Function like np.random.normal()?

[LI-explorer]

Hi, I have a question about Gaussian sampling. I want to generate 100 Gaussian-distributed points for each of 666 groups, ensuring that the points are different across groups. In NumPy, I can achieve this using:
np.random.normal(loc=0, scale=1, size=(666, 100))
Is there a way to implement the same functionality in Mitsuba or DrJit? Any guidance would be appreciated. Thanks!

[ziyi-zhang]

Hi,

You can warp a uniform sample to a gaussian sample using Box-Muller transformation (e.g., check square_to_std_normal).
Regarding the size, Drjit is not designed for such tensor-style operation. I would draw all 66600 samples in one pass, perform computation in a SIMD manner, and then group by 100 in the end using some reduction operation.

Best

[LI-explorer]

Thanks Ziyi for your response.
For your method, I’m wondering—if we draw all 66600 samples in one pass and then group them separately, will each group still follow the same distribution with mean = 0 and standard deviation = 1?

[ziyi-zhang]

Hi,
That should be fine.

I won't say "each group follows the N(0, 1) distribution" since it's a 1D distribution. More precisely, all 66600 samples are independent samples drawn from N(0, 1), with no correlation within or between group -- which is great for SIMD.

[ziyi-zhang]

Or think about it this way:
Mitsuba/Drjit supports ray-tracing rendering. For a 100×100 image with 4 samples per pixel, it samples and processes all 40000 rays simultaneously, estimating radiance for each in parallel. The results are then grouped by pixels for the final image. This is not that different from your task.