diff --git a/src/core/hle/D3D8/Direct3D9/Direct3D9.cpp b/src/core/hle/D3D8/Direct3D9/Direct3D9.cpp
index 550ff6c1a..2bcaa487e 100644
--- a/src/core/hle/D3D8/Direct3D9/Direct3D9.cpp
+++ b/src/core/hle/D3D8/Direct3D9/Direct3D9.cpp
@@ -43,6 +43,9 @@ namespace xboxkrnl
 #include "EmuShared.h"
 #include "gui\DbgConsole.h"
 #include "core\hle\D3D8\ResourceTracker.h"
+#include "core\hle\D3D8\XbVertexBuffer.h"
+#include "core\hle\D3D8\XbVertexShader.h"
+#include "core\hle\D3D8\XbPushBuffer.h"
 #include "core\kernel\memory-manager\VMManager.h" // for g_VMManager
 #include "core\kernel\support\EmuXTL.h"
 #include "Logging.h"
@@ -148,7 +151,7 @@ static DWORD                        g_VBLastSwap = 0;
 static XTL::D3DSWAPDATA				g_SwapData = {0};
 static DWORD						g_SwapLast = 0;
 
-static XTL::CxbxVertexBufferConverter VertexBufferConverter = {};
+static CxbxVertexBufferConverter VertexBufferConverter = {};
 
 // cached Direct3D state variable(s)
 static XTL::IDirect3DIndexBuffer   *pClosingLineLoopIndexBuffer = nullptr;
@@ -156,7 +159,7 @@ static XTL::IDirect3DIndexBuffer   *pClosingLineLoopIndexBuffer = nullptr;
 static XTL::IDirect3DIndexBuffer   *pQuadToTriangleD3DIndexBuffer = nullptr;
 static UINT                         QuadToTriangleD3DIndexBuffer_Size = 0; // = NrOfQuadVertices
 
-static XTL::INDEX16                *pQuadToTriangleIndexBuffer = nullptr;
+static INDEX16                     *pQuadToTriangleIndexBuffer = nullptr;
 static UINT                         QuadToTriangleIndexBuffer_Size = 0; // = NrOfQuadVertices
 
 static XTL::IDirect3DSurface       *g_DefaultHostDepthBufferSuface = NULL;
@@ -2525,7 +2528,7 @@ void CxbxUpdateActiveIndexBuffer
 		}
 
 		EmuLog(LOG_LEVEL::DEBUG, "CxbxUpdateActiveIndexBuffer: Copying %d indices (D3DFMT_INDEX16)", IndexCount);
-		memcpy(pData, pIndexData, IndexCount * sizeof(XTL::INDEX16));
+		memcpy(pData, pIndexData, IndexCount * sizeof(INDEX16));
 
 		indexBuffer.pHostIndexBuffer->Unlock();
 	}
@@ -3725,7 +3728,7 @@ HRESULT WINAPI XTL::EMUPATCH(D3DDevice_CreateVertexShader)
 	CxbxVertexShader* pCxbxVertexShader = (CxbxVertexShader*)calloc(1, sizeof(CxbxVertexShader));
 	D3DVERTEXELEMENT *pRecompiledDeclaration = nullptr;
 
-	pRecompiledDeclaration = XTL::EmuRecompileVshDeclaration((DWORD*)pDeclaration,
+	pRecompiledDeclaration = EmuRecompileVshDeclaration((DWORD*)pDeclaration,
                                                    /*bIsFixedFunction=*/pFunction == NULL,
                                                    &XboxDeclarationCount,
                                                    &HostDeclarationSize,
@@ -3749,7 +3752,7 @@ HRESULT WINAPI XTL::EMUPATCH(D3DDevice_CreateVertexShader)
 	{
 		bool bUseDeclarationOnly = false;
 
-		hRet = XTL::EmuRecompileVshFunction((DWORD*)pFunction,
+		hRet = EmuRecompileVshFunction((DWORD*)pFunction,
 			/*bNoReservedConstants=*/g_VertexShaderConstantMode == X_D3DSCM_NORESERVEDCONSTANTS,
 			pRecompiledDeclaration,
 			&bUseDeclarationOnly,
@@ -7354,11 +7357,11 @@ constexpr UINT QuadToTriangleVertexCount(UINT NrOfQuadVertices)
 
 // TODO : Move to own file
 bool WindingClockwise = true;
-constexpr unsigned int IndicesPerPage = PAGE_SIZE / sizeof(XTL::INDEX16);
+constexpr unsigned int IndicesPerPage = PAGE_SIZE / sizeof(INDEX16);
 constexpr unsigned int InputQuadsPerPage = ((IndicesPerPage * VERTICES_PER_QUAD) / VERTICES_PER_TRIANGLE) / TRIANGLES_PER_QUAD;
 
 // TODO : Move to own file
-XTL::INDEX16 *CxbxAssureQuadListIndexBuffer(UINT NrOfQuadVertices)
+INDEX16 *CxbxAssureQuadListIndexBuffer(UINT NrOfQuadVertices)
 {
 	if (QuadToTriangleIndexBuffer_Size < NrOfQuadVertices)
 	{
@@ -7369,10 +7372,10 @@ XTL::INDEX16 *CxbxAssureQuadListIndexBuffer(UINT NrOfQuadVertices)
 		if (pQuadToTriangleIndexBuffer != nullptr)
 			free(pQuadToTriangleIndexBuffer);
 
-		pQuadToTriangleIndexBuffer = (XTL::INDEX16 *)malloc(sizeof(XTL::INDEX16) * NrOfTriangleVertices);
+		pQuadToTriangleIndexBuffer = (INDEX16 *)malloc(sizeof(INDEX16) * NrOfTriangleVertices);
 
 		UINT i = 0;
-		XTL::INDEX16 j = 0;
+		INDEX16 j = 0;
 		while (i + 6 < NrOfTriangleVertices)
 		{
 			if (WindingClockwise) {
@@ -7423,7 +7426,7 @@ void CxbxAssureQuadListD3DIndexBuffer(UINT NrOfQuadVertices)
 		// Round the number of indices up so we'll allocate whole pages
 		QuadToTriangleD3DIndexBuffer_Size = RoundUp(NrOfQuadVertices, InputQuadsPerPage);
 		UINT NrOfTriangleVertices = QuadToTriangleVertexCount(QuadToTriangleD3DIndexBuffer_Size); // 4 > 6
-		UINT uiIndexBufferSize = sizeof(XTL::INDEX16) * NrOfTriangleVertices;
+		UINT uiIndexBufferSize = sizeof(INDEX16) * NrOfTriangleVertices;
 
 		// Create a new native index buffer of the above determined size :
 		if (pQuadToTriangleD3DIndexBuffer != nullptr) {
@@ -7445,7 +7448,7 @@ void CxbxAssureQuadListD3DIndexBuffer(UINT NrOfQuadVertices)
 			CxbxKrnlCleanup("CxbxAssureQuadListD3DIndexBuffer : IndexBuffer Create Failed!");
 
 		// Put quadlist-to-triangle-list index mappings into this buffer :
-		XTL::INDEX16* pIndexBufferData = nullptr;
+		INDEX16* pIndexBufferData = nullptr;
 		hRet = pQuadToTriangleD3DIndexBuffer->Lock(0, uiIndexBufferSize, (D3DLockData **)&pIndexBufferData, D3DLOCK_DISCARD);
 		DEBUG_D3DRESULT(hRet, "pQuadToTriangleD3DIndexBuffer->Lock");
 
@@ -7467,13 +7470,13 @@ void CxbxAssureQuadListD3DIndexBuffer(UINT NrOfQuadVertices)
 
 // TODO : Move to own file
 // Calls SetIndices with a separate index-buffer, that's populated with the supplied indices.
-void CxbxDrawIndexedClosingLine(XTL::INDEX16 LowIndex, XTL::INDEX16 HighIndex)
+void CxbxDrawIndexedClosingLine(INDEX16 LowIndex, INDEX16 HighIndex)
 {
 	LOG_INIT // Allows use of DEBUG_D3DRESULT
 
 	HRESULT hRet;
 
-	const UINT uiIndexBufferSize = sizeof(XTL::INDEX16) * 2; // 4 bytes needed for 2 indices
+	const UINT uiIndexBufferSize = sizeof(INDEX16) * 2; // 4 bytes needed for 2 indices
 	if (pClosingLineLoopIndexBuffer == nullptr) {
 		hRet = g_pD3DDevice->CreateIndexBuffer(
 			uiIndexBufferSize, 
@@ -7487,7 +7490,7 @@ void CxbxDrawIndexedClosingLine(XTL::INDEX16 LowIndex, XTL::INDEX16 HighIndex)
 			CxbxKrnlCleanup("Unable to create pClosingLineLoopIndexBuffer for D3DPT_LINELOOP emulation");
 	}
 
-	XTL::INDEX16 *pCxbxClosingLineLoopIndexBufferData = nullptr;
+	INDEX16 *pCxbxClosingLineLoopIndexBufferData = nullptr;
 	hRet = pClosingLineLoopIndexBuffer->Lock(0, uiIndexBufferSize, (D3DLockData **)(&pCxbxClosingLineLoopIndexBufferData), D3DLOCK_DISCARD);
 	DEBUG_D3DRESULT(hRet, "pClosingLineLoopIndexBuffer->Lock");
 
@@ -7514,11 +7517,11 @@ void CxbxDrawIndexedClosingLine(XTL::INDEX16 LowIndex, XTL::INDEX16 HighIndex)
 }
 
 // TODO : Move to own file
-void CxbxDrawIndexedClosingLineUP(XTL::INDEX16 LowIndex, XTL::INDEX16 HighIndex, void *pHostVertexStreamZeroData, UINT uiHostVertexStreamZeroStride)
+void CxbxDrawIndexedClosingLineUP(INDEX16 LowIndex, INDEX16 HighIndex, void *pHostVertexStreamZeroData, UINT uiHostVertexStreamZeroStride)
 {
 	LOG_INIT // Allows use of DEBUG_D3DRESULT
 
-	XTL::INDEX16 CxbxClosingLineIndices[2] = { LowIndex, HighIndex };
+	INDEX16 CxbxClosingLineIndices[2] = { LowIndex, HighIndex };
 
 	HRESULT hRet = g_pD3DDevice->DrawIndexedPrimitiveUP(
 		XTL::D3DPT_LINELIST,
@@ -7537,7 +7540,7 @@ void CxbxDrawIndexedClosingLineUP(XTL::INDEX16 LowIndex, XTL::INDEX16 HighIndex,
 
 // Requires assigned pIndexData
 // Called by D3DDevice_DrawIndexedVertices and EmuExecutePushBufferRaw (twice)
-void XTL::CxbxDrawIndexed(CxbxDrawContext &DrawContext)
+void CxbxDrawIndexed(CxbxDrawContext &DrawContext)
 {
 	LOG_INIT // Allows use of DEBUG_D3DRESULT
 
@@ -7548,7 +7551,7 @@ void XTL::CxbxDrawIndexed(CxbxDrawContext &DrawContext)
 	CxbxUpdateActiveIndexBuffer(DrawContext.pIndexData, DrawContext.dwVertexCount);
 	VertexBufferConverter.Apply(&DrawContext);
 
-	if (DrawContext.XboxPrimitiveType == X_D3DPT_QUADLIST) {
+	if (DrawContext.XboxPrimitiveType == XTL::X_D3DPT_QUADLIST) {
 		UINT uiStartIndex = 0;
 		int iNumVertices = (int)DrawContext.dwVertexCount;
 		// Indexed quadlist can be drawn using unpatched indexes via multiple draws of 2 'strip' triangles :
@@ -7564,7 +7567,7 @@ void XTL::CxbxDrawIndexed(CxbxDrawContext &DrawContext)
 
 			// Emulate a quad by drawing each as a fan of 2 triangles
 			HRESULT hRet = g_pD3DDevice->DrawIndexedPrimitive(
-				D3DPT_TRIANGLEFAN,
+				XTL::D3DPT_TRIANGLEFAN,
 				DrawContext.dwIndexBase,
 				LowIndex, // minIndex
 				(HighIndex - LowIndex) + 1, // NumVertices
@@ -7596,7 +7599,7 @@ void XTL::CxbxDrawIndexed(CxbxDrawContext &DrawContext)
 		DEBUG_D3DRESULT(hRet, "g_pD3DDevice->DrawIndexedPrimitive");
 
 		g_dwPrimPerFrame += DrawContext.dwHostPrimitiveCount;
-		if (DrawContext.XboxPrimitiveType == X_D3DPT_LINELOOP) {
+		if (DrawContext.XboxPrimitiveType == XTL::X_D3DPT_LINELOOP) {
 			// Close line-loops using a final single line, drawn from the end to the start vertex
 			LOG_TEST_CASE("X_D3DPT_LINELOOP");
 			// Read the end and start index from the supplied index data
@@ -7619,7 +7622,7 @@ void XTL::CxbxDrawIndexed(CxbxDrawContext &DrawContext)
 // TODO : Move to own file
 // Drawing function specifically for rendering Xbox draw calls supplying a 'User Pointer'.
 // Called by D3DDevice_DrawVerticesUP, EmuExecutePushBufferRaw and EmuFlushIVB
-void XTL::CxbxDrawPrimitiveUP(CxbxDrawContext &DrawContext)
+void CxbxDrawPrimitiveUP(CxbxDrawContext &DrawContext)
 {
 	LOG_INIT // Allows use of DEBUG_D3DRESULT
 
@@ -7629,7 +7632,7 @@ void XTL::CxbxDrawPrimitiveUP(CxbxDrawContext &DrawContext)
 	assert(DrawContext.dwIndexBase == 0); // No IndexBase under Draw*UP
 
 	VertexBufferConverter.Apply(&DrawContext);
-	if (DrawContext.XboxPrimitiveType == X_D3DPT_QUADLIST) {
+	if (DrawContext.XboxPrimitiveType == XTL::X_D3DPT_QUADLIST) {
 		// LOG_TEST_CASE("X_D3DPT_QUADLIST"); // X-Marbles and XDK Sample PlayField hits this case
 		// Draw quadlists using a single 'quad-to-triangle mapping' index buffer :
 		INDEX16 *pIndexData = CxbxAssureQuadListIndexBuffer(DrawContext.dwVertexCount);
@@ -7642,12 +7645,12 @@ void XTL::CxbxDrawPrimitiveUP(CxbxDrawContext &DrawContext)
 		INDEX16 HighIndex = (INDEX16)(DrawContext.dwVertexCount - 1);
 
 		HRESULT hRet = g_pD3DDevice->DrawIndexedPrimitiveUP(
-			D3DPT_TRIANGLELIST, // Draw indexed triangles instead of quads
+			XTL::D3DPT_TRIANGLELIST, // Draw indexed triangles instead of quads
 			LowIndex, // MinVertexIndex
 			(HighIndex - LowIndex) + 1, // NumVertexIndices
 			PrimitiveCount,
 			pIndexData,
-			D3DFMT_INDEX16, // IndexDataFormat
+			XTL::D3DFMT_INDEX16, // IndexDataFormat
 			DrawContext.pHostVertexStreamZeroData,
 			DrawContext.uiHostVertexStreamZeroStride
 		);
@@ -7666,7 +7669,7 @@ void XTL::CxbxDrawPrimitiveUP(CxbxDrawContext &DrawContext)
 		DEBUG_D3DRESULT(hRet, "g_pD3DDevice->DrawPrimitiveUP");
 
 		g_dwPrimPerFrame += DrawContext.dwHostPrimitiveCount;
-		if (DrawContext.XboxPrimitiveType == X_D3DPT_LINELOOP) {
+		if (DrawContext.XboxPrimitiveType == XTL::X_D3DPT_LINELOOP) {
 			// Note : XDK samples reaching this case : DebugKeyboard, Gamepad, Tiling, ShadowBuffer
 			// Since we can use pHostVertexStreamZeroData here, we can close the line simpler than
 			// via CxbxDrawIndexedClosingLine, by drawing two indices via DrawIndexedPrimitiveUP.
@@ -8324,9 +8327,9 @@ VOID WINAPI XTL::EMUPATCH(D3DDevice_SetPalette)
 	//    g_pD3DDevice9->SetCurrentTexturePalette(Stage, Stage);
 
 	if (Stage < XTL::X_D3DTS_STAGECOUNT) {
-		if (g_pCurrentPalette[Stage] != GetDataFromXboxResource(pPalette) && XTL::EmuD3DActiveTexture[Stage] != nullptr) {
+		if (g_pCurrentPalette[Stage] != GetDataFromXboxResource(pPalette) && EmuD3DActiveTexture[Stage] != nullptr) {
 			// If the palette for a texture has changed, we need to re-convert the texture
-			FreeHostResource(GetHostResourceKey(XTL::EmuD3DActiveTexture[Stage]));
+			FreeHostResource(GetHostResourceKey(EmuD3DActiveTexture[Stage]));
 		}
 
 		// Cache palette data and size
diff --git a/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.cpp b/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.cpp
index 65011bbf6..1555e1374 100644
--- a/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.cpp
+++ b/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.cpp
@@ -11,13 +11,13 @@
 // Walk an index buffer to find the minimum and maximum indices
 
 // Default implementation
-void WalkIndexBuffer_NoSIMD(XTL::INDEX16 & LowIndex, XTL::INDEX16 & HighIndex, XTL::INDEX16 * pIndexData, DWORD dwIndexCount)
+void WalkIndexBuffer_NoSIMD(INDEX16 & LowIndex, INDEX16 & HighIndex, INDEX16 * pIndexData, DWORD dwIndexCount)
 {
 	// Determine highest and lowest index in use 
 	LowIndex = pIndexData[0];
 	HighIndex = LowIndex;
 	for (unsigned int i = 1; i < dwIndexCount; i++) {
-		XTL::INDEX16 Index = pIndexData[i];
+		INDEX16 Index = pIndexData[i];
 		if (LowIndex > Index)
 			LowIndex = Index;
 		if (HighIndex < Index)
@@ -26,7 +26,7 @@ void WalkIndexBuffer_NoSIMD(XTL::INDEX16 & LowIndex, XTL::INDEX16 & HighIndex, X
 }
 
 //SSE 4.1 implementation
-void WalkIndexBuffer_SSE41(XTL::INDEX16 & LowIndex, XTL::INDEX16 & HighIndex, XTL::INDEX16 * pIndexData, DWORD dwIndexCount)
+void WalkIndexBuffer_SSE41(INDEX16 & LowIndex, INDEX16 & HighIndex, INDEX16 * pIndexData, DWORD dwIndexCount)
 {
 	// We can fit 8 ushorts into 128 bit SIMD registers
 	int iterations = dwIndexCount / 8;
@@ -58,8 +58,8 @@ void WalkIndexBuffer_SSE41(XTL::INDEX16 & LowIndex, XTL::INDEX16 & HighIndex, XT
 	max = _mm_minpos_epu16(max);
 
 	// Get the min and max out
-	LowIndex = (XTL::INDEX16) _mm_cvtsi128_si32(min);
-	HighIndex = (XTL::INDEX16) USHRT_MAX - _mm_cvtsi128_si32(max); // invert back
+	LowIndex = (INDEX16) _mm_cvtsi128_si32(min);
+	HighIndex = (INDEX16) USHRT_MAX - _mm_cvtsi128_si32(max); // invert back
 
 	// Compare with the remaining values that didn't fit neatly into the SIMD registers
 	for (DWORD i = dwIndexCount - remainder; i < dwIndexCount; i++) {
@@ -74,8 +74,8 @@ void WalkIndexBuffer_SSE41(XTL::INDEX16 & LowIndex, XTL::INDEX16 & HighIndex, XT
 // TODO AVX2, AVX512 implementations
 
 // Detect SSE support to select real implementation on first call
-void(*WalkIndexBuffer)(XTL::INDEX16 &, XTL::INDEX16 &, XTL::INDEX16 *, DWORD) =
-[](XTL::INDEX16 &LowIndex, XTL::INDEX16 &HighIndex, XTL::INDEX16 *pIndexData, DWORD dwIndexCount)
+void(*WalkIndexBuffer)(INDEX16 &, INDEX16 &, INDEX16 *, DWORD) =
+[](INDEX16 &LowIndex, INDEX16 &HighIndex, INDEX16 *pIndexData, DWORD dwIndexCount)
 {
 	SimdCaps supports;
 	if (supports.SSE41())
diff --git a/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.h b/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.h
index d064c7a4a..e3bd0cf2e 100644
--- a/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.h
+++ b/src/core/hle/D3D8/Direct3D9/WalkIndexBuffer.h
@@ -6,9 +6,9 @@
 
 extern void(*WalkIndexBuffer)
 (
-	XTL::INDEX16 &LowIndex,
-	XTL::INDEX16 &HighIndex,
-	XTL::INDEX16 *pIndexData,
+	INDEX16 &LowIndex,
+	INDEX16 &HighIndex,
+	INDEX16 *pIndexData,
 	DWORD dwIndexCount
 );
 
diff --git a/src/core/hle/D3D8/XbD3D8Types.h b/src/core/hle/D3D8/XbD3D8Types.h
index 908295912..8af691b67 100644
--- a/src/core/hle/D3D8/XbD3D8Types.h
+++ b/src/core/hle/D3D8/XbD3D8Types.h
@@ -1111,8 +1111,6 @@ const int X_D3DVSDT_PBYTE4      = 0x44; // xbox ext. 4D packed byte expanded to
 const int X_D3DVSDT_FLOAT2H     = 0x72; // xbox ext. 3D float that expands to (value, value, 0.0, value). Useful for projective texture coordinates.
 const int X_D3DVSDT_NONE        = 0x02; // xbox ext. nsp
 
-typedef WORD INDEX16;
-
 typedef enum _X_D3DVSD_TOKENTYPE
 {
 	X_D3DVSD_TOKEN_NOP = 0,           // NOP or extension
@@ -1210,96 +1208,6 @@ typedef DWORD NV2AMETHOD;
 //
 
 // Host vertex shader counts
-#define VSH_MAX_TEMPORARY_REGISTERS 32
-#define VSH_MAX_INTERMEDIATE_COUNT     1024 // The maximum number of intermediate format slots
-#define VSH_VS11_MAX_INSTRUCTION_COUNT 128
-#define VSH_VS2X_MAX_INSTRUCTION_COUNT 256
-#define VSH_VS30_MAX_INSTRUCTION_COUNT 512
-
 #define CXBX_D3DVS_CONSTREG_VERTEXDATA4F_BASE   (X_D3DVS_CONSTREG_COUNT + 1)
 
-typedef struct _CxbxPixelShader
-{
-	//IDirect3DPixelShader* pShader;
-	//ID3DXConstantTable *pConstantTable;
-	DWORD Handle;
-
-	BOOL bBumpEnvMap;
-
-	// constants
-	DWORD PSRealC0[8];
-	DWORD PSRealC1[8];
-	DWORD PSRealFC0;
-	DWORD PSRealFC1;
-
-	BOOL bConstantsInitialized;
-	BOOL bConstantsChanged;
-
-	DWORD dwStatus;
-	X_D3DPIXELSHADERDEF	PSDef;
-
-	DWORD dwStageMap[X_D3DTS_STAGECOUNT];
-
-}
-CxbxPixelShader;
-
-typedef struct _CxbxVertexShaderStreamElement
-{
-	UINT XboxType; // The stream element data types (xbox)
-	UINT HostByteSize; // The stream element data sizes (pc)
-}
-CxbxVertexShaderStreamElement;
-
-/* See host typedef struct _D3DVERTEXELEMENT9
-{
-	WORD    Stream;     // Stream index
-	WORD    Offset;     // Offset in the stream in bytes
-	BYTE    Type;       // Data type
-	BYTE    Method;     // Processing method
-	BYTE    Usage;      // Semantics
-	BYTE    UsageIndex; // Semantic index
-} D3DVERTEXELEMENT9, *LPD3DVERTEXELEMENT9;
-*/
-
-typedef struct _CxbxVertexShaderStreamInfo
-{
-	BOOL  NeedPatch;       // This is to know whether it's data which must be patched
-	BOOL DeclPosition;
-	WORD HostVertexStride;
-	DWORD NumberOfVertexElements;        // Number of the stream data types
-	WORD CurrentStreamNumber;
-	CxbxVertexShaderStreamElement VertexElements[X_VSH_MAX_ATTRIBUTES + 16]; // TODO : Why 16 extrahost additions?)
-}
-CxbxVertexShaderStreamInfo;
-
-typedef struct _CxbxVertexShaderInfo
-{
-	UINT                       NumberOfVertexStreams; // The number of streams the vertex shader uses
-	CxbxVertexShaderStreamInfo VertexStreams[X_VSH_MAX_STREAMS];
-}
-CxbxVertexShaderInfo;
-
-typedef struct _CxbxVertexShader
-{
-	// These are the parameters given by the XBE,
-	// we save them to be able to return them when necessary.
-	DWORD* pXboxDeclarationCopy;
-	DWORD                 XboxDeclarationCount; // Xbox's number of DWORD-sized X_D3DVSD* tokens
-	DWORD                 XboxFunctionSize;
-	DWORD* pXboxFunctionCopy;
-	UINT                  XboxNrAddressSlots;
-	DWORD                 XboxVertexShaderType;
-	// DWORD              XboxStatus; // Used by VshHandleIsValidShader()
-
-	// The resulting host variables
-	DWORD HostFVF; // Flexible Vertex Format (used when there's no host vertex shader)
-	XTL::IDirect3DVertexShader* pHostVertexShader; // if nullptr, use SetFVF(HostFVF);
-	XTL::IDirect3DVertexDeclaration* pHostVertexDeclaration;
-	DWORD                 HostDeclarationSize;
-
-	// Needed for dynamic stream patching
-	CxbxVertexShaderInfo  VertexShaderInfo;
-}
-CxbxVertexShader;
-
 #endif
diff --git a/src/core/hle/D3D8/XbPushBuffer.cpp b/src/core/hle/D3D8/XbPushBuffer.cpp
index 547585d4c..821124729 100644
--- a/src/core/hle/D3D8/XbPushBuffer.cpp
+++ b/src/core/hle/D3D8/XbPushBuffer.cpp
@@ -31,8 +31,11 @@
 
 #include "core\kernel\support\Emu.h"
 #include "core\kernel\support\EmuXTL.h"
-#include "XbD3D8Types.h" // For X_D3DFORMAT
+namespace XTL {
+	#include "core\hle\D3D8\XbD3D8Types.h" // For X_D3DFORMAT
+}
 #include "core\hle\D3D8\ResourceTracker.h"
+#include "core\hle\D3D8\XbPushBuffer.h"
 #include "devices/video/nv2a.h" // For g_NV2A, PGRAPHState
 #include "devices/video/nv2a_int.h" // For NV** defines
 #include "Logging.h"
@@ -47,7 +50,7 @@ static void DbgDumpMesh(WORD *pIndexData, DWORD dwCount);
 
 // Determine the size (in number of floating point texture coordinates) of the texture format (indexed 0 .. 3).
 // This is the reverse of the D3DFVF_TEXCOORDSIZE[0..3] macros.
-int XTL::DxbxFVF_GetNumberOfTextureCoordinates(DWORD dwFVF, int aTextureIndex)
+int DxbxFVF_GetNumberOfTextureCoordinates(DWORD dwFVF, int aTextureIndex)
 {
 	// See D3DFVF_TEXCOORDSIZE1()
 	switch ((dwFVF >> ((aTextureIndex * 2) + 16)) & 3) {
@@ -63,7 +66,7 @@ int XTL::DxbxFVF_GetNumberOfTextureCoordinates(DWORD dwFVF, int aTextureIndex)
 
 // Dxbx Note: This code appeared in EmuExecutePushBufferRaw and occured
 // in EmuFlushIVB too, so it's generalize in this single implementation.
-UINT XTL::DxbxFVFToVertexSizeInBytes(DWORD dwFVF, BOOL bIncludeTextures)
+UINT DxbxFVFToVertexSizeInBytes(DWORD dwFVF, BOOL bIncludeTextures)
 {
 /*
 	X_D3DFVF_POSITION_MASK    = $00E; // Dec  /2  #fl
@@ -115,12 +118,14 @@ UINT XTL::DxbxFVFToVertexSizeInBytes(DWORD dwFVF, BOOL bIncludeTextures)
 	return Result;
 }
 
-void XTL::EmuExecutePushBuffer
+void EmuExecutePushBuffer
 (
-    X_D3DPushBuffer       *pPushBuffer,
-    X_D3DFixup            *pFixup
+	XTL::X_D3DPushBuffer       *pPushBuffer,
+	XTL::X_D3DFixup            *pFixup
 )
 {
+	using namespace XTL;
+
 	//Check whether Fixup exists or not. 
 	if (pFixup != NULL) {
 		LOG_TEST_CASE("PushBuffer has fixups");
@@ -451,12 +456,14 @@ typedef union {
 	#define COMMAND_WORD_MASK_JUMP_LONG 0xFFFFFFFC /*  2 .. 28 */
 } nv_fifo_command;
 
-extern void XTL::EmuExecutePushBufferRaw
+extern void EmuExecutePushBufferRaw
 (
 	void *pPushData,
 	uint32_t uSizeInBytes
 )
 {
+	using namespace XTL; // for logging
+
 	HLE_init_pgraph_plugins(); // TODO : Move to more approriate spot
 
 	// Test-case : Azurik (see https://github.com/Cxbx-Reloaded/Cxbx-Reloaded/issues/360)
@@ -492,7 +499,7 @@ extern void XTL::EmuExecutePushBufferRaw
 	uint32_t *dma_put; // pushbuffer current end address
 	uint32_t *dma_get; //pushbuffer current read address
 	struct {
-		NV2AMETHOD mthd; // Current method
+		XTL::NV2AMETHOD mthd; // Current method
 		uint32_t subc; // :3 = Current subchannel
 		uint32_t mcnt; // :24 = Current method count
 		bool ni; // Current command's NI (non-increasing) flag
diff --git a/src/core/hle/D3D8/XbPushBuffer.h b/src/core/hle/D3D8/XbPushBuffer.h
index 8905d4327..ca2ac5111 100644
--- a/src/core/hle/D3D8/XbPushBuffer.h
+++ b/src/core/hle/D3D8/XbPushBuffer.h
@@ -24,6 +24,8 @@
 // ******************************************************************
 #ifndef XBPUSHBUFFER_H
 #define XBPUSHBUFFER_H
+
+#include "core/hle/D3D8/XbVertexBuffer.h" // for CxbxDrawContext
 
 extern int DxbxFVF_GetNumberOfTextureCoordinates(DWORD dwFVF, int aTextureIndex);
 extern UINT DxbxFVFToVertexSizeInBytes(DWORD dwFVF, BOOL bIncludeTextures);
@@ -33,8 +35,8 @@ extern void CxbxDrawPrimitiveUP(CxbxDrawContext &DrawContext);
 
 extern void EmuExecutePushBuffer
 (
-    X_D3DPushBuffer       *pPushBuffer,
-    X_D3DFixup            *pFixup
+	XTL::X_D3DPushBuffer       *pPushBuffer,
+	XTL::X_D3DFixup            *pFixup
 );
 
 extern void EmuExecutePushBufferRaw
diff --git a/src/core/hle/D3D8/XbVertexBuffer.cpp b/src/core/hle/D3D8/XbVertexBuffer.cpp
index 410ba5cfa..e770f3969 100644
--- a/src/core/hle/D3D8/XbVertexBuffer.cpp
+++ b/src/core/hle/D3D8/XbVertexBuffer.cpp
@@ -33,6 +33,8 @@
 #include "core\kernel\support\Emu.h"
 #include "core\kernel\support\EmuXTL.h"
 #include "core\hle\D3D8\ResourceTracker.h"
+#include "core\hle\D3D8\XbPushBuffer.h" // for DxbxFVF_GetNumberOfTextureCoordinates
+#include "core\hle\D3D8\XbVertexBuffer.h"
 
 #include <ctime>
 #include <chrono>
@@ -41,18 +43,18 @@
 #define MAX_STREAM_NOT_USED_TIME (2 * CLOCKS_PER_SEC) // TODO: Trim the not used time
 
 // Inline vertex buffer emulation
-extern XTL::X_D3DPRIMITIVETYPE XTL::g_InlineVertexBuffer_PrimitiveType = XTL::X_D3DPT_INVALID;
-extern DWORD                   XTL::g_InlineVertexBuffer_FVF = 0;
-extern struct XTL::_D3DIVB    *XTL::g_InlineVertexBuffer_Table = nullptr;
-extern UINT                    XTL::g_InlineVertexBuffer_TableLength = 0;
-extern UINT                    XTL::g_InlineVertexBuffer_TableOffset = 0;
+extern XTL::X_D3DPRIMITIVETYPE g_InlineVertexBuffer_PrimitiveType = XTL::X_D3DPT_INVALID;
+extern DWORD                   g_InlineVertexBuffer_FVF = 0;
+extern struct _D3DIVB         *g_InlineVertexBuffer_Table = nullptr;
+extern UINT                    g_InlineVertexBuffer_TableLength = 0;
+extern UINT                    g_InlineVertexBuffer_TableOffset = 0;
 
 FLOAT *g_InlineVertexBuffer_pData = nullptr;
 UINT   g_InlineVertexBuffer_DataSize = 0;
 
 extern XTL::X_D3DVertexBuffer      *g_pVertexBuffer = NULL;
 
-extern DWORD				XTL::g_dwPrimPerFrame = 0;
+extern DWORD				g_dwPrimPerFrame = 0;
 extern XTL::X_D3DVertexBuffer*g_D3DStreams[X_VSH_MAX_STREAMS];
 extern UINT g_D3DStreamStrides[X_VSH_MAX_STREAMS];
 extern XTL::X_D3DSurface* g_pXboxRenderTarget;
@@ -62,7 +64,7 @@ bool GetHostRenderTargetDimensions(DWORD* pHostWidth, DWORD* pHostHeight, XTL::I
 uint32_t GetPixelContainerWidth(XTL::X_D3DPixelContainer* pPixelContainer);
 uint32_t GetPixelContainerHeight(XTL::X_D3DPixelContainer* pPixelContainer);
 
-void XTL::CxbxPatchedStream::Activate(XTL::CxbxDrawContext *pDrawContext, UINT uiStream) const
+void CxbxPatchedStream::Activate(CxbxDrawContext *pDrawContext, UINT uiStream) const
 {
 	//LOG_INIT // Allows use of DEBUG_D3DRESULT
 
@@ -86,12 +88,12 @@ void XTL::CxbxPatchedStream::Activate(XTL::CxbxDrawContext *pDrawContext, UINT u
 	}
 }
 
-XTL::CxbxPatchedStream::CxbxPatchedStream()
+CxbxPatchedStream::CxbxPatchedStream()
 {
     isValid = false;
 }
 
-XTL::CxbxPatchedStream::~CxbxPatchedStream()
+CxbxPatchedStream::~CxbxPatchedStream()
 {
     if (bCachedHostVertexStreamZeroDataIsAllocated) {
         free(pCachedHostVertexStreamZeroData);
@@ -106,7 +108,7 @@ XTL::CxbxPatchedStream::~CxbxPatchedStream()
     }
 }
 
-XTL::CxbxVertexBufferConverter::CxbxVertexBufferConverter()
+CxbxVertexBufferConverter::CxbxVertexBufferConverter()
 {
     m_uiNbrStreams = 0;
     m_pVertexShaderInfo = nullptr;
@@ -145,9 +147,9 @@ int CountActiveD3DStreams()
 	return lastStreamIndex;
 }
 
-XTL::CxbxVertexShaderInfo *GetCxbxVertexShaderInfo(DWORD XboxVertexShaderHandle); // forward
+CxbxVertexShaderInfo *GetCxbxVertexShaderInfo(DWORD XboxVertexShaderHandle); // forward
 
-UINT XTL::CxbxVertexBufferConverter::GetNbrStreams(CxbxDrawContext *pDrawContext)
+UINT CxbxVertexBufferConverter::GetNbrStreams(CxbxDrawContext *pDrawContext)
 {
 	// Draw..Up always have one stream
 	if (pDrawContext->pXboxVertexStreamZeroData != xbnullptr) {
@@ -195,7 +197,7 @@ inline FLOAT ByteToFloat(const BYTE value)
 	return ((FLOAT)value) / 255.0f;
 }
 
-XTL::CxbxPatchedStream& XTL::CxbxVertexBufferConverter::GetPatchedStream(uint64_t key)
+CxbxPatchedStream& CxbxVertexBufferConverter::GetPatchedStream(uint64_t key)
 {
     // First, attempt to fetch an existing patched stream
     auto it = m_PatchedStreams.find(key);
@@ -222,7 +224,7 @@ XTL::CxbxPatchedStream& XTL::CxbxVertexBufferConverter::GetPatchedStream(uint64_
     return stream;
 }
 
-void XTL::CxbxVertexBufferConverter::PrintStats()
+void CxbxVertexBufferConverter::PrintStats()
 {
     printf("Vertex Buffer Cache Status: \n");
     printf("- Cache Size: %d\n", m_PatchedStreams.size());
@@ -230,7 +232,7 @@ void XTL::CxbxVertexBufferConverter::PrintStats()
     printf("- Misses: %d\n", m_TotalCacheMisses);
 }
 
-void XTL::CxbxVertexBufferConverter::ConvertStream
+void CxbxVertexBufferConverter::ConvertStream
 (
 	CxbxDrawContext *pDrawContext,
     UINT             uiStream
@@ -257,7 +259,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 			if (i + 1 <= dwTexN) {
 				pActivePixelContainer[i].NrTexCoords = DxbxFVF_GetNumberOfTextureCoordinates(XboxFVF, i);
 				// TODO : Use GetXboxBaseTexture()
-				X_D3DBaseTexture *pXboxBaseTexture = EmuD3DActiveTexture[i];
+				XTL::X_D3DBaseTexture *pXboxBaseTexture = XTL::EmuD3DActiveTexture[i];
 				if (pXboxBaseTexture != xbnullptr) {
 					extern XTL::X_D3DFORMAT GetXboxPixelContainerFormat(const XTL::X_D3DPixelContainer *pXboxPixelContainer); // TODO : Move to XTL-independent header file
 
@@ -297,7 +299,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 	UINT uiHostVertexStride = 0;
 	DWORD dwHostVertexDataSize = 0;
 	uint8_t *pHostVertexData = nullptr;
-	IDirect3DVertexBuffer *pNewHostVertexBuffer = nullptr;
+	XTL::IDirect3DVertexBuffer *pNewHostVertexBuffer = nullptr;
 
     if (pDrawContext->pXboxVertexStreamZeroData != xbnullptr) {
 		// There should only be one stream (stream zero) in this case
@@ -440,7 +442,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 				// Dxbx note : The following code handles only the D3DVSDT enums that need conversion;
 				// All other cases are catched by the memcpy in the default-block.
 				switch (pVertexShaderStreamInfo->VertexElements[uiElement].XboxType) {
-				case X_D3DVSDT_NORMSHORT1: { // 0x11:
+				case XTL::X_D3DVSDT_NORMSHORT1: { // 0x11:
 					// Test-cases : Halo - Combat Evolved
 					XboxElementByteSize = 1 * sizeof(SHORT);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_SHORT2N) {
@@ -456,7 +458,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_NORMSHORT2: { // 0x21:
+				case XTL::X_D3DVSDT_NORMSHORT2: { // 0x21:
 					// Test-cases : Baldur's Gate: Dark Alliance 2, F1 2002, Gun, Halo - Combat Evolved, Scrapland 
 					XboxElementByteSize = 2 * sizeof(SHORT);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_SHORT2N) {
@@ -476,7 +478,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_NORMSHORT3: { // 0x31:
+				case XTL::X_D3DVSDT_NORMSHORT3: { // 0x31:
 					// Test-cases : Cel Damage, Constantine, Destroy All Humans!
 					XboxElementByteSize = 3 * sizeof(SHORT);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_SHORT4N) {
@@ -495,7 +497,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_NORMSHORT4: { // 0x41:
+				case XTL::X_D3DVSDT_NORMSHORT4: { // 0x41:
 					// Test-cases : Judge Dredd: Dredd vs Death, NHL Hitz 2002, Silent Hill 2, Sneakers, Tony Hawk Pro Skater 4
 					XboxElementByteSize = 4 * sizeof(SHORT);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_SHORT4N) {
@@ -519,7 +521,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_NORMPACKED3: { // 0x16:
+				case XTL::X_D3DVSDT_NORMPACKED3: { // 0x16:
 					// Test-cases : Dashboard
 					XboxElementByteSize = 1 * sizeof(int32_t);
 					// Make it FLOAT3
@@ -539,14 +541,14 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					pHostVertexAsFloat[2] = PackedIntToFloat(NormPacked3.z, 511.0f, 512.f);
 					break;
 				}
-				case X_D3DVSDT_SHORT1: { // 0x15:
+				case XTL::X_D3DVSDT_SHORT1: { // 0x15:
 					XboxElementByteSize = 1 * sizeof(SHORT);
 					// Make it SHORT2 and set the second short to 0
 					pHostVertexAsShort[0] = pXboxVertexAsShort[0];
 					pHostVertexAsShort[1] = 0;
 					break;
 				}
-				case X_D3DVSDT_SHORT3: { // 0x35:
+				case XTL::X_D3DVSDT_SHORT3: { // 0x35:
 					// Test-cases : Turok
 					XboxElementByteSize = 3 * sizeof(SHORT);
 					// Make it a SHORT4 and set the fourth short to 1
@@ -556,7 +558,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					pHostVertexAsShort[3] = 1; // Turok verified (character disappears when this is 32767)
 					break;
 				}
-				case X_D3DVSDT_PBYTE1: { // 0x14:
+				case XTL::X_D3DVSDT_PBYTE1: { // 0x14:
 					XboxElementByteSize = 1 * sizeof(BYTE);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_UBYTE4N) {
 						// Make it UBYTE4N
@@ -572,7 +574,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_PBYTE2: { // 0x24:
+				case XTL::X_D3DVSDT_PBYTE2: { // 0x24:
 					XboxElementByteSize = 2 * sizeof(BYTE);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_UBYTE4N) {
 						// Make it UBYTE4N
@@ -589,7 +591,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_PBYTE3: { // 0x34:
+				case XTL::X_D3DVSDT_PBYTE3: { // 0x34:
 					// Test-cases : Turok
 					XboxElementByteSize = 3 * sizeof(BYTE);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_UBYTE4N) {
@@ -608,7 +610,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_PBYTE4: { // 0x44:
+				case XTL::X_D3DVSDT_PBYTE4: { // 0x44:
 					// Test-case : Jet Set Radio Future
 					XboxElementByteSize = 4 * sizeof(BYTE);
 					if (g_D3DCaps.DeclTypes & D3DDTCAPS_UBYTE4N) {
@@ -632,7 +634,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					}
 					break;
 				}
-				case X_D3DVSDT_FLOAT2H: { // 0x72:
+				case XTL::X_D3DVSDT_FLOAT2H: { // 0x72:
 					XboxElementByteSize = 3 * sizeof(FLOAT);
 					// Make it FLOAT4 and set the third float to 0.0
 					pHostVertexAsFloat[0] = pXboxVertexAsFloat[0];
@@ -641,7 +643,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 					pHostVertexAsFloat[3] = pXboxVertexAsFloat[2];
 					break;
 				}
-				case X_D3DVSDT_NONE: { // 0x02: // Skip it
+				case XTL::X_D3DVSDT_NONE: { // 0x02: // Skip it
 					// Test-case : WWE RAW2
 					LOG_TEST_CASE("X_D3DVSDT_NONE");
 					break;
@@ -675,7 +677,7 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 
 		// Locate texture coordinate offset in vertex structure.
 		if (bNeedTextureNormalization) {
-			uiTextureCoordinatesByteOffsetInVertex = XTL::DxbxFVFToVertexSizeInBytes(XboxFVF, /*bIncludeTextures=*/false);
+			uiTextureCoordinatesByteOffsetInVertex = DxbxFVFToVertexSizeInBytes(XboxFVF, /*bIncludeTextures=*/false);
 			if (bNeedVertexPatching) {
 				LOG_TEST_CASE("Potential xbox vs host texture-offset difference! (bNeedVertexPatching within bNeedTextureNormalization)");
 			}
@@ -784,9 +786,9 @@ void XTL::CxbxVertexBufferConverter::ConvertStream
 	patchedStream.Activate(pDrawContext, uiStream);
 }
 
-void XTL::CxbxVertexBufferConverter::Apply(CxbxDrawContext *pDrawContext)
+void CxbxVertexBufferConverter::Apply(CxbxDrawContext *pDrawContext)
 {
-	if ((pDrawContext->XboxPrimitiveType < X_D3DPT_POINTLIST) || (pDrawContext->XboxPrimitiveType > X_D3DPT_POLYGON))
+	if ((pDrawContext->XboxPrimitiveType < XTL::X_D3DPT_POINTLIST) || (pDrawContext->XboxPrimitiveType > XTL::X_D3DPT_POLYGON))
 		CxbxKrnlCleanup("Unknown primitive type: 0x%.02X\n", pDrawContext->XboxPrimitiveType);
 
     m_pVertexShaderInfo = nullptr;
@@ -812,7 +814,7 @@ void XTL::CxbxVertexBufferConverter::Apply(CxbxDrawContext *pDrawContext)
 		ConvertStream(pDrawContext, uiStream);
     }
 
-	if (pDrawContext->XboxPrimitiveType == X_D3DPT_QUADSTRIP) {
+	if (pDrawContext->XboxPrimitiveType == XTL::X_D3DPT_QUADSTRIP) {
 		// Quad strip is just like a triangle strip, but requires two vertices per primitive.
 		// A quadstrip starts with 4 vertices and adds 2 vertices per additional quad.
 		// This is much like a trianglestrip, which starts with 3 vertices and adds
@@ -823,12 +825,12 @@ void XTL::CxbxVertexBufferConverter::Apply(CxbxDrawContext *pDrawContext)
 		// handled by d3d :
 		// Test-case : XDK Samples (FocusBlur, MotionBlur, Trees, PaintEffect, PlayField)
 		// No need to set : pDrawContext->XboxPrimitiveType = X_D3DPT_TRIANGLESTRIP;
-		pDrawContext->dwHostPrimitiveCount = EmuD3DVertex2PrimitiveCount(X_D3DPT_TRIANGLESTRIP, pDrawContext->dwVertexCount);
+		pDrawContext->dwHostPrimitiveCount = EmuD3DVertex2PrimitiveCount(XTL::X_D3DPT_TRIANGLESTRIP, pDrawContext->dwVertexCount);
 	} else {
 		pDrawContext->dwHostPrimitiveCount = EmuD3DVertex2PrimitiveCount(pDrawContext->XboxPrimitiveType, pDrawContext->dwVertexCount);
 	}
 
-	if (pDrawContext->XboxPrimitiveType == X_D3DPT_POLYGON) {
+	if (pDrawContext->XboxPrimitiveType == XTL::X_D3DPT_POLYGON) {
 		// Convex polygon is the same as a triangle fan.
 		// No need to set : pDrawContext->XboxPrimitiveType = X_D3DPT_TRIANGLEFAN;
 		// Test-case : Panzer Dragoon ORTA (when entering in-game)
@@ -836,9 +838,9 @@ void XTL::CxbxVertexBufferConverter::Apply(CxbxDrawContext *pDrawContext)
 	}
 }
 
-VOID XTL::EmuFlushIVB()
+VOID EmuFlushIVB()
 {
-	CxbxUpdateNativeD3DResources();
+	XTL::CxbxUpdateNativeD3DResources();
 
     // Parse IVB table with current FVF shader if possible.
     bool bFVF = VshHandleIsFVF(g_CurrentXboxVertexShaderHandle);
diff --git a/src/core/hle/D3D8/XbVertexBuffer.h b/src/core/hle/D3D8/XbVertexBuffer.h
index 252a525e6..0ca8d37dd 100644
--- a/src/core/hle/D3D8/XbVertexBuffer.h
+++ b/src/core/hle/D3D8/XbVertexBuffer.h
@@ -29,9 +29,12 @@
 
 //#include <ctime> // Conflict with io.h
 
+//#include "core\kernel\support\EmuXTL.h"
+#include "core\hle\D3D8\XbVertexShader.h" // for CxbxVertexShaderInfo
+
 typedef struct _CxbxDrawContext
 {
-    IN     X_D3DPRIMITIVETYPE    XboxPrimitiveType;
+    IN     XTL::X_D3DPRIMITIVETYPE    XboxPrimitiveType;
     IN     DWORD                 dwVertexCount;
     IN     DWORD                 dwStartVertex; // Only D3DDevice_DrawVertices sets this (potentially higher than default 0)
     // The current vertex shader, used to identify the streams
@@ -54,7 +57,7 @@ class CxbxPatchedStream
 public:
     CxbxPatchedStream();
     ~CxbxPatchedStream();
-    void Activate(XTL::CxbxDrawContext *pDrawContext, UINT uiStream) const;
+    void Activate(CxbxDrawContext *pDrawContext, UINT uiStream) const;
     bool                    isValid = false;
     XTL::X_D3DPRIMITIVETYPE XboxPrimitiveType = XTL::X_D3DPT_NONE;
     PVOID                   pCachedXboxVertexData = xbnullptr;
@@ -88,7 +91,7 @@ class CxbxVertexBufferConverter
         std::list<CxbxPatchedStream> m_PatchedStreamUsageList;             // Linked list of vertex streams, least recently used is last in the list
         CxbxPatchedStream& GetPatchedStream(uint64_t);                     // Fetches (or inserts) a patched stream associated with the given key
 
-        XTL::CxbxVertexShaderInfo *m_pVertexShaderInfo;
+        CxbxVertexShaderInfo *m_pVertexShaderInfo;
 
         // Returns the number of streams of a patch
         UINT GetNbrStreams(CxbxDrawContext *pPatchDesc);
@@ -98,7 +101,7 @@ class CxbxVertexBufferConverter
 };
 
 // inline vertex buffer emulation
-extern X_D3DPRIMITIVETYPE      g_InlineVertexBuffer_PrimitiveType;
+extern XTL::X_D3DPRIMITIVETYPE      g_InlineVertexBuffer_PrimitiveType;
 extern DWORD                   g_InlineVertexBuffer_FVF;
 
 extern struct _D3DIVB
@@ -107,11 +110,11 @@ extern struct _D3DIVB
     FLOAT            Rhw;          // X_D3DVSDE_VERTEX (*)   > D3DFVF_XYZ / D3DFVF_XYZRHW
 	FLOAT			 Blend[4];	   // X_D3DVSDE_BLENDWEIGHT  > D3DFVF_XYZB1 (and 3 more up to D3DFVF_XYZB4)
     XTL::D3DXVECTOR3 Normal;       // X_D3DVSDE_NORMAL       > D3DFVF_NORMAL
-	D3DCOLOR         Diffuse;      // X_D3DVSDE_DIFFUSE      > D3DFVF_DIFFUSE
-	D3DCOLOR         Specular;     // X_D3DVSDE_SPECULAR     > D3DFVF_SPECULAR
+	XTL::D3DCOLOR         Diffuse;      // X_D3DVSDE_DIFFUSE      > D3DFVF_DIFFUSE
+	XTL::D3DCOLOR         Specular;     // X_D3DVSDE_SPECULAR     > D3DFVF_SPECULAR
 	FLOAT            Fog;          // X_D3DVSDE_FOG          > D3DFVF_FOG unavailable; TODO : How to handle?
-	D3DCOLOR         BackDiffuse;  // X_D3DVSDE_BACKDIFFUSE  > D3DFVF_BACKDIFFUSE unavailable; TODO : How to handle?
-	D3DCOLOR         BackSpecular; // X_D3DVSDE_BACKSPECULAR > D3DFVF_BACKSPECULAR unavailable; TODO : How to handle?
+	XTL::D3DCOLOR         BackDiffuse;  // X_D3DVSDE_BACKDIFFUSE  > D3DFVF_BACKDIFFUSE unavailable; TODO : How to handle?
+	XTL::D3DCOLOR         BackSpecular; // X_D3DVSDE_BACKSPECULAR > D3DFVF_BACKSPECULAR unavailable; TODO : How to handle?
     XTL::D3DXVECTOR4 TexCoord[4];  // X_D3DVSDE_TEXCOORD0    > D3DFVF_TEX1 (and 4 more up to D3DFVF_TEX4)
 
 	// (*) X_D3DVSDE_POSITION and X_D3DVSDE_VERTEX both set Position, but Rhw seems optional,
diff --git a/src/core/hle/D3D8/XbVertexShader.cpp b/src/core/hle/D3D8/XbVertexShader.cpp
index 78bba58c4..027ff0421 100644
--- a/src/core/hle/D3D8/XbVertexShader.cpp
+++ b/src/core/hle/D3D8/XbVertexShader.cpp
@@ -33,6 +33,8 @@
 #include "core\kernel\support\Emu.h"
 #include "core\kernel\support\EmuFS.h"
 #include "core\kernel\support\EmuXTL.h"
+#include "core\hle\D3D8\XbVertexShader.h"
+
 #include "XbD3D8Types.h" // For X_D3DVSDE_*
 #include <sstream>
 #include <unordered_map>
@@ -45,7 +47,7 @@
 	#define VSH_MAX_INSTRUCTION_COUNT VSH_VS2X_MAX_INSTRUCTION_COUNT // == 256
 #endif
 
-// Host Vertex Shader version (mustn't conflict with any VERSION_XVS*)
+// Internal Vertex Shader version (mustn't conflict with any VERSION_XVS*)
 #define VERSION_CXBX 0x7863 // 'cx' Cxbx vertex shader, not an official value, used in VshConvertShader() and VshWriteShader()
 
 #define DbgVshPrintf \
@@ -1743,8 +1745,8 @@ class XboxVertexDeclarationConverter
 {
 protected:
 	// Internal variables
-	XTL::CxbxVertexShaderInfo* pVertexShaderInfoToSet;
-	XTL::CxbxVertexShaderStreamInfo* pCurrentVertexShaderStreamInfo = nullptr;
+	CxbxVertexShaderInfo* pVertexShaderInfoToSet;
+	CxbxVertexShaderStreamInfo* pCurrentVertexShaderStreamInfo = nullptr;
 	DWORD temporaryCount;
 	bool IsFixedFunction;
 	XTL::D3DVERTEXELEMENT* pRecompiled;
@@ -2153,7 +2155,7 @@ private:
 		}
 
 		// save patching information
-		XTL::CxbxVertexShaderStreamElement *pCurrentElement = &(pCurrentVertexShaderStreamInfo->VertexElements[pCurrentVertexShaderStreamInfo->NumberOfVertexElements]);
+		CxbxVertexShaderStreamElement *pCurrentElement = &(pCurrentVertexShaderStreamInfo->VertexElements[pCurrentVertexShaderStreamInfo->NumberOfVertexElements]);
 		pCurrentElement->XboxType = XboxVertexElementDataType;
 		pCurrentElement->HostByteSize = HostVertexElementByteSize;
 		pCurrentVertexShaderStreamInfo->NumberOfVertexElements++;
@@ -2230,7 +2232,7 @@ private:
 	}
 
 public:
-	XTL::D3DVERTEXELEMENT *Convert(DWORD* pXboxDeclaration, bool bIsFixedFunction, XTL::CxbxVertexShaderInfo* pCxbxVertexShaderInfo)
+	XTL::D3DVERTEXELEMENT *Convert(DWORD* pXboxDeclaration, bool bIsFixedFunction, CxbxVertexShaderInfo* pCxbxVertexShaderInfo)
 	{
 		using namespace XTL;
 
@@ -2281,7 +2283,7 @@ public:
 	}
 };
 
-XTL::D3DVERTEXELEMENT *XTL::EmuRecompileVshDeclaration
+XTL::D3DVERTEXELEMENT *EmuRecompileVshDeclaration
 (
     DWORD                *pXboxDeclaration,
     bool                  bIsFixedFunction,
@@ -2301,16 +2303,18 @@ XTL::D3DVERTEXELEMENT *XTL::EmuRecompileVshDeclaration
 }
 
 // recompile xbox vertex shader function
-extern HRESULT XTL::EmuRecompileVshFunction
+extern HRESULT EmuRecompileVshFunction
 (
     DWORD        *pXboxFunction,
     bool          bNoReservedConstants,
-    D3DVERTEXELEMENT *pRecompiledDeclaration,
+	XTL::D3DVERTEXELEMENT *pRecompiledDeclaration,
     bool   		 *pbUseDeclarationOnly,
     DWORD        *pXboxFunctionSize,
-    LPD3DXBUFFER *ppRecompiledShader
+	XTL::LPD3DXBUFFER *ppRecompiledShader
 )
 {
+	using namespace XTL;
+
 	X_VSH_SHADER_HEADER   *pXboxVertexShaderHeader = (X_VSH_SHADER_HEADER*)pXboxFunction;
     DWORD               *pToken;
     boolean             EOI = false;
@@ -2437,7 +2441,7 @@ extern HRESULT XTL::EmuRecompileVshFunction
     return hRet;
 }
 
-extern void XTL::FreeVertexDynamicPatch(CxbxVertexShader *pVertexShader)
+extern void FreeVertexDynamicPatch(CxbxVertexShader *pVertexShader)
 {
     pVertexShader->VertexShaderInfo.NumberOfVertexStreams = 0;
 }
@@ -2448,7 +2452,7 @@ boolean VshHandleIsValidShader(DWORD XboxVertexShaderHandle)
 #if 0
 	//printf( "VS = 0x%.08X\n", XboxVertexShaderHandle );
 
-    XTL::CxbxVertexShader *pCxbxVertexShader = XTL::GetCxbxVertexShader(XboxVertexShaderHandle);
+    CxbxVertexShader *pCxbxVertexShader = GetCxbxVertexShader(XboxVertexShaderHandle);
     if (pCxbxVertexShader) {
         if (pCxbxVertexShader->XboxStatus != 0)
         {
@@ -2470,16 +2474,16 @@ boolean VshHandleIsValidShader(DWORD XboxVertexShaderHandle)
     return TRUE;
 }
 
-extern boolean XTL::IsValidCurrentShader(void)
+extern boolean IsValidCurrentShader(void)
 {
 	// Dxbx addition : There's no need to call
 	// XTL_EmuIDirect3DDevice_GetVertexShader, just check g_CurrentXboxVertexShaderHandle :
 	return VshHandleIsValidShader(g_CurrentXboxVertexShaderHandle);
 }
 
-XTL::CxbxVertexShaderInfo *GetCxbxVertexShaderInfo(DWORD XboxVertexShaderHandle)
+CxbxVertexShaderInfo *GetCxbxVertexShaderInfo(DWORD XboxVertexShaderHandle)
 {
-    XTL::CxbxVertexShader *pCxbxVertexShader = XTL::GetCxbxVertexShader(XboxVertexShaderHandle);
+    CxbxVertexShader *pCxbxVertexShader = GetCxbxVertexShader(XboxVertexShaderHandle);
 
     for (uint32_t i = 0; i < pCxbxVertexShader->VertexShaderInfo.NumberOfVertexStreams; i++)
     {
@@ -2491,9 +2495,9 @@ XTL::CxbxVertexShaderInfo *GetCxbxVertexShaderInfo(DWORD XboxVertexShaderHandle)
     return NULL;
 }
 
-std::unordered_map<DWORD, XTL::CxbxVertexShader*> g_CxbxVertexShaders;
+std::unordered_map<DWORD, CxbxVertexShader*> g_CxbxVertexShaders;
 
-XTL::CxbxVertexShader* XTL::GetCxbxVertexShader(DWORD XboxVertexShaderHandle)
+CxbxVertexShader* GetCxbxVertexShader(DWORD XboxVertexShaderHandle)
 {
 	if (VshHandleIsVertexShader(XboxVertexShaderHandle)) {
 		auto it = g_CxbxVertexShaders.find(XboxVertexShaderHandle);
@@ -2505,7 +2509,7 @@ XTL::CxbxVertexShader* XTL::GetCxbxVertexShader(DWORD XboxVertexShaderHandle)
 	return nullptr;
 }
 
-void XTL::SetCxbxVertexShader(DWORD XboxVertexShaderHandle, XTL::CxbxVertexShader* shader)
+void SetCxbxVertexShader(DWORD XboxVertexShaderHandle, CxbxVertexShader* shader)
 {
 	auto it = g_CxbxVertexShaders.find(XboxVertexShaderHandle);
 	if (it != g_CxbxVertexShaders.end() && it->second != nullptr && shader != nullptr) {
diff --git a/src/core/hle/D3D8/XbVertexShader.h b/src/core/hle/D3D8/XbVertexShader.h
index f5e3486e2..cf635a219 100644
--- a/src/core/hle/D3D8/XbVertexShader.h
+++ b/src/core/hle/D3D8/XbVertexShader.h
@@ -25,16 +25,86 @@
 #ifndef XBVERTEXSHADER_H
 #define XBVERTEXSHADER_H
 
-#include "Cxbx.h"
+//#include "Cxbx.h"
+//#include "core\hle\D3D8\XbVertexBuffer.h" // for CxbxVertexShaderStreamInfo
+namespace XTL {
+	#include "core\hle\D3D8\XbD3D8Types.h" // for X_VSH_MAX_ATTRIBUTES
+}
+
+// Host vertex shader counts
+#define VSH_MAX_TEMPORARY_REGISTERS 32
+#define VSH_MAX_INTERMEDIATE_COUNT     1024 // The maximum number of intermediate format slots
+#define VSH_VS11_MAX_INSTRUCTION_COUNT 128
+#define VSH_VS2X_MAX_INSTRUCTION_COUNT 256
+#define VSH_VS30_MAX_INSTRUCTION_COUNT 512
+
+typedef struct _CxbxVertexShaderStreamElement
+{
+	UINT XboxType; // The stream element data types (xbox)
+	UINT HostByteSize; // The stream element data sizes (pc)
+}
+CxbxVertexShaderStreamElement;
+
+/* See host typedef struct _D3DVERTEXELEMENT9
+{
+	WORD    Stream;     // Stream index
+	WORD    Offset;     // Offset in the stream in bytes
+	BYTE    Type;       // Data type
+	BYTE    Method;     // Processing method
+	BYTE    Usage;      // Semantics
+	BYTE    UsageIndex; // Semantic index
+} D3DVERTEXELEMENT9, *LPD3DVERTEXELEMENT9;
+*/
+
+typedef struct _CxbxVertexShaderStreamInfo
+{
+	BOOL  NeedPatch;       // This is to know whether it's data which must be patched
+	BOOL DeclPosition;
+	WORD HostVertexStride;
+	DWORD NumberOfVertexElements;        // Number of the stream data types
+	WORD CurrentStreamNumber;
+	CxbxVertexShaderStreamElement VertexElements[X_VSH_MAX_ATTRIBUTES + 16]; // TODO : Why 16 extrahost additions?)
+}
+CxbxVertexShaderStreamInfo;
+
+typedef struct _CxbxVertexShaderInfo
+{
+	UINT                       NumberOfVertexStreams; // The number of streams the vertex shader uses
+	CxbxVertexShaderStreamInfo VertexStreams[X_VSH_MAX_STREAMS];
+}
+CxbxVertexShaderInfo;
+
+typedef struct _CxbxVertexShader
+{
+	// These are the parameters given by the XBE,
+	// we save them to be able to return them when necessary.
+	DWORD* pXboxDeclarationCopy;
+	DWORD                 XboxDeclarationCount; // Xbox's number of DWORD-sized X_D3DVSD* tokens
+	DWORD                 XboxFunctionSize;
+	DWORD* pXboxFunctionCopy;
+	UINT                  XboxNrAddressSlots;
+	DWORD                 XboxVertexShaderType;
+	// DWORD              XboxStatus; // Used by VshHandleIsValidShader()
+
+	// The resulting host variables
+	DWORD HostFVF; // Flexible Vertex Format (used when there's no host vertex shader)
+	XTL::IDirect3DVertexShader* pHostVertexShader; // if nullptr, use SetFVF(HostFVF);
+	XTL::IDirect3DVertexDeclaration* pHostVertexDeclaration;
+	DWORD                 HostDeclarationSize;
+
+	// Needed for dynamic stream patching
+	CxbxVertexShaderInfo  VertexShaderInfo;
+}
+CxbxVertexShader;
 
 // recompile xbox vertex shader declaration
-extern D3DVERTEXELEMENT *EmuRecompileVshDeclaration
+extern XTL::D3DVERTEXELEMENT *EmuRecompileVshDeclaration
 (
     DWORD                *pXboxDeclaration,
     bool                  bIsFixedFunction,
     DWORD                *pXboxDeclarationCount,
     DWORD                *pHostDeclarationSize,
-    XTL::CxbxVertexShaderInfo *pCxbxVertexShaderInfo
+    CxbxVertexShaderInfo *pCxbxVertexShaderInfo
 );
 
 // recompile xbox vertex shader function
@@ -42,10 +112,10 @@ extern HRESULT EmuRecompileVshFunction
 (
     DWORD        *pXboxFunction,
     bool          bNoReservedConstants,
-    D3DVERTEXELEMENT *pRecompiledDeclaration,
+    XTL::D3DVERTEXELEMENT *pRecompiledDeclaration,
     bool   		 *pbUseDeclarationOnly,
     DWORD        *pXboxFunctionSize,
-    LPD3DXBUFFER *ppRecompiledShader
+	XTL::LPD3DXBUFFER *ppRecompiledShader
 );
 
 extern void FreeVertexDynamicPatch(CxbxVertexShader *pVertexShader);
@@ -55,7 +125,7 @@ extern boolean IsValidCurrentShader(void);
 
 inline boolean VshHandleIsVertexShader(DWORD Handle) { return (Handle & X_D3DFVF_RESERVED0) ? TRUE : FALSE; }
 inline boolean VshHandleIsFVF(DWORD Handle) { return !VshHandleIsVertexShader(Handle); }
-inline X_D3DVertexShader *VshHandleToXboxVertexShader(DWORD Handle) { return (X_D3DVertexShader *)(Handle & ~X_D3DFVF_RESERVED0);}
+inline XTL::X_D3DVertexShader *VshHandleToXboxVertexShader(DWORD Handle) { return (XTL::X_D3DVertexShader *)(Handle & ~X_D3DFVF_RESERVED0);}
 
 extern CxbxVertexShader* GetCxbxVertexShader(DWORD XboxVertexShaderHandle);
 extern void SetCxbxVertexShader(DWORD XboxVertexShaderHandle, CxbxVertexShader* shader);
diff --git a/src/core/kernel/support/Emu.h b/src/core/kernel/support/Emu.h
index 124e7e51b..2e3b543a3 100644
--- a/src/core/kernel/support/Emu.h
+++ b/src/core/kernel/support/Emu.h
@@ -77,6 +77,7 @@ typedef struct DUMMY_KERNEL
 	IMAGE_SECTION_HEADER SectionHeader;
 } *PDUMMY_KERNEL;
 
+typedef WORD INDEX16;
 
 extern bool g_DisablePixelShaders;
 extern bool g_UseAllCores;
diff --git a/src/core/kernel/support/EmuXTL.h b/src/core/kernel/support/EmuXTL.h
index 21793da6e..12a44f4cc 100644
--- a/src/core/kernel/support/EmuXTL.h
+++ b/src/core/kernel/support/EmuXTL.h
@@ -33,9 +33,6 @@ namespace XTL
     #include "core\hle\XAPI\Xapi.h"
     #include "core\hle\D3D8\Direct3D9/Direct3D9.h"
     #include "core\hle\D3D8\XbConvert.h"
-    #include "core\hle\D3D8\XbVertexBuffer.h"
-    #include "core\hle\D3D8\XbPushBuffer.h"
-    #include "core\hle\D3D8\XbVertexShader.h"
 	#include "core\hle\D3D8\XbPixelShader.h"
     #include "core\hle\D3D8\XbState.h"
     #include "core\hle\DSOUND\DirectSound\DirectSound.hpp"