Diffmpm.py crashes during the forward simulation

[increew]

Out-of-bounds Memory Access in the DiffMpm Example

In the example given in diffmpm.py, the code crashes during an iteration of my training process. I found that the there is an array-out-of-bounds access during the forward simulation.

How to reproduce

System specifications

• OS: Ubuntu 18.04 LTS
• Taichi version: 1.4.1
• Python version: 3.9.16

To reproduce the bug, only the forward simulation is needed.

The necessary files are given in https://www.dropbox.com/s/obhpljd9qp8vj10/bug_report.zip?dl=0. The value of parameters weights and bias when the crash happens is given in weights.npy and bias.npy, respectively.

To get a minimal example that reproduces the crash, run the file bug_report.py, which simply loads the value of weights and bias from their corresponding numpy files and run forward simulation for 1500 steps. You would observe the output saying:

[Taichi] version 1.4.1, llvm 15.0.4, commit e67c674e, linux, python 3.9.16
[Taichi] Starting on arch=x64
n_particles 2976
n_solid 2976
Traceback (most recent call last):
  File "/home/user/bug_report.py", line 419, in <module>
    reproduce_bug()
  File "/home/user/bug_report.py", line 364, in reproduce_bug
    forward(1500)
  File "/home/user/bug_report.py", line 254, in forward
    advance(s)
  File "/home/user/miniconda3/envs/tai/lib/python3.9/site-packages/taichi/ad/_ad.py", line 311, in decorated
    func(*args, **kwargs)
  File "/home/user/bug_report.py", line 234, in advance
    p2g(s)
  File "/home/user/miniconda3/envs/tai/lib/python3.9/site-packages/taichi/lang/kernel_impl.py", line 976, in wrapped
    raise type(e)('\n' + str(e)) from None
taichi.lang.exception.TaichiAssertionError:
(kernel=p2g_c82_0) Accessing field (S25place<f32>) of size (128, 128) with indices (-5980, -8119)

[erizmr]

Hi @increew , thanks for reporting this. I can reproduce the bug you mentioned and I think it is induced by simulation numerical overflow.