Shader Execution Reordering (SER) is an addition to DirectX Raytracing that enables application shader code inform hardware how to find coherency across rays so they can be sorted to execute better in parallel.

At GDC 2025 DXR 1.2 was announced including SER, and you can see it discussed in this: GDC DirectX State Of The Union YouTube Recording. In the video, Remedy showed the performance wins they achieve using a synergistic combination of Opacity Micromaps (OMMs) and Shader Execution Reordering in Alan Wake 2.

SER is designed such that when applications use it to give sorting hints to drivers, it is up to drivers/hardware to actually make use of the information and sort. SER support is a required feature in Shader Model 6.9, meaning all drivers must accept shader code using SER, but it is up to individual implementations as to whether they can actually make use of the information. So apps can write code once, at no disadvantage to not using SER, but the same code also runs better on implementations that are able to take advantage.

For now this is a preview feature, until Shader Model 6.9 is finalized.  NVIDIA and Intel developer preview driver support is available now, with AMD support coming soon.

The rest of this blog summarizes the feature, how to get bits, and highlights some sample code to help get started.

Parent blog for all other features in this release.

Overview

Because of the stochastic nature of many raytracing workloads, DXR applications often suffer from divergent shader execution and divergent data access. Tackling the problem with application-side logic has many downsides, both in terms of achievable performance and developer effort. The existing DXR API allows implementations to dynamically schedule shading work triggered by TraceRay and CallShader, but does not offer a way for the application to control scheduling in any way. Shader Execution Reordering (SER) fills this gap by introducing HLSL primitives that enable application-controlled reordering of work across the GPU for improved execution and data coherence.

Furthermore, the current TraceRay pipeline of traversal and ClosestHit/Miss shading is not always flexible enough. First, common code, such as vertex fetch and interpolation, must be duplicated in all ClosestHit shaders. Second, simple visibility rays must unnecessarily execute hit shaders in order to access basic information about the hit. To address these problems, the concept of a HitObject decouples raytracing traversal (including AnyHit shading and Intersection shading) from ClosestHit and Miss shading. This enables arbitrary RayGeneration code to execute between traversal, execution reordering, and ClosestHit/Miss handling, and allows ClosestHit/Miss dispatch starting from hit information from sources other than traversal, such as RayQuery.

The combination of HitObject and SER is particularly powerful and enables reordering for execution and data coherence using information in the HitObject and additional hints supplied by the user. The result is further improved coherence potential for hit/miss processing.

Specification (Docs)

For full documentation see the Shader Execution Reordering section of the DXR spec.

The DXR spec also has a section describing D3D12_RAYTRACING_TIER_1_2 including how SER fits in.

Availability

SER is a required part of Shader Model 6.9, currently in preview. This requires:

• AgilitySDK 1.717.0-preview available here.
• Preview Shader Model 6.9 support in DXC available here.

Device support:

NVIDIA: SER is already accelerated on all NVIDIA GeForce RTX™ 40 and 50 Series GPUs, access the driver here (requires an NVIDIA Developer Program account).

Intel: Support for SER is now available for Intel® Arc™ B-Series Graphics and Intel® Core™ Ultra Processors (Series 2) with the Intel® Arc™ Graphics developer preview driver available here.

AMD: AMD driver support for SER will be made available during Summer 2025.

WARP: The latest WARP software rasterizer preview supports DXR 1.2 including SER, available here.

To use the preview AgilitySDK with shader model 6.9, your machine needs to be in developer mode, and before calling D3D12CreateDevice() enable experimental shader models like this:

UUID Features[] = { D3D12ExperimentalShaderModels };
ThrowIfFailed(D3D12EnableExperimentalFeatures(_countof(Features), Features, nullptr, nullptr));

After creating a device, check for SER support by checking for Shader Model 6.9:

D3D12_FEATURE_DATA_SHADER_MODEL SM;
SM.HighestShaderModel = D3D_SHADER_MODEL_6_9;
m_dxrDevice->CheckFeatureSupport(D3D12_FEATURE_SHADER_MODEL, &SM, sizeof(SM));
ThrowIfFalse(SM.HighestShaderModel >= D3D_SHADER_MODEL_6_9,
    L"ERROR: Device doesn't support Shader Model 6.9.\n");

Also make sure raytracing is supporting by checking the raytracing tier (not shown here).  There is a D3D12_RAYTRACING_TIER_1_2 tier that can be queried, but that means all the features in this tier are supported: SER and Opacity Micromaps. If only SER is needed, just check for Shader Model 6.9 and D3D12_RAYTRACING_TIER_1_0/1_1 as needed.

Once Shader Model 6.9 goes out of preview there will also be a way to ask the device if it actually tries to do thread sorting requested by use of the SER feature, and it isn’t just a no-op. This would just be a convenience, for instance during development and testing particular devices, or if an app wanted to do its own manual sorting if SER wasn’t going to actually sort.

Content from NVIDIA

RTX Path Tracing is a code sample that strives to embody years of raytracing and neural graphics research and experience. It is intended as a starting point for a path tracer integration, as a reference for various integrated SDKs, and/or for learning and experimentation. This now has a DXR path with SER.

PIX

As usual SER comes with Day One PIX support. Please read the PIX blog post for more information.

Simple Microsoft SER Sample

The D3D12RaytracingHelloShaderExecutionReordering modifies the original D3D12RaytracingHelloWorld sample to minimally demonstrate various uses of Shader Execution Reordering and showing performance gains described below.

D3D12RaytracingHelloShaderExecutionReordering can be found in the DirectX-Graphics-Samples repo on github here.

This sample simply draws a fullscreen quad with triangle barycentrics used as the pixel color. Each ray does some artificial work when shading, and some proportion of rays do a heavier artificial workload, rendered white (vertical stripes). The Ray Generation Shader uses SER to tell the system which threads will be more expensive so it can try to sort similar threads to be together.

The shader file, Raytracing.hlsl contains some configuration options that can be tweaked before running the app, where the shader is compiled at launch. The options allow comparing the performance of ways of using SER, as well as not using SER at all. In fact the mechanics SER can be understood simply by playing with this shader file and running the app, ignoring the rest of the boilerplate C++ code in the sample.

Using SER with the settings below running on an NVIDIA RTX 4090 showed a 40% framerate increase versus not using SER, and a couple of configurations of Intel Arc B-Series GPUs each showed a 90% framerate increase.

//*********************************************************
// Configuration options
//*********************************************************

// TraceRay the old fashioned way
//#define USE_ORIGINAL_TRACERAY_NO_SER

// Call MaybeReorderThread(sortKey,1), sortKey is 1 bit 
// indicating if the thread has dummy work
#define REQUEST_REORDER

// Don't invoke ClosestHit or Miss shaders, use hitObject 
// properties in RayGen to shade
//#define SKIP_INVOKE_INSTEAD_SHADE_IN_RAYGEN

// Rays do loop a of artificial work in the 
// Closest Hit shader.  This setting makes 
// some rays looping more than others (a sort candidate):
#define USE_VARYING_ARTIFICIAL_WORK

// Number of iterations in the heavy artificial work loop
#define WORK_LOOP_ITERATIONS_HEAVY 5000

// Number of iterations in the light artificial work loop
#define WORK_LOOP_ITERATIONS_LIGHT 1000

// N, where 1/N is the proportion of rays that do the 
// heavy artificial work load
#define RAYS_WITH_HEAVY_WORK_FRACTION 4

// Put all the rays with dummy work on the left side
// #define SPATIALLY_SORTED

//*********************************************************

Below is the sample’s Ray Generation Shader illustrating various basic uses of SER via the above options. Notice that when SER is used, TraceRay returns a HitObject.

Depending on the config, the shader can call MaybeReorderThread(), in this case taking a shader defined sort key, though there’s another variant not shown that takes the hit object and sorts on its properties.

Finally, depending on the config, the shader can call HitObject::Invoke() to run Closest Hit or Miss Shader on the hit, or not bother calling Invoke() at all and do shading locally based on hit object properties. In this case shading is based on hit attributes (barycentrics) returned via hit.GetAttributes().

using namespace dx; // dx::HitObject and dx::MaybeReorderThread
[shader("raygeneration")]
void MyRaygenShader()
{
    RayDesc ray = 
        SetupRay(DispatchRaysIndex(), DispatchRaysDimensions()); 

    uint iterations = WORK_LOOP_ITERATIONS_LIGHT;

    #ifdef USE_VARYING_ARTIFICIAL_WORK

        #ifdef SPATIALLY_SORTED
            // Extra work is all on left side of screen
            if((origin.x + 1)/2.f <= 1.f/RAYS_WITH_HEAVY_WORK_FRACTION)
            {
                iterations = WORK_LOOP_ITERATIONS_HEAVY; 
            }
        #else
            // Extra work distributed in vertical bands
            if( (DispatchRaysIndex().x) % RAYS_WITH_HEAVY_WORK_FRACTION == 0 )
            {
                iterations = WORK_LOOP_ITERATIONS_HEAVY; 
            }
        #endif

    #endif

    RayPayload payload = { float4(0, 0, 0, 0), iterations };
    float4 color = float4(1,1,1,1);

    #ifdef USE_ORIGINAL_TRACERAY_NO_SER
        TraceRay(Scene, RAY_FLAG_NONE, ~0, 0, 1, 0, ray, payload);
        color = payload.color;
    #else

        HitObject hit = 
            HitObject::TraceRay(Scene, RAY_FLAG_NONE, ~0, 0, 1, 0, 
                                ray, payload);

        #ifdef REQUEST_REORDER
            int sortKey = iterations != WORK_LOOP_ITERATIONS_LIGHT ? 1:0;
            dx::MaybeReorderThread(sortKey, 1);

            // There's currently a DXC bug that causes "using namespace dx;" 
            // (at the top) to generate bad DXIL for MaybeReorderThread, 
            // so it's explicitly scoped here. The namespace works fine for 
            // HitObject
        #endif

        #ifdef SKIP_INVOKE_INSTEAD_SHADE_IN_RAYGEN
            if(hit.IsHit())
            {
                MyAttributes attr = hit.GetAttributes();
                color = ClosestHitWorker(attr,iterations);
            }
            else
            {
                color = MissWorker();
            }

        #else
            HitObject::Invoke(hit, payload);
            color = payload.color;
        #endif

    #endif

    // Write the raytraced color to the output texture.
    RenderTarget[DispatchRaysIndex().xy] = color;
}