Limit size of feature per sample

[qbhan]

Summary

I am trying to create a dataset for various Monte Carlo Denoisers.
For instance, to train a Sample-based Monte Carlo Denoising (Gharbi et al. 2019) with Path-Manifold Module (Cho et al.2021) I need about 51+36 = 87 channels of features per sample.

I successfully managed to get 51 channels for Sample-based Monte Carlo Denoising. However, I'm facing an error as I try to get more than 75 channels. Below are the printed 76 channels I want to get from each sample and the error message.

(mitsuba) kyubeom@793f1e47e7e1:~/mitsuba3_playground$ python integrators.py 
['normal:3f', 'depth:1f', 'albedo:3f', 'normal_diff:3f', 'depth_diff:1f', 'albedo_diff:3f', 'visibility_b0:1f', 'visibility_b1:1f', 'direction_b0:2f', 'direction_b1:2f', 'direction_b2:2f', 'direction_b3:2f', 'direction_b4:2f', 'direction_b5:2f', 'probability_b0:4f', 'probability_b1:4f', 'probability_b2:4f', 'probability_b3:4f', 'probability_b4:4f', 'probability_b5:4f', 'material_b0:1i', 'material_b1:1i', 'material_b2:1i', 'material_b3:1i', 'material_b4:1i', 'material_b5:1i', 'throughput_b0:3f', 'throughput_b1:3f', 'throughput_b2:3f', 'throughput_b3:3f', 'throughput_b4:3f', 'throughput_b5:3f']


Critical Dr.Jit compiler failure: jit_var(r3215): unknown variable!
Aborted (core dumped)

I'm wondering if there is any memory limit that a loop of a single sample can handle.
I would really appreciate some help on how to solve the problem.

System configuration

System information:

OS: Ubuntu 20.04
CPU: Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz
GPU: RTX3090
Python version: 3.11.0
LLVM version: None
CUDA version: 11.6
NVidia driver: 510.73.05

Dr.Jit version: 0.4.0
Mitsuba version: 3.2.0
Compiled with: pip install mitsuba
Variants compiled:
scalar_rgb
scalar_spectral
cuda_ad_rgb
llvm_ad_rgb

Description

I added a list of features as "aovs" on a simple path tracer with NEE and MIS. The name of features are given in the initialization stage.
The code below shows how the loop of sample() is configured.

def sample(self, scene: mi.Scene,
               sampler: mi.Sampler, 
               ray_: mi.RayDifferential3f, 
               medium: mi.Medium = None, 
               active: bool = True
               ):
        bsdf_ctx = mi.BSDFContext()


        # --------------------- Configure loop state ----------------------

        ray = mi.Ray3f(ray_)
        prev_si = dr.zeros(mi.SurfaceInteraction3f)

        depth = mi.UInt32(0)                            # Depth of current vertex
        L = mi.Spectrum(0.0)                            # Radiance accumulator
        β = mi.Spectrum(1.0)                            # Path throughput weight
        η = mi.Float(1)                                 # Index of refraction
        mis_em = mi.Float(1)                            # Emitter MIS weight
        active = mi.Bool(active)                        # Active SIMD lanes
        
        aovs = []
        for i in range(len(self.aov_list)):
            if '1f' in self.aov_list[i]: aovs.append(mi.Vector1f(0.0))
            elif '2f' in self.aov_list[i]: aovs.append(mi.Vector2f(0.0))
            elif '3f' in self.aov_list[i]: aovs.append(mi.Vector3f(0.0))
            elif '4f' in self.aov_list[i]: aovs.append(mi.Vector4f(0.0))
            elif '1i' in self.aov_list[i]: aovs.append(mi.Vector1i(0))


        direction = mi.Vector2f(0.0)

        first_non_specular = mi.Bool(False)
        depth_diff = mi.Float(0.0)
        β_diff = mi.Spectrum(1.0)
        L_diff = mi.Spectrum(0.0)

        loop = mi.Loop(name="Path Tracing", state=lambda: (
            sampler, ray, prev_si, depth, L, β, η, mis_em, active, aovs, first_non_specular, depth_diff, β_diff, L_diff))

        loop.set_max_iterations(self.max_depth)

Even though I assign dummy values like mi.Vector3f(1.0) still the error occurs.
I really appreciate the help!

[njroussel]

Hi @qbhan

Could you try compiling the project yourself rather than using the pip install version? I don't think there is any limit to the AOVs, but rather that you're encountering the same issue as here: #482. Compiling Mitsuba on your own using master might fix your problem.

[qbhan]

I just complied it manually and the error disappeared. Thank you!