This guide aims to provide insight and recommendations for integrating the Kickstart RT SDK into a game. It is expected that the reader has read the repo Readme and the SDK Reference Guide to familiarize themselves with the key algorithmic details of the SDK.
The SDK provides support for shadows, ambient occlusion (AO), global illumination (GI) and reflections.
The Kickstart RT SDK has 4 main phases of operation to generate raytraced output for your game.
-
BVH Building. Internally this updates the BLAS and TLAS for the scene.
-
Direct Light Injection. The game will provide the rasterized lighting and depth buffers along with view matrices which Kickstart RT uses to populate the internal lighting cache for each mesh.
-
Raytraced Output. Kickstart RT requires the gbuffer including depth, normals, roughness, specular and color buffers. These are used to generate the secondary rays that intersect with the BVHs to generate the raytraced reflections. Shadows will require lighting data.
-
Denoising. Raytracing is inherently noisy for non mirror surfaces, but Kickstart RT has NVIDIAs realtime denoiser (NRD) seamlessly integrated to denoise the output images if required. Kickstart RT requires the gbuffer again, along with the velocity buffer to provide correct denoising.
Throughout all of these phases or passes, the surface area of the integration should be relatively small and no shaders or material systems require any modifications. The only shaders that may be required will be a conversion pass to decode gbuffer image formats to Kickstart RT supported ones.
Each phase of operation will either work directly with the Kickstart RT execution context, or with a Kickstart RT render task and task container where the execution context maps to a DX12/Vulkan device, a task container maps to a command list and a task is a high level rendering operation.
The context manages the lifetime of objects (BLAS/TLAS) and the internal state. The tasks are scheduled for processing against a task container and the task container will then populate the provided command list when the tasks are built. It is important to highlight that the engine owns all command lists and command queues; Kickstart RT will record the command lists to be executed by the engine.
Kickstart RT will automatically build and maintain the BVH's for the scene, but the engine must provide the input vertex and index buffers with the transform matrices for each object that is to be visible to the raytracing engine.
Kickstart RT defines a Geometry to represent a mesh with an associated BLAS and an Instance to represent an instance of a Geometry with an associated lighting cache resource.
It is recommended that the engine creates a Kickstart RT geometry for all renderable game objects that are to appear in the raytraced outputs and the returned geometry handles should be associated with the game object.
// Build the BVHs for all geometries at startup
for (auto& geom : geometryList)
{
KickstartRT::D3D12::BVHTask::GeometryTask task;
// create a geometry handle
m_executeContext->CreateGeometryHandles(&task.handle, 1);
// store the handle for later use
prim->m_KSGeomHandle = task.handle;
task.taskOperation = SDK::D3D12::BVHTask::TaskOperation::Register;
task.input.allowUpdate = geom.isSkinnedMesh;
task.input.type = decltype(task.input)::Type::TrianglesIndexed;
task.input.surfelType = SurfelType::WarpedBarycentricStorage;
task.input.allowLightTransferTarget = geom.allowLodSwitching;
task.input.forceDirectTileMapping = geom.isHighpolygonModel;
task.input.tileUnitLength = systemDefinedUnitLength;
task.input.tileResolutionLimit = systemDefinedTileResolutionLimit;
// populate the index buffer and vertex buffer data
for (auto&& component : geom->components) {
decltype(task.input)::GeometryComponent cmp;
cmp.indexBuffer.resource = component->indexBuf;
cmp.indexBuffer.format = DXGI_FORMAT_R32_UINT;
cmp.indexBuffer.offsetInBytes = component->indexOffset * sizeof(uint32_t);
cmp.indexBuffer.count = component->numIndices;
cmp.vertexBuffer.resource = component->vertexBuf;
cmp.vertexBuffer.format = DXGI_FORMAT_R32G32B32_FLOAT;
cmp.vertexBuffer.offsetInBytes = component->vertexOffset * sizeof(float) * 3;
cmp.vertexBuffer.strideInBytes = sizeof(float) * 3;
cmp.vertexBuffer.count = component->numVertices;
cmp.useTransform = false;
task.input.components.push_back(cmp);
}
// schedule the task to be built against a task container.
m_taskContainer->ScheduleBVHTask(&task);
}
Multiple tasks can be scheduled at once for efficiency, but the above pseudo-code shows single tasks scheduled for simplicity.
When creating a Kickstart RT geometry, various default parameters can be over-ridden to help tune the Kickstart RT integration for your game. These are covered in the Optimisation and Tuning section. Normal guidance should be followed for efficient management of BVHs.
Each frame, the scene graph is walked and all relevant meshes that require a Kickstart RT instance/BLAS are processed. Kickstart RT will generate a TLAS that contains all known instances and it is the responsibility of the engine to determine which instances should be in the TLAS.
A game primitive's Kickstart RT instance can be in 1 of 4 states :
-
No longer in the TLAS : destroy instance if it exists
-
Newly visible in the TLAS : create the instance
-
Has moved/changed : update the instance
-
Static with no updates : nothing
All Kickstart RT instances should be checked for inclusion each frame and modified accordingly before the Kickstart TLAS task is scheduled and built.
// Gather the required instances and build the BLAS/TLAS
for (auto& prim : primitiveList)
{
if (prim->isVisible(extendedViewFrustum))
{
KickstartRT::D3D12::BVHTask::InstanceTask task;
// If the primitive has already been registered with KS and has an instance handle
if (prim->m_KSInstanceHandle != KickstartRT::D3D12::InstanceHandle::Null)
{
// If the prim hasn't moved/changed, continue
if (!prim->isDirty())
{
continue;
}
// Update the BLAS for this primitive.
// The default operation is KickstartRT::D3D12::BVHTask::TaskOperation::Register
task.taskOperation = KickstartRT::D3D12::BVHTask::TaskOperation::Update;
}
else
{
// Create an instance handle to hold the BLAS and lighting cache
m_12->m_executeContext->CreateInstanceHandles(&prim->m_KSInstanceHandle, 1);
}
// Associate the instance to update with the geometry handle
task.input.geomHandle = prim->m_KSGeomHandle;
task.handle = prim->m_KSInstanceHandle;
// Copy over the local to world space matrix
task.input.transform.f = prim->getLocalToWorldF();
// Schedule the task
m_taskContainer->ScheduleBVHTask(&task);
}
else if (prim->m_KSInstanceHandle != KickstartRT::D3D12::InstanceHandle::Null)
{
// Destroy the instance as it is no longer visible
m_executeContext->DestroyInstanceHandles(&prim->m_KSInstanceHandle, 1);
prim->m_KSInstanceHandle = KickstartRT::D3D12::InstanceHandle::Null;
}
}
// After generating all of the instances, build the TLAS
KickstartRT::D3D12::BVHTask::BVHBuildTask task;
// Set the number of BLASs to be built per frame to avoid large stutters
task.maxBlasBuildCount = c_MaxBlasBuildsPerFrame;
m_taskContainer->ScheduleBVHTask(&task);
It is important to note that the view frustum should be expanded in all directions when visibility culling the scene graph as raytracing needs to know about meshes behind the main camera as well as in front of it.
It is also important that the instances persist across frames (which may be different to native RT implementations) as the lighting cache is associated with each instance. storing the light injection values for the mesh.
All of this work can be done asynchronously, both on the GPU and whilst walking the scene graph on the CPU, so the costs should be hidden.
To confirm this works as intended, various debug visualization modes are supported by the Kickstart RT render tasks so the primary ray HitT values could be displayed, or the BVHs rendered with random colors.
enum class DebugOutputType : uint32_t {
Default = 0,
Debug_DirectLightingCache_PrimaryRays = 100,
Debug_RandomTileColor_PrimaryRays = 101,
Debug_RandomMeshColor_PrimaryRays = 102,
Debug_HitT_PrimaryRays = 103,
Debug_Barycentrics_PrimaryRays = 104,
};
Once the BVHs have been built, the current lighting values need to be baked, or injected, into the lighting cache of the Kickstart RT instances. To enable this, the engine needs to provide a depth buffer, the current direct lighting buffer and the set of current view transformations.
The direct lighting injection buffer is essentially the current rasterized frame. It should include everything that is to be reflected in the raytraced output and has an associated BVH to bake the lighting values into. This should include the lit shaded geometry and shadows, but not particles or transparent geometry if they do not have associated Kickstart RT instances.
// Light Injection
KickstartRT::D3D12::RenderTask::DirectLightingInjectionTask renderTask;
// Provide Depth buffer
renderTask.depth.tex = initShaderResourceTex(pDepthBuffer);
renderTask.depth.type = KickstartRT::D3D12::RenderTask::DepthType::R_ClipSpace;
// Provide color buffer
renderTask.directLighting = initShaderResourceTex(pDirectLightingBuffer);
// Viewport
renderTask.viewport = viewport;
// Matrices
renderTask.clipToViewMatrix = cameraView.getClipToViewF();
renderTask.viewToWorldMatrix = cameraView.getViewToWorldF();
// Tuning parameters
renderTask.averageWindow = c_AccumulationWindow;
renderTask.injectionResolutionStride = c_InjectionStride;
renderTask.useInlineRT = false;
// Schedule the task
m_taskContainer->ScheduleRenderTask(&renderTask);
To confirm that the injection works as intended, try to use the
Debug_DirectLightingCache_PrimaryRays
debug visualization mode. This
will render the BVH meshes with their lit colors baked in.
Kickstart RT can be used to generate various outputs, such as AO, GI, reflections and shadows. Apart from shadows, these all require similar gbuffer input in the form of depth, normal, roughness, specular buffers and various input matrices. The direct lighting buffer from the injection phase can also be passed back in for specular reflections and/or diffuse reflections (GI) to provide a higher quality reflection output.
Each type of output has its own task type, but there are a number of common parameters that have been put into a common structure.
KickstartRT::D3D12::RenderTask::TraceTaskCommon rtTaskCommon;
// Provide Depth buffer
rtTaskCommon.depth.tex = initShaderResourceTex(pDepthBuffer);
rtTaskCommon.depth.type = depthType;
// Provide Normal buffer
rtTaskCommon.normal.tex = initShaderResourceTex(pNormalBuffer);
rtTaskCommon.normal.type = normalType;
// Provide Roughness buffer
rtTaskCommon.roughness.tex = initShaderResourceTex(pRoughnessBuffer);
rtTaskCommon.roughness.globalRoughness = 1.0f;
rtTaskCommon.roughness.roughnessMask = { 0.0f, 0.0f, 0.0f, 1.0f }; // A Channel
// Provide Specular buffer
rtTaskCommon.specular.tex = initShaderResourceTex(pSpecularBuffer);
rtTaskCommon.specular.globalMetalness = 1.0f;
// Provide the direct lighting buffer used in the injection phase
rtTaskCommon.directLighting = initShaderResourceTex(pDirectLightingBuffer);
// Viewport
rtTaskCommon.viewport = viewport;
// Matrices
rtTaskCommon.viewToClipMatrix = cameraView.getViewToClipF();
rtTaskCommon.clipToViewMatrix = cameraView.getClipToViewF();
rtTaskCommon.viewToWorldMatrix = cameraView.getViewToWorldF();
rtTaskCommon.worldToViewMatrix = cameraView.getWorldToViewF();
// Tuning parameters
rtTaskCommon.useInlineRT = false;
rtTaskCommon.rayOffset.type = KickstartRT::D3D12::RenderTask::RayOffset::Type::e_CamDistance;
rtTaskCommon.enableBilinearSampling = true;
rtTaskCommon.halfResolutionMode = false; // Disable checkerboard
To schedule a specular reflection task :
// Set up the specular reflection render task.
KickstartRT::D3D12::RenderTask::TraceSpecularTask renderTask;
renderTask.common = rtTaskCommon;
renderTask.out = initUnorderedAccessTex(pOutputBuffer);
// Schedule the task
m_taskContainer->ScheduleRenderTask(&renderTask);
Or a diffuse task for GI :
// Set up the GI diffuserender task.
KickstartRT::D3D12::RenderTask::TraceDiffuseTask renderTask;
renderTask.common = rtTaskCommon;
renderTask.out = initUnorderedAccessTex(pOutputBuffer);
renderTask.diffuseBRDFType = KickstartRT::D3D12::RenderTask::DiffuseBRDFType::NormalizedDisney;
// Schedule the task
m_taskContainer->ScheduleRenderTask(&renderTask);
Or an AO task :
// Set up the AO task.
KickstartRT::D3D12::RenderTask::TraceAmbientOcclusionTask renderTask;
renderTask.common = rtTaskCommon;
renderTask.out = initUnorderedAccessTex(pOutputBuffer);
// Schedule the task
m_taskContainer->ScheduleRenderTask(&renderTask);
Various parameters are available at this stage, such as global metalness, various roughness modifiers, ray offset types, diffuse BRDF types etc as well as the opportunity to have a separate demodulated specular buffer returned that can be multiplied with the denoised specular buffer to provide a higher quality specular reflection.
Multiple tasks can be scheduled against a container before it is built and we would recommend a single task container to be used for the light injection, RT output and denoising.
Specular reflections are calculated using very common BRDF based on specular microfacet model with GG-X distribution and uses an efficient VNDF method for sampling.
There are two diffuse BRDFs to chose from. Default one is simple Lambertian BRDF, and second option is Frostbite's normalized version of disney diffuse BRDF.
The downside of using Lambertian BRDF is that energy conservation is not preserved when it is combined with our specular BRDF. This means, more light can be reflected from the surface than is received. Many rendering engines solve this when composing final image by modifying indirect diffuse and specular lighting to preserve energy conservation. If this is not the case, KickstartRT provides the option to use Frostbite's diffuse BRDF. Note that this BRDF requires the roughness texture to be passed into diffuse reflection pass as well. This ensures that energy conservation is preserved out of the box, and final pixel value can be simply calculated as a sum of direct lighting + indirect diffuse + indirect specular
.
Transparent reflections can be achieved by creating another specular reflection task that will take transparent surface properties as parameters. Kickstart RT will trace specular rays and shade them using the direct lighting cache. There is no support in Kickstart RT for light transmission. The feature works the best when there is a single transparent surface layer such as a regular clear class.
Perfectly smooth transparent reflections don't require denoising, but if the transparent surface has some roughness then a denoising pass might be required.
To render shadows the engine is required to convert its internal light representation to the Kickstart RT defined light structure LightInfo. The supported types are Directional, Point and Spot lights. Kickstart RT will trace a single ray towards each light source for each pixel, meaning that no stochastic selection process of the lights themselves are taking place, only the point on the light surface to trace against is randomized. For this reason it's advised to keep the number of active lights as low as possible.
There are two render tasks related to shadow rendering. TraceShadowTask and TraceMultiShadowTask. As implied the TraceMultiShadowTask is meant to be used when the number of light sources is greater than one.
The primary reason it's separated in two tasks relates to the output visibility representation. In multi light mode an additional destination UAV is required, and the data is not trivially discernible. In single light source mode the output is a single fp16+ value representing the closest hit distance, negative values indicate miss (=light is visible). When tracing multiple lights the visibility information for all lights is merged into a format tailored for NRD. Instead of hit distance a weighted combination of visibility, intensity, distance and depth is stored. The specifics of the format is not a secret, but it's purposely opaque to stay forward compatible with future NRD releases. DenoiseMultiShadow task understands this format and will return denoised visibility that the engine can read directly.
Both single and multi light sources can be passed to corresponding denoising tasks. The output is the same for shadow denoising, which is a single visibility value per pixel.
Noisy output is expected for non-mirror reflections and whilst game developers can write their own denoiser from scratch, Kickstart RT conveniently provides support for the NRD denoiser if desired. To simplify integration, the Kickstart RT API provides a simplified subset of parameters and options for controlling NRD. As source code is provided, these defaults can obviously be modified directly if needed.
NRD supports several denoisers depending on the input - Reblur, Relax and Sigma. To use the denoisers, an NRD context needs to be created for each denoiser when Kickstart RT is initialized. The relevant denoiser context should then be used when executing the denoising operations.
// Create the denoiser
KickstartRT::D3D12::DenoisingContextInput input;
input.maxWidth = maxWidth;
input.maxHeight = maxHeight;
input.denoisingMethod = KickstartRT::D3D12::DenoisingContextInput::DenoisingMethod::NRD_Relax;
input.signalType = KickstartRT::D3D12::DenoisingContextInput::SignalType::SpecularAndDiffuse;
m_KSDenoisingContext = m_executeContext->CreateDenoisingContextHandle(&input);
A denoising task is generated for each buffer to denoise, with the exception that specular reflection and diffuse GI can be denoised from a single combined task. Similar to the RT output passes, the depth, normal and roughness buffers need providing along with the velocity buffer to allow NRD to temporally denoise the various inputs correctly.
// Denoise
KickstartRT::D3D12::RenderTask::DenoisingTaskCommon dTaskCommon;
// Provide Depth buffer
dTaskCommon.depth.tex = initShaderResourceTex(pDepthBuffer);
dTaskCommon.depth.type = KickstartRT::D3D12::RenderTask::DepthType::R_ClipSpace;
// Provide Normal buffer
dTaskCommon.normal.tex = initShaderResourceTex(pNormalBuffer);
dTaskCommon.normal.type = KickstartRT::D3D12::RenderTask::NormalType::RGB_NormalizedVector; // RGB_Vector;
// Provide Velocity buffer
dTaskCommon.motion.tex = initShaderResourceTex(pVelocityBuffer);
dTaskCommon.motion.type = KickstartRT::D3D12::RenderTask::MotionType::RG_ViewSpace; // RG_ClipSpace; // RGB_WorldSpace;
// Provide Roughness buffer.
dTaskCommon.roughness.tex = initShaderResourceTex(pRoughnessBuffer);
dTaskCommon.roughness.globalRoughness = 1.0f;
dTaskCommon.roughness.roughnessMask = { 0.0f, 0.0f, 0.0f, 1.0f }; // A Channel
// Viewport
dTaskCommon.viewport = viewport;
// Matrices
dTaskCommon.viewToClipMatrix = cameraView.getViewToClipF();
dTaskCommon.clipToViewMatrix = cameraView.getClipToViewF();
dTaskCommon.clipToViewMatrixPrev = cameraView.getPreviousClipToViewF();
dTaskCommon.viewToWorldMatrix = cameraView.getViewToWorldF();
dTaskCommon.worldToViewMatrix = cameraView.getWorldToViewF();
dTaskCommon.worldToViewMatrixPrev = cameraView.getPreviousWorldToViewF();
// Tuning parameters
dTaskCommon.cameraJitter = { jitterX, jitterY };
dTaskCommon.halfResolutionMode = false; // Needs to match the RT output
// Denoise Specular and Diffuse buffers together
KickstartRT::D3D12::RenderTask::DenoiseSpecularAndDiffuseTask dTask;
dTask.common = dTaskCommon;
dTask.context = m_KSDenoisingContext;
// Provide Input Specular buffer
dTask.inSpecular = initShaderResourceTex(pInSpecularBuffer);
dTask.inDiffuse = initShaderResourceTex(pInDiffuseBuffer);
dTask.inOutSpecular = initCombinedAccessTex(pOutSpecularBuffer);
dTask.inOutDiffuse = initCombinedAccessTex(pOutDiffuseBuffer);
// Schedule the task
m_taskContainer->ScheduleRenderTask(&renderTask);
Once the output buffers have been denoised, they can be passed back into
the render pipeline for consumption by the remaining stages if they have been denoised with Relax or Sigma.
As for Reblur, the output color components are encoded in YCoCg color space in the current implementation, so applications need to decode them after fetching the texture. NRD SDK has provided a utility function,
REBLUR_BackEnd_UnpackRadianceAndNormHitDist()
to decode it to RGB color space, but it's essentialy a conversion from YCoCg to RGB, so you can do it inline with your own shader code. So, when using Reblur, it's goo to visit NRD's shader code once to make sure what you need to do with the output texture.
When a new object appears in the scene and a new Geometry and Instance are registered in the SDK, the Direct Lighting Cache is initialized with the specified values. To store light information in it, the SDK must perform Direct Lighting Injection over several frames with GBuffers drawn by the rasterizer.
Additionally, the Direct Lighting Cache is essentially tied to the object's polygon topology. If the topology is changed, it must be registered in the SDK as a separate Geometry and Instance. Therefore, if the topology of an object changes(for example due to a change of LoD) the information in the Direct Lighting Cache will be lost.
To avoid this data loss, the SDK provides a task to transfer the Direct Lighting Cache between objects that have different topologies. Before placing the transfer task, application must place the target geometry in the scene as usual(i.e. registering a new geometry and an instance) and then specify it as a target instance in the task.
KickstartRT::D3D12::RenderTask::DirectLightTransferTask transferTask;
// specify target instance to be transfered
transfer.target = targetInstanceHandle;
// Schedule the task
m_taskContainer->ScheduleRenderTask(&transferTask);
At first glance, this task looks pretty simple, but in fact, it requires careful control of the InstanceInclusionMask.
First, either the target or the source instance should not be visible in Reflection or GI rendering because they will be placed in the same space in general.
Therefore, the application needs to set InstanceInclusionMask::VisibleInRT
flags properly for those objects.
In addition, the object which will be the source of the transfer should set LightTransferSource
to be visible in the transfer operation. This flag is orthogonal to VisibleInRT
flags.
Here is one possible way to set a DirectLightTransferTask for switching LoD of an object.
To switch an object's LoD from 0 to 1,
1. Register geometries for LoD 0 and 1 to the SDK with `GeometryInput::allowLightTransferTarget` flag.
2. Register an instance for LoD 0.
3. Render the instance in the rasterizer and do sufficient light injection tasks for several frames.
~~~ Frame boundary ~~~~
~~~ Frame boundary ~~~~
When the rasterizer switches the LoD from 0 to 1,
4. Register an instance for LoD 1 with `DirectLightInjectionTarget` and `VisibleInRT` flags.
5. Update the instance for LoD 0 to clear `VisibleInRT` and `DirectLightInjectionTarget` flags, and set `LightTransferSource` flag.
6. Set a light transfer task. Its target is the instance for LoD 1.
7. Set Injection, Reflection, and GI tasks as needed (Since LoD 0 is invisible, it will not participate in the RT operations).
~~~ Frame boundary ~~~~
8. De-register the instance for LoD 0 (and the geometry too if it isn't needed anymore).
Another thing that needs to be careful is direct lighting cache instantiation.
Since the size of the direct lighting cache is calculated by a compute shader while the buffer allocation is done on the CPU side, it needs a readback from the GPU after running the compute shader after registering a new geometry. So, to make sure the direct lighting cache instantiation is correct, the application needs to run a GPU command list that contains a BVH task to register a geometry and make sure of the completion. Then call the MarkGPUTaskAsCompleted()
function to tell this to the SDK. (The GPU work scheduling is described by the following section).
After completing the geometry registering task, the direct lighting cache of the instance will be allocated and the application can set the instance of the geometry as a target of a direct lighting cache transfer task.
Another option is to register the geometry in advance of registering the instance which refers to the geometry. Direct lighting cache allocation is tied to an instance, but its buffer size is calculated when registering the referred geometry. Direct lighting cache is allocated immediately if the referred geometry has been registered in advance.
In contrast, it could be easy to manage the resource lifetime of the source instance of the direct lighting cache transfer. Application only needs to defer the destruction of the corresponding geometry and instance handles of the SDK. All other resources which are referred to when drawing the object in rasterization can be destructed as usual. SDK doesn't reference the resources such as vertex or index buffer except when registering the geometry in the SDK.
When registering a geometry that will be used s a target of the transfer task, GeometryInput::allowLightTransferTarget
must be set to true. Geometries with this flag will keep copies of the vertex and index buffers in the SDK even after its BLAS has been built, so it will use more VRAM than an object without this flag.
Finally, the mechanism of the transfer task is not a simple data copy operation by a compute shader. It's done by ray tracing. More specifically, two rays are traced from the center of each polygon of the target object, toward the direction of the normal and its reversed directions, and then the data is copied from the direct lighting cache of the hit surface that is closer to the target object. Therefore, if the objects are too far apart in shape or have a fine convex shape, the ray trace may not be able to reach the source surface to obtain the direct lighting cache information.
Preparing and scheduling tasks against a task container, as seen in the previous section, is the first part of the execution flow for Kickstart RT.
The next part is to record the command lists and then execute them. As Kickstart RT does not own the command lists, the game engine must manage the synchronization and dependencies as well as select which command queue the command list executes on. It is recommended to execute the BVH build on the asynchronous command queue early in the frame to allow it to overlap with other work and hide its cost.
// Record the task
KickstartRT::D3D12::BuildGPUTaskInput input = {};
KickstartRT::D3D12::GPUTaskHandle taskHandleTicket;
input.commandList = pD3DCommandList;
// Record the scheduled work to the command list
m_executeContext->BuildGPUTask(&taskHandleTicket, m_taskContainer, &input);
Calling BuildGPUTask
will process the tasks scheduled against the
specific task container and return a handle to the task, referred to as
a ticket in the above diagram. This ticket is used to track the lifetime
of the work and should be returned to Kickstart RT when the work has
been executed on the GPU so that resources can be recycled.
Kickstart RT supports a user configurable number of tickets in flight at any one time so it is important to track the progress of tickets with appropriate fences. A typical integration might set the number of task containers supported to be the product of the task containers needed per frame and the number of frames of run-ahead that the engine supports.
KickstartRT::D3D12::ExecuteContext_InitSettings settings = {};
settings.supportedWorkingsets = c_MaxTasksPerFrame * c_MaxFrameLookAhead; // 3 * 2
When a task handle ticket is no longer in use and the corresponding command list has finished executing on the GPU, it should be returned to the SDK.
// Return the taskHandle to the SDK to allow resource recycling
m_executeContext->MarkGPUTaskAsCompleted(taskHandleTicket);
Once the GPU task has been built, the task container will be invalid so a new one should be created.
// Re-create the task container after it has been built
m_taskContainer = m_executeContext->CreateTaskContainer();
Kickstart RT has been designed to integrate well with a native threading
implementation and render pipeline. The execute context is threadsafe
but task containers should only have work scheduled against them 1
thread a time. As BuildGPUTask
is a blocking operation and can have a
high cost depending on the tasks scheduled, it may be useful to execute
this on a worker thread, especially for any BVH builds.
Finally, the game engine is in full control of the execution order of the command lists and is expected to marshall their dependencies with fences as required. This also provides the flexibility to execute the command lists on different queues catering for async compute work providing fine grained control of when and where the work should be executed.
Integrating a complex API always provides opportunities for bugs and issues to occur. This section of the integration guide hopes to provide some tips and methods for debugging the integration as well as highlighting some of the common pitfalls.
Kickstart RT offers several built in debug features, such as the debug
output visualization modes that can be generated from a RenderTask
with
a provided output buffer and would be a useful addition to an
integration:
enum class DebugOutputType : uint32_t {
Default = 0,
Debug_DirectLightingCache_PrimaryRays = 100,
Debug_RandomTileColor_PrimaryRays = 101,
Debug_RandomMeshColor_PrimaryRays = 102,
Debug_HitT_PrimaryRays = 103,
Debug_Barycentrics_PrimaryRays = 104,
};
// Set the debug mode
KickstartRT::D3D12::RenderTask::TraceSpecularTask renderTask;
renderTask.debugParameters.debugOutputType = KickstartRT::D3D12::RenderTask::DebugParameters::DebugOutputType::Debug_RandomTileColor_PrimaryRays;
The meshes associated with the BVHs can be visualized by choosing
Debug_RandomMeshColor_PrimaryRays
as the debug output mode This can be
helpful to ensure the geometry is correct and matches the location of
the geometry in the game engine. If anything is incorrect at this stage,
it is worth checking if all of the geometry transform matrices are as
expected and the vertex and index buffer strides and layouts are
correct.
The direct lighting cache can be visualized by setting the debug output
type to be Debug_DirectLightingCache_PrimaryRays
.
It can be helpful to initialize the direct lighting cache for each geometry to a known fixed color, ie pink or cyan. Over a period of frames the debug colors should be replaced with the correct lit colors highlighting any areas that have not received any light injection.
// Set the initial tile cache to a known color
KickstartRT::D3D12::BVHTask::InstanceTask task;
task.input.initialTileColor[0] = 1.0f;
task.input.initialTileColor[1] = 0.0f;
task.input.initialTileColor[2] = 1.0f;
As a convenience, it is recommended to implement a way to clear the direct lighting cache to check the behavior of the cache at different times in the game level. As the cache is directly tied to the BVH of a mesh, this requires the engine to destroy all of the Kickstart RT geometries and instances and re-create them.
Once the BVH's are confirmed to be correct and the light injection is working, the next step is to get the specular reflections working. To start, the recommendation is to just use the depth and normal buffers and not the roughness or specular buffers. The global metalness can be set to 1.0f and the global roughness should be set to 0.0f to ensure maximum reflections and minimum noise. This should very quickly show up any issues in the implementation.
Any problems at this stage will be down to the depth buffer or normals, so check that the buffer format and encoding for both is compatible with Kickstart RT and has been correctly set.
enum class DepthType : uint32_t {
RGB_WorldSpace = 0, // RGBch represents a world position
R_ClipSpace = 1, // Rch represents depth value in viewport transformed clip space.
};
enum class NormalType : uint32_t {
RGB_Vector = 0, // RGBch represents a normal vector in XYZ.
RGB_NormalizedVector = 1, // RGBch represents normalized normal vector, (xyz) = (rgb) * 2.0 - 1.0.
RG_Octahedron = 2, // RGch represents octahedron encoded noamal vector (xyz) = ocd_decode(rg)
BA_Octahedron = 3, // BAch represents octahedron encoded noamal vector (xyz) = ocd_decode(ba)
RG_NormalizedOctahedron = 4,// RGch represents normalized octahedron encoded noamal vector (xyz) = ocd_decode((rg) * 2.0 - 1.0)
BA_NormalizedOctahedron = 5,// BAch represents normalized octahedron encoded noamal vector (xyz) = ocd_decode((ba) * 2.0 - 1.0)
};
A conversion pass may be required if your depth or normal formats are not supported.
Once the basic raytracing is working with global metalness and global roughness, the specular and roughness buffers from the gbuffer can be applied. If used, the global metalness and roughness over-rides should be set to 1.0f.
Another common issue is self intersection of the secondary rays at the point of reflection and an offset is typically applied to avoid this. Kickstart RT provides different options for the offset which should be tried to see which is more appropriate for the game. Issues with this can manifest as zero reflections in part of the image.
// This is a parameter set used to offset reflection rays when tracing rays.
struct RayOffset {
enum class Type {
e_Disabled = 0,
e_WorldPosition, // offset applied based on the magnitude of world position value.
e_CamDistance, // offset applied based on the distance from the camera.
} type = Type::e_Disabled;
struct ParametersForWorldPosition {
float threshold = 1.f / 32.f; // threshold to switch offsetting algorithm. Small value uses floating point offset, large value uses integer offset.
float floatScale = 1.f / 65536.f; // scaling factor for normal vector before adding to position value.
float intScale = 256.f; // after normal vector is scaled this factor, it is converted to integer value to make an offset in mantissa of position value.
} worldPosition;
struct ParametersForCamDistance {
float constant = 0.00174f;
float linear = -0.0001547f;
float quadratic = 0.0000996f;
} camDistance;
};
NRDs documentation should be referred to when trying to determine issues and when tuning for the game, but the inputs should be tested for correctness as the first step.
Assuming the gbuffer data passed to Kickstart RT works, some of the only differences may be with the velocity buffers, so these should be double checked. If no velocity buffer is provided, then NRD will try to create its own velocity vectors which can be a useful debug aide.
The motion vector buffer can be optionally left out, this way NRD will fall back to doing frame reprojection using the provided matrices instead which can be a good way to verify that matrices are correctly set up. This is not recommended for production as the reprojection will fail for moving objects, only use this mode for debugging purposes.
If motion vectors in the velocity buffer are in the correct scale and format, the stationary image quality should be equal when the velocity buffer is provided and when it's not.
Before any work can be done, there are some important initialisation settings that may require modification to support a large game. A typical setup may be :
Kickstart RT::D3D12::ExecuteContext_InitSettings settings = {};
settings.uploadHeapSizeForVolatileConstantBuffers = 4 * 1024 * 1024;
settings.descHeapSize = 128 * 1024;
settings.supportedWorkingsets = maxTasksPerFrame * maxFrameLookAhead; // 3 * 2 = 6
...
These may require tuning throughout the integration process, depending on the scale of the game data.
A common issue is that Kickstart RT can run out of descriptors when processing a lot of geometry. This can be modified by changing the value of m_unboundDescTableUpperbound
in the Kickstart RT SDK source code itself.
The direct lighting cache has 2 different layouts based on regular tiles or mesh colors. The mesh colors tend to use slightly more memory, but allow bilinear sampling of the direct lighting cache so generates a higher quality image and is the recommended surfel type.
The surfel type is selected when geometry is created.
enum class SurfelType : uint32_t {
WarpedBarycentricStorage,
MeshColors
};
KickstartRT::D3D12::BVHTask::GeometryTask task;
task.input.surfelType = KickstartRT::D3D12::BVHTask::GeometryInput::SurfelType::MeshColors;
Whether the surfel type is based on mesh colors or regular tiles, there are some modifiers to influence the granularity of subdivision that is used to create the direct lighting cache. A smaller surfel tile size per primitive results in a better quality lighting cache but at the cost of increased memory. They should be tuned to the game content.
KickstartRT::D3D12::BVHTask::GeometryTask task;
task.input.tileUnitLength = c_KickStartTileUnitLength;
task.input.tileResolutionLimit = c_KickStartTileResolutionLimit;
These are set when geometry is created, therefore any modifications for tuning requires the geometry to be destroyed and recreated.
struct DirectLightingInjectionTask : public Task {
DirectLightingInjectionTask() : Task(Task::Type::DirectLightInjection) {};
//< SDK accumulates direct lighting information into allocated tiles on the surfaces.
//< Longer average window will converge values slowly but more stable than shorter average window.
float averageWindow = 200.f;
//< The injection pass will trace and inject for every injectionResolutionStride'th pixel in directLighting texture
//< Note: must be non-zero
uint32_t injectionResolutionStride = 4;
...
};
When injecting the direct lighting into the cache, the average window parameter can be used to set the number of frames required for the image to converge to full brightness. A small number allows for fast reactions to lighting updates but may produce high frequency flickering. This can be reduced by increasing the average window to increase the temporal stability, but the right value needs tuning for the game.
Also, it is important to set an appropriate value to injectionResolutionStride. That value and the resolution of GBuffer are directly linked to the number of rays traced for injecting lighting to direct lighting cache. A ray for injecting lighting is traced each 4x4 pixel block of GBuffer by default. This is because the resolution of direct lighting cache is by far lower than the pixel resolution in general, so, it will be wasting to trace rays on each pixel for direct lighting injection. It is recommended to set a bigger stride especially when using high resolution GBuffer.
The quality of specular reflections can be improved by optionally binding the direct lighting input buffer to the common render task.
KickstartRT::D3D12::RenderTask::TraceTaskCommon rtTaskCommon;
rtTaskCommon.directLighting = initShaderResourceTex(pDirectLightingBuffer);
This could be perceived as a hybrid of raytraced reflections and screen space reflections. When a secondary ray hits a BVH, the provided direct lighting buffer and depth buffer are checked to see if they contain a sample for the hit point. If they do, then the direct lighting buffer is used directly, otherwise the value in the direct lighting cache is used.
This provides an obvious quality improvement when reflected secondary rays hit this buffer, but the transition between the direct lighting cache and the direct lighting input buffer can be obvious, so requires some subjective analysis to determine its use. The recommendation is to use the direct lighting input buffer.
One of the limitations of the direct lighting cache is that it can only cache the lighting values for surfaces that the camera has seen. Whilst this works well in many scenarios, it is easy to imagine scenes where this will break down, such as looking behind the player in a mirror where the camera has not yet looked; or seeing the reflections of ceiling lights in a reflective floor when the user has not yet looked up.
This can be improved by implementing a secondary camera (often referred to as a Lidar camera), at a reduced resolution, to inject supplementary direct lighting input buffers from different views into the direct lighting cache. Kickstart RT fully supports as many lighting input buffers as the game wants to provide and they do not need to be at the native resolution. We recommend 512x512 resolution as a good starting point for performance experimentation.
This would be game/engine specific, but could be set up to spiral around the player whilst rotating around the cameras X and Y axis to increase the visible area injected into the cache. The exact flight path and speed of this camera would need to be tailored for the environment, but has been seen to radically improve the quality of the direct lighting cache although there is a clear performance implication.
Combining the lidar camera with the previously discussed lighting cache debug initialisation colors is a powerful way to optimize the lidar camera path by highlighting unlit regions of the scene.
Whilst loading a new scene, it may be beneficial to warm the lighting cache using the Lidar camera mentioned above. This could be on a predetermined path, or stochastically sampled from within the scenes bounding volume.
An optimisation to reduce the cost of the RT output passes is to enable
checkerboard rendering which runs the reflection passes at a reduced
resolution. If using the NRD denoiser, this is as simple as setting a
flag on the reflection task and denoiser task so the halfResolutionMode
variable is enabled. This will run the reflection passes at a reduced
resolution and the NRD denoiser pass will resolve the images to the full
resolution. The engine does not need to modify any buffer sizes - it
just simply sets the flag.
Note : If the engine chooses to output demodulated specular buffers, then these are not denoised and so will not be resolved to full resolution automatically by NRD.
LODs are a useful optimisation technique in raytracing, as in normal rasterization and the LOD used in the BVH build should ideally match the LOD used in the raster pass. This is not a hard requirement, but using LODs will reduce the complexity of BVHs for geometry which will in turn speed up triangle intersection testing on the GPU.
If the engine switches LODs for a specific object, then the Kickstart RT instance for that LOD should be destroyed and a new instance created for the new LOD. This ensures there is only 1 LOD per object in the TLAS at a time.
The new LOD will have a cold direct lighting cache but this will be updated to full intensity as soon as it is visible in the input buffer.
Optionally, you can use the DirectLightTransfer task to transfer direct lighting information from the old mesh to the new one.
A lot of CPU work typically happens when walking the scene graph each frame looking for objects to raytrace and pass to Kickstart RT for the BVH build and update. The CPU cost of this work can typically be hidden by doing it in a worker thread before it is needed by the Kickstart RT lighting pass.
There is also an associated GPU cost of building the acceleration structures and, if supported, it is recommended to execute it early on the GPUs asynchronous compute queue.
The maxBlasBuildCount
parameter of the BVHBuildTask
can be modified to
limit the number of BVH builds per frame. This is a tradeoff as a small
maximum value will result in low overhead but creates a backlog of work
which can spread over multiple frames resulting in newly built BVH
geometry slowly popping into view. A large maximum value can cause a GPU
spike when there are many BVHs to build, resulting in a frame stutter
and increased GPU memory overhead from the temporary resources used in
the BVH build process.
As this value can be updated each frame, a sensible approach would be to find a high threshold for level load times (when lots of BVHs are built) and a lower threshold at runtime to handle newly visible or dynamic geometry.
A value of 100 is a good starting point for a large scene, but should be tailored to fit the data.
// After generating all of the instances, build the TLAS
KickstartRT::D3D12::BVHTask::BVHBuildTask task;
task.maxBlasBuildCount = c_MaxBlasBuildsPerFrame; // 100
m_taskContainer->ScheduleBVHTask(&task);