How to incorporate many coordination information in a Transform4f variable

[yangocean-sudo]

Hello,

We are trying to simulate the photon as one light ray in mitsuba. The simulation will contain millions of light rays, and after the discussion, using one single emitter is a better choice than using millions of emitters. In the emitter we used before, we hardcore the sample_ray function in spot.cpp, which fixed the local_dir to one value: (0.f, 0.f, 1.f). I noticed that our emitter is very similar to the spotlight emitter: our emitter is to emit rays from the different original points in different directions; while the spotlight emitter is to emit rays from one point to different directions based on the cone.
For now, the emitter reads the binary file which contains the photons' information, and converts them to Transform4f variables. My idea is to pass a list that contains the Transform4f elements to the sample_ray function. In spot.cpp, the sample_ray function returns Ray3f(si.p, m_to_world.value() * local_dir, time, wavelengths), depolarizer<Spectrum>(spec_weight * falloff / pdf_dir), the parameter d is constructed by m_to_world.value() * local_dir, while local_dir is a Vector3f element and it is made up of many elements of triplets. It does make sense because it determines the direction of the lights.
My question is how to create a Transform4f or Vector4f(?) variable like local_dir, that contains all the information of photons so that I can have "photons_tranfo * local_dir" as the return ray's direction parameter, where the local_dir is a fixed vector 3f. Is it possible?

Best wishes,
Yangyang

[njroussel]

Hi @yangocean-sudo

If I understood correctly, this is how you could instantiate a vectorized Transform4f from a std::vector which we'll assume you filled from your binary file.

if constexpr (dr::is_jit_v<Float>) {
    size_t N = 3;
    dr::Matrix<Float, 4> matrix = dr::zeros<dr::Matrix<Float, 4>>(N);

    std::vector<float> values{0, 1, 2};
    matrix.entry(0, 0) = dr::load<Float>(values.data(), N);
    matrix.entry(0, 1) = dr::load<Float>(values.data(), N);
    (...)

    Transform4f tf(matrix);
    std::cout << tf << std::endl;
}

(this snippet only works for JIT variants, making it work for scalar variants would require some more code)

Another important bit is that now local_dir should be a ScalarVector3f as it is no longer vectorized but rather a single vector.

[njroussel]

Yes, you can access a Transform's matrix directly. It's one of the struct's members.

[yangocean-sudo]

Sorry I haven't got the idea. I read the link and the drjit document about matrix for times but didn't find one way to modify the value (submatrix) of a Matrix4f variable. What I'm doing now is to get the Transform4f variables based on my binary file, and then create a Matrix4f variable, the size is equal to my photons' number. I want to replace the submatrix to these Transform4f variables, and then gather with the sample index, convert them to one Transform4f matrix, and finally get the direction vector.

Do I miss anything or which file do I also need to look at? Thank you for your time.

[njroussel]

I must still be misunderstanding what you're trying to achieve.

modify the value (submatrix) of a Matrix4f variable.

What do you mean by submatrix? The i-th matrix in the vectorized array of matrices?

[njroussel]

Here's some pseudo-code of what I think you should be doing in the constructor:

read std::vector<float> per coordinate for each origins, targets and ups from your file
load them each into separate Float variables with dr::load<Float>(std_vec_origins_x.data(), N)
create two Point3f and one Vector3f for origins, targets and ups from these Float variables
create Transform4f from these two Point3f and single Vector3f
get the matrix Matrix4f from the Transform

The sample_ray() can stay the same.

[yangocean-sudo]

Thank you very much! I found the variable type in Mitsuba is more potent than the standard one. With your pseudo-code and help, I finally created one emitter with a vectorized input.