Max Ray Tracing Depth

[hlavacs]

In rt_base.hpp
function
VkRayTracingPipelineCreateInfoKHR createRayTracingPipelineCreateInfo(std::span stages,
std::span shaderGroups,
uint32_t depth = 2)

The max depth is set to

    .maxPipelineRayRecursionDepth = std::max(2U, m_rtProperties.maxRayRecursionDepth),  // Ray depth

so parameter depth is ignored.

Should it be :

    .maxPipelineRayRecursionDepth = std::max(2U, std::min(depth, m_rtProperties.maxRayRecursionDepth)),  // Ray depth

?

[mklefrancois]

Thank you for pointing this out! You're absolutely right - the depth parameter was being completely ignored.

There is indeed an error and what we want is

.maxPipelineRayRecursionDepth = std::min(depth, m_rtProperties.maxRayRecursionDepth),  // Ray depth

The tutorial 06_reflection will be also be partially re-written to better describe the limitations of recursive TraceRay() calls and why the iterative approach is recommended for production use.

Thanks again for catching this!

[hlavacs]

Sure, and for good measure another one.

In 16_ray_query/shaders/ray_query.slang , in pathTrace() you have:

const float3 w = 1.0f; // Light sampling weight (simplified)
contrib += w * evalData.bsdf_diffuse; // Diffuse reflection
contrib += w * evalData.bsdf_glossy; // Specular reflection
contrib *= throughput; // Weight by path throughput

So no PDF weighing, and no true MIS, but basically NEE only for direct light sampling, and BSDF only for indirect light sampling.

[mklefrancois]

You're absolutely right! The implementation uses:

• NEE for direct lighting with w = 1.0
• BSDF sampling for indirect lighting only
• No explicit MIS between strategies

However, this is mathematically correct for delta light sources (directional lights/sun).
For delta distributions, the light PDF is infinite at the exact light direction, making the MIS weight:
mis_weight = pdf_light / (pdf_light + pdf_bsdf) = ∞ / (∞ + pdf_bsdf) = 1.0

Since BSDF sampling has zero probability of hitting a delta light exactly, there's no additional contribution to account for from that strategy.

The tutorial doesn't have area lights or environment map importance sampling, so full MIS isn't needed here. That said, bsdfEvaluateSimple() does compute evalData.pdf, which would enable proper MIS if area lights were added in the future.

I've added detailed comments to the shader explaining this. For reference, our production implementation with full MIS (handling area lights + environment) can be seen in gltf_pathtrace.slang at https://github.com/nvpro-samples/vk_gltf_renderer.

Thanks for the careful review!

[hlavacs]

Hallo, ah I see, thanks for the explanation! Its point lights, so weight is 1.