Merge branch 'vertexloader_cleanup' into vertex-loader-cleanup

Conflicts:
	Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp

Source/Core/VideoCommon/Src/DLCache.cpp

@@ -550,8 +550,7 @@ void CompileAndRunDisplayList(u32 address, u32 size, CachedDisplayList *dl)
 							cmd_byte & GX_VAT_MASK,   // Vertex loader index (0 - 7)
 							(cmd_byte & GX_PRIMITIVE_MASK) >> GX_PRIMITIVE_SHIFT,
 							numVertices);
-						u8* EndAddress = VertexManager::s_pCurBufferPointer;
+						u32 Vdatasize = (u32)(VertexManager::s_pCurBufferPointer - StartAddress);
-						u32 Vdatasize = (u32)(EndAddress - StartAddress);
 						if (Vdatasize > 0)
 						{
 							// Compile

Source/Core/VideoCommon/Src/VertexLoader.cpp

@@ -72,6 +72,10 @@ int colElements[2];
 float posScale;
 float tcScale[8];
 
+// bbox must read vertex position, so convert it to this buffer
+static float s_bbox_vertex_buffer[3];
+static u8 *s_bbox_pCurBufferPointer_orig;
+
 static const float fractionTable[32] = {
 	1.0f / (1U << 0), 1.0f / (1U << 1), 1.0f / (1U << 2), 1.0f / (1U << 3),
 	1.0f / (1U << 4), 1.0f / (1U << 5), 1.0f / (1U << 6), 1.0f / (1U << 7),
@@ -99,17 +103,32 @@ void LOADERDECL PosMtx_Write()
 	*VertexManager::s_pCurBufferPointer++ = 0;
 }
 
+void LOADERDECL UpdateBoundingBoxPrepare() 
+{
+	if (!PixelEngine::bbox_active)
+		return;
+	
+	// set our buffer as videodata buffer, so we will get a copy of the vertex positions
+	// this is a big hack, but so we can use the same converting function then without bbox
+	s_bbox_pCurBufferPointer_orig = VertexManager::s_pCurBufferPointer;
+	VertexManager::s_pCurBufferPointer = (u8*)s_bbox_vertex_buffer;
+}
+
 void LOADERDECL UpdateBoundingBox() 
 {
 	if (!PixelEngine::bbox_active)
 		return;
 	
-	// Truly evil hack, reading backwards from the write pointer. If we were writing to write-only
+	// reset videodata pointer
-	// memory like we might have been with a D3D vertex buffer, this would have been a bad idea.
+	VertexManager::s_pCurBufferPointer = s_bbox_pCurBufferPointer_orig;
-	float *data = (float *)(VertexManager::s_pCurBufferPointer - 12);
+	
+	// copy vertex pointers
+	memcpy(VertexManager::s_pCurBufferPointer, s_bbox_vertex_buffer, 12);
+	VertexManager::s_pCurBufferPointer += 12;
+
 	// We must transform the just loaded point by the current world and projection matrix - in software.
 	// Then convert to screen space and update the bounding box.
-	float p[3] = {data[0], data[1], data[2]};
+	float p[3] = {s_bbox_vertex_buffer[0], s_bbox_vertex_buffer[1], s_bbox_vertex_buffer[2]};
 
 	const float *world_matrix  = (float*)xfmem + MatrixIndexA.PosNormalMtxIdx * 4;
 	const float *proj_matrix = &g_fProjectionMatrix[0];
@@ -267,15 +286,16 @@ void VertexLoader::CompileVertexTranslator()
 	if (m_VtxDesc.Tex7MatIdx) {m_VertexSize += 1; m_NativeFmt->m_components |= VB_HAS_TEXMTXIDX7; WriteCall(TexMtx_ReadDirect_UByte); }
 
 	// Write vertex position loader
+	if(g_ActiveConfig.bUseBBox) {
+		WriteCall(UpdateBoundingBoxPrepare);
 		WriteCall(VertexLoader_Position::GetFunction(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements));
+		WriteCall(UpdateBoundingBox);
+	} else {
+		WriteCall(VertexLoader_Position::GetFunction(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements));
+	}
 	m_VertexSize += VertexLoader_Position::GetSize(m_VtxDesc.Position, m_VtxAttr.PosFormat, m_VtxAttr.PosElements);
 	nat_offset += 12;
 
-	// OK, so we just got a point. Let's go back and read it for the bounding box.
-	
-	if(g_ActiveConfig.bUseBBox)
-		WriteCall(UpdateBoundingBox);
-
 	// Normals
 	vtx_decl.num_normals = 0;
 	if (m_VtxDesc.Normal != NOT_PRESENT)
@@ -571,7 +591,6 @@ void VertexLoader::RunVertices(int vtx_attr_group, int primitive, int count)
 		if (remainingVerts < granularity) {
 			INCSTAT(stats.thisFrame.numBufferSplits);
 			// This buffer full - break current primitive and flush, to switch to the next buffer.
-			u8* plastptr = VertexManager::s_pCurBufferPointer;
 			if (v - startv > 0)
 				VertexManager::AddVertices(primitive, v - startv + extraverts);
 			VertexManager::Flush();
@@ -581,27 +600,28 @@ void VertexLoader::RunVertices(int vtx_attr_group, int primitive, int count)
 				case 3: // triangle strip, copy last two vertices
 					// a little trick since we have to keep track of signs
 					if (v & 1) {
-						memcpy_gc(VertexManager::s_pCurBufferPointer, plastptr-2*native_stride, native_stride);
+						g_pVideoData -= m_VertexSize*2;
-						memcpy_gc(VertexManager::s_pCurBufferPointer+native_stride, plastptr-native_stride*2, 2*native_stride);
+						ConvertVertices(1);
-						VertexManager::s_pCurBufferPointer += native_stride*3;
+						g_pVideoData -= m_VertexSize;
+						ConvertVertices(2);
 						extraverts = 3;
 					}
 					else {
-						memcpy_gc(VertexManager::s_pCurBufferPointer, plastptr-native_stride*2, native_stride*2);
+						g_pVideoData -= m_VertexSize*2;
-						VertexManager::s_pCurBufferPointer += native_stride*2;
+						ConvertVertices(2);
 						extraverts = 2;
 					}
 					break;
 				case 4: // tri fan, copy first and last vert
-					memcpy_gc(VertexManager::s_pCurBufferPointer, plastptr-native_stride*(v-startv+extraverts), native_stride);
+					g_pVideoData -= m_VertexSize*(v-startv+extraverts);
-					VertexManager::s_pCurBufferPointer += native_stride;
+					ConvertVertices(1);
-					memcpy_gc(VertexManager::s_pCurBufferPointer, plastptr-native_stride, native_stride);
+					g_pVideoData += m_VertexSize*(v-startv+extraverts-2);
-					VertexManager::s_pCurBufferPointer += native_stride;
+					ConvertVertices(1);
 					extraverts = 2;
 					break;
 				case 6: // line strip
-					memcpy_gc(VertexManager::s_pCurBufferPointer, plastptr-native_stride, native_stride);
+					g_pVideoData -= m_VertexSize*1;
-					VertexManager::s_pCurBufferPointer += native_stride;
+					ConvertVertices(1);
 					extraverts = 1;
 					break;
 				default:
@@ -615,13 +635,25 @@ void VertexLoader::RunVertices(int vtx_attr_group, int primitive, int count)
 		if (count - v < remainingVerts)
 			remainingVerts = count - v;
 
+		ConvertVertices(remainingVerts);
+
+		v += remainingVerts;
+	}
+
+	if (startv < count)
+		VertexManager::AddVertices(primitive, count - startv + extraverts);
+}
+
+
+void VertexLoader::ConvertVertices ( int count )
+{
 #ifdef USE_JIT
-		if (remainingVerts > 0) {
+	if (count > 0) {
-			loop_counter = remainingVerts;
+		loop_counter = count;
 		((void (*)())(void*)m_compiledCode)();
 	}
 #else
-		for (int s = 0; s < remainingVerts; s++)
+	for (int s = 0; s < count; s++)
 	{
 		tcIndex = 0;
 		colIndex = 0;
@@ -631,11 +663,6 @@ void VertexLoader::RunVertices(int vtx_attr_group, int primitive, int count)
 		PRIM_LOG("\n");
 	}
 #endif
-		v += remainingVerts;
-	}
-
-	if (startv < count)
-		VertexManager::AddVertices(primitive, count - startv + extraverts);
 }
 
 

Source/Core/VideoCommon/Src/VertexLoader.h

@@ -119,6 +119,7 @@ private:
 	void SetVAT(u32 _group0, u32 _group1, u32 _group2);
 
 	void CompileVertexTranslator();
+	void ConvertVertices(int count);
 
 	void WriteCall(TPipelineFunction);
 

Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp

@@ -31,6 +31,7 @@
 #include <tmmintrin.h>
 #endif
 
+// warning: mapping buffer should be disabled to use this
 #define LOG_NORM()  // PRIM_LOG("norm: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
 
 VertexLoader_Normal::Set VertexLoader_Normal::m_Table[NUM_NRM_TYPE][NUM_NRM_INDICES][NUM_NRM_ELEMENTS][NUM_NRM_FORMAT];

Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp

@@ -34,12 +34,14 @@ void LOG_TEX();
 template <>
 __forceinline void LOG_TEX<1>()
 {
+	// warning: mapping buffer should be disabled to use this
 	// PRIM_LOG("tex: %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0]);
 }
 
 template <>
 __forceinline void LOG_TEX<2>()
 {
+	// warning: mapping buffer should be disabled to use this
 	// PRIM_LOG("tex: %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0], ((float*)VertexManager::s_pCurBufferPointer)[1]);
 }
 
@@ -125,11 +127,9 @@ void LOADERDECL TexCoord_ReadIndex16_Float2_SSSE3()
 	const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
 	GC_ALIGNED128(const __m128i a = _mm_loadl_epi64((__m128i*)pData));
 	GC_ALIGNED128(const __m128i b = _mm_shuffle_epi8(a, kMaskSwap32));
-	u8* p = VertexManager::s_pCurBufferPointer;
+	_mm_storel_epi64((__m128i*)VertexManager::s_pCurBufferPointer, b);
-	_mm_storel_epi64((__m128i*)p, b);
 	LOG_TEX2();
-	p += 8;
+	VertexManager::s_pCurBufferPointer += 8;
-	VertexManager::s_pCurBufferPointer = p;
 	tcIndex++;
 }
 #endif

Source/Core/VideoCommon/Src/VideoCommon.h

@@ -90,7 +90,7 @@ struct TargetRectangle : public MathUtil::Rectangle<int>
 #define PRIM_LOG(...) DEBUG_LOG(VIDEO, ##__VA_ARGS__)
 #endif
 
-
+// warning: mapping buffer should be disabled to use this
 // #define LOG_VTX() DEBUG_LOG(VIDEO, "vtx: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[0], ((float*)VertexManager::s_pCurBufferPointer)[1], ((float*)VertexManager::s_pCurBufferPointer)[2]);
 
 #define LOG_VTX()