DepthTextureCopy

DepthTextureCopy copies an r32float projected-depth texture into a WebGPU render pass's depth attachment while preserving its color. Raster overlays drawn afterward can use ordinary depth testing to remain occluded by ray-traced geometry.

import {DepthTextureCopy} from '@luma.gl/anari';

const depthCopy = new DepthTextureCopy(graphicsDevice);
frame.render();
graphicsDevice.submit();
const channels = frame.captureChannels();
if (channels) {
  try {
    const pass = graphicsDevice.beginRenderPass({
      framebuffer, // The same color target used by frame.render(), with a depth attachment.
      clearColor: false,
      clearDepth: 1
    });
    depthCopy.draw(pass, {depthTexture: channels.projectedDepthTexture});
    // Draw raster labels/geometry here using depthCompare: 'less-equal'.
    // Use the same camera, viewport and zero-to-one clip-depth convention.
    pass.end();
    graphicsDevice.submit();
  } finally {
    channels.destroy();
  }
}
// On teardown:
depthCopy.destroy();

draw(renderPass, options)

The destination must have a depth attachment. The source depthTexture is borrowed, sampled r32float data containing projected device depth in the range zero to one. The helper uses unfiltered texel loads, always passes the depth test, writes depth, and disables color writes. It does not clear attachments or submit GPU commands.

viewportOrigin defaults to [0, 0]. Set it to the destination viewport's pixel origin for a source texture corresponding to an inset viewport. Pixels map one-to-one; the helper does not scale or resample. Pixels outside the source image are discarded, preserving existing destination depth. Clear destination depth to one when uncovered regions should represent background.

Submit a draw before changing options on the same helper instance or destroying its borrowed texture. destroy() releases the helper's model and uniform buffer only.

Ray-traced depth

ANARIFrame.captureChannels().projectedDepthTexture captures one centered primary intersection at full output resolution using the camera of the last rendered frame. Perspective and orthographic cameras are supported. Background is one; the other channels still use zero for misses. This depth is independent of radiance sampling, resolution scaling, and temporal history.

Depth conversion maps the scene camera's OpenGL clip range to WebGPU's zero-to-one attachment range. Raster overlays must use matching camera matrices with the same conversion, viewport dimensions, and standard near-zero/far-one depth convention. Reverse depth and custom depth ranges are not handled automatically. The capture uses opaque intersections; transparent transport and multisampled edge coverage remain separate work. The application owns overlay geometry, label layout, target selection, and command submission. This helper does not modify the temporal radiance history.