[Bug] drjit error when using batch of radiancemeters + llvm

nestor98 commented 4 months ago

Summary

When using a batch sensor containing radiancemeters, drjit throws an exception in llvm_spectral and llvm_ad_spectral. This doesn't happen if I use the scalar_spectral variant, or if I use perspective sensors instead of radiancemeters. The error is:

Caught a critical exception: drjit::detail::collect_indices(): encountered an uninitialized function argument while recording a virtual function call!

System configuration

System information:

OS: Windows-11 CPU: Intel64 Family 6 Model 151 Stepping 2, GenuineIntel GPU: NVIDIA GeForce GTX 1660 SUPER Python: 3.12.0 (tags/v3.12.0:0fb18b0, Oct 2 2023, 13:03:39) [MSC v.1935 64 bit (AMD64)] NVidia driver: 536.25 CUDA: 12.2.91 LLVM: 16.0.0

Dr.Jit: 0.4.6 (DEBUG) Mitsuba: 3.5.2 (DEBUG) Is custom build? False Compiled with: MSVC 19.37.32822.0 Variants: scalar_rgb llvm_mono scalar_spectral llvm_ad_spectral llvm_spectral

Description

I have narrowed down the scene to a minimal example with a single radiancemeter in the batch sensor. This error also does not happen when using a perspective sensor. The error

Caught a critical exception: drjit::detail::collect_indices(): encountered an uninitialized function argument while recording a virtual function call!

seems to happen by the end of the sample_ray_differential function in batch.cpp, just after calling the radiancemeter sample:

        auto [ray, spec] = sensor->sample_ray_differential(
            time, wavelength_sample, position_sample_2, aperture_sample,
            active);

Which is strange since it seems to finish executing the radiancemeter->sample_ray_differential function.

It may be some misconfiguration on my part, thanks for any help!

Steps to reproduce

This radiancemeter.xml scene will throw the error when executing mitsuba -m llvm_ad_spectral radiancemeter.xml or mitsuba -m llvm_spectral radiancemeter.xml:

<scene version="2.1.0">

<!-- Defaults, these can be set via the command line: -Darg=value -->

    <default name="resx" value="1"/>
    <default name="resy" value="1"/>

<!-- Camera and Rendering Parameters -->

    <integrator type="path" id="integrator" name="integrator">
        <integer name="max_depth" value="-1"/>
        <integer name="rr_depth" value="5000"/>
    </integrator>

    <!-- OPTION 1: -->
    <!-- Caught a critical exception: drjit::detail::collect_indices(): encountered an uninitialized function argument while recording a virtual function call!   -->
    <sensor type="batch" id="sensor" name="sensor">
        <film type="hdrfilm">
            <string name="pixel_format" value="rgba"/>
            <integer name="width" value="$resx"/>
            <integer name="height" value="$resy"/>
        </film>
        <sensor type="radiancemeter" name="sensor_0000">
            <transform name="to_world">
                <rotate x="1" angle="89.999995674289"/>
                <rotate z="1" angle="180.00000500895632"/>
                <translate value="0.000000 -0.100000 0.000000"/>
            </transform>
            <film type="hdrfilm">
                <string name="pixel_format" value="rgba"/>
                <integer name="width" value="$resx"/>
                <integer name="height" value="$resy"/>
            </film>
        </sensor>
    </sensor>
<!-- Materials -->

<!-- Emitters -->

    <shape type="rectangle" id="emitter" name="emitter">
        <transform name="to_world">
            <matrix value="10000.000000 0.000000 0.000000 0.000000 0.000000 0.000000 -10000.000000 0.000000 0.000000 10000.000000 0.000000 0.000000 0.000000 0.000000 0.000000 1.000000"/>
        </transform>
        <emitter type="area" name="emitter">
            <spectrum value="1000.000000" name="radiance"/>
        </emitter>
    </shape>

<!-- Shapes -->

<!-- Volumes -->

</scene>

Alternatives that do work

Executing the above scene with a scalar version: mitsuba -m scalar_spectral radiancemeter.xml

Using the same scene, but replacing the batch sensor with the inner radiancemeter:

<sensor type="radiancemeter" name="sensor_0000">
<transform name="to_world">
  <rotate x="1" angle="89.999995674289"/>
  <rotate z="1" angle="180.00000500895632"/>
  <translate value="0.000000 -0.100000 0.000000"/>
</transform>
<film type="hdrfilm">
    <string name="pixel_format" value="rgba"/>
    <integer name="width" value="$resx"/>
    <integer name="height" value="$resy"/>
</film>
</sensor>

Same scene, just replacing the inner radiancemeter with a perspective sensor:

<sensor type="batch" id="sensor" name="sensor">
<film type="hdrfilm">
    <string name="pixel_format" value="rgba"/>
    <integer name="width" value="$resx"/>
    <integer name="height" value="$resy"/>
</film>
  <sensor type="perspective" name="sensor_0000">
      <transform name="to_world">
          <rotate x="1" angle="89.999995674289"/>
          <rotate z="1" angle="180.00000500895632"/>
          <translate value="0.000000 -0.100000 0.000000"/>
      </transform>
      <film type="hdrfilm">
          <string name="pixel_format" value="rgba"/>
          <integer name="width" value="$resx"/>
          <integer name="height" value="$resy"/>
      </film>
  </sensor>
</sensor>

arwk commented 2 months ago

Hello,

Were you able to find the underlying problem? Today I ran into the same problem, but in my case with the CUDA backend. I was trying to put multiple "radiencemeter" sensors into a "batch" sensor group to get them rendered efficiently.

Kind regards arwk

nestor98 commented 2 months ago

Hi @arwk, unfortunately I didn't manage to solve it. I didn't look into it very in depth after I realized the huge changes that are going on right now in drjit, as they could possibly remove the issue in the near future.

In the meantime, I just render the radiancemeters in a loop, which is quite slower but at least it works.

Best Nestor

njroussel commented 2 months ago

Hi @nestor98

Sorry, I've somehow missed this entirely.

I'll try this out on the new upcoming Dr.JIT and report back :)

arwk commented 2 months ago

Hi @nestor98

Just wanted to share the workaround we came up with, which seems to work quite well. We are using an orthographic camera with a 1x1 pixel film, heavily scaled down to be almost point-like. This seems to approximate a radiencemeter quite well.

The following code defines such a sensor that is compatible with the batch-sensor-plugin


    "approximated-radiencemeter": {
        "type": "orthographic",
        "to_world": mi.ScalarTransform4f.look_at(
            origin=[0,0,3], target=[0, 0, 0], up=[0, 1, 0]
        ) @ mi.ScalarTransform4f.scale([.00001, .00001, 1]),
        # Scale the 1x1 Pixel down, to mimic the radiencemeter
        "film": {
            "type": "hdrfilm",
            "width": 1,
            "height": 1,
            "rfilter": {"type": "box"},
        },
    }

mitsuba-renderer / mitsuba3