Merge pull request #111 from MaSzyna-EU07/opengl-instancing

Some rendering improvements

CMakeLists.txt

@@ -411,6 +411,8 @@ if (${CMAKE_CXX_COMPILER_ID} STREQUAL MSVC)
 	set(CMAKE_CXX_FLAGS_RELWITHDEBINFO " /MD /Zi /O2 /DNDEBUG")
 	set(CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFO " /DEBUG /INCREMENTAL:NO /OPT:REF /OPT:ICF")
 
+	target_link_options(${PROJECT_NAME} PRIVATE "/MAP")
+
 	# /wd4996: disable "deprecation" warnings
 	# /wd4244: disable warnings for conversion with possible loss of data
 	# /wd5033: disable because it is all over Python headers

gl/shader.cpp

@@ -2,6 +2,7 @@
 
 #include <fstream>
 #include <sstream>
+#include <cstring>
 #include "shader.h"
 #include "glsl_common.h"
 #include "utilities/Logs.h"
@@ -13,6 +14,44 @@ inline bool strcend(std::string const &value, std::string const &ending)
     return std::equal(ending.rbegin(), ending.rend(), value.rbegin());
 }
 
+namespace {
+
+// The project's glad config only generates GLAD_GL_ARB_texture_filter_anisotropic
+// among the ARB/EXT extension flags -- GL_ARB_texture_gather (desktop) and
+// GL_EXT_gpu_shader5 (GLES 3.1) aren't compiled in, so we can't gate the
+// shadow textureGather() optimisation on a GLAD constant. Query the live
+// extension string instead. The first call walks the extension list once
+// (no extension count in the dozens is large enough to matter here) and
+// the result is cached in the static bool inside has_gl_extension(), so
+// every subsequent shader compile is a plain bool read.
+//
+// SHADERVALIDATOR_STANDALONE gates out the GL queries because the offline
+// shader validator tool links glad.c but never calls gladLoadGL(); the
+// function pointers stay null and a real call would crash. Returning
+// false there means the standalone validator simply compiles the original
+// 16-tap PCF fallback path in light_common.glsl, which is what we want.
+bool has_gl_extension(char const *name) {
+#ifdef SHADERVALIDATOR_STANDALONE
+    (void)name;
+    return false;
+#else
+    if (!glGetIntegerv || !glGetStringi) {
+        return false;
+    }
+    GLint count = 0;
+    glGetIntegerv(GL_NUM_EXTENSIONS, &count);
+    for (GLint i = 0; i < count; ++i) {
+        char const *ext = reinterpret_cast<char const *>(glGetStringi(GL_EXTENSIONS, i));
+        if (ext != nullptr && std::strcmp(ext, name) == 0) {
+            return true;
+        }
+    }
+    return false;
+#endif
+}
+
+} // anonymous namespace
+
 std::string gl::shader::read_file(const std::string &filename)
 {
     std::stringstream stream;
@@ -91,11 +130,33 @@ std::pair<GLuint, std::string> gl::shader::process_source(const std::string &fil
     if (!Global.gfx_usegles)
     {
         str += "#version 330 core\n";
+        // textureGather() on sampler2DArrayShadow is core in GLSL 4.0. On 3.30
+        // desktop it requires GL_ARB_gpu_shader5 -- the older
+        // GL_ARB_texture_gather (2010) only adds the non-shadow and the plain
+        // sampler2DShadow overloads, NOT the sampler2DArrayShadow one we need
+        // for cascaded shadow PCF. Some drivers advertise texture_gather but
+        // reject the shadow-array overload because the spec for it lives in
+        // gpu_shader5; so emit only when gpu_shader5 is advertised. When the
+        // extension is missing, calc_shadow() in light_common.glsl falls back
+        // to the original 16-tap hardware-PCF loop via #ifndef.
+        // (Project's glad config doesn't generate GLAD_GL_ARB_gpu_shader5,
+        //  so we query the live extension list -- see has_gl_extension above.)
+        static bool const have_gpu_shader5 = has_gl_extension("GL_ARB_gpu_shader5");
+        if (have_gpu_shader5)
+            str += "#extension GL_ARB_gpu_shader5 : enable\n";
     }
     else
     {
         if (GLAD_GL_ES_VERSION_3_1) {
             str += "#version 310 es\n";
+            // GLES 3.1 lacks textureGather on shadow samplers in core; the
+            // EXT_gpu_shader5 extension adds it. Only emit the directive when
+            // the driver advertises support (see desktop comment above).
+            // (Glad config doesn't generate GLAD_GL_EXT_gpu_shader5 -- query
+            //  the live extension string the same way.)
+            static bool const have_gpu_shader5 = has_gl_extension("GL_EXT_gpu_shader5");
+            if (have_gpu_shader5)
+                str += "#extension GL_EXT_gpu_shader5 : enable\n";
             if (type == GL_GEOMETRY_SHADER)
                 str += "#extension GL_EXT_geometry_shader : require\n";
         } else {

gl/ubo.h

@@ -81,7 +81,12 @@ namespace gl
         void set_modelview(const glm::mat4 &mv)
         {
             modelview = mv;
-            modelviewnormal = glm::mat3x4(glm::mat3(glm::transpose(glm::inverse(mv))));
+            // normal matrix = transpose(inverse(modelview)). The modelview is
+            // always affine, so its 3x3 normal matrix depends only on the
+            // upper-left 3x3 block; inverting that mat3 directly is markedly
+            // cheaper than a full mat4 inverse and yields an identical result
+            // (for affine M, mat3(inverse(M)) == inverse(mat3(M))).
+            modelviewnormal = glm::mat3x4(glm::transpose(glm::inverse(glm::mat3(mv))));
         }
     };
 

model/material.cpp

@@ -540,14 +540,30 @@ material_manager::create( std::string const &Filename, bool const Loadnow ) {
     }
 
 	if( false == material.name.empty() ) {
-		// if we have material name and shader it means resource was processed succesfully
+		// if we have material name and shader it means resource was processed succesfully.
-        materialhandle = m_materials.size();
+		//
-        m_materialmappings.emplace( material.name, materialhandle );
+		// IMPORTANT: capture the handle ONLY after emplace_back succeeds. The
+		// previous version pre-assigned `materialhandle = m_materials.size()`
+		// and then ran `finalize(Loadnow)` (which can throw shader_exception
+		// when a shader fails to compile) and `emplace_back` together inside
+		// the try. If finalize() threw, emplace_back never ran, but the
+		// handle was already set to size() -- pointing one past the actual
+		// last element. Subsequent `m_materials[handle]` lookups would then
+		// read garbage past end-of-vector, producing the classic 0x30-offset
+		// access violation seen in TSubModel::BinInit (vtable read on a
+		// bogus IMaterial pointer). Now we leave materialhandle as
+		// null_handle if anything throws, and the caller treats it as a
+		// failed-to-load material.
         try {
 			material.finalize(Loadnow);
+			materialhandle = m_materials.size();
 			m_materials.emplace_back( std::move(material) );
+			m_materialmappings.emplace( m_materials.back().name, materialhandle );
 		} catch (gl::shader_exception const &e) {
 			ErrorLog("invalid shader: " + std::string(e.what()));
+			// record the failure so subsequent Fetch_Material(filename) calls
+			// short-circuit without re-running the failing compile.
+			m_materialmappings.emplace( material.name, null_handle );
 		}
     }
     else {

rendering/opengl33renderer.cpp

@@ -2374,6 +2374,8 @@ 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
+	m_frame_instance_buckets.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));
@@ -2422,6 +2424,12 @@ void opengl33_renderer::Render(scene::basic_region *Region)
 			// at this stage the z-buffer is filled with only ground geometry
             Update_Mouse_Position();
 		}
+		// draw opaque cells front-to-back: with the depth test enabled this
+		// lets the GPU reject hidden fragments early, before the (expensive)
+		// lit fragment shader runs on them. Order is irrelevant to the final
+		// image for opaque geometry, so this is purely a fill-rate win.
+		std::sort( std::begin( m_cellqueue ), std::end( m_cellqueue ),
+			[]( distancecell_pair const &Left, distancecell_pair const &Right ) { return Left.first < Right.first; } );
 		Render(std::begin(m_cellqueue), std::end(m_cellqueue));
 		break;
 	}
@@ -2679,9 +2687,42 @@ void opengl33_renderer::Render(cell_sequence::iterator First, cell_sequence::ite
         case rendermode::shadows:
         {
             // TBD, TODO: refactor in to a method to reuse in branch below?
-			// opaque parts of instanced models -- batched path first
+			// opaque parts of instanced models -- accumulate this cell's buckets
-			for( auto const &bucket : cell->m_instancebuckets_opaque ) {
+			// into the frame-level map; the actual Render_Instanced() calls are
-				Render_Instanced( bucket.first.pModel, bucket.second );
+			// issued once per unique model after the cell loop (see flush below).
+			//
+			// Cell-level far-distance pre-cull: when the whole cell lies beyond
+			// the instance draw distance, every instance in it would fail the
+			// per-instance drawdistancethreshold test inside Render_Instanced(),
+			// so there is no point merging its buckets into the frame map. The
+			// test reproduces Render_Instanced()'s distance maths exactly -- same
+			// ZoomFactor / fDistanceFactor scaling, same +250 margin -- applied to
+			// the nearest point of the cell's bounding sphere. basic_cell::enclose_area
+			// guarantees m_area.radius >= |m_area.center - instance.location()| for
+			// every contained instance, so the nearest-point distance is a true
+			// lower bound on every instance's distance: the cull can never drop a
+			// cell that still holds a drawable instance. Shadows measure from the
+			// real (viewport) camera, matching Render_Instanced()'s shadow branch;
+			// every other gated mode measures from the pass camera. Non-instanced
+			// scenery and vehicles below keep their own per-node culling and are
+			// intentionally left untouched.
+			{
+				auto const &distancecamera = (
+					m_renderpass.draw_mode == rendermode::shadows
+						? m_renderpass.viewport_camera
+						: m_renderpass.pass_camera );
+				auto const cellcenterdistance { glm::length( cell->m_area.center - distancecamera.position() ) };
+				auto const cellnearestdistance { std::max( 0.0, cellcenterdistance - cell->m_area.radius ) };
+				auto const cellnearestdistancesquared {
+					( cellnearestdistance * cellnearestdistance )
+					/ ( static_cast<double>( Global.ZoomFactor ) * static_cast<double>( Global.ZoomFactor ) )
+					/ static_cast<double>( Global.fDistanceFactor ) };
+				if( cellnearestdistancesquared <= sq( static_cast<double>( m_renderpass.draw_range ) + 250.0 ) ) {
+					for( auto const &bucket : cell->m_instancebuckets_opaque ) {
+						auto &dest = m_frame_instance_buckets[ bucket.first ];
+						dest.insert( dest.end(), bucket.second.begin(), bucket.second.end() );
+					}
+				}
 			}
 			// remaining (non-instanceable) opaque instance nodes go through the per-node path
 			for (auto *instance : cell->m_instancesopaque)
@@ -2702,9 +2743,11 @@ void opengl33_renderer::Render(cell_sequence::iterator First, cell_sequence::ite
         case rendermode::reflections:
         {
             if( Global.reflectiontune.fidelity >= 1 ) {
-                // opaque parts of instanced models -- batched path first
+                // opaque parts of instanced models -- accumulate into the
+                // frame-level map; flushed once per unique model after the loop.
                 for( auto const &bucket : cell->m_instancebuckets_opaque ) {
-                    Render_Instanced( bucket.first.pModel, bucket.second );
+                    auto &dest = m_frame_instance_buckets[ bucket.first ];
+                    dest.insert( dest.end(), bucket.second.begin(), bucket.second.end() );
                 }
                 for( auto *instance : cell->m_instancesopaque ) {
                     if( instance->m_instanceable ) { continue; }
@@ -2744,6 +2787,17 @@ void opengl33_renderer::Render(cell_sequence::iterator First, cell_sequence::ite
 
 		++first;
 	}
+
+	// flush accumulated instance buckets: issue one Render_Instanced() per unique
+	// (TModel3d*, skins) key across every cell visited in this pass, instead of
+	// one call per cell. All per-instance frustum/distance/pixel-area culling
+	// 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 ) {
+		Render_Instanced( bucket.first.pModel, bucket.second );
+	}
+	m_frame_instance_buckets.clear();
 }
 
 void opengl33_renderer::Draw_Geometry(std::vector<gfx::geometrybank_handle>::iterator begin, std::vector<gfx::geometrybank_handle>::iterator end)
@@ -2962,8 +3016,13 @@ void opengl33_renderer::Render_Instanced( TModel3d *Model, std::vector<TAnimMode
 
 	// 1. Visibility / distance cull. Build parallel arrays of surviving
 	// instances and their precomputed camera-space root modelview matrices.
-	std::vector<glm::mat4> instance_modelviews;
+	// m_instance_modelviews is a persistent member reused across every
-	instance_modelviews.reserve( Instances.size() );
+	// Render_Instanced() call: clear() drops the contents but keeps the
+	// allocated capacity, so after the first few frames this stops calling
+	// malloc/free entirely (the reserve() below becomes a no-op once the
+	// buffer has grown to the largest batch encountered).
+	m_instance_modelviews.clear();
+	m_instance_modelviews.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
@@ -3028,22 +3087,22 @@ void opengl33_renderer::Render_Instanced( TModel3d *Model, std::vector<TAnimMode
 		if( scale.x != 1.0f || scale.y != 1.0f || scale.z != 1.0f ) {
 			mv = glm::scale( mv, scale );
 		}
-		instance_modelviews.emplace_back( mv );
+		m_instance_modelviews.emplace_back( mv );
 	}
 
-	if( instance_modelviews.empty() ) { return; }
+	if( m_instance_modelviews.empty() ) { return; }
 
 	// 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 = instance_modelviews.size();
+	std::size_t const total = m_instance_modelviews.size();
 	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].
 		instance_ubo->update(
-			reinterpret_cast<uint8_t const *>( instance_modelviews.data() + offset_idx ),
+			reinterpret_cast<uint8_t const *>( m_instance_modelviews.data() + offset_idx ),
 			0,
 			static_cast<int>( this_batch * sizeof( glm::mat4 ) ) );
 
@@ -3083,6 +3142,200 @@ void opengl33_renderer::Render_Instanced( TModel3d *Model, std::vector<TAnimMode
 	m_renderpass.draw_stats.models += static_cast<int>( total );
 }
 
+// Renders the per-track sleeper instances (TTrack::m_sleeper_local_transforms) using the
+// existing GPU-instanced submodel pipeline. The track owns a vector of pre-baked
+// local-space matrices; we compose each with `view * translate(track_origin - camera)`
+// to get a camera-space modelview, then issue batched glDrawElementsInstancedBaseVertex
+// calls -- one batch per MAX_INSTANCES_PER_BATCH sleepers.
+//
+// Skipped entirely when:
+//   - Global.SleeperDistance == 0 (sleeper rendering globally disabled)
+//   - the track has no sleepermodel
+//   - the track is farther than Global.SleeperDistance meters from the camera
+void opengl33_renderer::Render_Sleepers( TTrack *Track )
+{
+	if( Track == nullptr ) { return; }
+	if( false == Track->m_sleeper_enabled ) { return; }
+	if( Track->m_sleeper_model == nullptr ) { return; }
+	if( Track->m_sleeper_local_transforms.empty() ) { return; }
+	if( Global.SleeperDistance <= 0.f ) { return; }
+
+	// only the color and reflection passes draw sleepers; shadow/pick skip them on purpose
+	// (sleeper shadows would mostly fall back under the trackbed and pick already operates on
+	// the track itself).
+	switch( m_renderpass.draw_mode ) {
+	case rendermode::color:
+	case rendermode::reflections:
+		break;
+	default:
+		return;
+	}
+
+	// distance gate -- compare against Globals.SleeperDistance squared to avoid the sqrt
+	auto const camerapos = m_renderpass.pass_camera.position();
+	auto const trackpos  = Track->location();
+	auto const distsq    = glm::length2( trackpos - camerapos );
+	auto const cutoffsq  = static_cast<double>( Global.SleeperDistance ) * static_cast<double>( Global.SleeperDistance );
+	if( distsq > cutoffsq ) { return; }
+
+	// build camera-space modelview matrices.
+	// each sleeper's stored matrix is in track-local space (relative to Track->m_origin).
+	// Render_Sleepers is called from inside the per-cell origin push -- the cell's center
+	// already equals Track->m_origin (see basic_cell::insert), so the current GL_MODELVIEW
+	// is already view * translate(m_origin - camera). We just need to compose with each
+	// per-sleeper local transform to get the final modelview.
+	glm::mat4 const origin_mv = OpenGLMatrices.data( GL_MODELVIEW );
+
+	// per-sleeper frustum cull + LOD selection. The whole-track SleeperDistance
+	// gate above keeps or drops the track as a unit; everything from here on
+	// operates on individual sleepers.
+	//
+	// Each m_sleeper_local_transforms entry is
+	//   translate(world_pos - m_origin) * rotate(...) * translate(local_offset)
+	// so its translation column is the sleeper position relative to m_origin;
+	// adding m_origin back yields the sleeper's world-space position. The model's
+	// bounding radius (floored to a small minimum, in case it was never measured)
+	// serves both as the frustum test sphere and -- via the distance to that
+	// world position -- as the per-sleeper LOD distance.
+	auto const sleeperradius = std::max( Track->m_sleeper_model->bounding_radius(), 2.0f );
+
+	// Phase 1 -- frustum cull. For every sleeper that survives, store its
+	// camera-space modelview paired with the squared distance from the camera
+	// to ITS OWN world position (not the track / segment origin). That
+	// per-sleeper distance is what drives LOD selection below. A sleeper whose
+	// origin sits just off screen while its geometry still reaches into view is
+	// kept rather than wrongly culled.
+	std::vector<std::pair<float, glm::mat4>> survivors;
+	survivors.reserve( Track->m_sleeper_local_transforms.size() );
+	for( auto const &local : Track->m_sleeper_local_transforms ) {
+		glm::dvec3 const sleeperworldpos {
+			Track->m_origin + glm::dvec3( local[ 3 ].x, local[ 3 ].y, local[ 3 ].z ) };
+		if( false == m_renderpass.pass_camera.visible( scene::bounding_area{ sleeperworldpos, sleeperradius } ) ) {
+			continue;
+		}
+		auto const sleeperdistancesquared = static_cast<float>( glm::length2( sleeperworldpos - camerapos ) );
+		survivors.emplace_back( sleeperdistancesquared, origin_mv * local );
+	}
+
+	// every sleeper of this track was frustum-culled -- nothing left to draw
+	if( survivors.empty() ) { return; }
+
+	// Phase 2 -- sort survivors near-to-far. A submodel is drawn only while
+	// fSquareDist lies inside its [fSquareMinDist, fSquareMaxDist) range, so
+	// every distance between two consecutive range bounds selects an identical
+	// set of submodels -- i.e. the same LOD. Sorting turns each such LOD band
+	// into a contiguous run, letting the draw loop emit one instanced batch per
+	// band instead of one track-wide batch at a single distance.
+	std::sort( survivors.begin(), survivors.end(),
+		[]( std::pair<float, glm::mat4> const &Left, std::pair<float, glm::mat4> const &Right ) {
+			return Left.first < Right.first; } );
+
+	// contiguous copy of the sorted modelview matrices, for the UBO upload.
+	// Per-sleeper distances stay available as survivors[i].first, in lockstep.
+	std::vector<glm::mat4> instance_modelviews;
+	instance_modelviews.reserve( survivors.size() );
+	for( auto const &survivor : survivors ) {
+		instance_modelviews.emplace_back( survivor.second );
+	}
+
+	// collect the model's distinct LOD distance bounds. The sorted, de-duplicated
+	// set of every submodel's fSquareMinDist / fSquareMaxDist partitions distance
+	// into bands within which the selected LOD is constant. A model with no LOD
+	// yields a single band, and the draw loop below then behaves exactly like a
+	// single plain batched draw.
+	std::vector<float> lodbounds;
+	{
+		std::vector<TSubModel const *> pending;
+		if( Track->m_sleeper_model->Root != nullptr ) {
+			pending.push_back( Track->m_sleeper_model->Root );
+		}
+		while( false == pending.empty() ) {
+			auto const *submodel = pending.back();
+			pending.pop_back();
+			lodbounds.emplace_back( submodel->fSquareMinDist );
+			lodbounds.emplace_back( submodel->fSquareMaxDist );
+			if( submodel->Child != nullptr ) { pending.push_back( submodel->Child ); }
+			if( submodel->Next  != nullptr ) { pending.push_back( submodel->Next ); }
+		}
+	}
+	std::sort( lodbounds.begin(), lodbounds.end() );
+	lodbounds.erase( std::unique( lodbounds.begin(), lodbounds.end() ), lodbounds.end() );
+
+	// optional replacable skin: build a transient material_data so we can drive ReplacableSet
+	// the same way Render_Instanced does. when no skin is set we fall back to the model defaults.
+	material_data sleeper_material {};
+	bool const has_skin = ( Track->m_sleeper_skin != null_handle );
+	if( has_skin ) {
+		sleeper_material.replacable_skins[ 1 ] = Track->m_sleeper_skin;
+	}
+
+	// Phase 3 -- draw. Walk the sorted survivors one LOD band at a time. Each
+	// band is submitted as one or more instanced draws (split only when it
+	// exceeds MAX_INSTANCES_PER_BATCH); every draw in the band uses an in-band
+	// fSquareDist, so each sleeper renders at the LOD its own distance selects
+	// while instancing stays fully in effect.
+	auto *Model = Track->m_sleeper_model;
+	std::size_t const total = instance_modelviews.size();
+	std::size_t band_start = 0;
+	while( band_start < total ) {
+		// the band ends at the first sleeper distance that reaches the next LOD
+		// bound above the band's starting distance.
+		auto const upperbound = std::upper_bound(
+			lodbounds.begin(), lodbounds.end(), survivors[ band_start ].first );
+		float const band_limit = ( upperbound == lodbounds.end()
+			? std::numeric_limits<float>::max()
+			: *upperbound );
+		std::size_t band_end = band_start;
+		while( ( band_end < total ) && ( survivors[ band_end ].first < band_limit ) ) {
+			++band_end;
+		}
+		// every distance in the band selects the same LOD; the band's nearest
+		// sleeper is used as the representative fSquareDist.
+		float const band_distancesquared = survivors[ band_start ].first;
+
+		std::size_t offset_idx = band_start;
+		while( offset_idx < band_end ) {
+			std::size_t const this_batch = std::min<std::size_t>( band_end - offset_idx, gl::MAX_INSTANCES_PER_BATCH );
+
+			instance_ubo->update(
+				reinterpret_cast<uint8_t const *>( instance_modelviews.data() + offset_idx ),
+				0,
+				static_cast<int>( this_batch * sizeof( glm::mat4 ) ) );
+
+			::glPushMatrix();
+			::glLoadIdentity();
+
+			m_current_instance_count = this_batch;
+
+			Model->Root->fSquareDist = band_distancesquared;
+			auto alpha = ( has_skin ? sleeper_material.textures_alpha : 0x30300030 );
+			alpha ^= 0x0F0F000F;
+			Model->Root->ReplacableSet( ( has_skin ? sleeper_material.replacable_skins : nullptr ), alpha );
+			Model->Root->pRoot = Model;
+
+			Render( Model->Root );
+
+			m_current_instance_count = 0;
+
+			::glPopMatrix();
+
+			// restore instance_modelview[0] to identity so subsequent non-instanced draws
+			// continue to compute identity * modelview (mirroring Render_Instanced).
+			{
+				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;
+		}
+		band_start = band_end;
+	}
+
+	m_renderpass.draw_stats.instances += static_cast<int>( total );
+	m_renderpass.draw_stats.models    += static_cast<int>( total );
+}
+
 bool opengl33_renderer::Render(TDynamicObject *Dynamic)
 {
 	glDebug("Render TDynamicObject");
@@ -3792,6 +4045,24 @@ void opengl33_renderer::Render(scene::basic_cell::path_sequence::const_iterator
 		}
 	}
 
+	// fourth pass: per-track sleeper models (sleepermodel optional directive).
+	// drawn after rails/trackbeds so depth pre-pass culling is favourable, and only in passes
+	// where Render_Sleepers actually does work (it gates itself on draw mode / distance).
+	switch( m_renderpass.draw_mode ) {
+	case rendermode::color:
+	case rendermode::reflections: {
+		for( auto first { First }; first != Last; ++first ) {
+			auto *track = *first;
+			if( false == track->m_visible )       { continue; }
+			if( false == track->m_sleeper_enabled ) { continue; }
+			Render_Sleepers( track );
+		}
+		break;
+	}
+	default:
+		break;
+	}
+
 	// post-render reset
 	switch (m_renderpass.draw_mode)
 	{

rendering/opengl33renderer.h

@@ -282,6 +282,11 @@ class opengl33_renderer : public gfx_renderer {
 	void Render(TSubModel *Submodel);
 	void Render(TTrack *Track);
 	void Render(scene::basic_cell::path_sequence::const_iterator First, scene::basic_cell::path_sequence::const_iterator Last);
+	// renders the per-track sleeper instances (TTrack::m_sleeper_local_transforms) via GPU instancing.
+	// caller must already have the camera-relative world-space transform set on the matrix stack.
+	// no-op if the track has no sleepermodel, Global.SleeperDistance is 0, or the camera is beyond
+	// Global.SleeperDistance from the track origin.
+	void Render_Sleepers( TTrack *Track );
 	bool Render_cab(TDynamicObject const *Dynamic, float const Lightlevel, bool const Alpha = false);
     bool Render_interior( bool const Alpha = false );
     bool Render_lowpoly( TDynamicObject *Dynamic, float const Squaredistance, bool const Setup, bool const Alpha = false );
@@ -361,6 +366,12 @@ class opengl33_renderer : public gfx_renderer {
 	renderpass_config m_renderpass; // parameters for current render pass
 	section_sequence m_sectionqueue; // list of sections in current render pass
 	cell_sequence m_cellqueue;
+	// frame-level accumulation of per-cell opaque instance buckets. Each visited
+	// 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*).
+	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
 	std::vector<TSubModel const *> m_pickcontrolsitems;
@@ -403,6 +414,12 @@ class opengl33_renderer : public gfx_renderer {
 	// than a single regular draw. The vertex shader reads per-instance modelview
 	// from instance_ubo[gl_InstanceID]. Reset to 0 by Render_Instanced() on exit.
 	std::size_t m_current_instance_count { 0 };
+	// persistent scratch buffer for Render_Instanced(): holds the per-instance
+	// camera-space root modelview matrices for the batch currently being built.
+	// Kept as a member rather than a function-local so its heap allocation is
+	// 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;
 	gl::scene_ubs scene_ubs;
 	gl::model_ubs model_ubs;
 	gl::light_ubs light_ubs;

rendering/openglmatrixstack.h

@@ -20,67 +20,73 @@ class opengl_stack {
 
 public:
 // constructors:
-    opengl_stack() { m_stack.emplace(1.f); }
+    opengl_stack() {
+        // reserve generously up front: the matrix stack is pushed/popped once
+        // per submodel during scene traversal, and std::deque (the previous
+        // backing store) allocated a fresh heap block on every push for a
+        // 64-byte glm::mat4. A reserved vector never reallocates within this
+        // depth, so push/pop become allocation-free and references returned
+        // by data() stay valid across pushes, exactly as before.
+        m_stack.reserve( 256 );
+        m_stack.emplace_back( 1.f ); }
 
 // methods:
     glm::mat4 const &
         data() const {
-            return m_stack.top(); }
+            return m_stack.back(); }
     void
         push_matrix() {
-            m_stack.emplace( m_stack.top() ); }
+            glm::mat4 const top { m_stack.back() };
+            m_stack.emplace_back( top ); }
     void
         pop_matrix( bool const Upload = true ) {
             if( m_stack.size() > 1 ) {
-                m_stack.pop();
+                m_stack.pop_back();
                 if( Upload ) { upload(); } } }
     void
         load_identity( bool const Upload = true ) {
-            m_stack.top() = glm::mat4( 1.f );
+            m_stack.back() = glm::mat4( 1.f );
             if( Upload ) { upload(); } }
     void
         load_matrix( glm::mat4 const &Matrix, bool const Upload = true ) {
-            m_stack.top() = Matrix;
+            m_stack.back() = Matrix;
             if( Upload ) { upload(); } }
     void
         rotate( float const Angle, glm::vec3 const &Axis, bool const Upload = true ) {
-            m_stack.top() = glm::rotate( m_stack.top(), Angle, Axis );
+            m_stack.back() = glm::rotate( m_stack.back(), Angle, Axis );
             if( Upload ) { upload(); } }
     void
         translate( glm::vec3 const &Translation, bool const Upload = true ) {
-            m_stack.top() = glm::translate( m_stack.top(), Translation );
+            m_stack.back() = glm::translate( m_stack.back(), Translation );
             if( Upload ) { upload(); } }
     void
         scale( glm::vec3 const &Scale, bool const Upload = true ) {
-            m_stack.top() = glm::scale( m_stack.top(), Scale );
+            m_stack.back() = glm::scale( m_stack.back(), Scale );
             if( Upload ) { upload(); } }
     void
         multiply( glm::mat4 const &Matrix, bool const Upload = true ) {
-            m_stack.top() *= Matrix;
+            m_stack.back() *= Matrix;
             if( Upload ) { upload(); } }
     void
         ortho( float const Left, float const Right, float const Bottom, float const Top, float const Znear, float const Zfar, bool const Upload = true ) {
-            m_stack.top() *= glm::ortho( Left, Right, Bottom, Top, Znear, Zfar );
+            m_stack.back() *= glm::ortho( Left, Right, Bottom, Top, Znear, Zfar );
             if( Upload ) { upload(); } }
     void
         perspective( float const Fovy, float const Aspect, float const Znear, float const Zfar, bool const Upload = true ) {
-            m_stack.top() *= glm::perspective( Fovy, Aspect, Znear, Zfar );
+            m_stack.back() *= glm::perspective( Fovy, Aspect, Znear, Zfar );
             if( Upload ) { upload(); } }
     void
         look_at( glm::vec3 const &Eye, glm::vec3 const &Center, glm::vec3 const &Up, bool const Upload = true ) {
-            m_stack.top() *= glm::lookAt( Eye, Center, Up );
+            m_stack.back() *= glm::lookAt( Eye, Center, Up );
             if( Upload ) { upload(); } }
 
 private:
-// types:
-    typedef std::stack<glm::mat4> mat4_stack;
-
 // methods:
     void
-        upload() { ::glLoadMatrixf( glm::value_ptr( m_stack.top() ) ); }
+        upload() { ::glLoadMatrixf( glm::value_ptr( m_stack.back() ) ); }
 
 // members:
-    mat4_stack m_stack;
+    std::vector<glm::mat4> m_stack;
 };
 
 enum stack_mode { gl_modelview = 0, gl_projection = 1, gl_texture = 2 };

shaders/light_common.glsl

@@ -51,7 +51,6 @@ float calc_shadow()
 		
 		
 
-	float shadow = 0.0;
 	float bias = 0.00005f * float(cascade + 1U);
 	vec2 texel = vec2(1.0) / vec2(textureSize(shadowmap, 0));
 	//float radius = 1.0; f_light_pos[cascade].w; //0.5 + 2.0 * max(abs(2.0 * coords.x - 1.0), abs(2.0 * coords.y - 1.0));
@@ -63,37 +62,117 @@ float calc_shadow()
 		radius = mix(minradius, f_light_pos[cascade+1U].w/f_light_pos[cascade].w, dist_casc);
 	else
 		radius = 0.5;
-	
+
+#if defined(GL_ARB_gpu_shader5) || defined(GL_EXT_gpu_shader5) || __VERSION__ >= 400
+	// Fast path -- replace the original 4x4 grid of individual hardware-PCF
+	// lookups with 4 textureGather() calls. Each gather returns the 4 raw
+	// shadow comparisons of a 2x2 texel footprint, so 4 gathers laid out at
+	// (+-1, +-1) * radius * texel from the sample center cover the same 4x4
+	// sample area as the original kernel; summing all 16 comparisons and
+	// dividing by 16 reproduces the original loop's averaging. The cost on
+	// the TMUs drops from 16 hardware-PCF samples to 4 gathers (the gather
+	// path returns 4 values per fetch where the original needed 4 fetches),
+	// roughly a 4x reduction in shadow-sample work. The only thing dropped
+	// vs. the hardware-PCF path is the implicit bilinear blending inside
+	// each 2x2 footprint -- effectively turning a tent-weighted kernel into
+	// a box-weighted one of the same extent, which is imperceptible in
+	// motion. calc_shadow() is by far the heaviest piece of the lighting
+	// shader, so this is a measurable GPU saving on every shaded fragment.
+	float refz = coords.z + bias;
+	float layer = float(cascade);
+	vec2 off = radius * texel;
+	vec4 g0 = textureGather(shadowmap, vec3(coords.xy + vec2(-off.x, -off.y), layer), refz);
+	vec4 g1 = textureGather(shadowmap, vec3(coords.xy + vec2( off.x, -off.y), layer), refz);
+	vec4 g2 = textureGather(shadowmap, vec3(coords.xy + vec2(-off.x,  off.y), layer), refz);
+	vec4 g3 = textureGather(shadowmap, vec3(coords.xy + vec2( off.x,  off.y), layer), refz);
+	float shadow = dot(g0 + g1 + g2 + g3, vec4(1.0 / 16.0));
+	return shadow;
+#else
+	// Fallback for drivers without textureGather on shadow samplers
+	// (notably GLES 3.0 and any 3.3 desktop driver that doesn't expose
+	// GL_ARB_texture_gather). Identical to the previous implementation.
+	float shadow = 0.0;
 	for (float y = -1.5; y <= 1.5; y += 1.0)
 		for (float x = -1.5; x <= 1.5; x += 1.0)
 			shadow += texture(shadowmap, vec4(coords.xy + vec2(x, y) * radius * texel, cascade, coords.z + bias) );
 	shadow /= 16.0;
 
 	return shadow;
+#endif
 #else
 	return 0.0;
 #endif
 }
 
+// -----------------------------------------------------------------------
+// GGX Microfacet BRDF helpers (Cook-Torrance)
+// -----------------------------------------------------------------------
+
+// Trowbridge-Reitz (GGX) Normal Distribution Function
+//   D(N,H,α) = α⁴ / (π · ((NdotH)²·(α⁴−1)+1)²)
+//   α = roughness² (perceptual remapping so the slider feels linear)
+float D_GGX(float NdotH, float roughness)
+{
+    float a  = roughness * roughness;   // perceptual -> linear roughness
+    float a2 = a * a;
+    float d  = (NdotH * NdotH) * (a2 - 1.0) + 1.0;
+    return a2 / (3.14159265359 * d * d);
+}
+
+// Schlick-GGX single-term masking/shadowing (k remapped for direct lighting)
+float G_SchlickGGX(float NdotX, float roughness)
+{
+    float r = roughness + 1.0;
+    float k = (r * r) * (1.0 / 8.0);   // k_direct = (roughness+1)²/8
+    return NdotX / (NdotX * (1.0 - k) + k);
+}
+
+// Height-correlated Smith geometry term
+//   G(N,V,L) = G_SchlickGGX(NdotV) · G_SchlickGGX(NdotL)
+float G_Smith(float NdotV, float NdotL, float roughness)
+{
+    return G_SchlickGGX(NdotV, roughness) * G_SchlickGGX(NdotL, roughness);
+}
+
+// Returns vec2(diffuse, specular) for a single punctual light.
+//
+//   diffuse  – Lambert N·L (Fresnel-weighted diffuse is handled per-material
+//              in apply_lights, so we return raw N·L here).
+//   specular – Cook-Torrance GGX: D·G / (4·NdotL·NdotV).
+//              The Fresnel factor (F) is intentionally omitted here;
+//              apply_lights already carries a per-material Fresnel term
+//              that is applied to env reflections and can be routed to
+//              direct specular there.
+//
+// Roughness is derived identically to env_roughness in apply_lights so
+// that direct and indirect specular highlights read consistently.
 vec2 calc_light(vec3 light_dir, vec3 fragnormal)
 {
-	vec3 view_dir = normalize(vec3(0.0f, 0.0f, 0.0f) - f_pos.xyz);
+    vec3 N = fragnormal;
-	vec3 halfway_dir = normalize(light_dir + view_dir);
+    vec3 L = light_dir;
+    vec3 V = normalize(-f_pos.xyz);
+    vec3 H = normalize(L + V);
 
-	float diffuse_v = max(dot(fragnormal, light_dir), 0.0);
+    float NdotL = max(dot(N, L), 0.0);
+    float NdotV = max(dot(N, V), 1e-4);
+    float NdotH = max(dot(N, H), 0.0);
 
-	// Energy-conserving Blinn-Phong normalization:
+    float diffuse_v = NdotL;
-	// (n+8)/(8*pi) ensures the specular lobe integrates to the same
-	// total energy regardless of glossiness — low glossiness stays dim
-	// and spreads wide (blurry), high glossiness is bright and tight (sharp).
-	// Capped at 4.0 so very high glossiness (n>~92) does not produce pinhole
-	// highlights that blow past the tonemap shoulder and read as burnt white.
-	float n = max(glossiness, 0.01);
-	float normalization = min((n + 8.0) / (8.0 * 3.14159265), 4.0);
-	float NdotH = max(dot(fragnormal, halfway_dir), 0.0);
-	float specular_v = normalization * pow(NdotH, n);
 
-	return vec2(diffuse_v, specular_v);
+    // Mirror the env-map roughness derivation so direct and indirect lobes match.
+    // glossiness == param[1].w  →  roughness == 0.04 (near-mirror)
+    // glossiness == 0           →  roughness == 1.0  (fully diffuse)
+    float roughness = clamp(1.0 - glossiness / max(abs(param[1].w), 1.0), 0.04, 1.0);
+
+    // Cook-Torrance specular (no Fresnel — see above):
+    //   f_spec = D(N,H,α) · G(N,V,L,α) / (4 · NdotL · NdotV)
+    float D = D_GGX(NdotH, roughness);
+    float G = G_Smith(NdotV, NdotL, roughness);
+    float specular_v = (NdotL > 0.0)
+        ? (D * G) / max(4.0 * NdotL * NdotV, 1e-4)
+        : 0.0;
+
+    return vec2(diffuse_v, specular_v);
 }
 
 vec2 calc_point_light(light_s light, vec3 fragnormal)

utilities/Globals.cpp

@@ -582,6 +582,15 @@ bool global_settings::ConfigParseSimulation(cParser& Parser, const std::string&
         return true;
     }
 
+    if (token == "sleeperdistance")
+    {
+        float sleeperdistance = 0.f;
+        ParseOne(Parser, sleeperdistance);
+        // negative values disable the cap; we clamp at 0 so 0 means "do not render sleepers"
+        SleeperDistance = std::max(0.f, sleeperdistance);
+        return true;
+    }
+
     if (token == "createswitchtrackbeds")
     {
         ParseOne(Parser, CreateSwitchTrackbeds);
@@ -1567,6 +1576,7 @@ global_settings::export_as_text( std::ostream &Output ) const {
     export_as_text( Output, "gfx.smoke.fidelity", SmokeFidelity );
     export_as_text( Output, "smoothtraction", bSmoothTraction );
     export_as_text( Output, "splinefidelity", SplineFidelity );
+    export_as_text( Output, "sleeperdistance", SleeperDistance );
     export_as_text( Output, "rendercab", render_cab );
     export_as_text( Output, "createswitchtrackbeds", CreateSwitchTrackbeds );
     export_as_text( Output, "gfx.resource.sweep", ResourceSweep );

utilities/Globals.h

@@ -165,6 +165,7 @@ struct global_settings {
     GLint iMaxCabTextureSize{ 4096 }; // largest allowed texture in vehicle cab
     int iMultisampling{ 2 }; // tryb antyaliasingu: 0=brak,1=2px,2=4px,3=8px,4=16px
     float SplineFidelity{ 1.f }; // determines segment size during conversion of splines to geometry
+    float SleeperDistance{ 250.f }; // max distance (in meters) at which per-track sleeper models are still drawn; 0 disables sleeper rendering entirely
     bool Smoke{ true }; // toggles smoke simulation and visualization
     float SmokeFidelity{ 1.f }; // determines amount of generated smoke particles
     bool ResourceSweep{ true }; // gfx resource garbage collection

world/Segment.cpp

@@ -337,29 +337,29 @@ Math3D::vector3 TSegment::GetPoint(double const fDistance) const
     }
 };
 */
-// ustalenie pozycji osi na torze, przechyłki, pochylenia i kierunku jazdy
+
-void TSegment::RaPositionGet(double const fDistance, glm::dvec3 &p, glm::vec3 &a) const {
+void TSegment::RaPositionGet(double const fDistance, glm::dvec3 &position, glm::vec3 &rotation) const {
     if (bCurve) {
         // można by wprowadzić uproszczony wzór dla okręgów płaskich
         auto const t = GetTFromS(fDistance); // aproksymacja dystansu na krzywej Beziera na parametr (t)
-        p = FastGetPoint( t );
+        position = FastGetPoint( t );
         // przechyłka w danym miejscu (zmienia się liniowo)
-        a.x = std::lerp( fRoll1, fRoll2, t );
+        rotation.x = std::lerp( fRoll1, fRoll2, t );
         // pochodna jest 3*A*t^2+2*B*t+C
         auto const tangent = t * ( t * 3.0 * vA + vB + vB ) + vC;
         // pochylenie krzywej (w pionie)
-        a.y = std::atan( tangent.y );
+        rotation.y = std::atan( tangent.y );
         // kierunek krzywej w planie
-        a.z = -std::atan2( tangent.x, tangent.z );
+        rotation.z = -std::atan2( tangent.x, tangent.z );
     }
     else {
         // wyliczenie dla odcinka prostego jest prostsze
         auto const t = fDistance / fLength; // zerowych torów nie ma
-        p = FastGetPoint( t );
+        position = FastGetPoint( t );
         // przechyłka w danym miejscu (zmienia się liniowo)
-        a.x = std::lerp( fRoll1, fRoll2, t );
+        rotation.x = std::lerp( fRoll1, fRoll2, t );
-        a.y = fStoop; // pochylenie toru prostego
+        rotation.y = fStoop; // pochylenie toru prostego
-        a.z = fDirection; // kierunek toru w planie
+        rotation.z = fDirection; // kierunek toru w planie
     }
 };
 

world/Segment.h

@@ -91,7 +91,14 @@ public:
     Math3D::vector3
         GetPoint(double const fDistance) const;
 */
-	void RaPositionGet(double const fDistance, glm::dvec3 &p, glm::vec3 &a) const;
+
+	/// <summary>
+	/// ustalenie pozycji osi na torze, przechyłki, pochylenia i kierunku jazdy
+	/// </summary>
+	/// <param name="fDistance">Distance from p1</param>
+	/// <param name="position">Calculated position</param>
+	/// <param name="rotation">Calculated rotation</param>
+	void RaPositionGet(double const fDistance, glm::dvec3 &position, glm::vec3 &rotation) const;
 	glm::dvec3 FastGetPoint(double const t) const;
     inline
     glm::dvec3

world/Track.cpp

@@ -23,6 +23,8 @@ http://mozilla.org/MPL/2.0/.
 #include "vehicle/DynObj.h"
 #include "vehicle/Driver.h"
 #include "model/AnimModel.h"
+#include "model/MdlMngr.h"
+#include "model/Model3d.h"
 #include "utilities/Timer.h"
 #include "utilities/Logs.h"
 #include "rendering/renderer.h"
@@ -923,6 +925,37 @@ void TTrack::Load(cParser *parser, glm::dvec3 const &pOrigin)
             // memory cell holding friction value modifiers
             m_friction.first = parser->getToken<std::string>();
         }
+        else if( str == "sleepermodel" ) {
+            // sleepermodel <frequency> <model> <skin> <offsetX> <offsetY> <offsetZ> <ballastZ>
+            // - frequency: meters between consecutive sleeper instances (must be > 0)
+            // - model:     path to the .e3d sleeper model
+            // - skin:      replacable skin path, or "none" for the model's defaults
+            // - offset:    local-space offset applied per-instance (x=left/right, y=forward/back, z=up/down)
+            // - ballastZ:  vertical shift applied to the auto-generated trackbed (ballast). negative pushes ballast down.
+            float frequency { 0.f };
+            float offsetx { 0.f }, offsety { 0.f }, offsetz { 0.f };
+            float ballastz { 0.f };
+            parser->getTokens( 1, false ); *parser >> frequency;
+            auto modelpath { parser->getToken<std::string>( false ) };
+            auto skinpath  { parser->getToken<std::string>( false ) };
+            parser->getTokens( 3, false ); *parser >> offsetx >> offsety >> offsetz;
+            parser->getTokens( 1, false ); *parser >> ballastz;
+
+            if( frequency <= 0.01f ) {
+                ErrorLog( "Bad track: invalid sleepermodel frequency (" + std::to_string( frequency ) + ") for track \"" + m_name + "\"" );
+            }
+            else {
+                replace_slashes( modelpath );
+                m_sleeper_enabled = true;
+                m_sleeper_frequency = frequency;
+                m_sleeper_model_name = modelpath;
+                m_sleeper_skin_name = skinpath;
+                m_sleeper_offset = glm::vec3( offsetx, offsety, offsetz );
+                m_sleeper_ballast_z = ballastz;
+                // model and skin are resolved (and instance transforms baked) in build_sleeper_transforms,
+                // called after segment initialisation so the path geometry is final.
+            }
+        }
         else
             ErrorLog("Bad track: unknown property: \"" + str + "\" defined for track \"" + m_name + "\"");
         parser->getTokens();
@@ -942,6 +975,9 @@ void TTrack::Load(cParser *parser, glm::dvec3 const &pOrigin)
         + CurrentSegment()->FastGetPoint( 0.5 )
         + CurrentSegment()->FastGetPoint_1() )
         / 3.0 );
+    // sleeper transforms are baked later in create_geometry(), once the owning cell has
+    // assigned this track its m_origin (otherwise the local-space matrices would be relative
+    // to a stale origin and the renderer would draw sleepers in the wrong place).
 }
 
 bool TTrack::AssignEvents() {
@@ -1313,6 +1349,11 @@ glm::vec3 TTrack::get_nearest_point(const glm::dvec3 &point) const
 // wypełnianie tablic VBO
 void TTrack::create_geometry( gfx::geometrybank_handle const &Bank ) {
 	gfx::userdata_array empty_userdata;
+    // bake per-instance sleeper transforms now that the owning cell has assigned m_origin.
+    // safe to call here even if the track has no sleepermodel (early-outs internally).
+    if( m_sleeper_enabled && m_sleeper_local_transforms.empty() ) {
+        build_sleeper_transforms();
+    }
     switch (iCategoryFlag & 15)
     {
     case 1: // tor
@@ -1328,6 +1369,14 @@ void TTrack::create_geometry( gfx::geometrybank_handle const &Bank ) {
             { // podsypka z podkładami jest tylko dla zwykłego toru
                 gfx::vertex_array bpts1;
                 create_track_bed_profile( bpts1, trPrev, trNext );
+                // optional vertical shift of the auto-generated ballast (sleepermodel ballastZ).
+                // positive value raises the trackbed, negative pushes it down so a custom
+                // sleeper model placed on top can sit flush with the ballast surface.
+                if( m_sleeper_enabled && ( m_sleeper_ballast_z != 0.f ) ) {
+                    for( auto &v : bpts1 ) {
+                        v.position.y += m_sleeper_ballast_z;
+                    }
+                }
                 auto const texturelength { texture_length( m_material2 ) };
                 gfx::vertex_array vertices;
                 Segment->RenderLoft(vertices, m_origin, bpts1, iTrapezoid > 0, texturelength);
@@ -2360,6 +2409,17 @@ TTrack::export_as_text_( std::ostream &Output ) const {
     if( false == m_friction.first.empty() ) {
         Output << "friction " << m_friction.first << ' ';
     }
+    if( m_sleeper_enabled && ( false == m_sleeper_model_name.empty() ) ) {
+        Output
+            << "sleepermodel "
+            << m_sleeper_frequency << ' '
+            << m_sleeper_model_name << ' '
+            << ( m_sleeper_skin_name.empty() ? std::string{ "none" } : m_sleeper_skin_name ) << ' '
+            << m_sleeper_offset.x << ' '
+            << m_sleeper_offset.y << ' '
+            << m_sleeper_offset.z << ' '
+            << m_sleeper_ballast_z << ' ';
+    }
     // footer
     Output
         << "endtrack"
@@ -3620,3 +3680,156 @@ path_table::IsolatedBusy( std::string const &Name ) const {
     }
     multiplayer::WyslijString( Name, 10 ); // wolny (technically not found but, eh)
 }
+
+namespace {
+// Returns the list of segments to walk when laying out sleepers for a given track.
+// For plain tracks/turntables we just use the active Segment. For switches, crossings
+// and tributaries we use every initialised sub-path so the user gets sleepers covering
+// the full footprint of the junction (not just the currently-selected route).
+std::vector<TSegment *> sleeper_segments_for( TTrack const &Track )
+{
+    std::vector<TSegment *> out;
+    switch( Track.eType ) {
+    case tt_Switch:
+    case tt_Tributary: {
+        // both main + diverging branches
+        if( Track.SwitchExtension ) {
+            for( int i = 0; i < 2; ++i ) {
+                auto *seg = Track.SwitchExtension->Segments[ i ].get();
+                if( seg != nullptr ) { out.push_back( seg ); }
+            }
+        }
+        break;
+    }
+    case tt_Cross: {
+        // a road crossing potentially holds up to 6 connection segments; iterate them all,
+        // skipping zero-length / null entries.
+        if( Track.SwitchExtension ) {
+            for( int i = 0; i < 6; ++i ) {
+                auto *seg = Track.SwitchExtension->Segments[ i ].get();
+                if( seg == nullptr )    { continue; }
+                if( seg->GetLength() <= 0.0 ) { continue; }
+                out.push_back( seg );
+            }
+        }
+        break;
+    }
+    case tt_Normal:
+    case tt_Table:
+    default: {
+        if( Track.Segment ) { out.push_back( Track.Segment.get() ); }
+        break;
+    }
+    }
+    return out;
+}
+} // anonymous namespace
+
+// Resolves the sleeper model + (optional) replacable skin via the global model/material
+// managers, then walks every active sub-segment at the configured spacing and bakes a
+// local-space transform matrix (relative to m_origin) for every instance. The renderer
+// turns these into final camera-space modelview matrices at draw time.
+//
+// Per-instance orientation is built from an explicit tangent-based basis (right, up, forward)
+// rather than RPY-decomposed angles, so curves and switches stay aligned with the path
+// regardless of where they live in the parameter space.
+void TTrack::build_sleeper_transforms()
+{
+    m_sleeper_local_transforms.clear();
+    m_sleeper_model = nullptr;
+    m_sleeper_skin = 0;
+
+    if( false == m_sleeper_enabled )   { return; }
+    if( m_sleeper_model_name.empty() ) { return; }
+    if( m_sleeper_frequency <= 0.01f ) { return; }
+
+    auto const segments = sleeper_segments_for( *this );
+    if( segments.empty() ) { return; }
+
+    // resolve model
+    m_sleeper_model = TModelsManager::GetModel( m_sleeper_model_name, false );
+    if( m_sleeper_model == nullptr ) {
+        ErrorLog( "Bad track: sleepermodel model \"" + m_sleeper_model_name + "\" failed to load for track \"" + m_name + "\"" );
+        m_sleeper_enabled = false;
+        return;
+    }
+    // resolve replacable skin (optional)
+    if( ( false == m_sleeper_skin_name.empty() ) && ( m_sleeper_skin_name != "none" ) ) {
+        auto skinpath { m_sleeper_skin_name };
+        replace_slashes( skinpath );
+        m_sleeper_skin = GfxRenderer->Fetch_Material( skinpath );
+    }
+
+    auto const spacing = static_cast<double>( m_sleeper_frequency );
+    // small finite-difference epsilon used to extract the tangent from RaPositionGet.
+    // RaPositionGet's reported angles are correct in principle but for curves they're
+    // derived from the polynomial first derivative, which is sensitive to numerical noise
+    // at the segment endpoints. Sampling positions directly is robust for both straight
+    // segments and bezier curves, and it costs us two extra evaluations per sleeper.
+    double const eps = std::min( 0.1, spacing * 0.25 );
+    glm::vec3 const world_up { 0.f, 1.f, 0.f };
+
+    // user offset is (left/right, forward/back, up/down) in the local frame established by
+    // the basis below (x=right, y=up, z=forward). swap y<->z to match the documented axes.
+    glm::vec3 const local_offset { m_sleeper_offset.x, m_sleeper_offset.z, m_sleeper_offset.y };
+
+    for( auto *segment : segments ) {
+        auto const length = segment->GetLength();
+        if( length <= 0.0 ) { continue; }
+        // start half a frequency in so the first sleeper doesn't sit on the joint.
+        auto const start = std::min( spacing * 0.5, length * 0.5 );
+        auto const expected = static_cast<std::size_t>( std::max( 0.0, ( length - start ) / spacing ) ) + 1u;
+        m_sleeper_local_transforms.reserve( m_sleeper_local_transforms.size() + expected );
+
+        for( double s = start; s < length; s += spacing ) {
+            glm::dvec3 pos;
+            glm::vec3 angles;
+            segment->RaPositionGet( s, pos, angles );
+
+            // tangent direction via central difference (clamped to the segment endpoints)
+            auto const s_back = std::max( 0.0, s - eps );
+            auto const s_fwd  = std::min( length, s + eps );
+            glm::dvec3 p_back, p_fwd;
+            glm::vec3  dummy;
+            segment->RaPositionGet( s_back, p_back, dummy );
+            segment->RaPositionGet( s_fwd,  p_fwd,  dummy );
+            auto tangent = glm::vec3( p_fwd - p_back );
+            if( glm::length2( tangent ) < 1e-8f ) {
+                // degenerate sample (e.g. zero-length sub-segment); skip this position rather
+                // than emit a junk transform with NaN normals.
+                continue;
+            }
+            tangent = glm::normalize( tangent );
+
+            // build an orthonormal basis around the tangent. world up is the reference; if the
+            // track is almost vertical we fall back to world X so cross() doesn't collapse.
+            glm::vec3 up_ref = world_up;
+            if( std::abs( glm::dot( tangent, up_ref ) ) > 0.999f ) { up_ref = glm::vec3( 1.f, 0.f, 0.f ); }
+            glm::vec3 right = glm::normalize( glm::cross( up_ref, tangent ) );
+            glm::vec3 up    = glm::cross( tangent, right );
+
+            // apply track roll (banking) around the tangent / forward axis.
+            float const roll = angles.x;
+            if( roll != 0.f ) {
+                auto const roll_mat = glm::rotate( glm::mat4( 1.f ), roll, tangent );
+                right = glm::vec3( roll_mat * glm::vec4( right, 0.f ) );
+                up    = glm::vec3( roll_mat * glm::vec4( up,    0.f ) );
+            }
+
+            // assemble local transform: columns are (right, up, forward, translation).
+            // a sleeper modelled with X = sideways, Y = up, Z = along-track now ends up
+            // correctly oriented along the path tangent regardless of curve direction.
+            auto const localpos = glm::vec3( pos - m_origin );
+            glm::mat4 m { 1.f };
+            m[ 0 ] = glm::vec4( right,   0.f );
+            m[ 1 ] = glm::vec4( up,      0.f );
+            m[ 2 ] = glm::vec4( tangent, 0.f );
+            m[ 3 ] = glm::vec4( localpos, 1.f );
+
+            if( local_offset != glm::vec3( 0.f ) ) {
+                m = glm::translate( m, local_offset );
+            }
+            m_sleeper_local_transforms.emplace_back( m );
+        }
+    }
+}

world/Track.h

@@ -13,6 +13,9 @@ http://mozilla.org/MPL/2.0/.
 #include <vector>
 #include <deque>
 
+#include <glm/glm.hpp>
+#include <glm/mat4x4.hpp>
+
 #include "utilities/Classes.h"
 #include "world/Segment.h"
 #include "model/material.h"
@@ -183,6 +186,25 @@ public:
     std::vector<segment_data> m_paths; // source data for owned paths
 	int iterate_stamp = 0;
 
+    // sleepermodel optional parameter -------------------------------------------------
+    // Repeats a model along the path at fixed intervals (typically rail sleepers).
+    // Defined in scenery file as:
+    //   sleepermodel <frequency> <model> <skin> <offsetX> <offsetY> <offsetZ> <ballastZ>
+    // The renderer draws the model instances via GPU instancing and skips them entirely
+    // once the camera-to-track distance exceeds Global.SleeperDistance.
+    bool m_sleeper_enabled { false };
+    float m_sleeper_frequency { 0.6f }; // spacing along the path, in meters
+    std::string m_sleeper_model_name; // path to the e3d sleeper model (as written in the .scn)
+    std::string m_sleeper_skin_name;  // replacable skin path, or "none" for default
+    glm::vec3 m_sleeper_offset { 0.f, 0.f, 0.f }; // local offset from track centerline (x: left/right, y: forward/back, z: up/down)
+    float m_sleeper_ballast_z { 0.f }; // vertical offset applied to the trackbed (ballast) profile
+    TModel3d *m_sleeper_model { nullptr }; // resolved on init; nullptr means no model / failed to load
+    material_handle m_sleeper_skin { 0 }; // resolved replacable skin handle, 0 = use model defaults
+    // precomputed local-space transforms (relative to m_origin) for every sleeper instance along the path.
+    // Each matrix is translate(world_pos - m_origin) * rotate(direction, roll) * translate(local_offset).
+    // The renderer composes this with (view * translate(m_origin - camera_pos)) per draw.
+    std::vector<glm::mat4> m_sleeper_local_transforms;
+
 public:
     using dynamics_sequence = std::deque<TDynamicObject *>;
     using event_sequence = std::vector<std::pair<std::string, basic_event *> >;
@@ -346,6 +368,12 @@ private:
     void create_track_bed_profile( gfx::vertex_array &Output, TTrack const *Previous, TTrack const *Next );
     void create_road_profile( gfx::vertex_array &Output, bool const Forcetransition = false );
     void create_road_side_profile( gfx::vertex_array &Right, gfx::vertex_array &Left, gfx::vertex_array const &Road, bool const Forcetransition = false );
+    /// <summary>
+    /// resolves the sleeper model/skin via the model and material managers, and fills
+    /// m_sleeper_local_transforms by walking the active segment(s) at m_sleeper_frequency.
+    /// Safe to call multiple times; clears any previously cached transforms first.
+    /// </summary>
+    void build_sleeper_transforms();
 // members
     static profiles_array m_profiles; // shared database of path element profiles
     static profiles_map m_profilesmap;