Merge pull request #195 from MaSzyna-EU07/fix/instancing

Improve instancing rendering speed

model/AnimModel.cpp

@@ -403,8 +403,8 @@ namespace {
 // TSubModel tree, which would make batched rendering unsafe. Most animations
 // loaded from .t3d files are global functions of time (clocks, wind, sky) and
 // only transform the local modelview matrix — those are safe to share. Camera-
-// relative billboards also operate purely on the local matrix using whatever
+// relative billboards are not: they're computed on the cpu from the modelview of
-// modelview the caller pushed, which is exactly per-instance behaviour.
+// the drawn instance, which in a batch is only the submodel-local chain.
 // The runtime SetRotate/SetTranslate animations (at_Rotate / at_RotateXYZ /
 // at_Translate) are tied to per-instance iAnimOwner and are unsafe to share.
 // at_Undefined is the type assigned to .t3d submodels declared with `anim: true`
@@ -425,17 +425,40 @@ bool anim_type_unsafe_for_instancing( TAnimType a ) {
     }
 }
 
+// returns true if this animation type depends only on global state (simulation
+// time, wind, sky) and so gives every instance of the model the same local
+// transform. The batch evaluates it once and the per-instance placement comes
+// from the instance matrix, the result is the same as with per-instance drawing.
+bool anim_type_shared_by_instances( TAnimType a ) {
+    switch( a ) {
+    case TAnimType::at_SecondsJump:
+    case TAnimType::at_MinutesJump:
+    case TAnimType::at_HoursJump:
+    case TAnimType::at_Hours24Jump:
+    case TAnimType::at_Seconds:
+    case TAnimType::at_Minutes:
+    case TAnimType::at_Hours:
+    case TAnimType::at_Hours24:
+    case TAnimType::at_Wind:
+    case TAnimType::at_Sky:
+        return true;
+    default:
+        return false;
+    }
+}
+
 // recursively walks a submodel tree and returns true if any submodel declares
 // an animation type that's unsafe to batch, OR carries the runtime "needs
-// animation matrix" flag (iFlags bit 0x4000), which is set whenever the
+// animation matrix" flag (iFlags bit 0x4000) for anything else than an animation
-// submodel was tagged as animatable in the .t3d file or had WillBeAnimated()
+// shared by all instances. The flag is set whenever the submodel was tagged as
-// called on it during model load. Either signal means the submodel may receive
+// animatable in the .t3d file (any `anim:` other than `false`) or had
+// WillBeAnimated() called on it during model load. Such a submodel may receive
 // per-instance event-driven animation commands at runtime, which the GPU-
 // instanced path (one shared submodel tree across all instances) cannot serve.
 bool submodel_tree_blocks_instancing( TSubModel const *Sub ) {
     if( Sub == nullptr ) { return false; }
     if( anim_type_unsafe_for_instancing( Sub->b_Anim ) ) { return true; }
-    if( ( Sub->iFlags & 0x4000 ) != 0 ) { return true; }
+    if( ( ( Sub->iFlags & 0x4000 ) != 0 ) && ( false == anim_type_shared_by_instances( Sub->b_Anim ) ) ) { return true; }
     if( submodel_tree_blocks_instancing( Sub->Child ) ) { return true; }
     if( submodel_tree_blocks_instancing( Sub->Next ) ) { return true; }
     return false;
@@ -478,6 +501,9 @@ std::shared_ptr<TAnimContainer> TAnimModel::AddContainer(std::string const &Name
 		auto tmp = std::make_shared<TAnimContainer>();
         tmp->Init(tsb);
 		m_animlist.push_back(tmp);
+		// containers are added by events after the instance was loaded and bucketed;
+		// the instanced path skips RaAnimate(), so this instance has to be drawn on its own
+		m_instanceable = false;
 		return tmp;
     }
 	return nullptr;
@@ -638,6 +664,23 @@ void TAnimModel::RaPrepare()
 	}
 }
 
+glm::mat4 const &TAnimModel::rotation_scale() {
+
+    if( ( false == m_rotationscalevalid ) || ( vAngle != m_rotationscaleangles ) || ( m_scale != m_rotationscalescale ) ) {
+        // same order as the per-instance render path
+        glm::mat4 transform( 1.0f );
+        if( vAngle.y != 0.0f ) { transform = glm::rotate( transform, glm::radians( vAngle.y ), glm::vec3( 0.f, 1.f, 0.f ) ); }
+        if( vAngle.x != 0.0f ) { transform = glm::rotate( transform, glm::radians( vAngle.x ), glm::vec3( 1.f, 0.f, 0.f ) ); }
+        if( vAngle.z != 0.0f ) { transform = glm::rotate( transform, glm::radians( vAngle.z ), glm::vec3( 0.f, 0.f, 1.f ) ); }
+        if( m_scale != glm::vec3( 1.0f ) ) { transform = glm::scale( transform, m_scale ); }
+        m_rotationscale = transform;
+        m_rotationscaleangles = vAngle;
+        m_rotationscalescale = m_scale;
+        m_rotationscalevalid = true;
+    }
+    return m_rotationscale;
+}
+
 int TAnimModel::Flags()
 { // informacja dla TGround, czy ma być w Render, RenderAlpha, czy RenderMixed
     int i = pModel ? pModel->Flags() : 0; // pobranie flag całego modelu
@@ -688,6 +731,8 @@ std::optional<std::tuple<float, float, std::optional<glm::vec3>> > TAnimModel::L
 void TAnimModel::SkinSet( int const Index, material_handle const Material ) {
 
     m_materialdata.replacable_skins[ std::clamp( Index, 1, 4 ) ] = Material;
+    // the instance stays in the bucket of its old skin set, where it would get the textures of the batch
+    m_instanceable = false;
 }
 
 void TAnimModel::AnimUpdate(double dt)

model/AnimModel.h

@@ -172,6 +172,8 @@ public:
 public:
 // methods
     void RaPrepare(); // ustawienie animacji egzemplarza na wzorcu
+    // rotation and scale of the instance as a matrix: rotate y * rotate x * rotate z * scale
+    glm::mat4 const &rotation_scale();
     void RaAnimate( unsigned int const Framestamp ); // przeliczenie animacji egzemplarza
 
     // radius() subclass details, calculates node's bounding radius
@@ -190,6 +192,12 @@ public:
     TModel3d *pModel { nullptr };
     glm::vec3 vAngle; // bazowe obroty egzemplarza względem osi
     glm::vec3 m_scale { 1.0f, 1.0f, 1.0f }; // per-axis scale (see Scale() accessors above)
+    // rotation and scale part of the instance transform, cached for the instanced render path.
+    // rebuilt by rotation_scale() whenever the angles or the scale differ from the ones it was built from
+    glm::mat4 m_rotationscale { 1.0f };
+    glm::vec3 m_rotationscaleangles { 0.0f, 0.0f, 0.0f };
+    glm::vec3 m_rotationscalescale { 1.0f, 1.0f, 1.0f };
+    bool m_rotationscalevalid { false };
     material_data m_materialdata;
 
     std::string asText; // tekst dla wyświetlacza znakowego

rendering/opengl33geometrybank.cpp

@@ -215,13 +215,34 @@ opengl33_vaogeometrybank::draw_instanced_( gfx::geometry_handle const &Geometry,
     GLsizei const inst_count = static_cast<GLsizei>(InstanceCount);
 
     if( chunkrecord.index_count > 0 ) {
-        m_vao->bind();
+        // GLES 3.0 context doesn't provide the base vertex variant, same fallbacks as in draw_()
-        ::glDrawElementsInstancedBaseVertex(
+        if( glDrawElementsInstancedBaseVertex ) {
-            chunk.type,
+            m_vao->bind();
-            chunkrecord.index_count, GL_UNSIGNED_INT,
+            ::glDrawElementsInstancedBaseVertex(
-            reinterpret_cast<void const *>( chunkrecord.index_offset * sizeof( gfx::basic_index ) ),
+                chunk.type,
-            inst_count,
+                chunkrecord.index_count, GL_UNSIGNED_INT,
-            chunkrecord.vertex_offset );
+                reinterpret_cast<void const *>( chunkrecord.index_offset * sizeof( gfx::basic_index ) ),
+                inst_count,
+                chunkrecord.vertex_offset );
+        }
+        else if( glDrawElementsInstancedBaseVertexOES ) {
+            m_vao->bind();
+            ::glDrawElementsInstancedBaseVertexOES(
+                chunk.type,
+                chunkrecord.index_count, GL_UNSIGNED_INT,
+                reinterpret_cast<void const *>( chunkrecord.index_offset * sizeof( gfx::basic_index ) ),
+                inst_count,
+                chunkrecord.vertex_offset );
+        }
+        else {
+            setup_attrib( chunkrecord.vertex_offset );
+            m_vao->bind();
+            ::glDrawElementsInstanced(
+                chunk.type,
+                chunkrecord.index_count, GL_UNSIGNED_INT,
+                reinterpret_cast<void const *>( chunkrecord.index_offset * sizeof( gfx::basic_index ) ),
+                inst_count );
+        }
     }
     else {
         m_vao->bind();

rendering/opengl33renderer.cpp

@@ -2382,8 +2382,11 @@ void opengl33_renderer::Render(scene::basic_region *Region)
 
 	m_sectionqueue.clear();
 	m_cellqueue.clear();
-	// discard last pass's accumulated instance buckets before this pass starts
+	// discard last pass's accumulated instance buckets before this pass starts.
-	m_frame_instance_buckets.clear();
+	// the vectors themselves are kept, so their allocations get reused by the following passes
+	for( auto &bucket : m_frame_instance_buckets ) {
+		bucket.second.clear();
+	}
 	// build a list of region sections to render
 	glm::vec3 const cameraposition{m_renderpass.pass_camera.position()};
 	auto const camerax = static_cast<int>(std::floor(cameraposition.x / scene::EU07_SECTIONSIZE + scene::EU07_REGIONSIDESECTIONCOUNT / 2));
@@ -2802,10 +2805,11 @@ void opengl33_renderer::Render(cell_sequence::iterator First, cell_sequence::ite
 	// still happens inside Render_Instanced(), so correctness is unchanged -- only
 	// the number of calls and glBufferSubData round-trips drops sharply.
 	// (For pickscenery/pickcontrols modes the map stays empty, so this is a no-op.)
-	for( auto const &bucket : m_frame_instance_buckets ) {
+	for( auto &bucket : m_frame_instance_buckets ) {
+		if( bucket.second.empty() ) { continue; }
 		Render_Instanced( bucket.first.pModel, bucket.second );
+		bucket.second.clear();
 	}
-	m_frame_instance_buckets.clear();
 }
 
 void opengl33_renderer::Draw_Geometry(std::vector<gfx::geometrybank_handle>::iterator begin, std::vector<gfx::geometrybank_handle>::iterator end)
@@ -3022,15 +3026,46 @@ void opengl33_renderer::Render_Instanced( TModel3d *Model, std::vector<TAnimMode
 	if( Model == nullptr ) { return; }
 	if( Instances.empty() ) { return; }
 
-	// 1. Visibility / distance cull. Build parallel arrays of surviving
+	// LOD: a submodel is drawn only while fSquareDist lies inside its [fSquareMinDist, fSquareMaxDist)
-	// instances and their precomputed camera-space root modelview matrices.
+	// range, so the distinct range bounds split distance into bands with constant LOD. The bucket
-	// m_instance_modelviews is a persistent member reused across every
+	// collects instances from every visible cell, so a single fSquareDist for all of them would draw
-	// Render_Instanced() call: clear() drops the contents but keeps the
+	// far instances at the LOD of the nearest one. Each instance is assigned to its band and every
-	// allocated capacity, so after the first few frames this stops calling
+	// band is drawn with its own fSquareDist, like in Render_Sleepers(). The grouping is a counting
-	// malloc/free entirely (the reserve() below becomes a no-op once the
+	// sort, linear in the instance count.
-	// buffer has grown to the largest batch encountered).
+	// the same walk also checks whether any piece of the model can get into the shadow map. Models
-	m_instance_modelviews.clear();
+	// whose materials are all excluded by the shadow rank cutoff (e.g. grass) would otherwise go through
-	m_instance_modelviews.reserve( Instances.size() );
+	// the whole per-instance culling and the batched draws in every shadow pass, only for every submodel
+	// to be skipped in Render(TSubModel*)
+	auto const *skins { Instances.front() != nullptr && Instances.front()->Material() != nullptr ? Instances.front()->Material()->replacable_skins : nullptr };
+	bool castsshadow { m_renderpass.draw_mode != rendermode::shadows };
+	m_instance_lodbounds.clear();
+	m_instance_lodpending.clear();
+	m_instance_lodpending.push_back( Model->Root );
+	while( false == m_instance_lodpending.empty() ) {
+		auto const *submodel = m_instance_lodpending.back();
+		m_instance_lodpending.pop_back();
+		if( submodel == nullptr ) { continue; }
+		if( submodel->fSquareMinDist > 0.f ) { m_instance_lodbounds.emplace_back( submodel->fSquareMinDist ); }
+		if( submodel->fSquareMaxDist < std::numeric_limits<float>::max() ) { m_instance_lodbounds.emplace_back( submodel->fSquareMaxDist ); }
+		if( ( false == castsshadow ) && ( submodel->eType < TP_ROTATOR ) ) {
+			auto const material { submodel->m_material >= 0 ? submodel->m_material : ( skins != nullptr ? skins[ -submodel->m_material ] : null_handle ) };
+			castsshadow = ( Material( material ).shadow_rank <= Global.gfx_shadow_rank_cutoff );
+		}
+		m_instance_lodpending.push_back( submodel->Child );
+		m_instance_lodpending.push_back( submodel->Next );
+	}
+	if( false == castsshadow ) { return; }
+	std::sort( m_instance_lodbounds.begin(), m_instance_lodbounds.end() );
+	m_instance_lodbounds.erase( std::unique( m_instance_lodbounds.begin(), m_instance_lodbounds.end() ), m_instance_lodbounds.end() );
+	auto const bandcount { m_instance_lodbounds.size() + 1 };
+
+	// 1. Visibility / distance cull. Collect surviving instances with their precomputed
+	// camera-space root modelview matrices and lod bands. m_instance_survivors and the other
+	// scratch vectors are persistent members reused across every Render_Instanced() call:
+	// clear() drops the contents but keeps the allocated capacity, so after the first few
+	// frames this stops calling malloc/free entirely.
+	m_instance_survivors.clear();
+	m_instance_survivors.reserve( Instances.size() );
 
 	// Pull the current pass camera/view transform once. We use the current GL
 	// modelview matrix as the view matrix because at the point Render_Instanced
@@ -3039,11 +3074,14 @@ void opengl33_renderer::Render_Instanced( TModel3d *Model, std::vector<TAnimMode
 	glm::mat4 const view_matrix = OpenGLMatrices.data( GL_MODELVIEW );
 
 	bool prepared_shared_state = false; // RaPrepare() runs once per bucket
-	float closest_distancesquared = std::numeric_limits<float>::max();
 	material_data const *batch_material { nullptr };
+	float farthest_distancesquared { 0.f };
 
 	for( auto *Instance : Instances ) {
 		if( Instance == nullptr ) { continue; }
+		// instances which lost their instanceable status after being bucketed (animation or skin
+		// assigned by an event) are drawn by the per-node path
+		if( false == Instance->m_instanceable ) { continue; }
 		if( false == Instance->m_visible ) { continue; }
 		if( false == m_renderpass.pass_camera.visible( Instance->m_area ) ) { continue; }
 
@@ -3076,78 +3114,103 @@ void opengl33_renderer::Render_Instanced( TModel3d *Model, std::vector<TAnimMode
 			batch_material = Instance->Material();
 		}
 		Instance->m_framestamp = m_framestamp;
-		closest_distancesquared = std::min<float>( closest_distancesquared, static_cast<float>(distancesquared) );
 
 		// Build the camera-relative root modelview for this instance:
 		//   mv = view * translate(instance_pos - camera_pos) * rotate(instance_angles) * scale(m_scale)
-		// The scale is folded in here so the GPU-instanced path produces visually
+		// The rotation and scale part is cached by the instance, the translation goes into its last
-		// identical output to the regular per-instance path: each instance gets
+		// column. The scale is folded in so the GPU-instanced path produces visually identical output
-		// its own (translate × rotate × scale) baked into instance_modelview[i],
+		// to the regular per-instance path; the shader applies it via effective_modelview = instance_mv * model_local.
-		// which the shader applies via effective_modelview = instance_mv * model_local.
+		glm::mat4 root { Instance->rotation_scale() };
-		glm::dvec3 const offset = Instance->location() - m_renderpass.pass_camera.position();
+		root[ 3 ] = glm::vec4( glm::vec3( Instance->location() - m_renderpass.pass_camera.position() ), 1.f );
-		glm::mat4 mv = view_matrix;
+
-		mv = glm::translate( mv, glm::vec3( offset ) );
+		auto const distance { static_cast<float>( distancesquared ) };
-		auto const &angle = Instance->vAngle;
+		farthest_distancesquared = std::max( farthest_distancesquared, distance );
-		if( angle.y != 0.0f ) { mv = glm::rotate( mv, glm::radians( angle.y ), glm::vec3( 0.f, 1.f, 0.f ) ); }
+		m_instance_survivors.push_back( {
-		if( angle.x != 0.0f ) { mv = glm::rotate( mv, glm::radians( angle.x ), glm::vec3( 1.f, 0.f, 0.f ) ); }
+			view_matrix * root,
-		if( angle.z != 0.0f ) { mv = glm::rotate( mv, glm::radians( angle.z ), glm::vec3( 0.f, 0.f, 1.f ) ); }
+			distance,
-		auto const &scale = Instance->Scale();
+			static_cast<std::uint32_t>( std::upper_bound( m_instance_lodbounds.begin(), m_instance_lodbounds.end(), distance ) - m_instance_lodbounds.begin() ) } );
-		if( scale.x != 1.0f || scale.y != 1.0f || scale.z != 1.0f ) {
-			mv = glm::scale( mv, scale );
-		}
-		m_instance_modelviews.emplace_back( mv );
 	}
 
-	if( m_instance_modelviews.empty() ) { return; }
+	if( m_instance_survivors.empty() ) { return; }
+
+	// group the modelviews by lod band into a contiguous array, for the UBO upload. every distance in
+	// a band selects the same lod; the nearest one represents the band.
+	// within a band the instances are additionally ordered roughly near-to-far, in distance slots. dense
+	// alpha tested foliage then gets the most out of the early depth test, as the nearest instances fill
+	// the depth buffer first and hide the fragments of those behind them before they get shaded.
+	// both levels are done with a single counting sort, linear in the instance count
+	auto const distanceslots { 16 };
+	auto const keycount { bandcount * distanceslots };
+	auto const slotscale { distanceslots / std::sqrt( std::max( farthest_distancesquared, 1.f ) ) };
+	m_instance_bandstarts.assign( keycount + 1, 0 );
+	m_instance_banddistances.assign( bandcount, std::numeric_limits<float>::max() );
+	for( auto &survivor : m_instance_survivors ) {
+		m_instance_banddistances[ survivor.band ] = std::min( m_instance_banddistances[ survivor.band ], survivor.distancesquared );
+		auto const slot { std::min( distanceslots - 1, static_cast<int>( std::sqrt( survivor.distancesquared ) * slotscale ) ) };
+		survivor.band = static_cast<std::uint32_t>( survivor.band * distanceslots + slot ); // from now on it's the sort key
+		++m_instance_bandstarts[ survivor.band + 1 ];
+	}
+	for( std::size_t key = 1; key <= keycount; ++key ) {
+		m_instance_bandstarts[ key ] += m_instance_bandstarts[ key - 1 ];
+	}
+	m_instance_bandcursors.assign( m_instance_bandstarts.begin(), m_instance_bandstarts.end() - 1 );
+	m_instance_modelviews.resize( m_instance_survivors.size() );
+	for( auto const &survivor : m_instance_survivors ) {
+		m_instance_modelviews[ m_instance_bandcursors[ survivor.band ]++ ] = survivor.modelview;
+	}
+
+	auto alpha = ( batch_material != nullptr ? batch_material->textures_alpha : 0x30300030 );
+	alpha ^= 0x0F0F000F;
 
 	// 2. Walk the submodel tree once per sub-batch. The submodel-local matrix
 	// stack starts at identity; the per-instance camera transform comes from
 	// instance_modelview[gl_InstanceID] in the shader.
-	std::size_t const total = m_instance_modelviews.size();
+	for( std::size_t band = 0; band < bandcount; ++band ) {
-	std::size_t offset_idx = 0;
-	while( offset_idx < total ) {
-		std::size_t const this_batch = std::min<std::size_t>( total - offset_idx, gl::MAX_INSTANCES_PER_BATCH );
 
-		// 2a. Upload N modelviews to instance_ubo[0..N-1].
+		std::size_t offset_idx = m_instance_bandstarts[ band * distanceslots ];
-		instance_ubo->update(
+		std::size_t const band_end = m_instance_bandstarts[ ( band + 1 ) * distanceslots ];
-			reinterpret_cast<uint8_t const *>( m_instance_modelviews.data() + offset_idx ),
+		while( offset_idx < band_end ) {
-			0,
+			std::size_t const this_batch = std::min<std::size_t>( band_end - offset_idx, gl::MAX_INSTANCES_PER_BATCH );
-			static_cast<int>( this_batch * sizeof( glm::mat4 ) ) );
 
-		// 2b. Push identity onto the matrix stack so submodel transforms
+			// 2a. Upload N modelviews to instance_ubo[0..N-1].
-		// accumulate from identity, not from camera/instance space.
+			instance_ubo->update(
-		::glPushMatrix();
+				reinterpret_cast<uint8_t const *>( m_instance_modelviews.data() + offset_idx ),
-		::glLoadIdentity();
+				0,
+				static_cast<int>( this_batch * sizeof( glm::mat4 ) ) );
 
-		// 2c. Configure the existing TModel3d/TSubModel render path. Setting
+			// 2b. Push identity onto the matrix stack so submodel transforms
-		// m_current_instance_count routes draw(handle) calls to draw_instanced.
+			// accumulate from identity, not from camera/instance space.
-		m_current_instance_count = this_batch;
+			::glPushMatrix();
+			::glLoadIdentity();
 
-		Model->Root->fSquareDist = closest_distancesquared; // shared global, used by submodel LOD
+			// 2c. Configure the existing TModel3d/TSubModel render path. Setting
-		auto alpha = ( batch_material != nullptr ? batch_material->textures_alpha : 0x30300030 );
+			// m_current_instance_count routes draw(handle) calls to draw_instanced.
-		alpha ^= 0x0F0F000F;
+			m_current_instance_count = this_batch;
-		Model->Root->ReplacableSet( ( batch_material != nullptr ? batch_material->replacable_skins : nullptr ), alpha );
-		Model->Root->pRoot = Model;
 
-		Render( Model->Root );
+			Model->Root->fSquareDist = m_instance_banddistances[ band ]; // shared global, used by submodel LOD
+			Model->Root->ReplacableSet( ( batch_material != nullptr ? batch_material->replacable_skins : nullptr ), alpha );
+			Model->Root->pRoot = Model;
 
-		m_current_instance_count = 0;
+			Render( Model->Root );
 
-		::glPopMatrix();
+			m_current_instance_count = 0;
 
-		// 2d. Restore instance_modelview[0] to identity so subsequent
+			::glPopMatrix();
-		// non-instanced draws continue to compute identity * modelview.
+
-		{
+			// 2d. Restore instance_modelview[0] to identity so subsequent
-			glm::mat4 const identity( 1.0f );
+			// non-instanced draws continue to compute identity * modelview.
-			instance_ubo->update( reinterpret_cast<uint8_t const *>( &identity ), 0, sizeof( identity ) );
+			{
+				glm::mat4 const identity( 1.0f );
+				instance_ubo->update( reinterpret_cast<uint8_t const *>( &identity ), 0, sizeof( identity ) );
+			}
+
+			offset_idx += this_batch;
+			++m_renderpass.draw_stats.instanced_drawcalls;
 		}
-
-		offset_idx += this_batch;
-		++m_renderpass.draw_stats.instanced_drawcalls;
 	}
 
-	m_renderpass.draw_stats.instances += static_cast<int>( total );
+	m_renderpass.draw_stats.instances += static_cast<int>( m_instance_survivors.size() );
-	m_renderpass.draw_stats.models += static_cast<int>( total );
+	m_renderpass.draw_stats.models += static_cast<int>( m_instance_survivors.size() );
 }
 
 // Renders the per-track sleeper instances (TTrack::m_sleeper_local_transforms) using the

rendering/opengl33renderer.h

@@ -376,7 +376,7 @@ class opengl33_renderer : public gfx_renderer {
 	// cell's buckets are merged here keyed by (TModel3d*, skins), so that
 	// Render_Instanced() runs once per unique model across the whole pass instead
 	// of once per cell -- collapsing many tiny instanced draws into a few large
-	// batches. Reused every pass; cleared at the top of Render(scene::basic_region*).
+	// batches. Reused every pass; the vectors are emptied (not freed) at the top of Render(scene::basic_region*).
 	scene::basic_cell::instance_bucket_map m_frame_instance_buckets;
   renderpass_config m_colorpass; // parametrs of most recent color pass
 	std::array<renderpass_config, 3> m_shadowpass; // parametrs of most recent shadowmap pass for each of csm stages
@@ -426,6 +426,21 @@ class opengl33_renderer : public gfx_renderer {
 	// reused across calls -- clear() retains capacity, so once it has grown to
 	// the largest batch seen, steady-state frames perform no allocation here.
 	std::vector<glm::mat4> m_instance_modelviews;
+	// further persistent scratch buffers for Render_Instanced(): culled instances paired with their
+	// squared distance, and the lod bounds of the drawn model which split them into lod bands
+	struct instance_survivor {
+		glm::mat4 modelview;
+		float distancesquared;
+		std::uint32_t band;
+	};
+	std::vector<instance_survivor> m_instance_survivors;
+	std::vector<float> m_instance_lodbounds;
+	std::vector<TSubModel const *> m_instance_lodpending;
+	// per sort key (lod band and distance slot): first index in m_instance_modelviews (with end sentinel)
+	// and fill cursor; per lod band: nearest distance
+	std::vector<std::size_t> m_instance_bandstarts;
+	std::vector<std::size_t> m_instance_bandcursors;
+	std::vector<float> m_instance_banddistances;
 	gl::scene_ubs scene_ubs;
 	gl::model_ubs model_ubs;
 	gl::light_ubs light_ubs;

scene/scene.cpp

@@ -442,22 +442,18 @@ basic_cell::erase( TAnimModel *Instance ) {
                 [=]( TAnimModel *instance ) {
                     return instance == Instance; } ),
             std::end( m_instancesopaque ) );
-        // also remove from the per-(pModel, skins) instance bucket if present
+        // also remove from the per-(pModel, skins) instance bucket if present.
-        if( Instance->m_instanceable && Instance->Model() != nullptr ) {
+        // NOTE: searched by pointer rather than by key: the instance may have lost its instanceable
-            instance_bucket_key key;
+        // status or changed its skins since it was inserted, which would leave a dangling pointer behind
-            key.pModel = Instance->Model();
+        for( auto bucket = m_instancebuckets_opaque.begin(); bucket != m_instancebuckets_opaque.end(); ) {
-            auto const *mat = Instance->Material();
+            bucket->second.erase(
-            if( mat != nullptr ) {
+                std::remove( std::begin( bucket->second ), std::end( bucket->second ), Instance ),
-                for( int i = 0; i < 5; ++i ) { key.skins[i] = mat->replacable_skins[i]; }
+                std::end( bucket->second ) );
+            if( bucket->second.empty() ) {
+                bucket = m_instancebuckets_opaque.erase( bucket );
             }
-            auto bucket = m_instancebuckets_opaque.find( key );
+            else {
-            if( bucket != m_instancebuckets_opaque.end() ) {
+                ++bucket;
-                bucket->second.erase(
-                    std::remove( std::begin( bucket->second ), std::end( bucket->second ), Instance ),
-                    std::end( bucket->second ) );
-                if( bucket->second.empty() ) {
-                    m_instancebuckets_opaque.erase( bucket );
-                }
             }
         }
     }