Registered Extension Number




Extension and Version Dependencies


Extension Proposal

Other Extension Metadata

Last Modified Date


Interactions and External Dependencies
  • Christoph Kubisch, NVIDIA

  • Eric Werness, NVIDIA

  • Josh Barczak, Intel

  • Stu Smith, AMD


When adding adding transparency to a ray traced scene, an application can choose between further tessellating the geometry or using an any hit shader to allow the ray through specific parts of the geometry. These options have the downside of either significantly increasing memory consumption or adding runtime overhead to run shader code in the middle of traversal, respectively.

This extension adds the ability to add an opacity micromap to geometry when building an acceleration structure. The opacity micromap compactly encodes opacity information which can be read by the implementation to mark parts of triangles as opaque or transparent. The format is externally visible to allow the application to compress its internal geometry and surface representations into the compressed format ahead of time. The compressed format subdivides each triangle into a set of subtriangles, each of which can be assigned either two or four opacity values. These opacity values can control if a ray hitting that subtriangle is treated as an opaque hit, complete miss, or possible hit, depending on the controls described in Ray Opacity Micromap.

This extension provides:

New Object Types

New Commands

New Structures

New Enums

New Bitmasks

New Enum Constants



  • Extending VkAccessFlagBits2:



  • Extending VkBufferUsageFlagBits:



  • Extending VkBuildAccelerationStructureFlagBitsKHR:




  • Extending VkGeometryInstanceFlagBitsKHR:



  • Extending VkObjectType:


  • Extending VkPipelineCreateFlagBits:


  • Extending VkPipelineStageFlagBits2:


  • Extending VkQueryType:



  • Extending VkStructureType:











Reference code

uint32_t BarycentricsToSpaceFillingCurveIndex(float u, float v, uint32_t level)
    u = clamp(u, 0.0f, 1.0f);
    v = clamp(v, 0.0f, 1.0f);

    uint32_t iu, iv, iw;

    // Quantize barycentric coordinates
    float fu = u * (1u << level);
    float fv = v * (1u << level);

    iu = (uint32_t)fu;
    iv = (uint32_t)fv;

    float uf = fu - float(iu);
    float vf = fv - float(iv);

    if (iu >= (1u << level)) iu = (1u << level) - 1u;
    if (iv >= (1u << level)) iv = (1u << level) - 1u;

    uint32_t iuv = iu + iv;

    if (iuv >= (1u << level))
        iu -= iuv - (1u << level) + 1u;

    iw = ~(iu + iv);

    if (uf + vf >= 1.0f && iuv < (1u << level) - 1u) --iw;

    uint32_t b0 = ~(iu ^ iw);
    b0 &= ((1u << level) - 1u);
    uint32_t t = (iu ^ iv) & b0;

    uint32_t f = t;
    f ^= f >> 1u;
    f ^= f >> 2u;
    f ^= f >> 4u;
    f ^= f >> 8u;
    uint32_t b1 = ((f ^ iu) & ~b0) | t;

    // Interleave bits
    b0 = (b0 | (b0 << 8u)) & 0x00ff00ffu;
    b0 = (b0 | (b0 << 4u)) & 0x0f0f0f0fu;
    b0 = (b0 | (b0 << 2u)) & 0x33333333u;
    b0 = (b0 | (b0 << 1u)) & 0x55555555u;
    b1 = (b1 | (b1 << 8u)) & 0x00ff00ffu;
    b1 = (b1 | (b1 << 4u)) & 0x0f0f0f0fu;
    b1 = (b1 | (b1 << 2u)) & 0x33333333u;
    b1 = (b1 | (b1 << 1u)) & 0x55555555u;

    return b0 | (b1 << 1u);


(1) Is the build actually similar to an acceleration structure build?

  • Resolved: The build should be much lighter-weight than an acceleration structure build, but the infrastructure is similar enough that it makes sense to keep the concepts compatible.

(2) Why does VkMicromapUsageEXT not have type/pNext?

  • Resolved: There can be a very large number of these structures, so doubling the size of these can be significant memory consumption. Also, an application may be loading these directly from a file which is more compatible with it being a flat structure. The including structures are extensible and are probably a more suitable place to add extensibility.

(3) Why is there a SPIR-V extension?

  • Resolved: There is a ray flag. To be consistent with how the existing ray tracing extensions work that ray flag needs its own extension.

(4) Should there be indirect micromap build?

  • Resolved: Not for now. There is more in-depth usage metadata required and it seems less likely that something like a GPU culling system would need to change the counts for a micromap.

(5) Should micromaps have a micromap device address?

  • Resolved: There is no need right now (can just use the handle) but that is a bit different from acceleration structures, though the two are not completely parallel in their usage.

(6) Why are the alignment requirements defined as a mix of hardcoded values and caps?

  • Resolved: This is most parallel with the definition of VK_KHR_acceleration_structure and maintaining commonality makes it easier for applications to share memory.

Version History

  • Revision 2, 2022-06-22 (Eric Werness)

    • EXTify and clean up for discussion

  • Revision 1, 2022-01-01 (Eric Werness)

    • Initial revision

See Also

Document Notes

For more information, see the Vulkan Specification

This page is a generated document. Fixes and changes should be made to the generator scripts, not directly.

Copyright 2014-2023 The Khronos Group Inc.

SPDX-License-Identifier: CC-BY-4.0