feat: add forward+ rendering

2026-10-11 00:12:23 +02:00
parent c96b2282fb
commit 4bf87bd91b
26 changed files with 1346 additions and 191 deletions
@@ -5,6 +5,7 @@ set(SRC_FILES
        "src/Event.cpp"

        "src/Components/Animator.cpp"
+        "src/Components/LightComponent.cpp"
        "src/Components/MeshRendererComponent.cpp"
        "src/Components/Rotator.cpp"
        "src/Components/Spinner.cpp"
@@ -48,6 +49,7 @@ set(SRC_FILES
        "src/Graphics/Pipelines/LineRenderingPass.cpp"
        "src/Graphics/Pipelines/QuadRendererPass.cpp"
        "src/Graphics/Pipelines/SkyboxPipeline.cpp"
+        "src/Graphics/Pipelines/LightCullingPipeline.cpp"
        "src/Graphics/Pipelines/SkinningPipeline.cpp"
        "src/Graphics/Pipelines/ImguiPass.cpp"

@@ -0,0 +1,154 @@
+#version 460
+
+#extension GL_GOOGLE_include_directive : require
+#extension GL_EXT_buffer_reference : require
+#extension GL_EXT_scalar_block_layout : require
+
+#include "bindless.glsl"
+#include "scene_data.glsl"
+
+// Forward+ light culling.
+// One workgroup per LIGHT_TILE_SIZE x LIGHT_TILE_SIZE screen tile:
+//   1. find the min/max depth of the tile from the depth pre-pass
+//   2. build the tile frustum in view space
+//   3. test every light against it and write the visible ones to the tile's light list
+
+layout (push_constant, scalar) uniform constants
+{
+    SceneDataBuffer sceneData;
+    uvec2 screenSize;
+    uint depthTexID;
+    uint padding;
+} pcs;
+
+layout (local_size_x = LIGHT_TILE_SIZE, local_size_y = LIGHT_TILE_SIZE, local_size_z = 1) in;
+
+#define THREADS_PER_TILE (LIGHT_TILE_SIZE * LIGHT_TILE_SIZE)
+
+shared uint tileMinDepth;
+shared uint tileMaxDepth;
+shared uint tileLightCount;
+
+// left, right, bottom, top side planes (view space, through the camera origin)
+shared vec4 tilePlanes[4];
+// view space distances along -Z
+shared float tileNearDistance;
+shared float tileFarDistance;
+
+vec3 unprojectToView(vec2 ndc, float depth)
+{
+    vec4 p = pcs.sceneData.invProj * vec4(ndc, depth, 1.0);
+    return p.xyz / p.w;
+}
+
+// Plane through the camera origin and two points, normal facing towards `inside`
+vec4 makeSidePlane(vec3 a, vec3 b, vec3 inside)
+{
+    vec3 n = normalize(cross(a, b));
+    if (dot(n, inside) < 0.0) {
+        n = -n;
+    }
+    return vec4(n, 0.0);
+}
+
+bool sphereInsideTile(vec3 center, float radius)
+{
+    float distance = -center.z;
+    if (distance + radius < tileNearDistance || distance - radius > tileFarDistance) {
+        return false;
+    }
+
+    for (int i = 0; i < 4; i++) {
+        if (dot(tilePlanes[i].xyz, center) + tilePlanes[i].w < -radius) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void main()
+{
+    uint localIndex = gl_LocalInvocationIndex;
+    uvec2 tileID = gl_WorkGroupID.xy;
+    uint tileIndex = tileID.y * pcs.sceneData.tileCountX + tileID.x;
+    uint tileOffset = tileIndex * LIGHT_TILE_STRIDE;
+
+    if (localIndex == 0) {
+        tileMinDepth = 0xFFFFFFFFu;
+        tileMaxDepth = 0u;
+        tileLightCount = 0u;
+    }
+    memoryBarrierShared();
+    barrier();
+
+    // 1. min / max depth
+    uvec2 pixel = gl_GlobalInvocationID.xy;
+    if (all(lessThan(pixel, pcs.screenSize))) {
+        float depth = texelFetch(
+            sampler2D(textures[pcs.depthTexID], samplers[NEAREST_SAMPLER_ID]),
+            ivec2(pixel),
+            0).r;
+
+        // depth == 1.0 is the cleared far plane (sky), nothing there receives light.
+        // Depth is always positive, so its bits can be compared as uints.
+        if (depth < 1.0) {
+            uint depthBits = floatBitsToUint(depth);
+            atomicMin(tileMinDepth, depthBits);
+            atomicMax(tileMaxDepth, depthBits);
+        }
+    }
+    memoryBarrierShared();
+    barrier();
+
+    // 2. tile frustum
+    bool tileHasGeometry = tileMinDepth != 0xFFFFFFFFu;
+
+    if (localIndex == 0 && tileHasGeometry) {
+        vec2 screenSize = vec2(pcs.screenSize);
+        vec2 ndcMin = vec2(tileID * LIGHT_TILE_SIZE) / screenSize * 2.0 - 1.0;
+        vec2 ndcMax = vec2((tileID + 1) * LIGHT_TILE_SIZE) / screenSize * 2.0 - 1.0;
+        vec2 ndcCenter = (ndcMin + ndcMax) * 0.5;
+
+        vec3 corner00 = unprojectToView(vec2(ndcMin.x, ndcMin.y), 1.0);
+        vec3 corner10 = unprojectToView(vec2(ndcMax.x, ndcMin.y), 1.0);
+        vec3 corner11 = unprojectToView(vec2(ndcMax.x, ndcMax.y), 1.0);
+        vec3 corner01 = unprojectToView(vec2(ndcMin.x, ndcMax.y), 1.0);
+        vec3 center = unprojectToView(ndcCenter, 1.0);
+
+        tilePlanes[0] = makeSidePlane(corner01, corner00, center);
+        tilePlanes[1] = makeSidePlane(corner10, corner11, center);
+        tilePlanes[2] = makeSidePlane(corner00, corner10, center);
+        tilePlanes[3] = makeSidePlane(corner11, corner01, center);
+
+        tileNearDistance = -unprojectToView(ndcCenter, uintBitsToFloat(tileMinDepth)).z;
+        tileFarDistance = -unprojectToView(ndcCenter, uintBitsToFloat(tileMaxDepth)).z;
+    }
+    memoryBarrierShared();
+    barrier();
+
+    // 3. cull lights, every thread takes a slice of the light list
+    if (tileHasGeometry) {
+        for (uint i = localIndex; i < pcs.sceneData.numLights; i += THREADS_PER_TILE) {
+            Light light = pcs.sceneData.lights.data[i];
+
+            bool visible = light.type == TYPE_DIRECTIONAL_LIGHT;
+            if (!visible) {
+                vec3 viewPos = (pcs.sceneData.view * vec4(light.position, 1.0)).xyz;
+                visible = sphereInsideTile(viewPos, light.range);
+            }
+
+            if (visible) {
+                uint slot = atomicAdd(tileLightCount, 1u);
+                if (slot < MAX_LIGHTS_PER_TILE) {
+                    pcs.sceneData.tileLightIndices.data[tileOffset + 1 + slot] = i;
+                }
+            }
+        }
+    }
+    memoryBarrierShared();
+    barrier();
+
+    if (localIndex == 0) {
+        pcs.sceneData.tileLightIndices.data[tileOffset] = min(tileLightCount, uint(MAX_LIGHTS_PER_TILE));
+    }
+}
@@ -0,0 +1,96 @@
+#ifndef LIGHTING_GLSL
+#define LIGHTING_GLSL
+
+#include "lights.glsl"
+
+#define PI 3.14159265359
+
+// GGX / Trowbridge-Reitz normal distribution
+float distributionGGX(float NoH, float roughness)
+{
+    float a = roughness * roughness;
+    float a2 = a * a;
+    float d = NoH * NoH * (a2 - 1.0) + 1.0;
+    return a2 / (PI * d * d);
+}
+
+// Smith-Schlick GGX, k remapped for direct lighting
+float geometrySmithGGX(float NoV, float NoL, float roughness)
+{
+    float r = roughness + 1.0;
+    float k = (r * r) / 8.0;
+    float ggxV = NoV / (NoV * (1.0 - k) + k);
+    float ggxL = NoL / (NoL * (1.0 - k) + k);
+    return ggxV * ggxL;
+}
+
+vec3 fresnelSchlick(float VoH, vec3 f0)
+{
+    return f0 + (1.0 - f0) * pow(1.0 - VoH, 5.0);
+}
+
+// Inverse square falloff, windowed so it reaches exactly 0 at the light range.
+// The range is also the sphere the light culling uses, so no light gets cut off at a tile edge.
+float calculateDistanceAttenuation(float distance, float range)
+{
+    float window = clamp(1.0 - pow(distance / range, 4.0), 0.0, 1.0);
+    return (window * window) / max(distance * distance, 0.0001);
+}
+
+float calculateSpotAttenuation(Light light, vec3 l)
+{
+    float cd = dot(normalize(light.direction), -l);
+    return smoothstep(light.spotOuterCos, light.spotInnerCos, cd);
+}
+
+vec3 calculateLight(Light light, vec3 fragPos, vec3 n, vec3 v,
+                    vec3 diffuseColor, float roughness, float metallic, vec3 f0)
+{
+    vec3 l;
+    float attenuation = 1.0;
+
+    if (light.type == TYPE_DIRECTIONAL_LIGHT) {
+        l = normalize(-light.direction);
+    } else {
+        vec3 toLight = light.position - fragPos;
+        float distance = length(toLight);
+        l = toLight / max(distance, 0.0001);
+
+        attenuation = calculateDistanceAttenuation(distance, light.range);
+        if (light.type == TYPE_SPOT_LIGHT) {
+            attenuation *= calculateSpotAttenuation(light, l);
+        }
+    }
+
+    float NoL = clamp(dot(n, l), 0.0, 1.0);
+    if (NoL <= 0.0 || attenuation <= 0.0) {
+        return vec3(0.0);
+    }
+
+    vec3 h = normalize(v + l);
+    float NoV = clamp(dot(n, v), 0.0001, 1.0);
+    float NoH = clamp(dot(n, h), 0.0, 1.0);
+    float VoH = clamp(dot(v, h), 0.0, 1.0);
+
+    float D = distributionGGX(NoH, roughness);
+    float G = geometrySmithGGX(NoV, NoL, roughness);
+    vec3 F = fresnelSchlick(VoH, f0);
+
+    vec3 specular = (D * G * F) / (4.0 * NoV * NoL + 0.0001);
+    vec3 kd = (vec3(1.0) - F) * (1.0 - metallic);
+    vec3 diffuse = kd * diffuseColor / PI;
+
+    return (diffuse + specular) * light.color * light.intensity * attenuation * NoL;
+}
+
+// blue -> green -> red, used to visualise how many lights a tile has
+vec3 lightHeatmapColor(uint lightCount)
+{
+    float t = clamp(float(lightCount) / 16.0, 0.0, 1.0);
+    if (t < 0.5) {
+        return mix(vec3(0.0, 0.0, 1.0), vec3(0.0, 1.0, 0.0), t * 2.0);
+    }
+    return mix(vec3(0.0, 1.0, 0.0), vec3(1.0, 0.0, 0.0), (t - 0.5) * 2.0);
+}
+
+#endif // LIGHTING_GLSL
@@ -0,0 +1,41 @@
+#ifndef LIGHTS_GLSL
+#define LIGHTS_GLSL
+
+#extension GL_EXT_buffer_reference : require
+
+#define TYPE_DIRECTIONAL_LIGHT 0
+#define TYPE_POINT_LIGHT 1
+#define TYPE_SPOT_LIGHT 2
+
+// Forward+ tiling, must match LightCullingPipeline::TILE_SIZE / MAX_LIGHTS_PER_TILE
+#define LIGHT_TILE_SIZE 16
+#define MAX_LIGHTS_PER_TILE 255
+// Every tile stores its light count first, followed by MAX_LIGHTS_PER_TILE light indices
+#define LIGHT_TILE_STRIDE (MAX_LIGHTS_PER_TILE + 1)
+
+// Must match GPULight in Light.h
+struct Light {
+    vec3 position;
+    float range;
+
+    vec3 color;
+    float intensity;
+
+    // direction the light is pointing in (directional + spot lights)
+    vec3 direction;
+    uint type;
+
+    float spotInnerCos;
+    float spotOuterCos;
+    vec2 padding;
+};
+
+layout (buffer_reference, std430) readonly buffer LightsBuffer {
+    Light data[];
+};
+
+layout (buffer_reference, std430) buffer TileLightIndicesBuffer {
+    uint data[];
+};
+
+#endif // LIGHTS_GLSL
@@ -5,6 +5,7 @@
 #include "bindless.glsl"
 #include "scene_data.glsl"
 #include "mesh_pcs.glsl"
+#include "lighting.glsl"

 layout (location = 0) in vec3 inPos;
 layout (location = 1) in vec2 inUV;
@@ -20,16 +21,13 @@ void main()
    uint samplerID = pcs.sceneData.materials.data[pcs.materialID].textureFilteringMode;

    vec4 diffuse = sampleTexture2D(material.diffuseTex, samplerID, inUV) * material.baseColor;
-    outFragColor = diffuse;
-
+    vec3 baseColor = diffuse.rgb;

 //    if (diffuse.a < 0.1) {
 //        discard;
 //    }
-//
-//    vec3 baseColor = material.baseColor.rgb * diffuse.rgb;
-//
-//    vec3 normal = normalize(inNormal).rgb;
+
+    vec3 normal = normalize(inNormal);
 //    if (inTangent != vec4(0.0)) {
 //        // FIXME: sometimes Blender doesn't export tangents for some objects
 //        // for some reason. When we will start computing tangents manually,
@@ -39,91 +37,65 @@ void main()
 //        normal = inTBN * normalize(normal * 2.0 - 1.0);
 //        normal = normalize(normal);
 //    }
-//
-//    #ifdef PBR
-//    float metallicF = material.metallicRoughnessEmissive.r;
-//    float roughnessF = material.metallicRoughnessEmissive.g;
-//
-//    vec4 metallicRoughness = sampleTexture2DLinear(material.metallicRoughnessTex, inUV);
-//
-//    float roughness = roughnessF * metallicRoughness.g;
-//    // roughness *= roughness; // from perceptual to linear
-//    roughness = max(roughness, 1e-2); // 0.0 roughness leads to NaNs
-//
-//    float metallic = metallicF * metallicRoughness.b;
-//
-//    vec3 dielectricSpecular = vec3(0.04);
-//    vec3 black = vec3(0.0);
-//    vec3 diffuseColor = mix(baseColor * (1.0 - dielectricSpecular.r), black, metallic);
-//    vec3 f0 = mix(dielectricSpecular, baseColor, metallic);
-//    #else
-//    vec3 diffuseColor = baseColor;
-//    float metallic = 0.f;
-//    float roughness = 1.f;
-//    vec3 f0 = vec3(0.0);
-//    #endif
-//
-//    vec3 cameraPos = pcs.sceneData.cameraPos.xyz;
-//    vec3 fragPos = inPos.xyz;
-//    vec3 n = normal;
-//    vec3 v = normalize(cameraPos - fragPos);
-//
-//    // diffuseColor = vec3(1.0);
-//    // baseColor = vec3(1.0);
-//
-//    vec3 fragColor = vec3(0.0);
-//    for (int i = 0; i < pcs.sceneData.numLights; i++) {
-//        Light light = pcs.sceneData.lights.data[i];
-//
-//        vec3 l = light.direction;
-//        if (light.type != TYPE_DIRECTIONAL_LIGHT) {
-//            l = normalize(light.position - fragPos);
-//        }
-//        float NoL = clamp(dot(n, l), 0.0, 1.0);
-//
-//        float occlusion = 1.0;
-//        if (light.type == TYPE_DIRECTIONAL_LIGHT) {
-//            occlusion = calculateCSMOcclusion(
-//            fragPos, cameraPos, NoL,
-//            pcs.sceneData.csmShadowMapId,
-//            pcs.sceneData.cascadeFarPlaneZs,
-//            pcs.sceneData.csmLightSpaceTMs);
-//        } else if (light.type == TYPE_POINT_LIGHT && light.shadowMapID != 0) {
-//            occlusion = calculatePointShadow(
-//            fragPos, light.position, NoL,
-//            light.shadowMapID,
-//            pcs.sceneData.pointLightFarPlane);
-//        }
-//
-//        fragColor += calculateLight(light, fragPos, n, v, l,
-//        diffuseColor, roughness, metallic, f0, occlusion);
-//    }
-//
-//    // emissive
-//    float emissiveF = material.metallicRoughnessEmissive.b;
-//    vec3 emissiveColor = emissiveF * sampleTexture2DLinear(material.emissiveTex, inUV).rgb;
-//    fragColor += emissiveColor;
-//
-//    // ambient
-//    fragColor += baseColor * pcs.sceneData.ambientColor * pcs.sceneData.ambientIntensity;
-//
-//    #if 0
-//    // CSM DEBUG
-//    uint cascadeIndex = chooseCascade(inPos, cameraPos, pcs.sceneData.cascadeFarPlaneZs);
-//    fragColor *= debugShadowsFactor(cascadeIndex);
-//    #endif
-//
-//    #if 0
-//    // TANGENT DEBUG
-//    if (inTangent == vec4(0.0)) {
-//        fragColor = vec3(1.0f, 0.0f, 0.0f);
-//    }
-//    #endif
-//
-//    #if 1
-//    // NORMAL DEBUG
-//    // fragColor = normal;
-//    #endif

-//    outFragColor = vec4(1.0f, 0.0f, 0.0f, 1.0f);
+    float metallicF = material.metallicRoughnessEmissive.r;
+    float roughnessF = material.metallicRoughnessEmissive.g;
+
+    vec4 metallicRoughness = sampleTexture2DLinear(material.metallicRoughnessTex, inUV);
+
+    float roughness = roughnessF * metallicRoughness.g;
+    // roughness *= roughness; // from perceptual to linear
+    roughness = max(roughness, 1e-2); // 0.0 roughness leads to NaNs
+
+    float metallic = metallicF * metallicRoughness.b;
+
+    vec3 dielectricSpecular = vec3(0.04);
+    vec3 black = vec3(0.0);
+    vec3 diffuseColor = mix(baseColor * (1.0 - dielectricSpecular.r), black, metallic);
+    vec3 f0 = mix(dielectricSpecular, baseColor, metallic);
+
+    vec3 cameraPos = pcs.sceneData.cameraPos.xyz;
+    vec3 fragPos = inPos.xyz;
+    vec3 n = normal;
+    vec3 v = normalize(cameraPos - fragPos);
+
+    // forward+: only loop over the lights the culling pass put in this pixel's tile
+    uvec2 tileID = uvec2(gl_FragCoord.xy) / LIGHT_TILE_SIZE;
+    uint tileOffset = (tileID.y * pcs.sceneData.tileCountX + tileID.x) * LIGHT_TILE_STRIDE;
+    uint tileLightCount = pcs.sceneData.tileLightIndices.data[tileOffset];
+
+    vec3 fragColor = vec3(0.0);
+    for (uint i = 0; i < tileLightCount; i++) {
+        uint lightIndex = pcs.sceneData.tileLightIndices.data[tileOffset + 1 + i];
+        Light light = pcs.sceneData.lights.data[lightIndex];
+
+        fragColor += calculateLight(light, fragPos, n, v,
+            diffuseColor, roughness, metallic, f0);
+    }
+
+    // emissive
+    float emissiveF = material.metallicRoughnessEmissive.b;
+    vec3 emissiveColor = emissiveF * sampleTexture2DLinear(material.emissiveTex, inUV).rgb;
+    fragColor += emissiveColor;
+
+    // ambient
+    fragColor += baseColor * pcs.sceneData.ambientColor * pcs.sceneData.ambientIntensity;
+
+    if (pcs.sceneData.debugLightHeatmap != 0) {
+        fragColor = mix(fragColor, lightHeatmapColor(tileLightCount), 0.5);
+    }
+
+    #if 0
+    // TANGENT DEBUG
+    if (inTangent == vec4(0.0)) {
+        fragColor = vec3(1.0f, 0.0f, 0.0f);
+    }
+    #endif
+
+    #if 0
+    // NORMAL DEBUG
+    fragColor = normal;
+    #endif
+
+    outFragColor = vec4(fragColor, diffuse.a);
 }
@@ -10,6 +10,10 @@ layout (location = 2) out vec3 outNormal;
 layout (location = 3) out vec4 outTangent;
 layout (location = 4) out mat3 outTBN;

+// The depth pre-pass reuses this shader, positions have to match bit for bit
+// so the main pass can depth test against it with LESS_OR_EQUAL
+invariant gl_Position;
+
 void main()
 {
    Vertex v = pcs.vertexBuffer.vertices[gl_VertexIndex];
@@ -6,11 +6,14 @@
 #extension GL_EXT_buffer_reference : require

 #include "materials.glsl"
+#include "lights.glsl"

+// Must match GameRenderer::GPUSceneData
 layout (buffer_reference, scalar) readonly buffer SceneDataBuffer {
    // camera
    mat4 view;
    mat4 proj;
+    mat4 invProj;
    mat4 viewProj;
    vec4 cameraPos;

@@ -23,6 +26,14 @@ layout (buffer_reference, scalar) readonly buffer SceneDataBuffer {
    float fogDensity;

    MaterialsBuffer materials;
+
+    // forward+
+    LightsBuffer lights;
+    TileLightIndicesBuffer tileLightIndices;
+    uint numLights;
+    uint tileCountX;
+    uint debugLightHeatmap;
+    uint padding;
 } sceneDataBuffer;

 #endif // SCENE_DATA_GLSL
@@ -10,7 +10,6 @@ layout(location = 0) out vec4 outColor;
 layout(push_constant) uniform SkyboxPC {
    mat4 invViewProj;
    vec4 skyboxRot[3];
-    vec3 cameraPos;
    uint skyboxTextureId;
 } pcs;

@@ -22,7 +21,8 @@ void main()
    vec4 world = pcs.invViewProj * ndc;
    vec3 worldPos = world.xyz / world.w;

-    vec3 dir = normalize(worldPos - pcs.cameraPos);
+    // invViewProj has no translation, so the camera sits at the origin
+    vec3 dir = normalize(worldPos);
    dir.y *= -1.0;

    mat3 skyboxRot = mat3(
@@ -0,0 +1,51 @@
+#ifndef LIGHTCOMPONENT_H
+#define LIGHTCOMPONENT_H
+
+#include <destrum/ObjectModel/Component.h>
+#include <destrum/Graphics/Light.h>
+
+#include <glm/glm.hpp>
+
+class LightComponent final : public Component {
+public:
+    explicit LightComponent(GameObject& parent, LightType type = LightType::Point);
+
+    LightComponent(const LightComponent& other) = delete;
+    LightComponent(LightComponent&& other) noexcept = delete;
+    LightComponent& operator=(const LightComponent& other) = delete;
+    LightComponent& operator=(LightComponent&& other) noexcept = delete;
+
+    void Render(const RenderContext& ctx) override;
+
+    std::string GetTypeName() const override { return "LightComponent"; }
+
+    nlohmann::json Serialize() const override;
+    void Deserialize(const nlohmann::json& data) override;
+    void ImGuiInspector() override;
+
+    void SetType(LightType type) { m_Type = type; }
+    [[nodiscard]] LightType GetType() const { return m_Type; }
+
+    void SetColor(const glm::vec3& color) { m_Color = color; }
+    [[nodiscard]] const glm::vec3& GetColor() const { return m_Color; }
+
+    void SetIntensity(float intensity);
+    [[nodiscard]] float GetIntensity() const { return m_Intensity; }
+
+    void SetRange(float range);
+    [[nodiscard]] float GetRange() const { return m_Range; }
+
+    void SetConeAngles(float innerConeAngle, float outerConeAngle);
+    [[nodiscard]] float GetInnerConeAngle() const { return m_InnerConeAngle; }
+    [[nodiscard]] float GetOuterConeAngle() const { return m_OuterConeAngle; }
+
+private:
+    LightType m_Type{LightType::Point};
+    glm::vec3 m_Color{1.0f};
+    float m_Intensity{1.0f};
+    float m_Range{10.0f};
+    float m_InnerConeAngle{0.35f};
+    float m_OuterConeAngle{0.5f};
+};
+
+#endif // LIGHTCOMPONENT_H
@@ -0,0 +1,47 @@
+#ifndef LIGHT_H
+#define LIGHT_H
+
+#include <cstdint>
+
+#include <glm/vec2.hpp>
+#include <glm/vec3.hpp>
+
+enum class LightType : std::uint32_t {
+    Directional = 0,
+    Point = 1,
+    Spot = 2,
+};
+
+struct Light {
+    LightType type{LightType::Point};
+
+    glm::vec3 position{0.0f};
+    glm::vec3 direction{0.0f, -1.0f, 0.0f};
+
+    glm::vec3 color{1.0f};
+    float intensity{1.0f};
+
+    float range{10.0f};
+
+    float innerConeAngle{0.35f};
+    float outerConeAngle{0.5f};
+};
+
+struct GPULight {
+    glm::vec3 position;
+    float range;
+
+    glm::vec3 color;
+    float intensity;
+
+    glm::vec3 direction;
+    std::uint32_t type;
+
+    float spotInnerCos;
+    float spotOuterCos;
+    glm::vec2 padding;
+};
+
+static_assert(sizeof(GPULight) == 64, "GPULight must match the struct in lights.glsl");
+
+#endif //LIGHT_H
@@ -0,0 +1,78 @@
+#ifndef LIGHTCULLINGPIPELINE_H
+#define LIGHTCULLINGPIPELINE_H
+
+#include <vulkan/vulkan.h>
+#include <array>
+#include <memory>
+#include <glm/glm.hpp>
+
+#include <destrum/Graphics/ComputePipeline.h>
+#include <destrum/Graphics/Light.h>
+#include <destrum/Graphics/ids.h>
+#include <destrum/Graphics/Resources/AppendableBuffer.h>
+#include <destrum/Graphics/Resources/Buffer.h>
+
+class GfxDevice;
+class RenderResources;
+
+class LightCullingPipeline final {
+public:
+
+    static constexpr std::uint32_t TILE_SIZE = 16;
+    static constexpr std::uint32_t MAX_LIGHTS_PER_TILE = 255;
+    static constexpr std::uint32_t TILE_STRIDE = MAX_LIGHTS_PER_TILE + 1;
+
+    static constexpr std::size_t MAX_LIGHTS = 1024;
+
+    void init(GfxDevice& gfxDevice, RenderResources& resources, glm::ivec2 drawImageSize);
+    void cleanup(GfxDevice& gfxDevice);
+
+    void resize(GfxDevice& gfxDevice, glm::ivec2 drawImageSize);
+
+    void beginDrawing(std::size_t frameIndex);
+    void appendLight(const Light& light);
+
+    void flushCurrentFrame(VkCommandBuffer cmd, GfxDevice& gfxDevice);
+
+    void cullLights(
+        VkCommandBuffer cmd,
+        const RenderResources& resources,
+        const GPUBuffer& sceneDataBuffer,
+        ImageID depthImageId,
+        VkExtent2D renderExtent);
+
+    [[nodiscard]] VkDeviceAddress getLightsBufferAddress() const;
+    [[nodiscard]] VkDeviceAddress getTileLightIndicesBufferAddress() const { return tileLightIndicesBuffer.address; }
+
+    [[nodiscard]] std::uint32_t getNumLights() const;
+    [[nodiscard]] glm::uvec2 getTileCount() const { return tileCount; }
+
+private:
+    void createTileBuffer(GfxDevice& gfxDevice, glm::ivec2 drawImageSize);
+
+    VkPipelineLayout m_pipelineLayout{VK_NULL_HANDLE};
+    std::unique_ptr<ComputePipeline> lightCullingPipeline;
+
+    struct PushConstants {
+        VkDeviceAddress sceneDataBuffer;
+        glm::uvec2 screenSize;
+        std::uint32_t depthTexId;
+        std::uint32_t padding;
+    };
+
+    struct PerFrameData {
+        AppendableBuffer<GPULight> lightsBuffer;
+    };
+
+    std::array<PerFrameData, FRAMES_IN_FLIGHT> framesData;
+    std::size_t currentFrameIndex{0};
+    bool warnedAboutLightLimit{false};
+
+    GPUBuffer tileLightIndicesBuffer;
+    glm::uvec2 tileCount{0u, 0u};
+
+    PerFrameData& getCurrentFrameData();
+    const PerFrameData& getCurrentFrameData() const;
+};
+
+#endif //LIGHTCULLINGPIPELINE_H
@@ -27,6 +27,15 @@ public:
        VkFormat drawImageFormat,
        VkFormat depthImageFormat);

+    void drawDepthPrepass(
+        VkCommandBuffer cmd,
+        VkExtent2D renderExtent,
+        const RenderResources& resources,
+        const Camera& camera,
+        const GPUBuffer& sceneDataBuffer,
+        const std::vector<MeshDrawCommand>& drawCommands,
+        const std::vector<std::size_t>& sortedDrawCommands);
+
    void draw(
        VkCommandBuffer cmd,
        VkExtent2D renderExtent,
@@ -55,9 +64,20 @@ private:
        std::uint32_t padding;
    };

+    void drawMeshes(
+        VkCommandBuffer cmd,
+        const Pipeline& pipeline,
+        VkExtent2D renderExtent,
+        const RenderResources& resources,
+        const Camera& camera,
+        const GPUBuffer& sceneDataBuffer,
+        const std::vector<MeshDrawCommand>& drawCommands,
+        const std::vector<std::size_t>& sortedDrawCommands);
+
 private:
    VkPipelineLayout m_pipelineLayout{VK_NULL_HANDLE};
    std::unique_ptr<Pipeline> m_pipeline;
+    std::unique_ptr<Pipeline> m_depthPrepassPipeline;

    bool m_renderWireframe{false};
 };
@@ -35,7 +35,6 @@ private:
    struct SkyboxPushConstants {
        glm::mat4  invViewProj;
        glm::vec4  skyboxRot[3];
-        glm::vec3  cameraPos;
        std::uint32_t skyboxTextureId;
    };

@@ -57,7 +57,10 @@ public:
    BindlessSetManager& getBindlessSetManager();
    VkDescriptorSetLayout getBindlessDescSetLayout() const;
    const VkDescriptorSet& getBindlessDescSet() const;
-    void bindBindlessDescSet(VkCommandBuffer cmd, VkPipelineLayout layout) const;
+    void bindBindlessDescSet(
+        VkCommandBuffer cmd,
+        VkPipelineLayout layout,
+        VkPipelineBindPoint bindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS) const;

    MeshCache& meshes() { return *meshCache; }
    const MeshCache& meshes() const { return *meshCache; }
@@ -5,6 +5,8 @@
 #include <glm/vec3.hpp>

 #include <destrum/Graphics/Camera.h>
+#include <destrum/Graphics/Light.h>
+#include <destrum/Graphics/Pipelines/LightCullingPipeline.h>
 #include <destrum/Graphics/Pipelines/LineRenderingPass.h>
 #include <destrum/Graphics/Pipelines/MeshPipeline.h>
 #include <destrum/Graphics/Pipelines/QuadRendererPass.h>
@@ -20,9 +22,11 @@

 #include "Pipelines/SkyboxPipeline.h"

-class GameRenderer {
+class GameRenderer
+{
 public:
-    struct SceneData {
+    struct SceneData
+    {
        const Camera& camera;
        glm::vec3 ambientColor;
        float ambientIntensity;
@@ -40,11 +44,15 @@ public:
    void cleanup(GfxDevice& gfxDevice);

    void drawMesh(MeshID id, const glm::mat4& transform, MaterialID materialId);
-    void drawSkinnedMesh(MeshID id, const glm::mat4& transform, MaterialID materialId, SkinnedMesh* skinnedMesh, std::size_t jointMatricesStartIndex);
+    void drawSkinnedMesh(MeshID id, const glm::mat4& transform, MaterialID materialId, SkinnedMesh* skinnedMesh,
+                         std::size_t jointMatricesStartIndex);
+    void submitLight(const Light& light);
    const GPUImage& getDrawImage() const;

-    void resize(GfxDevice& gfxDevice, const glm::ivec2& newSize) {
+    void resize(GfxDevice& gfxDevice, const glm::ivec2& newSize)
+    {
        createDrawImage(gfxDevice, newSize, false);
+        lightCullingPipeline->resize(gfxDevice, newSize);
    }

    Material& getMaterialMutable(MaterialID id);
@@ -54,18 +62,31 @@ public:

    void flushMaterialUpdates(GfxDevice& gfxDevice);

-    [[nodiscard]] SkinningPipeline& getSkinningPipeline() const {
+    [[nodiscard]] SkinningPipeline& getSkinningPipeline() const
+    {
        return *skinningPipeline;
    }

-    void setRenderWireframe(bool wireframe) {
+    void setRenderWireframe(bool wireframe)
+    {
        meshPipeline->setRenderWireframe(wireframe);
    }

-    bool getRenderWireframe() {
+    bool getRenderWireframe()
+    {
        return meshPipeline->getRenderWireframe();
    }

+    void setRenderLightHeatmap(bool heatmap)
+    {
+        renderLightHeatmap = heatmap;
+    }
+
+    bool getRenderLightHeatmap() const
+    {
+        return renderLightHeatmap;
+    }
+
    RenderResources* getResources()
    {
        return resources;
@@ -96,33 +117,41 @@ private:
    VkSampleCountFlagBits samples{VK_SAMPLE_COUNT_1_BIT};


-    struct GPUSceneData {
-        // camera
+    struct GPUSceneData
+    {
        glm::mat4 view;
        glm::mat4 proj;
+        glm::mat4 invProj;
        glm::mat4 viewProj;
        glm::vec4 cameraPos;

-        // ambient
        glm::vec3 ambientColor;
        float ambientIntensity;

-        // fog
        glm::vec3 fogColor;
        float fogDensity;

        VkDeviceAddress materialsBuffer;
+
+        VkDeviceAddress lightsBuffer;
+        VkDeviceAddress tileLightIndicesBuffer;
+        std::uint32_t numLights;
+        std::uint32_t tileCountX;
+        std::uint32_t debugLightHeatmap;
+        std::uint32_t padding;
    };

    NBuffer sceneDataBuffer;

-
    std::unique_ptr<MeshPipeline> meshPipeline;
    std::unique_ptr<SkyboxPipeline> skyboxPipeline;
    std::unique_ptr<LineRenderingPass> lineRenderingPass;
    std::unique_ptr<QuadRendererPass> quadRendererPass;

    std::unique_ptr<SkinningPipeline> skinningPipeline;
+    std::unique_ptr<LightCullingPipeline> lightCullingPipeline;
+
+    bool renderLightHeatmap{false};
    bool initialized{false};
 };

@@ -1,9 +1,8 @@
 #ifndef DESTRUM_ENGINECOMPONENTLIST_H
 #define DESTRUM_ENGINECOMPONENTLIST_H

-// Keep the component manifest in one place. Consumers define X before
-// expanding one of these lists to generate registration or other metadata.
 #include <destrum/Components/Animator.h>
+#include <destrum/Components/LightComponent.h>
 #include <destrum/Components/MeshRendererComponent.h>
 #include <destrum/Components/OrbitAndSpin.h>
 #include <destrum/Components/Rotator.h>
@@ -15,6 +14,7 @@

 #define DESTRUM_ENGINE_COMPONENTS(X) \
    X(MeshRendererComponent) \
+    X(LightComponent) \
    X(Rotator) \
    X(Spinner) \
    X(OrbitAndSpin) \
@@ -22,11 +22,8 @@ public:
    static bool Load(Scene& scene, const std::filesystem::path& path,
                     std::function<void(float)> progressCallback = nullptr);

-    // Thread-safe: parse and validate a scene file on any thread.
-    // No engine state is touched during parsing.
    static LoadResult LoadSceneFile(const std::filesystem::path& path);

-    // Main-thread-only: construct a scene from pre-validated JSON.
    static bool ConstructScene(Scene& scene, LoadResult& result,
                               std::function<void(float)> progressCallback = nullptr);

@@ -0,0 +1,126 @@
+#include <destrum/Components/LightComponent.h>
+
+#include <destrum/Graphics/RenderContext.h>
+#include <destrum/ObjectModel/Transform.h>
+
+#include <glm/gtc/constants.hpp>
+#include <glm/gtc/quaternion.hpp>
+#include <glm/gtx/quaternion.hpp>
+
+#include <algorithm>
+#include <array>
+
+#include <imgui.h>
+
+LightComponent::LightComponent(GameObject& parent, LightType type)
+    : Component(parent, "LightComponent")
+    , m_Type(type)
+{
+}
+
+void LightComponent::SetIntensity(float intensity)
+{
+    m_Intensity = std::max(intensity, 0.0f);
+}
+
+void LightComponent::SetRange(float range)
+{
+    m_Range = std::max(range, 0.01f);
+}
+
+void LightComponent::SetConeAngles(float innerConeAngle, float outerConeAngle)
+{
+    m_InnerConeAngle = std::clamp(innerConeAngle, 0.0f, glm::half_pi<float>());
+    m_OuterConeAngle = std::clamp(outerConeAngle, m_InnerConeAngle, glm::half_pi<float>());
+}
+
+void LightComponent::Render(const RenderContext& ctx)
+{
+    auto& transform = GetTransform();
+
+    ctx.renderer.submitLight(Light{
+        .type = m_Type,
+        .position = transform.GetWorldPosition(),
+        .direction = transform.GetWorldRotation() * glm::vec3{0.0f, 0.0f, -1.0f},
+        .color = m_Color,
+        .intensity = m_Intensity,
+        .range = m_Range,
+        .innerConeAngle = m_InnerConeAngle,
+        .outerConeAngle = m_OuterConeAngle,
+    });
+}
+
+nlohmann::json LightComponent::Serialize() const {
+    return {
+        {"type", static_cast<std::uint32_t>(m_Type)},
+        {"color", {m_Color.x, m_Color.y, m_Color.z}},
+        {"intensity", m_Intensity},
+        {"range", m_Range},
+        {"innerConeAngle", m_InnerConeAngle},
+        {"outerConeAngle", m_OuterConeAngle}
+    };
+}
+
+void LightComponent::Deserialize(const nlohmann::json& data) {
+    if (data.contains("type")) {
+        const auto type = data.at("type").get<std::uint32_t>();
+        if (type <= static_cast<std::uint32_t>(LightType::Spot)) {
+            m_Type = static_cast<LightType>(type);
+        }
+    }
+    if (data.contains("color")) {
+        const auto& value = data.at("color");
+        m_Color = {value.at(0).get<float>(), value.at(1).get<float>(), value.at(2).get<float>()};
+    }
+    if (data.contains("intensity")) {
+        SetIntensity(data.at("intensity").get<float>());
+    }
+    if (data.contains("range")) {
+        SetRange(data.at("range").get<float>());
+    }
+
+    float innerConeAngle = m_InnerConeAngle;
+    float outerConeAngle = m_OuterConeAngle;
+    if (data.contains("innerConeAngle")) {
+        innerConeAngle = data.at("innerConeAngle").get<float>();
+    }
+    if (data.contains("outerConeAngle")) {
+        outerConeAngle = data.at("outerConeAngle").get<float>();
+    }
+    SetConeAngles(innerConeAngle, outerConeAngle);
+}
+
+void LightComponent::ImGuiInspector() {
+    constexpr std::array typeNames{"Directional", "Point", "Spot"};
+
+    int type = static_cast<int>(m_Type);
+    if (ImGui::Combo("Type", &type, typeNames.data(), static_cast<int>(typeNames.size()))) {
+        SetType(static_cast<LightType>(type));
+    }
+
+    ImGui::ColorEdit3("Color", &m_Color.x);
+
+    float intensity = m_Intensity;
+    if (ImGui::DragFloat("Intensity", &intensity, 0.05f, 0.0f)) {
+        SetIntensity(intensity);
+    }
+
+    if (m_Type == LightType::Directional) {
+        return;
+    }
+
+    float range = m_Range;
+    if (ImGui::DragFloat("Range", &range, 0.05f, 0.01f)) {
+        SetRange(range);
+    }
+
+    if (m_Type == LightType::Spot) {
+        float innerConeAngle = m_InnerConeAngle;
+        float outerConeAngle = m_OuterConeAngle;
+        bool changed = ImGui::SliderAngle("Inner Cone", &innerConeAngle, 0.0f, 90.0f);
+        changed |= ImGui::SliderAngle("Outer Cone", &outerConeAngle, 0.0f, 90.0f);
+        if (changed) {
+            SetConeAngles(innerConeAngle, outerConeAngle);
+        }
+    }
+}
@@ -0,0 +1,271 @@
+#include <destrum/Graphics/Pipelines/LightCullingPipeline.h>
+
+#include "destrum/FS/AssetFS.h"
+#include "destrum/Graphics/GfxDevice.h"
+#include "destrum/Graphics/RenderResources.h"
+#include "destrum/Graphics/Util.h"
+
+#include <array>
+#include <cassert>
+#include <cmath>
+#include <cstdint>
+#include <stdexcept>
+
+#include "volk.h"
+#include "spdlog/spdlog.h"
+
+void LightCullingPipeline::init(GfxDevice& gfxDevice, RenderResources& resources, glm::ivec2 drawImageSize)
+{
+    try
+    {
+        const auto& device = gfxDevice.getDevice();
+
+        const auto pushConstant = VkPushConstantRange{
+            .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+            .offset = 0,
+            .size = sizeof(PushConstants),
+        };
+
+        const auto pushConstants = std::array{pushConstant};
+
+        const auto layouts = std::array{
+            resources.getBindlessDescSetLayout()
+        };
+
+        VkPipelineLayoutCreateInfo pipelineLayoutInfo{};
+        pipelineLayoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
+        pipelineLayoutInfo.setLayoutCount = static_cast<std::uint32_t>(layouts.size());
+        pipelineLayoutInfo.pSetLayouts = layouts.data();
+        pipelineLayoutInfo.pushConstantRangeCount = static_cast<std::uint32_t>(pushConstants.size());
+        pipelineLayoutInfo.pPushConstantRanges = pushConstants.data();
+
+        if (vkCreatePipelineLayout(device, &pipelineLayoutInfo, nullptr, &m_pipelineLayout) != VK_SUCCESS)
+        {
+            throw std::runtime_error("Could not make light culling pipeline layout");
+        }
+
+        ComputePipelineConfigInfo pipelineConfig{};
+        pipelineConfig.name = "light culling compute pipeline";
+        pipelineConfig.pipelineLayout = m_pipelineLayout;
+
+        const auto lightCullingShaderPath = AssetFS::GetInstance().GetCookedPathForFile(
+            "engine://shaders/light_culling.comp");
+
+        lightCullingPipeline = std::make_unique<ComputePipeline>(
+            gfxDevice,
+            lightCullingShaderPath.generic_string(),
+            pipelineConfig
+        );
+
+        for (std::size_t i = 0; i < FRAMES_IN_FLIGHT; ++i)
+        {
+            auto& lightsBuffer = framesData[i].lightsBuffer;
+            lightsBuffer.capacity = MAX_LIGHTS;
+            lightsBuffer.buffer = gfxDevice.createBuffer(
+                MAX_LIGHTS * sizeof(GPULight),
+                VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT,
+                VMA_MEMORY_USAGE_CPU_TO_GPU);
+            vkutil::addDebugLabel(device, lightsBuffer.buffer.buffer, "lights");
+        }
+
+        createTileBuffer(gfxDevice, drawImageSize);
+    }
+    catch (...)
+    {
+        cleanup(gfxDevice);
+        throw;
+    }
+}
+
+void LightCullingPipeline::cleanup(GfxDevice& gfxDevice)
+{
+    for (auto& frame : framesData)
+    {
+        gfxDevice.destroyBuffer(frame.lightsBuffer.buffer);
+        frame.lightsBuffer.buffer = {};
+        frame.lightsBuffer.size = 0;
+    }
+
+    gfxDevice.destroyBuffer(tileLightIndicesBuffer);
+    tileLightIndicesBuffer = {};
+    tileCount = glm::uvec2{0u, 0u};
+
+    if (m_pipelineLayout != VK_NULL_HANDLE && gfxDevice.getDevice() != VK_NULL_HANDLE)
+    {
+        vkDestroyPipelineLayout(gfxDevice.getDevice(), m_pipelineLayout, nullptr);
+    }
+    m_pipelineLayout = VK_NULL_HANDLE;
+    lightCullingPipeline.reset();
+}
+
+void LightCullingPipeline::resize(GfxDevice& gfxDevice, glm::ivec2 drawImageSize)
+{
+    gfxDevice.waitIdle();
+
+    gfxDevice.destroyBuffer(tileLightIndicesBuffer);
+    tileLightIndicesBuffer = {};
+    createTileBuffer(gfxDevice, drawImageSize);
+}
+
+void LightCullingPipeline::createTileBuffer(GfxDevice& gfxDevice, glm::ivec2 drawImageSize)
+{
+    assert(drawImageSize.x > 0 && drawImageSize.y > 0);
+
+    tileCount = glm::uvec2{
+        (static_cast<std::uint32_t>(drawImageSize.x) + TILE_SIZE - 1) / TILE_SIZE,
+        (static_cast<std::uint32_t>(drawImageSize.y) + TILE_SIZE - 1) / TILE_SIZE,
+    };
+
+    const std::size_t numTiles = static_cast<std::size_t>(tileCount.x) * tileCount.y;
+    tileLightIndicesBuffer = gfxDevice.createBuffer(
+        numTiles * TILE_STRIDE * sizeof(std::uint32_t),
+        VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT);
+    vkutil::addDebugLabel(gfxDevice.getDevice(), tileLightIndicesBuffer.buffer, "tile light indices");
+
+    spdlog::debug("Forward+ light grid: {}x{} tiles", tileCount.x, tileCount.y);
+}
+
+void LightCullingPipeline::beginDrawing(std::size_t frameIndex)
+{
+    currentFrameIndex = frameIndex;
+    getCurrentFrameData().lightsBuffer.clear();
+}
+
+void LightCullingPipeline::appendLight(const Light& light)
+{
+    auto& lightsBuffer = getCurrentFrameData().lightsBuffer;
+    if (lightsBuffer.size >= lightsBuffer.capacity)
+    {
+        if (!warnedAboutLightLimit)
+        {
+            spdlog::warn("Light limit of {} reached, extra lights are ignored", MAX_LIGHTS);
+            warnedAboutLightLimit = true;
+        }
+        return;
+    }
+
+    const float directionLength = glm::length(light.direction);
+    const glm::vec3 direction = directionLength > 1e-6f
+                                    ? light.direction / directionLength
+                                    : glm::vec3{0.0f, -1.0f, 0.0f};
+
+    const auto gpuLight = GPULight{
+        .position = light.position,
+        .range = light.range,
+        .color = light.color,
+        .intensity = light.intensity,
+        .direction = direction,
+        .type = static_cast<std::uint32_t>(light.type),
+        .spotInnerCos = std::cos(light.innerConeAngle),
+        .spotOuterCos = std::cos(light.outerConeAngle),
+        .padding = glm::vec2{0.0f},
+    };
+    lightsBuffer.append(gpuLight);
+}
+
+void LightCullingPipeline::flushCurrentFrame(VkCommandBuffer cmd, GfxDevice& gfxDevice)
+{
+    auto& buffer = getCurrentFrameData().lightsBuffer.buffer;
+    gfxDevice.getMemoryManager().flushAllocation(buffer, 0, VK_WHOLE_SIZE);
+    vkutil::bufferHostWriteToShaderReadBarrier(
+        cmd,
+        buffer.buffer,
+        0,
+        static_cast<VkDeviceSize>(MAX_LIGHTS * sizeof(GPULight)));
+}
+
+void LightCullingPipeline::cullLights(
+    VkCommandBuffer cmd,
+    const RenderResources& resources,
+    const GPUBuffer& sceneDataBuffer,
+    ImageID depthImageId,
+    VkExtent2D renderExtent)
+{
+    if (!lightCullingPipeline)
+    {
+        return;
+    }
+
+    {
+        // previous frame's mesh pass might still be reading the tile lists
+        const VkBufferMemoryBarrier2 tileBufferBarrier{
+            .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
+            .srcStageMask = VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT,
+            .srcAccessMask = VK_ACCESS_2_SHADER_STORAGE_READ_BIT,
+            .dstStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
+            .dstAccessMask = VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .buffer = tileLightIndicesBuffer.buffer,
+            .offset = 0,
+            .size = VK_WHOLE_SIZE,
+        };
+
+        const VkDependencyInfo dependencyInfo{
+            .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+            .bufferMemoryBarrierCount = 1,
+            .pBufferMemoryBarriers = &tileBufferBarrier,
+        };
+        vkCmdPipelineBarrier2(cmd, &dependencyInfo);
+    }
+
+    lightCullingPipeline->bind(cmd);
+    resources.bindBindlessDescSet(cmd, m_pipelineLayout, VK_PIPELINE_BIND_POINT_COMPUTE);
+
+    const auto cs = PushConstants{
+        .sceneDataBuffer = sceneDataBuffer.address,
+        .screenSize = glm::uvec2{renderExtent.width, renderExtent.height},
+        .depthTexId = static_cast<std::uint32_t>(depthImageId),
+        .padding = 0,
+    };
+
+    vkCmdPushConstants(cmd,
+                       m_pipelineLayout,
+                       VK_SHADER_STAGE_COMPUTE_BIT,
+                       0,
+                       sizeof(PushConstants),
+                       &cs);
+
+    // one workgroup per tile
+    vkCmdDispatch(cmd, tileCount.x, tileCount.y, 1);
+
+    const VkBufferMemoryBarrier2 tileBufferBarrier{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
+        .srcStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
+        .srcAccessMask = VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT,
+        .dstStageMask = VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT,
+        .dstAccessMask = VK_ACCESS_2_SHADER_STORAGE_READ_BIT,
+        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .buffer = tileLightIndicesBuffer.buffer,
+        .offset = 0,
+        .size = VK_WHOLE_SIZE,
+    };
+
+    const VkDependencyInfo dependencyInfo{
+        .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
+        .bufferMemoryBarrierCount = 1,
+        .pBufferMemoryBarriers = &tileBufferBarrier,
+    };
+    vkCmdPipelineBarrier2(cmd, &dependencyInfo);
+}
+
+VkDeviceAddress LightCullingPipeline::getLightsBufferAddress() const
+{
+    return getCurrentFrameData().lightsBuffer.buffer.address;
+}
+
+std::uint32_t LightCullingPipeline::getNumLights() const
+{
+    return static_cast<std::uint32_t>(getCurrentFrameData().lightsBuffer.size);
+}
+
+LightCullingPipeline::PerFrameData& LightCullingPipeline::getCurrentFrameData()
+{
+    return framesData[currentFrameIndex % FRAMES_IN_FLIGHT];
+}
+
+const LightCullingPipeline::PerFrameData& LightCullingPipeline::getCurrentFrameData() const
+{
+    return framesData[currentFrameIndex % FRAMES_IN_FLIGHT];
+}
@@ -23,66 +23,122 @@ void MeshPipeline::init(
    VkFormat drawImageFormat,
    VkFormat depthImageFormat)
 {
-    try {
-    const auto vertexShader =
-        AssetFS::GetInstance().GetCookedPathForFile("engine://shaders/mesh.vert");
+    try
+    {
+        const auto vertexShader =
+            AssetFS::GetInstance().GetCookedPathForFile("engine://shaders/mesh.vert");

-    const auto fragShader =
-        AssetFS::GetInstance().GetCookedPathForFile("engine://shaders/mesh.frag");
+        const auto fragShader =
+            AssetFS::GetInstance().GetCookedPathForFile("engine://shaders/mesh.frag");

-    const auto bufferRange = VkPushConstantRange{
-        .stageFlags = VK_SHADER_STAGE_VERTEX_BIT | VK_SHADER_STAGE_FRAGMENT_BIT,
-        .offset = 0,
-        .size = sizeof(PushConstants),
-    };
+        const auto bufferRange = VkPushConstantRange{
+            .stageFlags = VK_SHADER_STAGE_VERTEX_BIT | VK_SHADER_STAGE_FRAGMENT_BIT,
+            .offset = 0,
+            .size = sizeof(PushConstants),
+        };

-    const auto pushConstantRanges = std::array{bufferRange};
+        const auto pushConstantRanges = std::array{bufferRange};

-    const auto layouts = std::array{
-        resources.getBindlessDescSetLayout()
-    };
+        const auto layouts = std::array{
+            resources.getBindlessDescSetLayout()
+        };

-    VkPipelineLayoutCreateInfo pipelineLayoutInfo{};
-    pipelineLayoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
-    pipelineLayoutInfo.setLayoutCount = static_cast<std::uint32_t>(layouts.size());
-    pipelineLayoutInfo.pSetLayouts = layouts.data();
-    pipelineLayoutInfo.pushConstantRangeCount =
-        static_cast<std::uint32_t>(pushConstantRanges.size());
-    pipelineLayoutInfo.pPushConstantRanges = pushConstantRanges.data();
+        VkPipelineLayoutCreateInfo pipelineLayoutInfo{};
+        pipelineLayoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
+        pipelineLayoutInfo.setLayoutCount = static_cast<std::uint32_t>(layouts.size());
+        pipelineLayoutInfo.pSetLayouts = layouts.data();
+        pipelineLayoutInfo.pushConstantRangeCount =
+            static_cast<std::uint32_t>(pushConstantRanges.size());
+        pipelineLayoutInfo.pPushConstantRanges = pushConstantRanges.data();

-    if (vkCreatePipelineLayout(
+        if (vkCreatePipelineLayout(
            gfxDevice.getDevice(),
            &pipelineLayoutInfo,
            nullptr,
            &m_pipelineLayout) != VK_SUCCESS)
-    {
-        throw std::runtime_error("Could not create mesh pipeline layout");
+        {
+            throw std::runtime_error("Could not create mesh pipeline layout");
+        }
+
+        {
+            PipelineConfigInfo pipelineConfig{};
+            Pipeline::DefaultPipelineConfigInfo(pipelineConfig);
+
+            pipelineConfig.name = "Mesh Depth Prepass Pipeline";
+            pipelineConfig.pipelineLayout = m_pipelineLayout;
+
+            pipelineConfig.vertexAttributeDescriptions = {};
+            pipelineConfig.vertexBindingDescriptions = {};
+
+            pipelineConfig.colorAttachments = {};
+            pipelineConfig.colorBlendInfo.attachmentCount = 0;
+            pipelineConfig.depthAttachment = depthImageFormat;
+
+            m_depthPrepassPipeline = std::make_unique<Pipeline>(
+                gfxDevice,
+                vertexShader.string(),
+                "",
+                pipelineConfig
+            );
+        }
+
+        {
+            PipelineConfigInfo pipelineConfig{};
+            Pipeline::DefaultPipelineConfigInfo(pipelineConfig);
+
+            pipelineConfig.name = "Mesh Pipeline";
+            pipelineConfig.pipelineLayout = m_pipelineLayout;
+
+            pipelineConfig.vertexAttributeDescriptions = {};
+            pipelineConfig.vertexBindingDescriptions = {};
+
+            pipelineConfig.colorAttachments = {drawImageFormat};
+            pipelineConfig.depthAttachment = depthImageFormat;
+
+            // depth was already written by the pre-pass, only shade the visible surface
+            pipelineConfig.depthStencilInfo.depthWriteEnable = VK_FALSE;
+            pipelineConfig.depthStencilInfo.depthCompareOp = VK_COMPARE_OP_LESS_OR_EQUAL;
+
+            m_pipeline = std::make_unique<Pipeline>(
+                gfxDevice,
+                vertexShader.string(),
+                fragShader.string(),
+                pipelineConfig
+            );
+        }
    }
-
-    PipelineConfigInfo pipelineConfig{};
-    Pipeline::DefaultPipelineConfigInfo(pipelineConfig);
-
-    pipelineConfig.name = "Mesh Pipeline";
-    pipelineConfig.pipelineLayout = m_pipelineLayout;
-
-    pipelineConfig.vertexAttributeDescriptions = {};
-    pipelineConfig.vertexBindingDescriptions = {};
-
-    pipelineConfig.colorAttachments = {drawImageFormat};
-    pipelineConfig.depthAttachment = depthImageFormat;
-
-        m_pipeline = std::make_unique<Pipeline>(
-        gfxDevice,
-        vertexShader.string(),
-        fragShader.string(),
-        pipelineConfig
-        );
-    } catch (...) {
+    catch (...)
+    {
        cleanup(gfxDevice.getDevice());
        throw;
    }
 }

+void MeshPipeline::drawDepthPrepass(
+    VkCommandBuffer cmd,
+    VkExtent2D renderExtent,
+    const RenderResources& resources,
+    const Camera& camera,
+    const GPUBuffer& sceneDataBuffer,
+    const std::vector<MeshDrawCommand>& drawCommands,
+    const std::vector<std::size_t>& sortedDrawCommands)
+{
+    if (!m_depthPrepassPipeline)
+    {
+        return;
+    }
+
+    drawMeshes(
+        cmd,
+        *m_depthPrepassPipeline,
+        renderExtent,
+        resources,
+        camera,
+        sceneDataBuffer,
+        drawCommands,
+        sortedDrawCommands);
+}
+
 void MeshPipeline::draw(
    VkCommandBuffer cmd,
    VkExtent2D renderExtent,
@@ -92,13 +148,35 @@ void MeshPipeline::draw(
    const std::vector<MeshDrawCommand>& drawCommands,
    const std::vector<std::size_t>& sortedDrawCommands)
 {
-    if (!m_pipeline) {
+    if (!m_pipeline)
+    {
        return;
    }

+    drawMeshes(
+        cmd,
+        *m_pipeline,
+        renderExtent,
+        resources,
+        camera,
+        sceneDataBuffer,
+        drawCommands,
+        sortedDrawCommands);
+}
+
+void MeshPipeline::drawMeshes(
+    VkCommandBuffer cmd,
+    const Pipeline& pipeline,
+    VkExtent2D renderExtent,
+    const RenderResources& resources,
+    const Camera& camera,
+    const GPUBuffer& sceneDataBuffer,
+    const std::vector<MeshDrawCommand>& drawCommands,
+    const std::vector<std::size_t>& sortedDrawCommands)
+{
    const MeshCache& meshCache = resources.meshes();

-    m_pipeline->bind(cmd);
+    pipeline.bind(cmd);

    resources.bindBindlessDescSet(cmd, m_pipelineLayout);

@@ -130,8 +208,10 @@ void MeshPipeline::draw(

    const auto frustum = edge::createFrustumFromCamera(camera);

-    const auto drawOne = [&](const MeshDrawCommand& dc) {
-        if (!edge::isInFrustum(frustum, dc.worldBoundingSphere)) {
+    const auto drawOne = [&](const MeshDrawCommand& dc)
+    {
+        if (!edge::isInFrustum(frustum, dc.worldBoundingSphere))
+        {
            return;
        }

@@ -139,7 +219,8 @@ void MeshPipeline::draw(

        const auto& mesh = meshCache.getMesh(dc.meshId);

-        if (dc.meshId != prevMeshId) {
+        if (dc.meshId != prevMeshId)
+        {
            prevMeshId = dc.meshId;

            vkCmdBindIndexBuffer(
@@ -156,9 +237,9 @@ void MeshPipeline::draw(
            .transform = dc.transformMatrix,
            .sceneDataBuffer = sceneDataBuffer.address,
            .vertexBuffer =
-                dc.skinnedMesh != nullptr
-                    ? dc.skinnedMesh->skinnedVertexBuffer.address
-                    : mesh.vertexBuffer.address,
+            dc.skinnedMesh != nullptr
+                ? dc.skinnedMesh->skinnedVertexBuffer.address
+                : mesh.vertexBuffer.address,
            .materialId = dc.materialId,
            .padding = 0,
        };
@@ -175,26 +256,30 @@ void MeshPipeline::draw(
        vkCmdDrawIndexed(cmd, mesh.numIndices, 1, 0, 0, 0);
    };

-    if (!sortedDrawCommands.empty()) {
-        for (std::size_t drawIndex : sortedDrawCommands) {
+    if (!sortedDrawCommands.empty())
+    {
+        for (std::size_t drawIndex : sortedDrawCommands)
+        {
            assert(drawIndex < drawCommands.size());
            drawOne(drawCommands[drawIndex]);
        }
-    } else {
-        for (const MeshDrawCommand& dc : drawCommands) {
+    }
+    else
+    {
+        for (const MeshDrawCommand& dc : drawCommands)
+        {
            drawOne(dc);
        }
    }
-
-    // Optional:
-    // spdlog::debug("Actual mesh draw calls: {}", actualDrawCalls);
 }

 void MeshPipeline::cleanup(VkDevice device)
 {
    m_pipeline.reset();
+    m_depthPrepassPipeline.reset();

-    if (m_pipelineLayout != VK_NULL_HANDLE && device != VK_NULL_HANDLE) {
+    if (m_pipelineLayout != VK_NULL_HANDLE && device != VK_NULL_HANDLE)
+    {
        vkDestroyPipelineLayout(device, m_pipelineLayout, nullptr);
    }
    m_pipelineLayout = VK_NULL_HANDLE;
@@ -119,14 +119,16 @@ void SkyboxPipeline::draw(

    const glm::mat3 r = glm::mat3(skyboxRotation);

+    // rotation only, translation in the inverse makes the sky jitter far from the origin
+    const glm::mat4 viewRotation = glm::mat4(glm::mat3(camera.GetViewMatrix()));
+
    const auto pcs = SkyboxPushConstants{
-        .invViewProj = glm::inverse(camera.GetViewProjectionMatrix()),
+        .invViewProj = glm::inverse(camera.GetProjectionMatrix() * viewRotation),
        .skyboxRot = {
            glm::vec4(r[0], 0.f),
            glm::vec4(r[1], 0.f),
            glm::vec4(r[2], 0.f)
        },
-        .cameraPos = camera.GetPosition(),
        .skyboxTextureId = static_cast<std::uint32_t>(skyboxTextureId),
    };

@@ -266,10 +266,13 @@ const VkDescriptorSet& RenderResources::getBindlessDescSet() const {
    return imageCache->bindlessSetManager.getDescSet();
 }

-void RenderResources::bindBindlessDescSet(VkCommandBuffer cmd, VkPipelineLayout layout) const {
+void RenderResources::bindBindlessDescSet(
+    VkCommandBuffer cmd,
+    VkPipelineLayout layout,
+    VkPipelineBindPoint bindPoint) const {
    vkCmdBindDescriptorSets(
        cmd,
-        VK_PIPELINE_BIND_POINT_GRAPHICS,
+        bindPoint,
        layout,
        0,
        1,
@@ -21,7 +21,8 @@ GameRenderer::GameRenderer()

 void GameRenderer::init(GfxDevice& gfxDevice, RenderResources& _resources, glm::ivec2 drawImageSize)
 {
-    try {
+    try
+    {
        resources = &_resources;
        sceneDataBuffer.init(
            gfxDevice,
@@ -34,12 +35,15 @@ void GameRenderer::init(GfxDevice& gfxDevice, RenderResources& _resources, glm::
        meshPipeline = std::make_unique<MeshPipeline>();
        meshPipeline->init(gfxDevice, _resources, drawImageFormat, depthImageFormat);

-skyboxPipeline = std::make_unique<SkyboxPipeline>();
+        skyboxPipeline = std::make_unique<SkyboxPipeline>();
        skyboxPipeline->init(gfxDevice, _resources, drawImageFormat, depthImageFormat);

        skinningPipeline = std::make_unique<SkinningPipeline>();
        skinningPipeline->init(gfxDevice);

+        lightCullingPipeline = std::make_unique<LightCullingPipeline>();
+        lightCullingPipeline->init(gfxDevice, _resources, drawImageSize);
+
        lineRenderingPass = std::make_unique<LineRenderingPass>();
        lineRenderingPass->init(gfxDevice, drawImageFormat);

@@ -48,7 +52,9 @@ skyboxPipeline = std::make_unique<SkyboxPipeline>();

        GameState::GetInstance().SetRenderer(this);
        initialized = true;
-    } catch (...) {
+    }
+    catch (...)
+    {
        cleanup(gfxDevice);
        throw;
    }
@@ -59,6 +65,7 @@ void GameRenderer::beginDrawing(GfxDevice& gfxDevice)
    flushMaterialUpdates(gfxDevice);
    meshDrawCommands.clear();
    skinningPipeline->beginDrawing(gfxDevice.getCurrentFrameIndex());
+    lightCullingPipeline->beginDrawing(gfxDevice.getCurrentFrameIndex());
    QuadRenderingManager::GetInstance().ClearQuads();
 }

@@ -70,13 +77,16 @@ void GameRenderer::endDrawing()
    std::stable_sort(
        sortedMeshDrawCommands.begin(),
        sortedMeshDrawCommands.end(),
-        [this](std::size_t left, std::size_t right) {
+        [this](std::size_t left, std::size_t right)
+        {
            const auto& lhs = meshDrawCommands[left];
            const auto& rhs = meshDrawCommands[right];
-            if (lhs.meshId != rhs.meshId) {
+            if (lhs.meshId != rhs.meshId)
+            {
                return lhs.meshId < rhs.meshId;
            }
-            if (lhs.materialId != rhs.materialId) {
+            if (lhs.materialId != rhs.materialId)
+            {
                return lhs.materialId < rhs.materialId;
            }
            return (lhs.skinnedMesh != nullptr) < (rhs.skinnedMesh != nullptr);
@@ -106,9 +116,17 @@ void GameRenderer::draw(VkCommandBuffer cmd, GfxDevice& gfxDevice, const Camera&
        }
    }

+    {
+        TracyVkZone(gfxDevice.getTracyVkCtx(), cmd, "Lights Buffer Flush");
+
+        lightCullingPipeline->flushCurrentFrame(cmd, gfxDevice);
+    }
+
+    // TODO: there has to be a better way to do this
    const auto gpuSceneData = GPUSceneData{
        .view = sceneData.camera.GetViewMatrix(),
        .proj = sceneData.camera.GetProjectionMatrix(),
+        .invProj = glm::inverse(sceneData.camera.GetProjectionMatrix()),
        .viewProj = sceneData.camera.GetViewProjectionMatrix(),
        .cameraPos = glm::vec4{sceneData.camera.m_position, 1.f},
        .ambientColor = {sceneData.ambientColor},
@@ -116,6 +134,12 @@ void GameRenderer::draw(VkCommandBuffer cmd, GfxDevice& gfxDevice, const Camera&
        .fogColor = {sceneData.fogColor},
        .fogDensity = sceneData.fogDensity,
        .materialsBuffer = resources->materials().getMaterialDataBufferAddress(),
+        .lightsBuffer = lightCullingPipeline->getLightsBufferAddress(),
+        .tileLightIndicesBuffer = lightCullingPipeline->getTileLightIndicesBufferAddress(),
+        .numLights = lightCullingPipeline->getNumLights(),
+        .tileCountX = lightCullingPipeline->getTileCount().x,
+        .debugLightHeatmap = renderLightHeatmap ? 1u : 0u,
+        .padding = 0,
    };

    {
@@ -165,12 +189,60 @@ void GameRenderer::draw(VkCommandBuffer cmd, GfxDevice& gfxDevice, const Camera&
        depthImage.setLayout(VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL);
    }

+    {
+        TracyVkZone(gfxDevice.getTracyVkCtx(), cmd, "Depth Prepass");
+
+        auto depthPrepassInfo = vkutil::createRenderingInfo({
+            .renderExtent = drawImage.getExtent2D(),
+            .depthImageView = depthImage.imageView,
+            .depthImageClearValue = 1.f,
+        });
+        depthPrepassInfo.renderingInfo.pDepthAttachment = &depthPrepassInfo.depthAttachment;
+
+        vkCmdBeginRendering(cmd, &depthPrepassInfo.renderingInfo);
+
+        meshPipeline->drawDepthPrepass(
+            cmd,
+            drawImage.getExtent2D(),
+            *resources,
+            camera,
+            sceneDataBuffer.getBuffer(),
+            meshDrawCommands,
+            sortedMeshDrawCommands);
+
+        vkCmdEndRendering(cmd);
+    }
+
+    {
+        TracyVkZone(gfxDevice.getTracyVkCtx(), cmd, "Light Culling");
+
+        vkutil::transitionImage(
+            cmd,
+            depthImage.image,
+            depthImage.layout,
+            VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+        depthImage.setLayout(VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+
+        lightCullingPipeline->cullLights(
+            cmd,
+            *resources,
+            sceneDataBuffer.getBuffer(),
+            depthImageId,
+            drawImage.getExtent2D());
+
+        vkutil::transitionImage(
+            cmd,
+            depthImage.image,
+            depthImage.layout,
+            VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL);
+        depthImage.setLayout(VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL);
+    }
+
    auto renderInfo = vkutil::createRenderingInfo({
        .renderExtent = drawImage.getExtent2D(),
        .colorImageView = drawImage.imageView,
        .colorImageClearValue = glm::vec4{0.f, 0.f, 0.f, 1.f},
        .depthImageView = depthImage.imageView,
-        .depthImageClearValue = 1.f,
    });
    renderInfo.renderingInfo.pColorAttachments = &renderInfo.colorAttachment;
    renderInfo.renderingInfo.pDepthAttachment = &renderInfo.depthAttachment;
@@ -226,11 +298,11 @@ void GameRenderer::draw(VkCommandBuffer cmd, GfxDevice& gfxDevice, const Camera&
            *resources,
            drawImage,
            glm::ortho(0.0f,
-                static_cast<float>(drawImage.extent.width),
-                static_cast<float>(drawImage.extent.height),
-                0.0f, -1.0f, 1.0f));
+                       static_cast<float>(drawImage.extent.width),
+                       static_cast<float>(drawImage.extent.height),
+                       0.0f, -1.0f, 1.0f));
    }
-// vkutil::cmdEndLabel(cmd);
+    // vkutil::cmdEndLabel(cmd);
 }

 void GameRenderer::drawQuads(
@@ -246,11 +318,12 @@ void GameRenderer::cleanup(GfxDevice& gfxDevice)
 {
    VkDevice device = gfxDevice.getDevice();

-    if (device != VK_NULL_HANDLE) {
+    if (device != VK_NULL_HANDLE)
+    {
        vkDeviceWaitIdle(device);
    }

-if (skinningPipeline)
+    if (skinningPipeline)
        skinningPipeline->cleanup(gfxDevice);

    if (lineRenderingPass)
@@ -259,6 +332,9 @@ if (skinningPipeline)
    if (quadRendererPass)
        quadRendererPass->cleanup(gfxDevice);

+    if (lightCullingPipeline)
+        lightCullingPipeline->cleanup(gfxDevice);
+
    if (skyboxPipeline)
        skyboxPipeline->cleanup(device);

@@ -281,6 +357,7 @@ if (skinningPipeline)
    skinningPipeline.reset();
    lineRenderingPass.reset();
    quadRendererPass.reset();
+    lightCullingPipeline.reset();
    pendingMaterialUploads.clear();
    meshDrawCommands.clear();
    sortedMeshDrawCommands.clear();
@@ -323,6 +400,11 @@ void GameRenderer::drawSkinnedMesh(MeshID id,
    });
 }

+void GameRenderer::submitLight(const Light& light)
+{
+    lightCullingPipeline->appendLight(light);
+}
+
 const GPUImage& GameRenderer::getDrawImage() const
 {
    assert(resources);
@@ -349,7 +431,8 @@ void GameRenderer::setSkyboxTexture(ImageID skyboxImageId)

 void GameRenderer::flushMaterialUpdates(GfxDevice& gfxDevice)
 {
-    if (!pendingMaterialUploads.empty()) {
+    if (!pendingMaterialUploads.empty())
+    {
        // Material data is currently shared by all frame contexts. Wait before
        // mutating it so an older in-flight frame cannot read the overwritten
        // range. This can later be replaced with per-frame material buffers.