Mitsuba 3 Spectral Rendering: setting the number of spectral samples

[juanraul8]

Dear Mitsuba 3 team,

Is there a way to define the number of spectral samples which are going to be used for the render? Similar to the setting parameter SPECTRUM_SAMPLES in Mitsuba 0.6.

I tried to compile the scalar_spectral with the following configuration:

 "scalar_spectral": {
        "float": "float",
        "spectrum": "Color<Float, 10>"
    },

Unfortunately, I got the following error:

ImportError: pybind11::detail::get_type_info: unable to find type info for "drjit::Array<float, 10ul>"

Best regards,
Juan Raul Padron Griffe

[njroussel]

Hi @juanraul8

In my limited experience with spectral rendering and Mitsuba 0.6, my understanding is the following:

In Mitsuba 3 (and Mitsuba 2), the wavelengths are defined at runtime and are just an additional dimension to sample and integrate over. Specifically, during the ray generation we make use of Endpoint.sample_ray() which has as its second argument a "uniformly distributed 1D value that is used to sample the spectral dimension of the emission profile.". In fact, usually this means that its second argument is used to make a call to Endpoint.sample_wavelengths().

Now for your compilation issue, I might have misunderstood what you did, but you should just have to add "scalar_spectral" to the "enabled" list of variants if it is not already in it.
The snippet you posted is the template definition that will then be "injected" into Mitsuba. You can learn a bit more about this here.

For spectral variants, you'll see that the template definition is usually something like this:

    "*_spectral": {
        "float": ...,
        "spectrum": "Spectrum<Float, 4>"
    },

The number 4 defines the number of wavelengths per ray, and it is somewhat of an arbitrary choice. You should be able to increase/decrease the number depending on whether or not a single ray (and a single path) should carry more/less spectral information in your scene/setup.

[juanraul8]

Hi @njroussel

Thanks for the explanation!. I actually wrote the code snippet wrong in the message above. The idea was tracking more than 4 wavelengths, lets say 10. Below you can find the mitsuba.conf file I tried, in case you want to reproduce the issue.

{
...

"enabled": [ "scalar_rgb", "scalar_spectral" ],

# The variant below will automatically be set when importing the Python
# mitsuba module. If the value is "", the desired variant will need to be
# specified using "mitsuba.set_variant(..)" before importing any packages.

"python-default": "",

# -------------------------------------------------------------
# The remainder of this file defines the variants and does not
# need to be edited (unless you wish to create new variants).
# -------------------------------------------------------------

# Scalar variant definitions
....

"scalar_spectral": {
    "float": "float",
    "spectrum": "Spectrum<Float, 10>"
},

"scalar_spectral_double": {
    "float": "double",
    "spectrum": "Spectrum<Float, 10>"
},

"scalar_spectral_polarized": {
    "float": "float",
    "spectrum": "MuellerMatrix<Spectrum<Float, 10>>"
},

"scalar_spectral_polarized_double": {
    "float": "double",
    "spectrum": "MuellerMatrix<Spectrum<Float, 10>>"
},

# LLVM variant definitions
...

}

If I tried to compile Mitsuba 3 with this configuration, I got the following error:

[ImportError: pybind11::detail::get_type_info: unable to find type info for "drjit::Array<float, 10ul>"](url)

If I write 3 o 4 wavelengths, the code compiles fine!

[Speierers]

This is probably because Dr.Jit only implements Python bindings for a limited set of drjit::Array (e.g. Array<float/int, 3/4>). You will probably need to add those bindings to get this to work.

Take a look at drjit/src/python/bind.h, it defines the macro DRJIT_BIND_ARRAY_TYPES that generates bindings for the arrays. You could add the following line to add binding for the drjit::Array<float, 10> type:

    DRJIT_BIND_ARRAY_TYPES_DIM(Module, Guide, Scalar, 10)

Although this might not be enough.