb.audio_addr.cur_addr);
-	AcceleratorSetup(&pb, &cur_addr);
-
 	// If DSP DROM coefficients are available, support polyphase resampling.
-	if (coeffs && pb.src_type == SRCTYPE_POLYPHASE)
+	if (coeffs && srctype == SRCTYPE_POLYPHASE)
 	{
 		s16 temp[4];
 		u32 idx = 0;
 
-		u32 ratio = HILO_TO_32(pb.src.ratio);
-		u32 curr_pos = pb.src.cur_addr_frac;
+		temp[idx++ & 3] = last_samples[0];
+		temp[idx++ & 3] = last_samples[1];
+		temp[idx++ & 3] = last_samples[2];
+		temp[idx++ & 3] = last_samples[3];
 
-		temp[idx++ & 3] = pb.src.last_samples[0];
-		temp[idx++ & 3] = pb.src.last_samples[1];
-		temp[idx++ & 3] = pb.src.last_samples[2];
-		temp[idx++ & 3] = pb.src.last_samples[3];
-
-		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
+		for (u32 i = 0; i < count; ++i)
 		{
 			curr_pos += ratio;
 			while (curr_pos >= 0x10000)
 			{
-				temp[idx++ & 3] = AcceleratorGetSample();
+				temp[idx++ & 3] = input_callback(curr_pos >> 16);
 				curr_pos -= 0x10000;
 			}
 
 			u16 curr_pos_frac = ((curr_pos & 0xFFFF) >> 9) << 2;
-			const s16* c = &coeffs[pb.coef_select * 0x200 + curr_pos_frac];
+			const s16* c = &coeffs[curr_pos_frac];
 
 			s64 t0 = temp[idx++ & 3];
 			s64 t1 = temp[idx++ & 3];
@@ -283,42 +297,28 @@ void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
 
 			s64 samp = (t0 * c[0] + t1 * c[1] + t2 * c[2] + t3 * c[3]) >> 15;
 
-			samples[i] = (s16)samp;
+			output[i] = (s16)samp;
 		}
 
-		pb.src.last_samples[3] = temp[--idx & 3];
-		pb.src.last_samples[2] = temp[--idx & 3];
-		pb.src.last_samples[1] = temp[--idx & 3];
-		pb.src.last_samples[0] = temp[--idx & 3];
-		pb.src.cur_addr_frac = curr_pos & 0xFFFF;
+		last_samples[3] = temp[--idx & 3];
+		last_samples[2] = temp[--idx & 3];
+		last_samples[1] = temp[--idx & 3];
+		last_samples[0] = temp[--idx & 3];
 	}
-	else if (pb.src_type == SRCTYPE_LINEAR || (!coeffs && pb.src_type == SRCTYPE_POLYPHASE))
+	else if (srctype == SRCTYPE_LINEAR || (!coeffs && srctype == SRCTYPE_POLYPHASE))
 	{
-		// Convert the input to a higher or lower sample rate using a linear
-		// interpolation algorithm. The input to output ratio is set in
-		// pb.src.ratio, which is a floating point num stored as a 32b integer:
-		//  * Upper 16 bits of the ratio are the integer part
-		//  * Lower 16 bits are the decimal part
-		u32 ratio = HILO_TO_32(pb.src.ratio);
-
-		// We start getting samples not from sample 0, but 0.<cur_addr_frac>.
-		// This avoids discontinuties in the audio stream, especially with very
-		// low ratios which interpolate a lot of values between two "real"
-		// samples.
-		u32 curr_pos = pb.src.cur_addr_frac;
-
 		// This is the circular buffer containing samples to use for the
 		// interpolation. It is initialized with the values from the PB, and it
 		// will be stored back to the PB at the end.
 		s16 temp[4];
 		u32 idx = 0;
 
-		temp[idx++ & 3] = pb.src.last_samples[0];
-		temp[idx++ & 3] = pb.src.last_samples[1];
-		temp[idx++ & 3] = pb.src.last_samples[2];
-		temp[idx++ & 3] = pb.src.last_samples[3];
+		temp[idx++ & 3] = last_samples[0];
+		temp[idx++ & 3] = last_samples[1];
+		temp[idx++ & 3] = last_samples[2];
+		temp[idx++ & 3] = last_samples[3];
 
-		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
+		for (u32 i = 0; i < count; ++i)
 		{
 			curr_pos += ratio;
 
@@ -326,7 +326,7 @@ void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
 			// circular buffer.
 			while (curr_pos >= 0x10000)
 			{
-				temp[idx++ & 3] = AcceleratorGetSample();
+				temp[idx++ & 3] = input_callback(curr_pos >> 16);
 				curr_pos -= 0x10000;
 			}
 
@@ -352,27 +352,43 @@ void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
 				idx += 3;
 			}
 
-			samples[i] = sample;
+			output[i] = sample;
 		}
 
-		// Update the four last_samples values in the PB as well as the current
-		// position.
-		pb.src.last_samples[3] = temp[--idx & 3];
-		pb.src.last_samples[2] = temp[--idx & 3];
-		pb.src.last_samples[1] = temp[--idx & 3];
-		pb.src.last_samples[0] = temp[--idx & 3];
-		pb.src.cur_addr_frac = curr_pos & 0xFFFF;
+		// Update the four last_samples values.
+		last_samples[3] = temp[--idx & 3];
+		last_samples[2] = temp[--idx & 3];
+		last_samples[1] = temp[--idx & 3];
+		last_samples[0] = temp[--idx & 3];
 	}
 	else // SRCTYPE_NEAREST
 	{
 		// No sample rate conversion here: simply read samples from the
 		// accelerator to the output buffer.
-		for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
-			samples[i] = AcceleratorGetSample();
+		for (u32 i = 0; i < count; ++i)
+			output[i] = input_callback(i);
 
-		memcpy(pb.src.last_samples, samples + SAMPLES_PER_FRAME - 4, 4 * sizeof (u16));
+		memcpy(last_samples, output + count - 4, 4 * sizeof (u16));
 	}
 
+	return curr_pos;
+}
+
+// Read SAMPLES_PER_FRAME input samples from ARAM, decoding and converting rate
+// if required.
+void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
+{
+	u32 cur_addr = HILO_TO_32(pb.audio_addr.cur_addr);
+	AcceleratorSetup(&pb, &cur_addr);
+
+	if (coeffs)
+		coeffs += pb.coef_select * 0x200;
+	u32 curr_pos = ResampleAudio([](u32) { return AcceleratorGetSample(); },
+	                             samples, SAMPLES_PER_FRAME, pb.src.last_samples,
+	                             pb.src.cur_addr_frac, HILO_TO_32(pb.src.ratio),
+	                             pb.src_type, coeffs);
+	pb.src.cur_addr_frac = (curr_pos & 0xFFFF);
+
 	// Update current position in the PB.
 	pb.audio_addr.cur_addr_hi = (u16)(cur_addr >> 16);
 	pb.audio_addr.cur_addr_lo = (u16)(cur_addr & 0xFFFF);
@@ -478,28 +494,15 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, con
 	// Wiimote mixing.
 	if (pb.remote)
 	{
-		// Interpolate 18 samples from the 96 samples we read before. The real
-		// DSP code does it using a polyphase interpolation, we just use a
-		// linear interpolation here.
+		// Interpolate 18 samples from the 96 samples we read before.
 		s16 wm_samples[18];
 
-		s16 curr0 = pb.remote_src.last_samples[2];
-		s16 curr1 = pb.remote_src.last_samples[3];
-
-		u32 ratio = 0x55555; // about 96/18 = 5.33333
-		u32 curr_pos = pb.remote_src.cur_addr_frac;
-		for (u32 i = 0; i < 18; ++i)
-		{
-			s32 curr_frac_pos = curr_pos & 0xFFFF;
-			s16 sample = curr0 + (s16)(((curr1 - curr0) * (s32)curr_frac_pos) >> 16);
-			wm_samples[i] = sample;
-
-			curr_pos += ratio;
-			curr0 = curr1;
-			curr1 = samples[curr_pos >> 16];
-		}
-		pb.remote_src.last_samples[2] = curr0;
-		pb.remote_src.last_samples[3] = curr1;
+		// We use ratio 0x55555 == (5 * 65536 + 21845) / 65536 == 5.3333 which
+		// is the nearest we can get to 96/18
+		u32 curr_pos = ResampleAudio([&samples](u32 i) { return samples[i]; },
+		                             wm_samples, 18, pb.remote_src.last_samples,
+		                             pb.remote_src.cur_addr_frac, 0x55555,
+		                             SRCTYPE_POLYPHASE, coeffs);
 		pb.remote_src.cur_addr_frac = curr_pos & 0xFFFF;
 
 		// Mix to main[0-3] and aux[0-3]

From a997824f680aaf9cc86adfd56e8bea865e5614dd Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 29 Mar 2013 13:51:52 +0100
Subject: [PATCH 13/28] Add missing <functional> header include

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 7bf060b0d4..64a01ca8fe 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -30,6 +30,8 @@
 #include "UCode_AXStructs.h"
 #include "../../DSP.h"
 
+#include <functional>
+
 #ifdef AX_GC
 # define PB_TYPE AXPB
 # define SAMPLES_PER_FRAME 32

From e9b236be051bf0b4c4e06e5d6add0d946ed86554 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 29 Mar 2013 13:55:55 +0100
Subject: [PATCH 14/28] OSX sucks at c++11

---
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h   | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 64a01ca8fe..0b5bf4a022 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -30,7 +30,14 @@
 #include "UCode_AXStructs.h"
 #include "../../DSP.h"
 
-#include <functional>
+// I hate OSX.
+#if defined(__APPLE__)
+# include <tr1/functional>
+using std::tr1::function;
+#else
+# include <functional>
+using std::function;
+#endif
 
 #ifdef AX_GC
 # define PB_TYPE AXPB
@@ -265,7 +272,7 @@ u16 AcceleratorGetSample()
 // We start getting samples not from sample 0, but 0.<curr_pos_frac>. This
 // avoids discontinuties in the audio stream, especially with very low ratios
 // which interpolate a lot of values between two "real" samples.
-u32 ResampleAudio(std::function<s16(u32)> input_callback, s16* output, u32 count,
+u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
                   s16* last_samples, u32 curr_pos, u32 ratio, int srctype,
                   const s16* coeffs)
 {

From ef501137be4a019a18719e1ce802ec42ca68ab03 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 29 Mar 2013 21:02:27 +0100
Subject: [PATCH 15/28] Fix audio glitching at the end of a voice because of
 bad non-looping sound handling in AXWii

---
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 75 +++++++++++--------
 1 file changed, 45 insertions(+), 30 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 0b5bf4a022..87c0d51f46 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -147,6 +147,7 @@ void DumpPB(const PB_TYPE& pb)
 static u32 acc_loop_addr, acc_end_addr;
 static u32* acc_cur_addr;
 static PB_TYPE* acc_pb;
+static bool acc_end_reached;
 
 // Sets up the simulated accelerator.
 void AcceleratorSetup(PB_TYPE* pb, u32* cur_addr)
@@ -155,6 +156,7 @@ void AcceleratorSetup(PB_TYPE* pb, u32* cur_addr)
 	acc_loop_addr = HILO_TO_32(pb->audio_addr.loop_addr);
 	acc_end_addr = HILO_TO_32(pb->audio_addr.end_addr);
 	acc_cur_addr = cur_addr;
+	acc_end_reached = false;
 }
 
 // Reads a sample from the simulated accelerator. Also handles looping and
@@ -164,6 +166,49 @@ u16 AcceleratorGetSample()
 {
 	u16 ret;
 
+	// Have we reached the end address?
+	//
+	// On real hardware, this would raise an interrupt that is handled by the
+	// UCode. We simulate what this interrupt does here.
+	if ((*acc_cur_addr & ~1) == (acc_end_addr & ~1))
+	{
+		// loop back to loop_addr.
+		*acc_cur_addr = acc_loop_addr;
+
+		if (acc_pb->audio_addr.looping)
+		{
+			// Set the ADPCM infos to continue processing at loop_addr.
+			//
+			// For some reason, yn1 and yn2 aren't set if the voice is not of
+			// stream type. This is what the AX UCode does and I don't really
+			// know why.
+			acc_pb->adpcm.pred_scale = acc_pb->adpcm_loop_info.pred_scale;
+			if (!acc_pb->is_stream)
+			{
+				acc_pb->adpcm.yn1 = acc_pb->adpcm_loop_info.yn1;
+				acc_pb->adpcm.yn2 = acc_pb->adpcm_loop_info.yn2;
+			}
+		}
+		else
+		{
+			// Non looping voice reached the end -> running = 0.
+			acc_pb->running = 0;
+
+#ifdef AX_WII
+			// One of the few meaningful differences between AXGC and AXWii:
+			// while AXGC handles non looping voices ending by having 0000
+			// samples at the loop address, AXWii has the 0000 samples
+			// internally in DRAM and use an internal pointer to it (loop addr
+			// does not contain 0000 samples on AXWii!).
+			acc_end_reached = true;
+#endif
+		}
+	}
+
+	// See above for explanations about acc_end_reached.
+	if (acc_end_reached)
+		return 0;
+
 	switch (acc_pb->audio_addr.sample_format)
 	{
 		case 0x00:	// ADPCM
@@ -219,36 +264,6 @@ u16 AcceleratorGetSample()
 			return 0;
 	}
 
-	// Have we reached the end address?
-	//
-	// On real hardware, this would raise an interrupt that is handled by the
-	// UCode. We simulate what this interrupt does here.
-	if ((*acc_cur_addr & ~1) == (acc_end_addr & ~1))
-	{
-		// loop back to loop_addr.
-		*acc_cur_addr = acc_loop_addr;
-
-		if (acc_pb->audio_addr.looping)
-		{
-			// Set the ADPCM infos to continue processing at loop_addr.
-			//
-			// For some reason, yn1 and yn2 aren't set if the voice is not of
-			// stream type. This is what the AX UCode does and I don't really
-			// know why.
-			acc_pb->adpcm.pred_scale = acc_pb->adpcm_loop_info.pred_scale;
-			if (!acc_pb->is_stream)
-			{
-				acc_pb->adpcm.yn1 = acc_pb->adpcm_loop_info.yn1;
-				acc_pb->adpcm.yn2 = acc_pb->adpcm_loop_info.yn2;
-			}
-		}
-		else
-		{
-			// Non looping voice reached the end -> running = 0.
-			acc_pb->running = 0;
-		}
-	}
-
 	return ret;
 }
 

From 4b09f525f6e63a1e3686d84f12463ba127eeda36 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Fri, 29 Mar 2013 22:22:24 +0100
Subject: [PATCH 16/28] Fix AUX volume mixing in AXWii: implement volume
 ramping and MixAdd properly. Home menu sounds now work properly.

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 28 +++++++++++++++++--
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  4 +++
 2 files changed, 29 insertions(+), 3 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 1fde3f74b4..8381564f3f 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -29,7 +29,8 @@
 
 
 CUCode_AXWii::CUCode_AXWii(DSPHLE *dsp_hle, u32 l_CRC)
-	: CUCode_AX(dsp_hle, l_CRC)
+	: CUCode_AX(dsp_hle, l_CRC),
+	  m_last_aux_volume(0x8000)
 {
 	WARN_LOG(DSPHLE, "Instantiating CUCode_AXWii");
 }
@@ -269,6 +270,22 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 
 void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume)
 {
+	u16 prev_volume = m_last_aux_volume;
+
+	// Generate a volume ramp going from prev_volume to volume.
+	//
+	// The AXWii UCode uses integer arithmetic with 2 multipliers because it
+	// can't get enough precision to represent 1 / 96. We can use floating
+	// point arithmetic, so we will - it makes the code more readable and more
+	// precise.
+	u16 volume_ramp[96];
+	float curr_volume = prev_volume;
+	for (int i = 0; i < 96; ++i)
+	{
+		volume_ramp[i] = (u16)curr_volume;
+		curr_volume += (volume - prev_volume) / 96.0;
+	}
+
 	int* buffers[3] = { 0 };
 	int* main_buffers[3] = {
 		m_samples_left,
@@ -311,9 +328,12 @@ void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16
 	for (u32 i = 0; i < 3; ++i)
 		for (u32 j = 0; j < 3 * 32; ++j)
 		{
-			s64 new_val = main_buffers[i][j] + Common::swap32(*ptr++);
-			main_buffers[i][j] = (new_val * volume) >> 15;
+			s64 sample = (s64)(s32)Common::swap32(*ptr++);
+			sample *= volume_ramp[j];
+			main_buffers[i][j] += (s32)(sample >> 15);
 		}
+
+	m_last_aux_volume = volume;
 }
 
 void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
@@ -397,4 +417,6 @@ void CUCode_AXWii::DoState(PointerWrap &p)
 	p.Do(m_samples_aux1);
 	p.Do(m_samples_aux2);
 	p.Do(m_samples_aux3);
+
+	p.Do(m_last_aux_volume);
 }
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index 6ace48abc1..031971a2a4 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -29,6 +29,7 @@ public:
 	virtual void DoState(PointerWrap &p);
 
 protected:
+	// Additional AUX buffers
 	int m_samples_auxC_left[32 * 3];
 	int m_samples_auxC_right[32 * 3];
 	int m_samples_auxC_surround[32 * 3];
@@ -43,6 +44,9 @@ protected:
 	int m_samples_wm3[6 * 3];
 	int m_samples_aux3[6 * 3];
 
+	// Last MixAUXSamples volume value. Used to generate volume ramps.
+	u16 m_last_aux_volume;
+
 	// Convert a mixer_control bitfield to our internal representation for that
 	// value. Required because that bitfield has a different meaning in some
 	// versions of AX.

From c271082ec50bda616233fbf883ad0c58cda8fcd0 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sat, 30 Mar 2013 00:55:55 +0100
Subject: [PATCH 17/28] Add volume ramping for MAIN output, separate old volume
 values for each AUX channel and refactor

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 43 ++++++++++---------
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   | 11 ++++-
 2 files changed, 32 insertions(+), 22 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 8381564f3f..df9a6b5409 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -30,8 +30,10 @@
 
 CUCode_AXWii::CUCode_AXWii(DSPHLE *dsp_hle, u32 l_CRC)
 	: CUCode_AX(dsp_hle, l_CRC),
-	  m_last_aux_volume(0x8000)
+	  m_last_main_volume(0x8000)
 {
+	for (int i = 0; i < 3; ++i)
+		m_last_aux_volumes[i] = 0x8000;
 	WARN_LOG(DSPHLE, "Instantiating CUCode_AXWii");
 }
 
@@ -228,6 +230,16 @@ AXMixControl CUCode_AXWii::ConvertMixerControl(u32 mixer_control)
 	return (AXMixControl)ret;
 }
 
+void CUCode_AXWii::GenerateVolumeRamp(u16* output, u16 vol1, u16 vol2, size_t nvals)
+{
+	float curr = vol1;
+	for (size_t i = 0; i < nvals; ++i)
+	{
+		curr += (vol2 - vol1) / (float)nvals;
+		output[i] = curr;
+	}
+}
+
 void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 {
 	AXPBWii pb;
@@ -270,21 +282,9 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 
 void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume)
 {
-	u16 prev_volume = m_last_aux_volume;
-
-	// Generate a volume ramp going from prev_volume to volume.
-	//
-	// The AXWii UCode uses integer arithmetic with 2 multipliers because it
-	// can't get enough precision to represent 1 / 96. We can use floating
-	// point arithmetic, so we will - it makes the code more readable and more
-	// precise.
 	u16 volume_ramp[96];
-	float curr_volume = prev_volume;
-	for (int i = 0; i < 96; ++i)
-	{
-		volume_ramp[i] = (u16)curr_volume;
-		curr_volume += (volume - prev_volume) / 96.0;
-	}
+	GenerateVolumeRamp(volume_ramp, m_last_aux_volumes[aux_id], volume, 96);
+	m_last_aux_volumes[aux_id] = volume;
 
 	int* buffers[3] = { 0 };
 	int* main_buffers[3] = {
@@ -332,12 +332,14 @@ void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16
 			sample *= volume_ramp[j];
 			main_buffers[i][j] += (s32)(sample >> 15);
 		}
-
-	m_last_aux_volume = volume;
 }
 
 void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
 {
+	u16 volume_ramp[96];
+	GenerateVolumeRamp(volume_ramp, m_last_main_volume, volume, 96);
+	m_last_main_volume = volume;
+
 	int surround_buffer[3 * 32] = { 0 };
 
 	for (u32 i = 0; i < 3 * 32; ++i)
@@ -353,8 +355,8 @@ void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
 		int right = m_samples_right[i];
 
 		// Apply global volume. Cast to s64 to avoid overflow.
-		left = ((s64)left * volume) >> 15;
-		right = ((s64)right * volume) >> 15;
+		left = ((s64)left * volume_ramp[i]) >> 15;
+		right = ((s64)right * volume_ramp[i]) >> 15;
 
 		if (left < -32767)  left = -32767;
 		if (left > 32767)   left = 32767;
@@ -418,5 +420,6 @@ void CUCode_AXWii::DoState(PointerWrap &p)
 	p.Do(m_samples_aux2);
 	p.Do(m_samples_aux3);
 
-	p.Do(m_last_aux_volume);
+	p.Do(m_last_main_volume);
+	p.Do(m_last_aux_volumes);
 }
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index 031971a2a4..c1d281d9c3 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -44,14 +44,21 @@ protected:
 	int m_samples_wm3[6 * 3];
 	int m_samples_aux3[6 * 3];
 
-	// Last MixAUXSamples volume value. Used to generate volume ramps.
-	u16 m_last_aux_volume;
+	// Last volume values for MAIN and AUX. Used to generate volume ramps to
+	// interpolate nicely between old and new volume values.
+	u16 m_last_main_volume;
+	u16 m_last_aux_volumes[3];
 
 	// Convert a mixer_control bitfield to our internal representation for that
 	// value. Required because that bitfield has a different meaning in some
 	// versions of AX.
 	AXMixControl ConvertMixerControl(u32 mixer_control);
 
+	// Generate a volume ramp from vol1 to vol2, interpolating n volume values.
+	// Uses floating point arithmetic, which isn't exactly what the UCode does,
+	// but this gives better precision and nicer code.
+	void GenerateVolumeRamp(u16* output, u16 vol1, u16 vol2, size_t nvals);
+
 	virtual void HandleCommandList();
 
 	void SetupProcessing(u32 init_addr);

From e3b0a2c9bf3a73016d641788687aad8b49e04d97 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sat, 30 Mar 2013 14:02:30 +0100
Subject: [PATCH 18/28] Add an option to run the AX processing on the CPU
 thread. Fixes timing issues causing audio glitches on Wii, and should improve
 the overall stability of AX HLE.

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp    | 43 +++++++++++++++----
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h | 11 ++++-
 Source/Core/DolphinWX/Src/ConfigMain.cpp      |  4 +-
 3 files changed, 47 insertions(+), 11 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
index 3675970917..ef7f2b61ca 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
@@ -18,27 +18,35 @@
 #include "UCode_AX.h"
 #include "../../DSP.h"
 #include "FileUtil.h"
+#include "ConfigManager.h"
 
 #define AX_GC
 #include "UCode_AX_Voice.h"
 
 CUCode_AX::CUCode_AX(DSPHLE* dsp_hle, u32 crc)
 	: IUCode(dsp_hle, crc)
+	, m_work_available(false)
 	, m_cmdlist_size(0)
-	, m_axthread(&SpawnAXThread, this)
+	, m_run_on_thread(SConfig::GetInstance().m_LocalCoreStartupParameter.bDSPThread)
 {
 	WARN_LOG(DSPHLE, "Instantiating CUCode_AX: crc=%08x", crc);
 	m_rMailHandler.PushMail(DSP_INIT);
 	DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
 
 	LoadResamplingCoefficients();
+
+	if (m_run_on_thread)
+		m_axthread = std::thread(SpawnAXThread, this);
 }
 
 CUCode_AX::~CUCode_AX()
 {
-	m_cmdlist_size = (u16)-1;	// Special value to signal end
-	NotifyAXThread();
-	m_axthread.join();
+	if (m_run_on_thread)
+	{
+		m_cmdlist_size = (u16)-1;	// Special value to signal end
+		NotifyAXThread();
+		m_axthread.join();
+	}
 
 	m_rMailHandler.Clear();
 }
@@ -85,20 +93,32 @@ void CUCode_AX::AXThread()
 		m_processing.lock();
 		HandleCommandList();
 		m_cmdlist_size = 0;
-
-		// Signal end of processing
-		m_rMailHandler.PushMail(DSP_YIELD);
-		DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
+		SignalWorkEnd();
 		m_processing.unlock();
 	}
 }
 
+void CUCode_AX::SignalWorkEnd()
+{
+	// Signal end of processing
+	m_rMailHandler.PushMail(DSP_YIELD);
+	DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
+}
+
 void CUCode_AX::NotifyAXThread()
 {
 	std::unique_lock<std::mutex> lk(m_cmdlist_mutex);
 	m_cmdlist_cv.notify_one();
 }
 
+void CUCode_AX::StartWorking()
+{
+	if (m_run_on_thread)
+		NotifyAXThread();
+	else
+		m_work_available = true;
+}
+
 void CUCode_AX::HandleCommandList()
 {
 	// Temp variables for addresses computation
@@ -620,6 +640,7 @@ void CUCode_AX::HandleMail(u32 mail)
 	if (next_is_cmdlist)
 	{
 		CopyCmdList(mail, cmdlist_size);
+		StartWorking();
 		NotifyAXThread();
 	}
 	else if (m_UploadSetupInProgress)
@@ -688,6 +709,12 @@ void CUCode_AX::Update(int cycles)
 		m_rMailHandler.PushMail(DSP_RESUME);
 		DSP::GenerateDSPInterruptFromDSPEmu(DSP::INT_DSP);
 	}
+	else if (m_work_available)
+	{
+		HandleCommandList();
+		m_cmdlist_size = 0;
+		SignalWorkEnd();
+	}
 }
 
 void CUCode_AX::DoAXState(PointerWrap& p)
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
index 68248fef60..fbecac1990 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
@@ -102,11 +102,15 @@ protected:
 	int m_samples_auxB_right[32 * 5];
 	int m_samples_auxB_surround[32 * 5];
 
+	// This flag is set if there is anything to process.
+	bool m_work_available;
+
 	// Volatile because it's set by HandleMail and accessed in
 	// HandleCommandList, which are running in two different threads.
 	volatile u16 m_cmdlist[512];
 	volatile u32 m_cmdlist_size;
 
+	bool m_run_on_thread;
 	std::thread m_axthread;
 
 	// Sync objects
@@ -130,13 +134,18 @@ protected:
 	// versions of AX.
 	AXMixControl ConvertMixerControl(u32 mixer_control);
 
-	// Send a notification to the AX thread to tell him a new cmdlist addr is
+	// Signal that we should start handling a command list. Dispatches to the
+	// AX thread if using a thread, else just sets a boolean flag.
+	void StartWorking();
+
+	// Send a notification to the AX thread to tell it a new cmdlist addr is
 	// available for processing.
 	void NotifyAXThread();
 
 	void AXThread();
 
 	virtual void HandleCommandList();
+	void SignalWorkEnd();
 
 	void SetupProcessing(u32 init_addr);
 	void DownloadAndMixWithVolume(u32 addr, u16 vol_main, u16 vol_auxa, u16 vol_auxb);
diff --git a/Source/Core/DolphinWX/Src/ConfigMain.cpp b/Source/Core/DolphinWX/Src/ConfigMain.cpp
index cbd780e6aa..c55b75f044 100644
--- a/Source/Core/DolphinWX/Src/ConfigMain.cpp
+++ b/Source/Core/DolphinWX/Src/ConfigMain.cpp
@@ -515,7 +515,7 @@ void CConfigMain::InitializeGUITooltips()
 	InterfaceLang->SetToolTip(_("Change the language of the user interface.\nRequires restart."));
 
 	// Audio tooltips
-	DSPThread->SetToolTip(_("Run DSP LLE on a dedicated thread (not recommended)."));
+	DSPThread->SetToolTip(_("Run DSP HLE and LLE on a dedicated thread (not recommended: might cause audio glitches with HLE and freezes with LLE)."));
 	BackendSelection->SetToolTip(_("Changing this will have no effect while the emulator is running!"));
 
 	// Gamecube - Devices
@@ -657,7 +657,7 @@ void CConfigMain::CreateGUIControls()
 	// Audio page
 	DSPEngine = new wxRadioBox(AudioPage, ID_DSPENGINE, _("DSP Emulator Engine"),
 				wxDefaultPosition, wxDefaultSize, arrayStringFor_DSPEngine, 0, wxRA_SPECIFY_ROWS);
-	DSPThread = new wxCheckBox(AudioPage, ID_DSPTHREAD, _("DSP LLE on Thread"));
+	DSPThread = new wxCheckBox(AudioPage, ID_DSPTHREAD, _("DSP on Dedicated Thread"));
 	DumpAudio = new wxCheckBox(AudioPage, ID_DUMP_AUDIO, _("Dump Audio"),
 				wxDefaultPosition, wxDefaultSize, 0, wxDefaultValidator);
 	DPL2Decoder = new wxCheckBox(AudioPage, ID_DPL2DECODER, _("Dolby Pro Logic II decoder"));

From a813f9e13c697bc866ad7f0402dfbfeb3e4c0fee Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sat, 30 Mar 2013 14:38:14 +0100
Subject: [PATCH 19/28] Support loading polyphase resampling coeffs from User
 and Sys in HLE

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp    | 25 ++++++++++++++++---
 1 file changed, 21 insertions(+), 4 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
index ef7f2b61ca..b21e2f66f4 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
@@ -55,12 +55,29 @@ void CUCode_AX::LoadResamplingCoefficients()
 {
 	m_coeffs_available = false;
 
-	std::string filename = File::GetUserPath(D_GCUSER_IDX) + "dsp_coef.bin";
-	if (!File::Exists(filename))
+	std::string filenames[] = {
+		File::GetUserPath(D_GCUSER_IDX) + "dsp_coef.bin",
+		File::GetSysDirectory() + "/GC/dsp_coef.bin"
+	};
+
+	size_t fidx;
+	std::string filename;
+	for (fidx = 0; fidx < sizeof (filenames) / sizeof (filenames[0]); ++fidx)
+	{
+		filename = filenames[fidx];
+		if (!File::Exists(filename))
+			continue;
+
+		if (File::GetSize(filename) != 0x1000)
+			continue;
+
+		break;
+	}
+
+	if (fidx >= sizeof (filenames) / sizeof (filenames[0]))
 		return;
 
-	if (File::GetSize(filename) != 0x1000)
-		return;
+	WARN_LOG(DSPHLE, "Loading polyphase resampling coeffs from %s", filename.c_str());
 
 	FILE* fp = fopen(filename.c_str(), "rb");
 	fread(m_coeffs, 1, 0x1000, fp);

From 04f9c6793bea3b5bd7a57e21074e69ceca79e31d Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sat, 30 Mar 2013 14:39:59 +0100
Subject: [PATCH 20/28] Fix the argument to the samples reading callback in
 voice processing - should fix issues with wiimote audio, untested

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index 87c0d51f46..f36f5760e7 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -291,6 +291,8 @@ u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
                   s16* last_samples, u32 curr_pos, u32 ratio, int srctype,
                   const s16* coeffs)
 {
+	int read_samples_count = 0;
+
 	// If DSP DROM coefficients are available, support polyphase resampling.
 	if (coeffs && srctype == SRCTYPE_POLYPHASE)
 	{
@@ -307,7 +309,7 @@ u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
 			curr_pos += ratio;
 			while (curr_pos >= 0x10000)
 			{
-				temp[idx++ & 3] = input_callback(curr_pos >> 16);
+				temp[idx++ & 3] = input_callback(read_samples_count++);
 				curr_pos -= 0x10000;
 			}
 
@@ -350,7 +352,7 @@ u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
 			// circular buffer.
 			while (curr_pos >= 0x10000)
 			{
-				temp[idx++ & 3] = input_callback(curr_pos >> 16);
+				temp[idx++ & 3] = input_callback(read_samples_count++);
 				curr_pos -= 0x10000;
 			}
 

From 79c031624351c7e0d93879cc92c00a3c5e1c328e Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sat, 30 Mar 2013 16:59:06 +0100
Subject: [PATCH 21/28] Disable the polyphase resampler - it causes audio
 glitches with non integer ratios

---
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index f36f5760e7..a93f69fb2c 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -293,8 +293,11 @@ u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
 {
 	int read_samples_count = 0;
 
+	// TODO(delroth): find out why the polyphase resampling algorithm causes
+	// audio glitches in Wii games with non integral ratios.
+
 	// If DSP DROM coefficients are available, support polyphase resampling.
-	if (coeffs && srctype == SRCTYPE_POLYPHASE)
+	if (0) // if (coeffs && srctype == SRCTYPE_POLYPHASE)
 	{
 		s16 temp[4];
 		u32 idx = 0;
@@ -331,7 +334,7 @@ u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
 		last_samples[1] = temp[--idx & 3];
 		last_samples[0] = temp[--idx & 3];
 	}
-	else if (srctype == SRCTYPE_LINEAR || (!coeffs && srctype == SRCTYPE_POLYPHASE))
+	else if (srctype == SRCTYPE_LINEAR || srctype == SRCTYPE_POLYPHASE)
 	{
 		// This is the circular buffer containing samples to use for the
 		// interpolation. It is initialized with the values from the PB, and it

From 276c457bed3e9737e769cc1224467f57642f4661 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sat, 30 Mar 2013 22:22:57 +0100
Subject: [PATCH 22/28] Basic framework to support the old AXWii version used
 in Wii Sports and Excite Truck

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp    | 20 +++---
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h |  3 +
 .../Src/HW/DSPHLE/UCodes/UCode_AXStructs.h    | 46 -------------
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 65 +++++++++++++++++-
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  6 ++
 .../Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h     | 67 ++++++++++---------
 6 files changed, 118 insertions(+), 89 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
index b21e2f66f4..9c7d57f813 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.cpp
@@ -307,19 +307,18 @@ void CUCode_AX::HandleCommandList()
 	}
 }
 
-static void ApplyUpdatesForMs(AXPB& pb, int curr_ms)
+void CUCode_AX::ApplyUpdatesForMs(int curr_ms, u16* pb, u16* num_updates, u16* updates)
 {
 	u32 start_idx = 0;
 	for (int i = 0; i < curr_ms; ++i)
-		start_idx += pb.updates.num_updates[i];
+		start_idx += num_updates[i];
 
-	u32 update_addr = HILO_TO_32(pb.updates.data);
-	for (u32 i = start_idx; i < start_idx + pb.updates.num_updates[curr_ms]; ++i)
+	for (u32 i = start_idx; i < start_idx + num_updates[curr_ms]; ++i)
 	{
-		u16 update_off = HLEMemory_Read_U16(update_addr + 4 * i);
-		u16 update_val = HLEMemory_Read_U16(update_addr + 4 * i + 2);
+		u16 update_off = Common::swap16(updates[2 * i]);
+		u16 update_val = Common::swap16(updates[2 * i + 1]);
 
-		((u16*)&pb)[update_off] = update_val;
+		pb[update_off] = update_val;
 	}
 }
 
@@ -462,11 +461,14 @@ void CUCode_AX::ProcessPBList(u32 pb_addr)
 		if (!ReadPB(pb_addr, pb))
 			break;
 
+		u32 updates_addr = HILO_TO_32(pb.updates.data);
+		u16* updates = (u16*)HLEMemory_Get_Pointer(updates_addr);
+
 		for (int curr_ms = 0; curr_ms < 5; ++curr_ms)
 		{
-			ApplyUpdatesForMs(pb, curr_ms);
+			ApplyUpdatesForMs(curr_ms, (u16*)&pb, pb.updates.num_updates, updates);
 
-			ProcessVoice(pb, buffers, ConvertMixerControl(pb.mixer_control),
+			ProcessVoice(pb, buffers, spms, ConvertMixerControl(pb.mixer_control),
 			             m_coeffs_available ? m_coeffs : NULL);
 
 			// Forward the buffers
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
index fbecac1990..71f195bf33 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX.h
@@ -134,6 +134,9 @@ protected:
 	// versions of AX.
 	AXMixControl ConvertMixerControl(u32 mixer_control);
 
+	// Apply updates to a PB. Generic, used in AX GC and AX Wii.
+	void ApplyUpdatesForMs(int curr_ms, u16* pb, u16* num_updates, u16* updates);
+
 	// Signal that we should start handling a command list. Dispatches to the
 	// AX thread if using a thread, else just sets a boolean flag.
 	void StartWorking();
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
index 5fc4806e26..1161847bc8 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXStructs.h
@@ -324,52 +324,6 @@ struct AXPBWii
 	u16 pad[12]; // align us, captain! (32B)
 };
 
-// Seems like nintendo used an early version of AXWii and forgot to remove the update functionality ;p
-struct PBUpdatesWiiSports
-{
-	u16 num_updates[3];
-	u16 data_hi;
-	u16 data_lo;
-};
-
-struct AXPBWiiSports
-{
-	u16 next_pb_hi;
-	u16 next_pb_lo;
-	u16 this_pb_hi;
-	u16 this_pb_lo;
-
-	u16 src_type;		// Type of sample rate converter (none, 4-tap, linear)
-	u16 coef_select;	// coef for the 4-tap src
-	u32 mixer_control;
-
-	u16 running;		// 1=RUN   0=STOP
-	u16 is_stream;		// 1 = stream, 0 = one shot
-
-	PBMixerWii mixer;
-	PBInitialTimeDelay initial_time_delay;
-	PBUpdatesWiiSports updates;
-	PBDpopWii dpop;
-	PBVolumeEnvelope vol_env;
-	PBAudioAddr audio_addr;
-	PBADPCMInfo adpcm;
-	PBSampleRateConverter src;
-	PBADPCMLoopInfo adpcm_loop_info;
-	PBLowPassFilter lpf;
-	PBBiquadFilter biquad;
-
-	// WIIMOTE :D
-	u16 remote;
-	u16 remote_mixer_control;
-
-	PBMixerWM remote_mixer;
-	PBDpopWM remote_dpop;
-	PBSampleRateConverterWM remote_src;
-	PBInfImpulseResponseWM remote_iir;
-
-	u16 pad[7]; // align us, captain! (32B)
-};
-
 // TODO: All these enums have changed a lot for wii
 enum {
     AUDIOFORMAT_ADPCM = 0,
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index df9a6b5409..64d27ffbd1 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -35,6 +35,8 @@ CUCode_AXWii::CUCode_AXWii(DSPHLE *dsp_hle, u32 l_CRC)
 	for (int i = 0; i < 3; ++i)
 		m_last_aux_volumes[i] = 0x8000;
 	WARN_LOG(DSPHLE, "Instantiating CUCode_AXWii");
+
+	m_old_axwii = (l_CRC == 0xfa450138);
 }
 
 CUCode_AXWii::~CUCode_AXWii()
@@ -100,6 +102,8 @@ void CUCode_AXWii::HandleCommandList()
 			case CMD_UNK_08: curr_idx += 13; break;
 			case CMD_UNK_09: curr_idx += 13; break;
 
+			// TODO(delroth): figure this one out, it's used by almost every
+			// game I've tested so far.
 			case CMD_UNK_0A: curr_idx += 4; break;
 
 			case CMD_OUTPUT:
@@ -240,6 +244,43 @@ void CUCode_AXWii::GenerateVolumeRamp(u16* output, u16 vol1, u16 vol2, size_t nv
 	}
 }
 
+bool CUCode_AXWii::ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* updates)
+{
+	u16* pb_mem = (u16*)&pb;
+
+	if (!m_old_axwii)
+		return false;
+
+	// Copy the num_updates field.
+	memcpy(num_updates, pb_mem + 41, 6);
+
+	// Get the address of the updates data
+	u16 addr_hi = pb_mem[44];
+	u16 addr_lo = pb_mem[45];
+	u32 addr = HILO_TO_32(addr);
+	u16* ptr = (u16*)HLEMemory_Get_Pointer(addr);
+
+	// Copy the updates data and change the offset to match a PB without
+	// updates data.
+	u32 updates_count = num_updates[0] + num_updates[1] + num_updates[2];
+	for (u32 i = 0; i < updates_count; ++i)
+	{
+		u16 update_off = Common::swap16(ptr[2 * i]);
+		u16 update_val = Common::swap16(ptr[2 * i + 1]);
+
+		if (update_off > 45)
+			update_off -= 5;
+
+		updates[2 * i] = update_off;
+		updates[2 * i + 1] = update_val;
+	}
+
+	// Remove the updates data from the PB
+	memmove(pb_mem + 41, pb_mem + 45, sizeof (pb) - 2 * 45);
+
+	return true;
+}
+
 void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 {
 	AXPBWii pb;
@@ -272,8 +313,28 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 		if (!ReadPB(pb_addr, pb))
 			break;
 
-		ProcessVoice(pb, buffers, ConvertMixerControl(HILO_TO_32(pb.mixer_control)),
-		             m_coeffs_available ? m_coeffs : NULL);
+		u16 num_updates[3];
+		u16 updates[1024];
+		if (ExtractUpdatesFields(pb, num_updates, updates))
+		{
+			for (int curr_ms = 0; curr_ms < 3; ++curr_ms)
+			{
+				ApplyUpdatesForMs(curr_ms, (u16*)&pb, num_updates, updates);
+				ProcessVoice(pb, buffers, 32,
+				             ConvertMixerControl(HILO_TO_32(pb.mixer_control)),
+				             m_coeffs_available ? m_coeffs : NULL);
+
+				// Forward the buffers
+				for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
+					buffers.ptrs[i] += 32;
+			}
+		}
+		else
+		{
+			ProcessVoice(pb, buffers, 96,
+			             ConvertMixerControl(HILO_TO_32(pb.mixer_control)),
+			             m_coeffs_available ? m_coeffs : NULL);
+		}
 
 		WritePB(pb_addr, pb);
 		pb_addr = HILO_TO_32(pb.next_pb);
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index c1d281d9c3..ce09b14674 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -44,11 +44,17 @@ protected:
 	int m_samples_wm3[6 * 3];
 	int m_samples_aux3[6 * 3];
 
+	// Are we implementing an old version of AXWii which still has updates?
+	bool m_old_axwii;
+
 	// Last volume values for MAIN and AUX. Used to generate volume ramps to
 	// interpolate nicely between old and new volume values.
 	u16 m_last_main_volume;
 	u16 m_last_aux_volumes[3];
 
+	// If needed, extract the updates related fields from a PB.
+	bool ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* updates);
+
 	// Convert a mixer_control bitfield to our internal representation for that
 	// value. Required because that bitfield has a different meaning in some
 	// versions of AX.
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
index a93f69fb2c..e364e42888 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AX_Voice.h
@@ -41,10 +41,10 @@ using std::function;
 
 #ifdef AX_GC
 # define PB_TYPE AXPB
-# define SAMPLES_PER_FRAME 32
+# define MAX_SAMPLES_PER_FRAME 32
 #else
 # define PB_TYPE AXPBWii
-# define SAMPLES_PER_FRAME 96
+# define MAX_SAMPLES_PER_FRAME 96
 #endif
 
 // Put all of that in an anonymous namespace to avoid stupid compilers merging
@@ -403,9 +403,9 @@ u32 ResampleAudio(function<s16(u32)> input_callback, s16* output, u32 count,
 	return curr_pos;
 }
 
-// Read SAMPLES_PER_FRAME input samples from ARAM, decoding and converting rate
+// Read <count> input samples from ARAM, decoding and converting rate
 // if required.
-void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
+void GetInputSamples(PB_TYPE& pb, s16* samples, u16 count, const s16* coeffs)
 {
 	u32 cur_addr = HILO_TO_32(pb.audio_addr.cur_addr);
 	AcceleratorSetup(&pb, &cur_addr);
@@ -413,7 +413,7 @@ void GetInputSamples(PB_TYPE& pb, s16* samples, const s16* coeffs)
 	if (coeffs)
 		coeffs += pb.coef_select * 0x200;
 	u32 curr_pos = ResampleAudio([](u32) { return AcceleratorGetSample(); },
-	                             samples, SAMPLES_PER_FRAME, pb.src.last_samples,
+	                             samples, count, pb.src.last_samples,
 	                             pb.src.cur_addr_frac, HILO_TO_32(pb.src.ratio),
 	                             pb.src_type, coeffs);
 	pb.src.cur_addr_frac = (curr_pos & 0xFFFF);
@@ -459,18 +459,18 @@ s16 LowPassFilter(s16* samples, u32 count, s16 yn1, u16 a0, u16 b0)
 
 // Process 1ms of audio (for AX GC) or 3ms of audio (for AX Wii) from a PB and
 // mix it to the output buffers.
-void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, const s16* coeffs)
+void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, u16 count, AXMixControl mctrl, const s16* coeffs)
 {
 	// If the voice is not running, nothing to do.
 	if (!pb.running)
 		return;
 
 	// Read input samples, performing sample rate conversion if needed.
-	s16 samples[SAMPLES_PER_FRAME];
-	GetInputSamples(pb, samples, coeffs);
+	s16 samples[MAX_SAMPLES_PER_FRAME];
+	GetInputSamples(pb, samples, count, coeffs);
 
 	// Apply a global volume ramp using the volume envelope parameters.
-	for (u32 i = 0; i < SAMPLES_PER_FRAME; ++i)
+	for (u32 i = 0; i < count; ++i)
 	{
 		samples[i] = ((s32)samples[i] * pb.vol_env.cur_volume) >> 15;
 		pb.vol_env.cur_volume += pb.vol_env.cur_volume_delta;
@@ -478,39 +478,39 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, con
 
 	// Optionally, execute a low pass filter
 	if (pb.lpf.enabled)
-		pb.lpf.yn1 = LowPassFilter(samples, SAMPLES_PER_FRAME, pb.lpf.yn1, pb.lpf.a0, pb.lpf.b0);
+		pb.lpf.yn1 = LowPassFilter(samples, count, pb.lpf.yn1, pb.lpf.a0, pb.lpf.b0);
 
 	// Mix LRS, AUXA and AUXB depending on mixer_control
 	// TODO: Handle DPL2 on AUXB.
 
 	if (mctrl & MIX_L)
-		MixAdd(buffers.left, samples, SAMPLES_PER_FRAME, &pb.mixer.left, &pb.dpop.left, mctrl & MIX_L_RAMP);
+		MixAdd(buffers.left, samples, count, &pb.mixer.left, &pb.dpop.left, mctrl & MIX_L_RAMP);
 	if (mctrl & MIX_R)
-		MixAdd(buffers.right, samples, SAMPLES_PER_FRAME, &pb.mixer.right, &pb.dpop.right, mctrl & MIX_R_RAMP);
+		MixAdd(buffers.right, samples, count, &pb.mixer.right, &pb.dpop.right, mctrl & MIX_R_RAMP);
 	if (mctrl & MIX_S)
-		MixAdd(buffers.surround, samples, SAMPLES_PER_FRAME, &pb.mixer.surround, &pb.dpop.surround, mctrl & MIX_S_RAMP);
+		MixAdd(buffers.surround, samples, count, &pb.mixer.surround, &pb.dpop.surround, mctrl & MIX_S_RAMP);
 
 	if (mctrl & MIX_AUXA_L)
-		MixAdd(buffers.auxA_left, samples, SAMPLES_PER_FRAME, &pb.mixer.auxA_left, &pb.dpop.auxA_left, mctrl & MIX_AUXA_L_RAMP);
+		MixAdd(buffers.auxA_left, samples, count, &pb.mixer.auxA_left, &pb.dpop.auxA_left, mctrl & MIX_AUXA_L_RAMP);
 	if (mctrl & MIX_AUXA_R)
-		MixAdd(buffers.auxA_right, samples, SAMPLES_PER_FRAME, &pb.mixer.auxA_right, &pb.dpop.auxA_right, mctrl & MIX_AUXA_R_RAMP);
+		MixAdd(buffers.auxA_right, samples, count, &pb.mixer.auxA_right, &pb.dpop.auxA_right, mctrl & MIX_AUXA_R_RAMP);
 	if (mctrl & MIX_AUXA_S)
-		MixAdd(buffers.auxA_surround, samples, SAMPLES_PER_FRAME, &pb.mixer.auxA_surround, &pb.dpop.auxA_surround, mctrl & MIX_AUXA_S_RAMP);
+		MixAdd(buffers.auxA_surround, samples, count, &pb.mixer.auxA_surround, &pb.dpop.auxA_surround, mctrl & MIX_AUXA_S_RAMP);
 
 	if (mctrl & MIX_AUXB_L)
-		MixAdd(buffers.auxB_left, samples, SAMPLES_PER_FRAME, &pb.mixer.auxB_left, &pb.dpop.auxB_left, mctrl & MIX_AUXB_L_RAMP);
+		MixAdd(buffers.auxB_left, samples, count, &pb.mixer.auxB_left, &pb.dpop.auxB_left, mctrl & MIX_AUXB_L_RAMP);
 	if (mctrl & MIX_AUXB_R)
-		MixAdd(buffers.auxB_right, samples, SAMPLES_PER_FRAME, &pb.mixer.auxB_right, &pb.dpop.auxB_right, mctrl & MIX_AUXB_R_RAMP);
+		MixAdd(buffers.auxB_right, samples, count, &pb.mixer.auxB_right, &pb.dpop.auxB_right, mctrl & MIX_AUXB_R_RAMP);
 	if (mctrl & MIX_AUXB_S)
-		MixAdd(buffers.auxB_surround, samples, SAMPLES_PER_FRAME, &pb.mixer.auxB_surround, &pb.dpop.auxB_surround, mctrl & MIX_AUXB_S_RAMP);
+		MixAdd(buffers.auxB_surround, samples, count, &pb.mixer.auxB_surround, &pb.dpop.auxB_surround, mctrl & MIX_AUXB_S_RAMP);
 
 #ifdef AX_WII
 	if (mctrl & MIX_AUXC_L)
-		MixAdd(buffers.auxC_left, samples, SAMPLES_PER_FRAME, &pb.mixer.auxC_left, &pb.dpop.auxC_left, mctrl & MIX_AUXC_L_RAMP);
+		MixAdd(buffers.auxC_left, samples, count, &pb.mixer.auxC_left, &pb.dpop.auxC_left, mctrl & MIX_AUXC_L_RAMP);
 	if (mctrl & MIX_AUXC_R)
-		MixAdd(buffers.auxC_right, samples, SAMPLES_PER_FRAME, &pb.mixer.auxC_right, &pb.dpop.auxC_right, mctrl & MIX_AUXC_R_RAMP);
+		MixAdd(buffers.auxC_right, samples, count, &pb.mixer.auxC_right, &pb.dpop.auxC_right, mctrl & MIX_AUXC_R_RAMP);
 	if (mctrl & MIX_AUXC_S)
-		MixAdd(buffers.auxC_surround, samples, SAMPLES_PER_FRAME, &pb.mixer.auxC_surround, &pb.dpop.auxC_surround, mctrl & MIX_AUXC_S_RAMP);
+		MixAdd(buffers.auxC_surround, samples, count, &pb.mixer.auxC_surround, &pb.dpop.auxC_surround, mctrl & MIX_AUXC_S_RAMP);
 #endif
 
 	// Optionally, phase shift left or right channel to simulate 3D sound.
@@ -523,13 +523,16 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, con
 	// Wiimote mixing.
 	if (pb.remote)
 	{
-		// Interpolate 18 samples from the 96 samples we read before.
+		// Old AXWii versions process ms per ms.
+		u16 wm_count = count == 96 ? 18 : 6;
+
+		// Interpolate at most 18 samples from the 96 samples we read before.
 		s16 wm_samples[18];
 
 		// We use ratio 0x55555 == (5 * 65536 + 21845) / 65536 == 5.3333 which
 		// is the nearest we can get to 96/18
 		u32 curr_pos = ResampleAudio([&samples](u32 i) { return samples[i]; },
-		                             wm_samples, 18, pb.remote_src.last_samples,
+		                             wm_samples, wm_count, pb.remote_src.last_samples,
 		                             pb.remote_src.cur_addr_frac, 0x55555,
 		                             SRCTYPE_POLYPHASE, coeffs);
 		pb.remote_src.cur_addr_frac = curr_pos & 0xFFFF;
@@ -539,21 +542,21 @@ void ProcessVoice(PB_TYPE& pb, const AXBuffers& buffers, AXMixControl mctrl, con
 #define WMCHAN_MIX_RAMP(n) ((pb.remote_mixer_control >> (2 * n)) & 2)
 
 		if (WMCHAN_MIX_ON(0))
-			MixAdd(buffers.wm_main0, wm_samples, 18, &pb.remote_mixer.main0, &pb.remote_dpop.main0, WMCHAN_MIX_RAMP(0));
+			MixAdd(buffers.wm_main0, wm_samples, wm_count, &pb.remote_mixer.main0, &pb.remote_dpop.main0, WMCHAN_MIX_RAMP(0));
 		if (WMCHAN_MIX_ON(1))
-			MixAdd(buffers.wm_aux0, wm_samples, 18, &pb.remote_mixer.aux0, &pb.remote_dpop.aux0, WMCHAN_MIX_RAMP(1));
+			MixAdd(buffers.wm_aux0, wm_samples, wm_count, &pb.remote_mixer.aux0, &pb.remote_dpop.aux0, WMCHAN_MIX_RAMP(1));
 		if (WMCHAN_MIX_ON(2))
-			MixAdd(buffers.wm_main1, wm_samples, 18, &pb.remote_mixer.main1, &pb.remote_dpop.main1, WMCHAN_MIX_RAMP(2));
+			MixAdd(buffers.wm_main1, wm_samples, wm_count, &pb.remote_mixer.main1, &pb.remote_dpop.main1, WMCHAN_MIX_RAMP(2));
 		if (WMCHAN_MIX_ON(3))
-			MixAdd(buffers.wm_aux1, wm_samples, 18, &pb.remote_mixer.aux1, &pb.remote_dpop.aux1, WMCHAN_MIX_RAMP(3));
+			MixAdd(buffers.wm_aux1, wm_samples, wm_count, &pb.remote_mixer.aux1, &pb.remote_dpop.aux1, WMCHAN_MIX_RAMP(3));
 		if (WMCHAN_MIX_ON(4))
-			MixAdd(buffers.wm_main2, wm_samples, 18, &pb.remote_mixer.main2, &pb.remote_dpop.main2, WMCHAN_MIX_RAMP(4));
+			MixAdd(buffers.wm_main2, wm_samples, wm_count, &pb.remote_mixer.main2, &pb.remote_dpop.main2, WMCHAN_MIX_RAMP(4));
 		if (WMCHAN_MIX_ON(5))
-			MixAdd(buffers.wm_aux2, wm_samples, 18, &pb.remote_mixer.aux2, &pb.remote_dpop.aux2, WMCHAN_MIX_RAMP(5));
+			MixAdd(buffers.wm_aux2, wm_samples, wm_count, &pb.remote_mixer.aux2, &pb.remote_dpop.aux2, WMCHAN_MIX_RAMP(5));
 		if (WMCHAN_MIX_ON(6))
-			MixAdd(buffers.wm_main3, wm_samples, 18, &pb.remote_mixer.main3, &pb.remote_dpop.main3, WMCHAN_MIX_RAMP(6));
+			MixAdd(buffers.wm_main3, wm_samples, wm_count, &pb.remote_mixer.main3, &pb.remote_dpop.main3, WMCHAN_MIX_RAMP(6));
 		if (WMCHAN_MIX_ON(7))
-			MixAdd(buffers.wm_aux3, wm_samples, 18, &pb.remote_mixer.aux3, &pb.remote_dpop.aux3, WMCHAN_MIX_RAMP(7));
+			MixAdd(buffers.wm_aux3, wm_samples, wm_count, &pb.remote_mixer.aux3, &pb.remote_dpop.aux3, WMCHAN_MIX_RAMP(7));
 	}
 #endif
 }

From 60b43eb8d34d0267732b686f9e02094f68a8ec48 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Sun, 31 Mar 2013 00:25:00 +0100
Subject: [PATCH 23/28] Support the old AXWii version used in games like Wii
 Sports or Excite Truck

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 250 +++++++++++++-----
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  32 ++-
 2 files changed, 206 insertions(+), 76 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 64d27ffbd1..1e966d2d0a 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -50,8 +50,10 @@ void CUCode_AXWii::HandleCommandList()
 	u16 addr2_hi, addr2_lo;
 	u16 volume;
 
+	u32 pb_addr = 0;
+
 //	WARN_LOG(DSPHLE, "Command list:");
-//	for (u32 i = 0; m_cmdlist[i] != CMD_END; ++i)
+//	for (u32 i = 0; m_cmdlist[i] != CMD_END_OLD; ++i)
 //		WARN_LOG(DSPHLE, "%04x", m_cmdlist[i]);
 //	WARN_LOG(DSPHLE, "-------------");
 
@@ -61,78 +63,160 @@ void CUCode_AXWii::HandleCommandList()
 	{
 		u16 cmd = m_cmdlist[curr_idx++];
 
-		switch (cmd)
+		if (m_old_axwii)
 		{
-			// Some of these commands are unknown, or unused in this AX HLE.
-			// We still need to skip their arguments using "curr_idx += N".
-
-			case CMD_SETUP:
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				SetupProcessing(HILO_TO_32(addr));
-				break;
-
-			case CMD_ADD_TO_LR:
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				AddToLR(HILO_TO_32(addr));
-				break;
-
-			case CMD_UNK_02: curr_idx += 2; break;
-			case CMD_UNK_03: curr_idx += 2; break;
-
-			case CMD_PROCESS:
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				ProcessPBList(HILO_TO_32(addr));
-				break;
-
-			case CMD_MIX_AUXA:
-			case CMD_MIX_AUXB:
-			case CMD_MIX_AUXC:
-				volume = m_cmdlist[curr_idx++];
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				addr2_hi = m_cmdlist[curr_idx++];
-				addr2_lo = m_cmdlist[curr_idx++];
-				MixAUXSamples(cmd - CMD_MIX_AUXA, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
-				break;
-
-			// These two go together and manipulate some AUX buffers.
-			case CMD_UNK_08: curr_idx += 13; break;
-			case CMD_UNK_09: curr_idx += 13; break;
-
-			// TODO(delroth): figure this one out, it's used by almost every
-			// game I've tested so far.
-			case CMD_UNK_0A: curr_idx += 4; break;
-
-			case CMD_OUTPUT:
-				volume = m_cmdlist[curr_idx++];
-				addr_hi = m_cmdlist[curr_idx++];
-				addr_lo = m_cmdlist[curr_idx++];
-				addr2_hi = m_cmdlist[curr_idx++];
-				addr2_lo = m_cmdlist[curr_idx++];
-				OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), volume);
-				break;
-
-			case CMD_UNK_0C: curr_idx += 5; break;
-
-			case CMD_WM_OUTPUT:
+			switch (cmd)
 			{
-				u32 addresses[4] = {
-					(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
-					(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
-					(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
-					(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
-				};
-				curr_idx += 8;
-				OutputWMSamples(addresses);
-				break;
-			}
+				// Some of these commands are unknown, or unused in this AX HLE.
+				// We still need to skip their arguments using "curr_idx += N".
 
-			case CMD_END:
-				end = true;
-				break;
+				case CMD_SETUP_OLD:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					SetupProcessing(HILO_TO_32(addr));
+					break;
+
+				case CMD_ADD_TO_LR_OLD:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					AddToLR(HILO_TO_32(addr));
+					break;
+
+				case CMD_UNK_02_OLD: curr_idx += 2; break;
+				case CMD_UNK_03_OLD: curr_idx += 2; break;
+
+				case CMD_PB_ADDR_OLD:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					pb_addr = HILO_TO_32(addr);
+					break;
+
+				case CMD_PROCESS_OLD:
+					ProcessPBList(pb_addr);
+					break;
+
+				case CMD_MIX_AUXA_OLD:
+				case CMD_MIX_AUXB_OLD:
+				case CMD_MIX_AUXC_OLD:
+					volume = m_cmdlist[curr_idx++];
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					addr2_hi = m_cmdlist[curr_idx++];
+					addr2_lo = m_cmdlist[curr_idx++];
+					MixAUXSamples(cmd - CMD_MIX_AUXA_OLD, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
+					break;
+
+				// These two go together and manipulate some AUX buffers.
+				case CMD_UNK_09_OLD: curr_idx += 13; break;
+				case CMD_UNK_0A_OLD: curr_idx += 13; break;
+
+				// TODO(delroth): figure this one out, it's used by almost every
+				// game I've tested so far.
+				case CMD_UNK_0B_OLD: curr_idx += 4; break;
+
+				case CMD_OUTPUT_OLD:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					addr2_hi = m_cmdlist[curr_idx++];
+					addr2_lo = m_cmdlist[curr_idx++];
+					OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), 0x8000);
+					break;
+
+				case CMD_UNK_0D_OLD: curr_idx += 5; break;
+
+				case CMD_WM_OUTPUT_OLD:
+				{
+					u32 addresses[4] = {
+						(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
+						(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
+						(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
+						(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
+					};
+					curr_idx += 8;
+					OutputWMSamples(addresses);
+					break;
+				}
+
+				case CMD_END_OLD:
+					end = true;
+					break;
+			}
+		}
+		else
+		{
+			switch (cmd)
+			{
+				// Some of these commands are unknown, or unused in this AX HLE.
+				// We still need to skip their arguments using "curr_idx += N".
+
+				case CMD_SETUP:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					SetupProcessing(HILO_TO_32(addr));
+					break;
+
+				case CMD_ADD_TO_LR:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					AddToLR(HILO_TO_32(addr));
+					break;
+
+				case CMD_UNK_02: curr_idx += 2; break;
+				case CMD_UNK_03: curr_idx += 2; break;
+
+				case CMD_PROCESS:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					ProcessPBList(HILO_TO_32(addr));
+					break;
+
+				case CMD_MIX_AUXA:
+				case CMD_MIX_AUXB:
+				case CMD_MIX_AUXC:
+					volume = m_cmdlist[curr_idx++];
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					addr2_hi = m_cmdlist[curr_idx++];
+					addr2_lo = m_cmdlist[curr_idx++];
+					MixAUXSamples(cmd - CMD_MIX_AUXA, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
+					break;
+
+				// These two go together and manipulate some AUX buffers.
+				case CMD_UNK_08: curr_idx += 13; break;
+				case CMD_UNK_09: curr_idx += 13; break;
+
+				// TODO(delroth): figure this one out, it's used by almost every
+				// game I've tested so far.
+				case CMD_UNK_0A: curr_idx += 4; break;
+
+				case CMD_OUTPUT:
+					volume = m_cmdlist[curr_idx++];
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					addr2_hi = m_cmdlist[curr_idx++];
+					addr2_lo = m_cmdlist[curr_idx++];
+					OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), volume);
+					break;
+
+				case CMD_UNK_0C: curr_idx += 5; break;
+
+				case CMD_WM_OUTPUT:
+				{
+					u32 addresses[4] = {
+						(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
+						(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
+						(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
+						(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
+					};
+					curr_idx += 8;
+					OutputWMSamples(addresses);
+					break;
+				}
+
+				case CMD_END:
+					end = true;
+					break;
+			}
 		}
 	}
 }
@@ -244,7 +328,8 @@ void CUCode_AXWii::GenerateVolumeRamp(u16* output, u16 vol1, u16 vol2, size_t nv
 	}
 }
 
-bool CUCode_AXWii::ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* updates)
+bool CUCode_AXWii::ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* updates,
+                                        u32* updates_addr)
 {
 	u16* pb_mem = (u16*)&pb;
 
@@ -260,6 +345,8 @@ bool CUCode_AXWii::ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* upda
 	u32 addr = HILO_TO_32(addr);
 	u16* ptr = (u16*)HLEMemory_Get_Pointer(addr);
 
+	*updates_addr = addr;
+
 	// Copy the updates data and change the offset to match a PB without
 	// updates data.
 	u32 updates_count = num_updates[0] + num_updates[1] + num_updates[2];
@@ -276,11 +363,26 @@ bool CUCode_AXWii::ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* upda
 	}
 
 	// Remove the updates data from the PB
-	memmove(pb_mem + 41, pb_mem + 45, sizeof (pb) - 2 * 45);
+	memmove(pb_mem + 41, pb_mem + 46, sizeof (pb) - 2 * 46);
 
 	return true;
 }
 
+void CUCode_AXWii::ReinjectUpdatesFields(AXPBWii& pb, u16* num_updates, u32 updates_addr)
+{
+	u16* pb_mem = (u16*)&pb;
+
+	// Make some space
+	memmove(pb_mem + 46, pb_mem + 41, sizeof (pb) - 2 * 46);
+
+	// Reinsert previous values
+	pb_mem[41] = num_updates[0];
+	pb_mem[42] = num_updates[1];
+	pb_mem[43] = num_updates[2];
+	pb_mem[44] = updates_addr >> 16;
+	pb_mem[45] = updates_addr & 0xFFFF;
+}
+
 void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 {
 	AXPBWii pb;
@@ -315,7 +417,8 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 
 		u16 num_updates[3];
 		u16 updates[1024];
-		if (ExtractUpdatesFields(pb, num_updates, updates))
+		u32 updates_addr;
+		if (ExtractUpdatesFields(pb, num_updates, updates, &updates_addr))
 		{
 			for (int curr_ms = 0; curr_ms < 3; ++curr_ms)
 			{
@@ -328,6 +431,7 @@ void CUCode_AXWii::ProcessPBList(u32 pb_addr)
 				for (u32 i = 0; i < sizeof (buffers.ptrs) / sizeof (buffers.ptrs[0]); ++i)
 					buffers.ptrs[i] += 32;
 			}
+			ReinjectUpdatesFields(pb, num_updates, updates_addr);
 		}
 		else
 		{
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index ce09b14674..e5ba319787 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -52,8 +52,11 @@ protected:
 	u16 m_last_main_volume;
 	u16 m_last_aux_volumes[3];
 
-	// If needed, extract the updates related fields from a PB.
-	bool ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* updates);
+	// If needed, extract the updates related fields from a PB. We need to
+	// reinject them afterwards so that the correct PB typs is written to RAM.
+	bool ExtractUpdatesFields(AXPBWii& pb, u16* num_updates, u16* updates,
+	                          u32* updates_addr);
+	void ReinjectUpdatesFields(AXPBWii& pb, u16* num_updates, u32 updates_addr);
 
 	// Convert a mixer_control bitfield to our internal representation for that
 	// value. Required because that bitfield has a different meaning in some
@@ -91,7 +94,30 @@ private:
 		CMD_OUTPUT = 0x0B,
 		CMD_UNK_0C = 0x0C,
 		CMD_WM_OUTPUT = 0x0D,
-		CMD_END = 0x0E
+		CMD_END = 0x0E,
+	};
+
+	// A lot of these are similar to the new version, but there is an offset in
+	// the command ids due to the PB_ADDR command (which was removed from the
+	// new AXWii).
+	enum CmdTypeOld
+	{
+		CMD_SETUP_OLD = 0x00,
+		CMD_ADD_TO_LR_OLD = 0x01,
+		CMD_UNK_02_OLD = 0x02,
+		CMD_UNK_03_OLD = 0x03,
+		CMD_PB_ADDR_OLD = 0x04,
+		CMD_PROCESS_OLD = 0x05,
+		CMD_MIX_AUXA_OLD = 0x06,
+		CMD_MIX_AUXB_OLD = 0x07,
+		CMD_MIX_AUXC_OLD = 0x08,
+		CMD_UNK_09_OLD = 0x09,
+		CMD_UNK_0A_OLD = 0x0A,
+		CMD_UNK_0B_OLD = 0x0B,
+		CMD_OUTPUT_OLD = 0x0C,	// no volume!
+		CMD_UNK_0D_OLD = 0x0D,
+		CMD_WM_OUTPUT_OLD = 0x0E,
+		CMD_END_OLD = 0x0F
 	};
 };
 

From 5c67a0bcf681e8cfb86e368e06b093b616574297 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Tue, 2 Apr 2013 02:53:19 +0200
Subject: [PATCH 24/28] Partial fix for the EA Wii games having no sound -
 implement the second OUTPUT command in AXWii. Sound in videos and musics
 still don't play.

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 35 +++++++++++++------
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  7 ++--
 2 files changed, 28 insertions(+), 14 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 1e966d2d0a..8e19a63d4c 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -53,7 +53,7 @@ void CUCode_AXWii::HandleCommandList()
 	u32 pb_addr = 0;
 
 //	WARN_LOG(DSPHLE, "Command list:");
-//	for (u32 i = 0; m_cmdlist[i] != CMD_END_OLD; ++i)
+//	for (u32 i = 0; m_cmdlist[i] != CMD_END; ++i)
 //		WARN_LOG(DSPHLE, "%04x", m_cmdlist[i]);
 //	WARN_LOG(DSPHLE, "-------------");
 
@@ -115,15 +115,15 @@ void CUCode_AXWii::HandleCommandList()
 				case CMD_UNK_0B_OLD: curr_idx += 4; break;
 
 				case CMD_OUTPUT_OLD:
+				case CMD_OUTPUT_DPL2_OLD:
 					addr_hi = m_cmdlist[curr_idx++];
 					addr_lo = m_cmdlist[curr_idx++];
 					addr2_hi = m_cmdlist[curr_idx++];
 					addr2_lo = m_cmdlist[curr_idx++];
-					OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), 0x8000);
+					OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), 0x8000,
+					              cmd == CMD_OUTPUT_DPL2_OLD);
 					break;
 
-				case CMD_UNK_0D_OLD: curr_idx += 5; break;
-
 				case CMD_WM_OUTPUT_OLD:
 				{
 					u32 addresses[4] = {
@@ -190,16 +190,16 @@ void CUCode_AXWii::HandleCommandList()
 				case CMD_UNK_0A: curr_idx += 4; break;
 
 				case CMD_OUTPUT:
+				case CMD_OUTPUT_DPL2:
 					volume = m_cmdlist[curr_idx++];
 					addr_hi = m_cmdlist[curr_idx++];
 					addr_lo = m_cmdlist[curr_idx++];
 					addr2_hi = m_cmdlist[curr_idx++];
 					addr2_lo = m_cmdlist[curr_idx++];
-					OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), volume);
+					OutputSamples(HILO_TO_32(addr2), HILO_TO_32(addr), volume,
+					              cmd == CMD_OUTPUT_DPL2);
 					break;
 
-				case CMD_UNK_0C: curr_idx += 5; break;
-
 				case CMD_WM_OUTPUT:
 				{
 					u32 addresses[4] = {
@@ -499,17 +499,26 @@ void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16
 		}
 }
 
-void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
+void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume,
+                                 bool upload_auxc)
 {
 	u16 volume_ramp[96];
 	GenerateVolumeRamp(volume_ramp, m_last_main_volume, volume, 96);
 	m_last_main_volume = volume;
 
-	int surround_buffer[3 * 32] = { 0 };
+	int upload_buffer[3 * 32] = { 0 };
 
 	for (u32 i = 0; i < 3 * 32; ++i)
-		surround_buffer[i] = Common::swap32(m_samples_surround[i]);
-	memcpy(HLEMemory_Get_Pointer(surround_addr), surround_buffer, sizeof (surround_buffer));
+		upload_buffer[i] = Common::swap32(m_samples_surround[i]);
+	memcpy(HLEMemory_Get_Pointer(surround_addr), upload_buffer, sizeof (upload_buffer));
+
+	if (upload_auxc)
+	{
+		surround_addr += sizeof (upload_buffer);
+		for (u32 i = 0; i < 3 * 32; ++i)
+			upload_buffer[i] = Common::swap32(m_samples_auxC_left[i]);
+		memcpy(HLEMemory_Get_Pointer(surround_addr), upload_buffer, sizeof (upload_buffer));
+	}
 
 	short buffer[3 * 32 * 2];
 
@@ -539,6 +548,10 @@ void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume)
 	}
 
 	memcpy(HLEMemory_Get_Pointer(lr_addr), buffer, sizeof (buffer));
+
+	// There should be a DSP_SYNC message sent here. However, it looks like not
+	// sending it does not cause any issue, and sending it actually causes some
+	// sounds to go at half speed. I have no idea why.
 }
 
 void CUCode_AXWii::OutputWMSamples(u32* addresses)
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index e5ba319787..b6e163541d 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -74,7 +74,8 @@ protected:
 	void AddToLR(u32 val_addr);
 	void ProcessPBList(u32 pb_addr);
 	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
-	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume);
+	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume,
+	                   bool upload_auxc);
 	void OutputWMSamples(u32* addresses);	// 4 addresses
 
 private:
@@ -92,7 +93,7 @@ private:
 		CMD_UNK_09 = 0x09,
 		CMD_UNK_0A = 0x0A,
 		CMD_OUTPUT = 0x0B,
-		CMD_UNK_0C = 0x0C,
+		CMD_OUTPUT_DPL2 = 0x0C,
 		CMD_WM_OUTPUT = 0x0D,
 		CMD_END = 0x0E,
 	};
@@ -115,7 +116,7 @@ private:
 		CMD_UNK_0A_OLD = 0x0A,
 		CMD_UNK_0B_OLD = 0x0B,
 		CMD_OUTPUT_OLD = 0x0C,	// no volume!
-		CMD_UNK_0D_OLD = 0x0D,
+		CMD_OUTPUT_DPL2_OLD = 0x0D,
 		CMD_WM_OUTPUT_OLD = 0x0E,
 		CMD_END_OLD = 0x0F
 	};

From 74dee41b874b1509b2680806d36c1276676349e4 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Tue, 2 Apr 2013 03:10:29 +0200
Subject: [PATCH 25/28] Add implementation for command 02: SUB_TO_LR

---
 .../Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp   | 12 +++++++-----
 Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h  |  6 +++---
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 8e19a63d4c..810194eed9 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -77,12 +77,12 @@ void CUCode_AXWii::HandleCommandList()
 					break;
 
 				case CMD_ADD_TO_LR_OLD:
+				case CMD_SUB_TO_LR_OLD:
 					addr_hi = m_cmdlist[curr_idx++];
 					addr_lo = m_cmdlist[curr_idx++];
-					AddToLR(HILO_TO_32(addr));
+					AddToLR(HILO_TO_32(addr), cmd == CMD_SUB_TO_LR_OLD);
 					break;
 
-				case CMD_UNK_02_OLD: curr_idx += 2; break;
 				case CMD_UNK_03_OLD: curr_idx += 2; break;
 
 				case CMD_PB_ADDR_OLD:
@@ -156,12 +156,12 @@ void CUCode_AXWii::HandleCommandList()
 					break;
 
 				case CMD_ADD_TO_LR:
+				case CMD_SUB_TO_LR:
 					addr_hi = m_cmdlist[curr_idx++];
 					addr_lo = m_cmdlist[curr_idx++];
-					AddToLR(HILO_TO_32(addr));
+					AddToLR(HILO_TO_32(addr), cmd == CMD_SUB_TO_LR);
 					break;
 
-				case CMD_UNK_02: curr_idx += 2; break;
 				case CMD_UNK_03: curr_idx += 2; break;
 
 				case CMD_PROCESS:
@@ -278,12 +278,14 @@ void CUCode_AXWii::SetupProcessing(u32 init_addr)
 	}
 }
 
-void CUCode_AXWii::AddToLR(u32 val_addr)
+void CUCode_AXWii::AddToLR(u32 val_addr, bool neg)
 {
 	int* ptr = (int*)HLEMemory_Get_Pointer(val_addr);
 	for (int i = 0; i < 32 * 3; ++i)
 	{
 		int val = (int)Common::swap32(*ptr++);
+		if (neg)
+			val = -val;
 
 		m_samples_left[i] += val;
 		m_samples_right[i] += val;
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index b6e163541d..9c49f35a81 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -71,7 +71,7 @@ protected:
 	virtual void HandleCommandList();
 
 	void SetupProcessing(u32 init_addr);
-	void AddToLR(u32 val_addr);
+	void AddToLR(u32 val_addr, bool neg);
 	void ProcessPBList(u32 pb_addr);
 	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
 	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume,
@@ -83,7 +83,7 @@ private:
 	{
 		CMD_SETUP = 0x00,
 		CMD_ADD_TO_LR = 0x01,
-		CMD_UNK_02 = 0x02,
+		CMD_SUB_TO_LR = 0x02,
 		CMD_UNK_03 = 0x03,
 		CMD_PROCESS = 0x04,
 		CMD_MIX_AUXA = 0x05,
@@ -105,7 +105,7 @@ private:
 	{
 		CMD_SETUP_OLD = 0x00,
 		CMD_ADD_TO_LR_OLD = 0x01,
-		CMD_UNK_02_OLD = 0x02,
+		CMD_SUB_TO_LR_OLD = 0x02,
 		CMD_UNK_03_OLD = 0x03,
 		CMD_PB_ADDR_OLD = 0x04,
 		CMD_PROCESS_OLD = 0x05,

From 0220fd194072f596092cde1c76be0712d80af786 Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Tue, 2 Apr 2013 03:23:48 +0200
Subject: [PATCH 26/28] Implemented command 03: ADD_SUB_TO_LR

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 27 +++++++++++++++++--
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  5 ++--
 2 files changed, 28 insertions(+), 4 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index 810194eed9..ec640ca038 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -83,7 +83,11 @@ void CUCode_AXWii::HandleCommandList()
 					AddToLR(HILO_TO_32(addr), cmd == CMD_SUB_TO_LR_OLD);
 					break;
 
-				case CMD_UNK_03_OLD: curr_idx += 2; break;
+				case CMD_ADD_SUB_TO_LR_OLD:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					AddSubToLR(HILO_TO_32(addr));
+					break;
 
 				case CMD_PB_ADDR_OLD:
 					addr_hi = m_cmdlist[curr_idx++];
@@ -162,7 +166,11 @@ void CUCode_AXWii::HandleCommandList()
 					AddToLR(HILO_TO_32(addr), cmd == CMD_SUB_TO_LR);
 					break;
 
-				case CMD_UNK_03: curr_idx += 2; break;
+				case CMD_ADD_SUB_TO_LR:
+					addr_hi = m_cmdlist[curr_idx++];
+					addr_lo = m_cmdlist[curr_idx++];
+					AddSubToLR(HILO_TO_32(addr));
+					break;
 
 				case CMD_PROCESS:
 					addr_hi = m_cmdlist[curr_idx++];
@@ -292,6 +300,21 @@ void CUCode_AXWii::AddToLR(u32 val_addr, bool neg)
 	}
 }
 
+void CUCode_AXWii::AddSubToLR(u32 val_addr)
+{
+	int* ptr = (int*)HLEMemory_Get_Pointer(val_addr);
+	for (int i = 0; i < 32 * 3; ++i)
+	{
+		int val = (int)Common::swap32(*ptr++);
+		m_samples_left[i] += val;
+	}
+	for (int i = 0; i < 32 * 3; ++i)
+	{
+		int val = (int)Common::swap32(*ptr++);
+		m_samples_right[i] -= val;
+	}
+}
+
 AXMixControl CUCode_AXWii::ConvertMixerControl(u32 mixer_control)
 {
 	u32 ret = 0;
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index 9c49f35a81..8952961004 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -72,6 +72,7 @@ protected:
 
 	void SetupProcessing(u32 init_addr);
 	void AddToLR(u32 val_addr, bool neg);
+	void AddSubToLR(u32 val_addr);
 	void ProcessPBList(u32 pb_addr);
 	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
 	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume,
@@ -84,7 +85,7 @@ private:
 		CMD_SETUP = 0x00,
 		CMD_ADD_TO_LR = 0x01,
 		CMD_SUB_TO_LR = 0x02,
-		CMD_UNK_03 = 0x03,
+		CMD_ADD_SUB_TO_LR = 0x03,
 		CMD_PROCESS = 0x04,
 		CMD_MIX_AUXA = 0x05,
 		CMD_MIX_AUXB = 0x06,
@@ -106,7 +107,7 @@ private:
 		CMD_SETUP_OLD = 0x00,
 		CMD_ADD_TO_LR_OLD = 0x01,
 		CMD_SUB_TO_LR_OLD = 0x02,
-		CMD_UNK_03_OLD = 0x03,
+		CMD_ADD_SUB_TO_LR_OLD = 0x03,
 		CMD_PB_ADDR_OLD = 0x04,
 		CMD_PROCESS_OLD = 0x05,
 		CMD_MIX_AUXA_OLD = 0x06,

From 6e708005dfd673e305bc568fcc8e4b4e5ac54adf Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Tue, 2 Apr 2013 04:12:17 +0200
Subject: [PATCH 27/28] Implement AXWii commands 08/09, aka. "upload AUXA/AUXB
 and use it as a temp buffer to mix to MAIN L/R/S and AUXC L", aka. "what the
 fuck were they thinking?!"

---
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp | 82 +++++++++++++++++--
 .../Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h   |  9 +-
 2 files changed, 81 insertions(+), 10 deletions(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
index ec640ca038..41d782a01a 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.cpp
@@ -110,9 +110,22 @@ void CUCode_AXWii::HandleCommandList()
 					MixAUXSamples(cmd - CMD_MIX_AUXA_OLD, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
 					break;
 
-				// These two go together and manipulate some AUX buffers.
-				case CMD_UNK_09_OLD: curr_idx += 13; break;
-				case CMD_UNK_0A_OLD: curr_idx += 13; break;
+				case CMD_UPL_AUXA_MIX_LRSC_OLD:
+				case CMD_UPL_AUXB_MIX_LRSC_OLD:
+				{
+					volume = m_cmdlist[curr_idx++];
+					u32 addresses[6] = {
+						(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
+						(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
+						(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
+						(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
+						(u32)(m_cmdlist[curr_idx + 8] << 16) | m_cmdlist[curr_idx + 9],
+						(u32)(m_cmdlist[curr_idx + 10] << 16) | m_cmdlist[curr_idx + 11],
+					};
+					curr_idx += 12;
+					UploadAUXMixLRSC(cmd == CMD_UPL_AUXB_MIX_LRSC_OLD, addresses, volume);
+					break;
+				}
 
 				// TODO(delroth): figure this one out, it's used by almost every
 				// game I've tested so far.
@@ -189,9 +202,22 @@ void CUCode_AXWii::HandleCommandList()
 					MixAUXSamples(cmd - CMD_MIX_AUXA, HILO_TO_32(addr), HILO_TO_32(addr2), volume);
 					break;
 
-				// These two go together and manipulate some AUX buffers.
-				case CMD_UNK_08: curr_idx += 13; break;
-				case CMD_UNK_09: curr_idx += 13; break;
+				case CMD_UPL_AUXA_MIX_LRSC:
+				case CMD_UPL_AUXB_MIX_LRSC:
+				{
+					volume = m_cmdlist[curr_idx++];
+					u32 addresses[6] = {
+						(u32)(m_cmdlist[curr_idx + 0] << 16) | m_cmdlist[curr_idx + 1],
+						(u32)(m_cmdlist[curr_idx + 2] << 16) | m_cmdlist[curr_idx + 3],
+						(u32)(m_cmdlist[curr_idx + 4] << 16) | m_cmdlist[curr_idx + 5],
+						(u32)(m_cmdlist[curr_idx + 6] << 16) | m_cmdlist[curr_idx + 7],
+						(u32)(m_cmdlist[curr_idx + 8] << 16) | m_cmdlist[curr_idx + 9],
+						(u32)(m_cmdlist[curr_idx + 10] << 16) | m_cmdlist[curr_idx + 11],
+					};
+					curr_idx += 12;
+					UploadAUXMixLRSC(cmd == CMD_UPL_AUXB_MIX_LRSC, addresses, volume);
+					break;
+				}
 
 				// TODO(delroth): figure this one out, it's used by almost every
 				// game I've tested so far.
@@ -524,6 +550,50 @@ void CUCode_AXWii::MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16
 		}
 }
 
+void CUCode_AXWii::UploadAUXMixLRSC(int aux_id, u32* addresses, u16 volume)
+{
+	int* aux_left = aux_id ? m_samples_auxB_left : m_samples_auxA_left;
+	int* aux_right = aux_id ? m_samples_auxB_right : m_samples_auxA_right;
+	int* aux_surround = aux_id ? m_samples_auxB_surround : m_samples_auxA_surround;
+	int* auxc_buffer = aux_id ? m_samples_auxC_surround : m_samples_auxC_right;
+
+	int* upload_ptr = (int*)HLEMemory_Get_Pointer(addresses[0]);
+	for (u32 i = 0; i < 96; ++i)
+		*upload_ptr++ = Common::swap32(aux_left[i]);
+	for (u32 i = 0; i < 96; ++i)
+		*upload_ptr++ = Common::swap32(aux_right[i]);
+	for (u32 i = 0; i < 96; ++i)
+		*upload_ptr++ = Common::swap32(aux_surround[i]);
+
+	upload_ptr = (int*)HLEMemory_Get_Pointer(addresses[1]);
+	for (u32 i = 0; i < 96; ++i)
+		*upload_ptr++ = Common::swap32(auxc_buffer[i]);
+
+	u16 volume_ramp[96];
+	GenerateVolumeRamp(volume_ramp, m_last_aux_volumes[aux_id], volume, 96);
+	m_last_aux_volumes[aux_id] = volume;
+
+	int* mix_dest[4] = {
+		m_samples_left,
+		m_samples_right,
+		m_samples_surround,
+		m_samples_auxC_left
+	};
+	for (u32 mix_i = 0; mix_i < 4; ++mix_i)
+	{
+		int* dl_ptr = (int*)HLEMemory_Get_Pointer(addresses[2 + mix_i]);
+		for (u32 i = 0; i < 96; ++i)
+			aux_left[i] = Common::swap32(dl_ptr[i]);
+
+		for (u32 i = 0; i < 96; ++i)
+		{
+			s64 sample = (s64)(s32)aux_left[i];
+			sample *= volume_ramp[i];
+			mix_dest[mix_i][i] += (s32)(sample >> 15);
+		}
+	}
+}
+
 void CUCode_AXWii::OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume,
                                  bool upload_auxc)
 {
diff --git a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
index 8952961004..6e4eb7ad4e 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
+++ b/Source/Core/Core/Src/HW/DSPHLE/UCodes/UCode_AXWii.h
@@ -75,6 +75,7 @@ protected:
 	void AddSubToLR(u32 val_addr);
 	void ProcessPBList(u32 pb_addr);
 	void MixAUXSamples(int aux_id, u32 write_addr, u32 read_addr, u16 volume);
+	void UploadAUXMixLRSC(int aux_id, u32* addresses, u16 volume);
 	void OutputSamples(u32 lr_addr, u32 surround_addr, u16 volume,
 	                   bool upload_auxc);
 	void OutputWMSamples(u32* addresses);	// 4 addresses
@@ -90,8 +91,8 @@ private:
 		CMD_MIX_AUXA = 0x05,
 		CMD_MIX_AUXB = 0x06,
 		CMD_MIX_AUXC = 0x07,
-		CMD_UNK_08 = 0x08,
-		CMD_UNK_09 = 0x09,
+		CMD_UPL_AUXA_MIX_LRSC = 0x08,
+		CMD_UPL_AUXB_MIX_LRSC = 0x09,
 		CMD_UNK_0A = 0x0A,
 		CMD_OUTPUT = 0x0B,
 		CMD_OUTPUT_DPL2 = 0x0C,
@@ -113,8 +114,8 @@ private:
 		CMD_MIX_AUXA_OLD = 0x06,
 		CMD_MIX_AUXB_OLD = 0x07,
 		CMD_MIX_AUXC_OLD = 0x08,
-		CMD_UNK_09_OLD = 0x09,
-		CMD_UNK_0A_OLD = 0x0A,
+		CMD_UPL_AUXA_MIX_LRSC_OLD = 0x09,
+		CMD_UPL_AUXB_MIX_LRSC_OLD = 0x0a,
 		CMD_UNK_0B_OLD = 0x0B,
 		CMD_OUTPUT_OLD = 0x0C,	// no volume!
 		CMD_OUTPUT_DPL2_OLD = 0x0D,

From 986013729142651df2f511d60f2877104589666e Mon Sep 17 00:00:00 2001
From: Pierre Bourdon <delroth@gmail.com>
Date: Tue, 2 Apr 2013 20:42:07 +0200
Subject: [PATCH 28/28] Initialize the audiocommon mixer on the first write to
 DSPCR, even if DSP is not enabled. Fixes issue 6192.

---
 Source/Core/Core/Src/HW/DSPHLE/DSPHLE.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Source/Core/Core/Src/HW/DSPHLE/DSPHLE.cpp b/Source/Core/Core/Src/HW/DSPHLE/DSPHLE.cpp
index 6d2a1f9097..0232605f91 100644
--- a/Source/Core/Core/Src/HW/DSPHLE/DSPHLE.cpp
+++ b/Source/Core/Core/Src/HW/DSPHLE/DSPHLE.cpp
@@ -273,7 +273,7 @@ u16 DSPHLE::DSP_WriteControlRegister(unsigned short _Value)
 	UDSPControl Temp(_Value);
 	if (!m_InitMixer)
 	{
-		if (!Temp.DSPHalt && Temp.DSPInit)
+		if (!Temp.DSPHalt)
 		{
 			InitMixer();
 		}