Closed free7187 closed 7 months ago
Speierers
commented
8 months ago Hi @free7187,
Could you maybe provide a simple reproducer for this issue? It is otherwise difficult to guess what's going wrong here.
Also, could you clarify whether those are mesh spheres or analytical spheres?
learner-shx
commented
8 months ago Hi @free7187,
Could you maybe provide a simple reproducer for this issue? It is otherwise difficult to guess what's going wrong here.
Also, could you clarify whether those are mesh spheres or analytical spheres?
Hi @Speierers, Sorry for unclear description, here is an example code to reproduce the issue.
import sys
sys.path.insert(0, "./mitsuba3/build/Release/python")
import mitsuba as mi
import drjit as dr
import time
import random
class SceneDescFactory:
def __init__(self, spp, res, max_depth, integrator):
self.scene_desc = {}
# init scene
self.init_scene_desc(spp, res, max_depth, integrator)
def init_scene_desc(self, spp, res, max_depth, integrator):
self.scene_desc = {
"type": "scene",
"sensor": {
"type": "perspective",
"fov_axis": "smaller",
"near_clip": 0.001,
"far_clip": 100,
"focus_distance": 1000,
"fov": 39.3077,
"to_world": mi.ScalarTransform4f.look_at(
origin =(0, 0, 4),
target =(0, 0, 0),
up =(0, 1, 0)
),
"sampler":{
"type": "independent",
"sample_count": spp
},
"film": {
"type": "hdrfilm",
"width": res,
"height": res,
"rfilter": {
"type": "tent"
},
"pixel_format": "rgb",
"component_format": "float32"
}
},
"integrator": {
"type": "direct",
}
}
def merge_desc(self, *desc_list):
merged_desc = {}
for desc in desc_list:
merged_desc = {**merged_desc, **desc}
return merged_desc
def get_debug_light_desc(self):
light_desc = {
"light": {
"type": "obj",
"filename": r".\mitsuba3\tutorials\scenes\meshes\cbox_luminaire.obj",
"to_world": mi.ScalarTransform4f.translate((0, -0.01, 0)),
"bsdf": {
"type": "diffuse",
"reflectance": {
"type": "rgb",
"value": [1, 1, 1]
}
},
"emitter": {
"type": "area",
"radiance": {
"type": "rgb",
"value": [18.387, 13.9873, 6.75357]
}
}
}
}
return light_desc
def get_sphere_desc(self, index):
sphere_desc = {
f"sphere_{index}": {
"type": "sphere",
"bsdf": {
"type": "diffuse",
"reflectance": {
"type": "rgb",
"value": [0.105421, 0.37798, 0.076425]
}
}
}
}
return sphere_desc
def get_scene_desc(self):
return self.scene_desc
def add_sphere(self, index):
self.scene_desc[f"shpere_{index}"] = self.get_sphere_desc(index)[f"sphere_{index}"]
def add_debug_light(self):
self.scene_desc = self.merge_desc(self.scene_desc,
self.get_debug_light_desc())
mi.set_variant('cuda_ad_rgb')
sdf = SceneDescFactory(512, 512, 5, "prb")
start_time = time.time()
print(time.time() - start_time, " ----- editting sdf")
for i in range(10000):
sdf.add_sphere(i)
sdf.add_debug_light()
dic = sdf.get_scene_desc()
print(time.time() - start_time, " ----- loading scene desc")
scene = mi.load_dict(sdf.get_scene_desc())
print(time.time() - start_time, " ----- traversing scene")
params = mi.traverse(scene)
print(time.time() - start_time, " ----- rendering")
image_ref = mi.render(scene, spp=512)
print(time.time() - start_time, " ----- end initial render" )
mi.Bitmap(image_ref[:, :, :]).write("test_sphere_ref.exr")
Place the file in mitsuba3's parent directory to assure it can find the relative paths correctly. Then it can be run. Thank you!
Speierers
commented
8 months ago Thanks for the reproducer. Before trying to run it next week, can you check whether this issue persist if all the spheres don't overlap perfectly? Also does this issue also occur on LLVM mode?
learner-shx
commented
8 months ago Thanks for the reproducer. Before trying to run it next week, can you check whether this issue persist if all the spheres don't overlap perfectly? Also does this issue also occur on LLVM mode?
I tried set to_world with random translate, the problem was gone, thank you.
njroussel
commented
7 months ago I can't reproduce this on either Embree/OptiX. I'll close this for now.
It's still quite likely that there is an underlying issue, if anyone comes across something similar we'll need to dig into this further.
Summary
illegal memory access when large amount of spheres.
System configuration
System information:
OS: Windows-10 CPU: Intel64 Family 6 Model 183 Stepping 1, GenuineIntel GPU: NVIDIA GeForce RTX 4090 Python: 3.10.9 | packaged by Anaconda, Inc. | (main, Mar 1 2023, 18:18:15) [MSC v.1916 64 bit (AMD64)] NVidia driver: 536.25 CUDA: 11.8.89 LLVM: 15.-1.-1
Dr.Jit: 0.4.4 Mitsuba: 3.5.0 Is custom build? False Compiled with: MSVC 19.38.33133.0 Variants: scalar_rgb scalar_spectral cuda_ad_rgb llvm_ad_rgb
Description
When I use python mitsuba to load a dict that contains about 10000 spheres, and render, I get
with path integrator, 512 resolution and 16 spp, use cuda_ad_rgb.
Steps to reproduce