Aligned index buffers.

3 files changed, 29 insertions, 49 deletions

diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index a50eb7211c..0f3c900d2b 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -453,6 +453,7 @@ LLVertexBuffer::LLVertexBuffer(U32 typemask, S32 usage) :
 	mTypeMask = typemask;
 	mStride = stride;
 	mAlignedOffset = 0;
+	mAlignedIndexOffset = 0;
 
 	sCount++;
 }
@@ -642,16 +643,20 @@ void LLVertexBuffer::createGLIndices()
 
 	mEmpty = TRUE;
 
+	//pad by 16 bytes for aligned copies
+	size += 16;
+
 	if (useVBOs())
 	{
+		//pad by another 16 bytes for VBO pointer adjustment
+		size += 16;
 		mMappedIndexData = NULL;
 		genIndices();
 		mResized = TRUE;
 	}
 	else
 	{
-		mMappedIndexData = new U8[size];
-		memset(mMappedIndexData, 0, size);
+		mMappedIndexData = (U8*) _mm_malloc(size, 16);
 		static int gl_buffer_idx = 0;
 		mGLIndices = ++gl_buffer_idx;
 	}
@@ -699,7 +704,7 @@ void LLVertexBuffer::destroyGLIndices()
 		}
 		else
 		{
-			delete [] mMappedIndexData;
+			_mm_free(mMappedIndexData);
 			mMappedIndexData = NULL;
 			mEmpty = TRUE;
 		}
@@ -836,26 +841,10 @@ void LLVertexBuffer::resizeBuffer(S32 newnverts, S32 newnindices)
 			}
 			else
 			{
-				//delete old buffer, keep GL buffer for now
 				if (!useVBOs())
 				{
-					U8* old = mMappedData;
-					mMappedData = new U8[newsize];
-					if (old)
-					{	
-						memcpy(mMappedData, old, llmin(newsize, oldsize));
-						if (newsize > oldsize)
-						{
-							memset(mMappedData+oldsize, 0, newsize-oldsize);
-						}
-
-						delete [] old;
-					}
-					else
-					{
-						memset(mMappedData, 0, newsize);
-						mEmpty = TRUE;
-					}
+					_mm_free(mMappedData);
+					mMappedData = (U8*) _mm_malloc(newsize, 16);
 				}
 				mResized = TRUE;
 			}
@@ -875,24 +864,8 @@ void LLVertexBuffer::resizeBuffer(S32 newnverts, S32 newnindices)
 			{
 				if (!useVBOs())
 				{
-					//delete old buffer, keep GL buffer for now
-					U8* old = mMappedIndexData;
-					mMappedIndexData = new U8[new_index_size];
-					
-					if (old)
-					{	
-						memcpy(mMappedIndexData, old, llmin(new_index_size, old_index_size));
-						if (new_index_size > old_index_size)
-						{
-							memset(mMappedIndexData+old_index_size, 0, new_index_size - old_index_size);
-						}
-						delete [] old;
-					}
-					else
-					{
-						memset(mMappedIndexData, 0, new_index_size);
-						mEmpty = TRUE;
-					}
+					_mm_free(mMappedIndexData);
+					mMappedIndexData = (U8*) _mm_malloc(new_index_size, 16);
 				}
 				mResized = TRUE;
 			}
@@ -958,7 +931,10 @@ U8* LLVertexBuffer::mapBuffer(S32 access)
 		}
 		{
 			LLMemType mt_v(LLMemType::MTYPE_VERTEX_MAP_BUFFER_INDICES);
-			mMappedIndexData = (U8*) glMapBufferARB(GL_ELEMENT_ARRAY_BUFFER_ARB, GL_WRITE_ONLY_ARB);
+			U8* src = (U8*) glMapBufferARB(GL_ELEMENT_ARRAY_BUFFER_ARB, GL_WRITE_ONLY_ARB);
+			mMappedIndexData = LL_NEXT_ALIGNED_ADDRESS<U8>(src);
+			mAlignedIndexOffset = mMappedIndexData - src;
+
 			stop_glerror();
 		}
 
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index 03799af978..c6fd0a9e3c 100644
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -187,7 +187,7 @@ public:
 	S32 getRequestedVerts() const			{ return mRequestedNumVerts; }
 	S32 getRequestedIndices() const			{ return mRequestedNumIndices; }
 
-	U8* getIndicesPointer() const			{ return useVBOs() ? NULL : mMappedIndexData; }
+	U8* getIndicesPointer() const			{ return useVBOs() ? (U8*) mAlignedIndexOffset : mMappedIndexData; }
 	U8* getVerticesPointer() const			{ return useVBOs() ? NULL : mMappedData; }
 	S32 getStride() const					{ return mStride; }
 	U32 getTypeMask() const					{ return mTypeMask; }
@@ -214,6 +214,7 @@ protected:
 	S32		mRequestedNumIndices;  // Number of indices requested
 
 	ptrdiff_t mAlignedOffset;
+	ptrdiff_t mAlignedIndexOffset;
 	S32		mStride;
 	U32		mTypeMask;
 	S32		mUsage;			// GL usage
diff --git a/indra/newview/lldrawpoolavatar.cpp b/indra/newview/lldrawpoolavatar.cpp
index 1e9053239d..4fb8f5266e 100644
--- a/indra/newview/lldrawpoolavatar.cpp
+++ b/indra/newview/lldrawpoolavatar.cpp
@@ -1542,27 +1542,30 @@ LLVertexBufferAvatar::LLVertexBufferAvatar()
 
 void LLVertexBufferAvatar::setupVertexBuffer(U32 data_mask) const
 {
-/*	if (sRenderingSkinned)
+	if (sRenderingSkinned)
 	{
 		U8* base = useVBOs() ? NULL : mMappedData;
 
-		glVertexPointer(3,GL_FLOAT, mStride, (void*)(base + 0));
-		glNormalPointer(GL_FLOAT, mStride, (void*)(base + mOffsets[TYPE_NORMAL]));
-		glTexCoordPointer(2,GL_FLOAT, mStride, (void*)(base + mOffsets[TYPE_TEXCOORD0]));
+		glVertexPointer(3,GL_FLOAT, LLVertexBuffer::sTypeOffsets[LLVertexBuffer::TYPE_VERTEX], (void*)(base + 0));
+		glNormalPointer(GL_FLOAT, LLVertexBuffer::sTypeOffsets[LLVertexBuffer::TYPE_NORMAL], (void*)(base + mOffsets[TYPE_NORMAL]));
+		glTexCoordPointer(2,GL_FLOAT, LLVertexBuffer::sTypeOffsets[LLVertexBuffer::TYPE_TEXCOORD0], (void*)(base + mOffsets[TYPE_TEXCOORD0]));
 		
-		set_vertex_weights(LLDrawPoolAvatar::sVertexProgram->mAttribute[LLViewerShaderMgr::AVATAR_WEIGHT], mStride, (F32*)(base + mOffsets[TYPE_WEIGHT]));
+		set_vertex_weights(LLDrawPoolAvatar::sVertexProgram->mAttribute[LLViewerShaderMgr::AVATAR_WEIGHT], 
+						LLVertexBuffer::sTypeOffsets[LLVertexBuffer::TYPE_WEIGHT], (F32*)(base + mOffsets[TYPE_WEIGHT]));
 
 		if (sShaderLevel >= LLDrawPoolAvatar::SHADER_LEVEL_BUMP)
 		{
-			set_binormals(LLDrawPoolAvatar::sVertexProgram->mAttribute[LLViewerShaderMgr::BINORMAL], mStride, (LLVector3*)(base + mOffsets[TYPE_BINORMAL]));
+			set_binormals(LLDrawPoolAvatar::sVertexProgram->mAttribute[LLViewerShaderMgr::BINORMAL],
+				LLVertexBuffer::sTypeOffsets[LLVertexBuffer::TYPE_BINORMAL], (LLVector3*)(base + mOffsets[TYPE_BINORMAL]));
 		}
 	
 		if (sShaderLevel >= LLDrawPoolAvatar::SHADER_LEVEL_CLOTH)
 		{
-			set_vertex_clothing_weights(LLDrawPoolAvatar::sVertexProgram->mAttribute[LLViewerShaderMgr::AVATAR_CLOTHING], mStride, (LLVector4*)(base + mOffsets[TYPE_CLOTHWEIGHT]));
+			set_vertex_clothing_weights(LLDrawPoolAvatar::sVertexProgram->mAttribute[LLViewerShaderMgr::AVATAR_CLOTHING], 
+				LLVertexBuffer::sTypeOffsets[LLVertexBuffer::TYPE_CLOTHWEIGHT], (LLVector4*)(base + mOffsets[TYPE_CLOTHWEIGHT]));
 		}
 	}
-	else*/
+	else
 	{
 		LLVertexBuffer::setupVertexBuffer(data_mask);
 	}