SH-1682 Use GL_ARB_map_buffer_range to mitigate impact of mapping larger vertex buffer objects. Limit number of textures per batch to 6 (prevents frame stalls on NVIDIA).

7 files changed, 118 insertions, 88 deletions

diff --git a/indra/newview/app_settings/settings.xml b/indra/newview/app_settings/settings.xml
index bed7e4705e..b7ba9e3607 100644
--- a/indra/newview/app_settings/settings.xml
+++ b/indra/newview/app_settings/settings.xml
@@ -7126,7 +7126,7 @@
     <string>Vector3</string>
     <key>Value</key>
     <array>
-      <real>0.05</real>
+      <real>0.1</real>
       <real>0.0</real>
       <real>0.0</real>
     </array>
@@ -7154,7 +7154,7 @@
     <string>Vector3</string>
     <key>Value</key>
     <array>
-      <real>0.25</real>
+      <real>0.01</real>
       <real>0.0</real>
       <real>0.0</real>
     </array>
@@ -7513,6 +7513,17 @@
       <key>Value</key>
       <integer>0</integer>
     </map>
+  <key>RenderMaxTextureIndex</key>
+  <map>
+    <key>Comment</key>
+    <string>Maximum texture index to use for indexed texture rendering.</string>
+    <key>Persist</key>
+    <integer>1</integer>
+    <key>Type</key>
+    <string>U32</string>
+    <key>Value</key>
+    <integer>6</integer>
+  </map>
     <key>RenderDebugTextureBind</key>
     <map>
       <key>Comment</key>
@@ -9011,7 +9022,7 @@
       <key>Type</key>
       <string>Boolean</string>
       <key>Value</key>
-      <integer>1</integer>
+      <integer>0</integer>
     </map>
   <key>RenderUseStreamVBO</key>
   <map>
diff --git a/indra/newview/lldrawpool.cpp b/indra/newview/lldrawpool.cpp
index d92d9e92a1..f5483d969d 100644
--- a/indra/newview/lldrawpool.cpp
+++ b/indra/newview/lldrawpool.cpp
@@ -191,7 +191,7 @@ void LLDrawPool::renderPostDeferred(S32 pass)
 //virtual
 void LLDrawPool::endRenderPass( S32 pass )
 {
-	for (U32 i = 1; i < gGLManager.mNumTextureImageUnits; i++)
+	for (U32 i = 0; i < gGLManager.mNumTextureImageUnits; i++)
 	{ //dummy cleanup of any currently bound textures
 		if (gGL.getTexUnit(i)->getCurrType() != LLTexUnit::TT_NONE)
 		{
diff --git a/indra/newview/lldrawpoolsimple.cpp b/indra/newview/lldrawpoolsimple.cpp
index 6963fea26b..5dbb27cabb 100644
--- a/indra/newview/lldrawpoolsimple.cpp
+++ b/indra/newview/lldrawpoolsimple.cpp
@@ -71,14 +71,7 @@ void LLDrawPoolGlow::renderPostDeferred(S32 pass)
 void LLDrawPoolGlow::endPostDeferredPass(S32 pass)
 {
 	gDeferredFullbrightProgram.unbind();
-	for (U32 i = 0; i < 8; i++)
-	{
-		if (gGL.getTexUnit(i)->getCurrType() != LLTexUnit::TT_NONE)
-		{
-			gGL.getTexUnit(i)->unbind(gGL.getTexUnit(i)->getCurrType());
-			gGL.getTexUnit(i)->disable();
-		}
-	}
+	LLRenderPass::endRenderPass(pass);
 }
 
 void LLDrawPoolGlow::render(S32 pass)
@@ -225,15 +218,6 @@ void LLDrawPoolSimple::endDeferredPass(S32 pass)
 	LLRenderPass::endRenderPass(pass);
 
 	gDeferredDiffuseProgram.unbind();
-
-	for (U32 i = 0; i < 8; i++)
-	{
-		if (gGL.getTexUnit(i)->getCurrType() != LLTexUnit::TT_NONE)
-		{
-			gGL.getTexUnit(i)->unbind(gGL.getTexUnit(i)->getCurrType());
-			gGL.getTexUnit(i)->disable();
-		}
-	}
 }
 
 void LLDrawPoolSimple::renderDeferred(S32 pass)
@@ -368,14 +352,7 @@ void LLDrawPoolFullbright::renderPostDeferred(S32 pass)
 void LLDrawPoolFullbright::endPostDeferredPass(S32 pass)
 {
 	gDeferredFullbrightProgram.unbind();
-	for (U32 i = 0; i < 8; i++)
-	{
-		if (gGL.getTexUnit(i)->getCurrType() != LLTexUnit::TT_NONE)
-		{
-			gGL.getTexUnit(i)->unbind(gGL.getTexUnit(i)->getCurrType());
-			gGL.getTexUnit(i)->disable();
-		}
-	}
+	LLRenderPass::endRenderPass(pass);
 }
 
 void LLDrawPoolFullbright::beginRenderPass(S32 pass)
diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp
index e30522d380..540ed054e9 100644
--- a/indra/newview/llface.cpp
+++ b/indra/newview/llface.cpp
@@ -429,11 +429,11 @@ U16 LLFace::getGeometryAvatar(
 
 	if (mVertexBuffer.notNull())
 	{
-		mVertexBuffer->getVertexStrider      (vertices, mGeomIndex);
-		mVertexBuffer->getNormalStrider      (normals, mGeomIndex);
-		mVertexBuffer->getTexCoord0Strider    (tex_coords, mGeomIndex);
-		mVertexBuffer->getWeightStrider(vertex_weights, mGeomIndex);
-		mVertexBuffer->getClothWeightStrider(clothing_weights, mGeomIndex);
+		mVertexBuffer->getVertexStrider      (vertices, mGeomIndex, mGeomCount);
+		mVertexBuffer->getNormalStrider      (normals, mGeomIndex, mGeomCount);
+		mVertexBuffer->getTexCoord0Strider    (tex_coords, mGeomIndex, mGeomCount);
+		mVertexBuffer->getWeightStrider(vertex_weights, mGeomIndex, mGeomCount);
+		mVertexBuffer->getClothWeightStrider(clothing_weights, mGeomIndex, mGeomCount);
 	}
 
 	return mGeomIndex;
@@ -446,17 +446,17 @@ U16 LLFace::getGeometry(LLStrider<LLVector3> &vertices, LLStrider<LLVector3> &no
 	
 	if (mVertexBuffer.notNull())
 	{
-		mVertexBuffer->getVertexStrider(vertices,   mGeomIndex);
+		mVertexBuffer->getVertexStrider(vertices,   mGeomIndex, mGeomCount);
 		if (mVertexBuffer->hasDataType(LLVertexBuffer::TYPE_NORMAL))
 		{
-			mVertexBuffer->getNormalStrider(normals,    mGeomIndex);
+			mVertexBuffer->getNormalStrider(normals,    mGeomIndex, mGeomCount);
 		}
 		if (mVertexBuffer->hasDataType(LLVertexBuffer::TYPE_TEXCOORD0))
 		{
-			mVertexBuffer->getTexCoord0Strider(tex_coords, mGeomIndex);
+			mVertexBuffer->getTexCoord0Strider(tex_coords, mGeomIndex, mGeomCount);
 		}
 
-		mVertexBuffer->getIndexStrider(indicesp, mIndicesIndex);
+		mVertexBuffer->getIndexStrider(indicesp, mIndicesIndex, mIndicesCount);
 	}
 	
 	return mGeomIndex;
@@ -1092,27 +1092,6 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 	const LLTextureEntry *tep = mVObjp->getTE(f);
 	const U8 bump_code = tep ? tep->getBumpmap() : 0;
 
-	if (rebuild_pos)
-	{
-		mVertexBuffer->getVertexStrider(vert, mGeomIndex);
-		vertices = (LLVector4a*) vert.get();
-	}
-	if (rebuild_normal)
-	{
-		mVertexBuffer->getNormalStrider(norm, mGeomIndex);
-		normals = (LLVector4a*) norm.get();
-	}
-	if (rebuild_binormal)
-	{
-		mVertexBuffer->getBinormalStrider(binorm, mGeomIndex);
-		binormals = (LLVector4a*) binorm.get();
-	}
-	if (rebuild_weights)
-	{
-		mVertexBuffer->getWeight4Strider(wght, mGeomIndex);
-		weights = (LLVector4a*) wght.get();
-	}
-	
 	F32 tcoord_xoffset = 0.f ;
 	F32 tcoord_yoffset = 0.f ;
 	F32 tcoord_xscale = 1.f ;
@@ -1121,12 +1100,6 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 	if (rebuild_tcoord)
 	{
-		mVertexBuffer->getTexCoord0Strider(tex_coords, mGeomIndex);
-		if (bump_code && mVertexBuffer->hasDataType(LLVertexBuffer::TYPE_TEXCOORD1))
-		{
-			mVertexBuffer->getTexCoord1Strider(tex_coords2, mGeomIndex);
-		}
-
 		in_atlas = isAtlasInUse() ;
 		if(in_atlas)
 		{
@@ -1139,11 +1112,7 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 			tcoord_yscale = tmp->mV[1] ;	
 		}
 	}
-	if (rebuild_color)
-	{	
-		mVertexBuffer->getColorStrider(colors, mGeomIndex);
-	}
-
+	
 	BOOL is_static = mDrawablep->isStatic();
 	BOOL is_global = is_static;
 
@@ -1182,7 +1151,8 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 	// INDICES
 	if (full_rebuild)
 	{
-		mVertexBuffer->getIndexStrider(indicesp, mIndicesIndex);
+		mVertexBuffer->getIndexStrider(indicesp, mIndicesIndex, mIndicesCount, true);
+
 		__m128i* dst = (__m128i*) indicesp.get();
 		__m128i* src = (__m128i*) vf.mIndices;
 		__m128i offset = _mm_set1_epi16(index_offset);
@@ -1199,6 +1169,8 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 		{
 			indicesp[i] = vf.mIndices[i]+index_offset;
 		}
+
+		mVertexBuffer->setBuffer(0);
 	}
 	
 	LLMatrix4a mat_normal;
@@ -1344,6 +1316,8 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 
 		if (!in_atlas && !do_bump)
 		{ //not in atlas or not bump mapped, might be able to do a cheap update
+			mVertexBuffer->getTexCoord0Strider(tex_coords, mGeomIndex, mGeomCount);
+
 			if (texgen != LLTextureEntry::TEX_GEN_PLANAR)
 			{
 				if (!do_tex_mat)
@@ -1416,9 +1390,15 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 					}
 				}
 			}
+
+			mVertexBuffer->setBuffer(0);
 		}
 		else
 		{ //either bump mapped or in atlas, just do the whole expensive loop
+			mVertexBuffer->getTexCoord0Strider(tex_coords, mGeomIndex, mGeomCount, true);
+
+			std::vector<LLVector2> bump_tc;
+		
 			for (S32 i = 0; i < num_vertices; i++)
 			{	
 				LLVector2 tc(vf.mTexCoords[i]);
@@ -1549,8 +1529,20 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 				
 
 				*tex_coords++ = tc;
-				
-				if (bump_code && mVertexBuffer->hasDataType(LLVertexBuffer::TYPE_TEXCOORD1))
+				if (do_bump)
+				{
+					bump_tc.push_back(tc);
+				}
+			}
+
+			mVertexBuffer->setBuffer(0);
+
+
+			if (do_bump)
+			{
+				mVertexBuffer->getTexCoord1Strider(tex_coords2, mGeomIndex, mGeomCount, true);
+		
+				for (S32 i = 0; i < num_vertices; i++)
 				{
 					LLVector4a tangent;
 					tangent.setCross3(vf.mBinormals[i], vf.mNormals[i]);
@@ -1572,16 +1564,22 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 					}
 
 					binormal.normalize3fast();
+					LLVector2 tc = bump_tc[i];
 					tc += LLVector2( bump_s_primary_light_ray.dot3(tangent).getF32(), bump_t_primary_light_ray.dot3(binormal).getF32() );
 					
 					*tex_coords2++ = tc;
-				}	
+				}
+
+				mVertexBuffer->setBuffer(0);
 			}
 		}
 	}
 
 	if (rebuild_pos)
 	{
+		mVertexBuffer->getVertexStrider(vert, mGeomIndex, mGeomCount, true);
+		vertices = (LLVector4a*) vert.get();
+	
 		LLMatrix4a mat_vert;
 		mat_vert.loadu(mat_vert_in);
 
@@ -1607,11 +1605,15 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 			index_dst += 4;
 		}
 		while (index_dst < index_end);
-		
+
+		mVertexBuffer->setBuffer(0);
 	}
 		
 	if (rebuild_normal)
 	{
+		mVertexBuffer->getNormalStrider(norm, mGeomIndex, mGeomCount, true);
+		normals = (LLVector4a*) norm.get();
+	
 		for (S32 i = 0; i < num_vertices; i++)
 		{	
 			LLVector4a normal;
@@ -1619,10 +1621,15 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 			normal.normalize3fast();
 			normals[i] = normal;
 		}
+
+		mVertexBuffer->setBuffer(0);
 	}
 		
 	if (rebuild_binormal)
 	{
+		mVertexBuffer->getBinormalStrider(binorm, mGeomIndex, mGeomCount, true);
+		binormals = (LLVector4a*) binorm.get();
+		
 		for (S32 i = 0; i < num_vertices; i++)
 		{	
 			LLVector4a binormal;
@@ -1630,15 +1637,22 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 			binormal.normalize3fast();
 			binormals[i] = binormal;
 		}
+
+		mVertexBuffer->setBuffer(0);
 	}
 	
 	if (rebuild_weights && vf.mWeights)
 	{
+		mVertexBuffer->getWeight4Strider(wght, mGeomIndex, mGeomCount, true);
+		weights = (LLVector4a*) wght.get();
 		LLVector4a::memcpyNonAliased16((F32*) weights, (F32*) vf.mWeights, num_vertices*4*sizeof(F32));
+		mVertexBuffer->setBuffer(0);
 	}
 
 	if (rebuild_color)
 	{
+		mVertexBuffer->getColorStrider(colors, mGeomIndex, mGeomCount, true);
+
 		LLVector4a src;
 
 		U32 vec[4];
@@ -1657,6 +1671,8 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
 		{	
 			dst[i] = src;
 		}
+
+		mVertexBuffer->setBuffer(0);
 	}
 
 	if (rebuild_tcoord)
@@ -2073,13 +2089,13 @@ S32 LLFace::getColors(LLStrider<LLColor4U> &colors)
 	}
 	
 	// llassert(mGeomIndex >= 0);
-	mVertexBuffer->getColorStrider(colors, mGeomIndex);
+	mVertexBuffer->getColorStrider(colors, mGeomIndex, mGeomCount);
 	return mGeomIndex;
 }
 
 S32	LLFace::getIndices(LLStrider<U16> &indicesp)
 {
-	mVertexBuffer->getIndexStrider(indicesp, mIndicesIndex);
+	mVertexBuffer->getIndexStrider(indicesp, mIndicesIndex, mIndicesCount);
 	llassert(indicesp[0] != indicesp[1]);
 	return mIndicesIndex;
 }
diff --git a/indra/newview/llspatialpartition.h b/indra/newview/llspatialpartition.h
index ae5d4fa463..db8a0c2992 100644
--- a/indra/newview/llspatialpartition.h
+++ b/indra/newview/llspatialpartition.h
@@ -209,7 +209,7 @@ public:
 	typedef std::vector<LLPointer<LLDrawInfo> > drawmap_elem_t; 
 	typedef std::map<U32, drawmap_elem_t > draw_map_t;	
 	typedef std::vector<LLPointer<LLVertexBuffer> > buffer_list_t;
-	typedef std::map<LLPointer<LLViewerTexture>, buffer_list_t> buffer_texture_map_t;
+	typedef std::map<LLFace*, buffer_list_t> buffer_texture_map_t;
 	typedef std::map<U32, buffer_texture_map_t> buffer_map_t;
 
 	typedef LLOctreeListener<LLDrawable>	BaseType;
@@ -401,7 +401,7 @@ protected:
 
 public:
 	bridge_list_t mBridgeList;
-	buffer_map_t mBufferMap; //used by volume buffers to store unique buffers per texture
+	buffer_map_t mBufferMap; //used by volume buffers to attempt to reuse vertex buffers
 
 	F32 mBuilt;
 	OctreeNode* mOctreeNode;
diff --git a/indra/newview/llviewercontrol.cpp b/indra/newview/llviewercontrol.cpp
index 61200db2b9..87ca80260f 100644
--- a/indra/newview/llviewercontrol.cpp
+++ b/indra/newview/llviewercontrol.cpp
@@ -591,6 +591,7 @@ void settings_setup_listeners()
 	gSavedSettings.getControl("OctreeMaxNodeCapacity")->getSignal()->connect(boost::bind(&handleRepartition, _2));
 	gSavedSettings.getControl("OctreeAlphaDistanceFactor")->getSignal()->connect(boost::bind(&handleRepartition, _2));
 	gSavedSettings.getControl("OctreeAttachmentSizeFactor")->getSignal()->connect(boost::bind(&handleRepartition, _2));
+	gSavedSettings.getControl("RenderMaxTextureIndex")->getSignal()->connect(boost::bind(&handleResetVertexBuffersChanged, _2));
 	gSavedSettings.getControl("RenderUseTriStrips")->getSignal()->connect(boost::bind(&handleResetVertexBuffersChanged, _2));
 	gSavedSettings.getControl("RenderAnimateTrees")->getSignal()->connect(boost::bind(&handleResetVertexBuffersChanged, _2));
 	gSavedSettings.getControl("RenderAvatarVP")->getSignal()->connect(boost::bind(&handleSetShaderChanged, _2));
diff --git a/indra/newview/llvovolume.cpp b/indra/newview/llvovolume.cpp
index 125283d474..58e38ea9b1 100644
--- a/indra/newview/llvovolume.cpp
+++ b/indra/newview/llvovolume.cpp
@@ -4426,6 +4426,33 @@ void LLVolumeGeometryManager::rebuildMesh(LLSpatialGroup* group)
 	llassert(!group || !group->isState(LLSpatialGroup::NEW_DRAWINFO));
 }
 
+struct CompareBatchBreakerModified
+{
+	bool operator()(const LLFace* const& lhs, const LLFace* const& rhs)
+	{
+		const LLTextureEntry* lte = lhs->getTextureEntry();
+		const LLTextureEntry* rte = rhs->getTextureEntry();
+
+		if (lte->getBumpmap() != rte->getBumpmap())
+		{
+			return lte->getBumpmap() < rte->getBumpmap();
+		}
+		else if (lte->getFullbright() != rte->getFullbright())
+		{
+			return lte->getFullbright() < rte->getFullbright();
+		}
+		else  if (lte->getGlow() != rte->getGlow())
+		{
+			return lte->getGlow() < rte->getGlow();
+		}
+		else
+		{
+			return lhs->getTexture() < rhs->getTexture();
+		}
+		
+	}
+};
+
 void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::vector<LLFace*>& faces, BOOL distance_sort, BOOL batch_textures)
 {
 	//calculate maximum number of vertices to store in a single buffer
@@ -4435,7 +4462,7 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 	if (!distance_sort)
 	{
 		//sort faces by things that break batches
-		std::sort(faces.begin(), faces.end(), LLFace::CompareBatchBreaker());
+		std::sort(faces.begin(), faces.end(), CompareBatchBreakerModified());
 	}
 	else
 	{
@@ -4462,6 +4489,8 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 		texture_index_channels = gDeferredAlphaProgram.mFeatures.mIndexedTextureChannels;
 	}
 
+	texture_index_channels = llmin(texture_index_channels, (S32) gSavedSettings.getU32("RenderMaxTextureIndex"));
+	
 
 	while (face_iter != faces.end())
 	{
@@ -4587,7 +4616,7 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 	
 		//create/delete/resize vertex buffer if needed
 		LLVertexBuffer* buffer = NULL;
-		LLSpatialGroup::buffer_texture_map_t::iterator found_iter = group->mBufferMap[mask].find(tex);
+		LLSpatialGroup::buffer_texture_map_t::iterator found_iter = group->mBufferMap[mask].find(*face_iter);
 		
 		if (found_iter != group->mBufferMap[mask].end())
 		{
@@ -4618,7 +4647,7 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 			}
 		}
 
-		buffer_map[mask][tex].push_back(buffer);
+		buffer_map[mask][*face_iter].push_back(buffer);
 
 		//add face geometry
 
@@ -4649,12 +4678,8 @@ void LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, std::
 
 					U32 te_idx = facep->getTEOffset();
 
-					if (facep->getGeometryVolume(*volume, te_idx, 
-						vobj->getRelativeXform(), vobj->getRelativeXformInvTrans(), index_offset))
-					{
-						buffer->markDirty(facep->getGeomIndex(), facep->getGeomCount(), 
-							facep->getIndicesStart(), facep->getIndicesCount());
-					}
+					facep->getGeometryVolume(*volume, te_idx, 
+						vobj->getRelativeXform(), vobj->getRelativeXformInvTrans(), index_offset);
 				}
 			}