[Bug]: LLVM Fatal Error: Cannot select FrameIndex<4> In function: func_narrow_phase_convex_specializations_c414_0_kernel_0_serial on AMD GPU

[ZJLi2013]

Bug Description

hi, genesis team

many thanks for this great work. I am recently verifing genesis on AMD GPUs, in a simple case with box-box-collision=true, and it has the following runtime errors:

  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/genesis/engine/scene.py", line 857, in build
    self._sim.step()
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/genesis/engine/simulator.py", line 282, in step
    self.rigid_solver.substep(self.cur_substep_local)
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/genesis/engine/solvers/rigid/rigid_solver.py", line 947, in substep
    self._func_constraint_force()
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/genesis/engine/solvers/rigid/rigid_solver.py", line 1029, in _func_constraint_force
    self.collider.detection()
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/genesis/engine/solvers/rigid/collider/collider.py", line 520, in detection
    func_narrow_phase_convex_specializations(
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/quadrants/lang/_quadrants_callable.py", line 96, in __call__
    return self.wrapper.__call__(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/quadrants/lang/kernel_impl.py", line 141, in wrapped_func
    return primal(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/quadrants/lang/kernel.py", line 575, in __call__
    ret = self.launch_kernel(key, kernel_cpp, compiled_kernel_data, *py_args)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/quadrants/lang/kernel.py", line 509, in launch_kernel
    raise e from None
  File "/opt/conda/envs/py_3.12/lib/python3.12/site-packages/quadrants/lang/kernel.py", line 504, in launch_kernel
    prog.launch_kernel(compiled_kernel_data, launch_ctx)
RuntimeError: [llvm_context.cpp:operator()@83] LLVM Fatal Error: Cannot select: 0x562cfe212020: i64 = FrameIndex<4>
In function: func_narrow_phase_convex_specializations_c414_0_kernel_0_serial

trace as following

scene.build()
  → sim.step()                           # scene.py:857
    → rigid_solver.substep()             # rigid_solver.py:947
      → _func_constraint_force()         # rigid_solver.py:1029
        → collider.detection()           # collider.py:520
          → func_narrow_phase_convex_specializations()  # narrowphase.py

wonder is it a knonw issue ?

Thanks
David

Steps to Reproduce

If possible, provide a script triggering the bug, e.g.

# code snippet triggering the bug
import genesis as gs
gs.init()
scene = gs.Scene(show_viewer=False)
broken = scene.add_entity(
    gs.morphs.MJCF(file='attachement.xml'),
)
scene.build()

Make sure to attached any needed assets (here attachement.xml)!

Expected Behavior

should work right on AMD GPUs

Screenshots/Videos

No response

Relevant log output

Environment

• OS: [e.g. Ubuntu 24.04, Windows 11 24H2]
• GPU/CPU [e.g. A100, RTX 4090, M3pr, Intel I9-9900k, Ryzen 5900x] (N/A if no GPU/CPU)
• GPU-driver version (N/A if no GPU)
• CUDA / CUDA-toolkit version (N/A if non-Nvidia)

Release version or Commit ID

geneiss 0.4.1 & 0.4.3

Additional Context

No response

[v01dXYZ]

@ZJLi2013 Can you add attachement.xml to your description to let us inspect the issue?

[ZJLi2013]

@ZJLi2013 Can you add attachement.xml to your description to let us inspect the issue?

thanks for replying, it's (file="xml/franka_emika_panda/panda.xml"), actually from the genesis repo

[v01dXYZ]

@ZJLi2013 next time you report such a bug, could you write a full failure reproducer script and your environment specs? It would be easier for me to catch the failure.

I tested with different versions using a AMD Instinct MI300X VF (I didn't test 0.4.5 as it uses cuda.bindings).

• master (e807698): unable to reproduce failure (quadrants: 0.5.2)
• 0.4.4: unable to reproduce failure (quadrants: 0.4.5)
• 0.4.3: same failure as described (quadrants: 0.4.4)
• 0.4.2: same failure as described (quadrants: 0.4.4)
• 0.4.1: same failure as described (quadrants: 0.4.3)

Using this code snippet:

# code snippet triggering the bug
import genesis as gs
gs.init()
scene = gs.Scene(
    show_viewer=False,
    rigid_options=gs.options.RigidOptions(box_box_detection=True),
)
broken = scene.add_entity(
    gs.morphs.MJCF(file='Genesis/genesis/assets/xml/franka_emika_panda/panda.xml'),
)
scene.build()

[v01dXYZ]

@hughperkins Should I keep trying to investigate or is it enough to close the issue?

[duburcqa]

The bug in Quadrants is most likely not fixed. It just happens that Genesis’ source code has sufficiently changed to not trigger this bug anymore. It would be great to truly fix it.

[v01dXYZ]

Below is the log file I'm going to analyze. It contains the verbose output of LLVM for the run of the snippet.

fail_logs_xz9.txt

[ZJLi2013]

@ZJLi2013 next time you report such a bug, could you write a full failure reproducer script and your environment specs? It would be easier for me to catch the failure.

I tested with different versions using a AMD Instinct MI300X VF (I didn't test 0.4.5 as it uses cuda.bindings).

• master (e807698): unable to reproduce failure (quadrants: 0.5.2)
• 0.4.4: unable to reproduce failure (quadrants: 0.4.5)
• 0.4.3: same failure as described (quadrants: 0.4.4)
• 0.4.2: same failure as described (quadrants: 0.4.4)
• 0.4.1: same failure as described (quadrants: 0.4.3)

Using this code snippet:

code snippet triggering the bug

import genesis as gs
gs.init()
scene = gs.Scene(
show_viewer=False,
rigid_options=gs.options.RigidOptions(box_box_detection=True),
)
broken = scene.add_entity(
gs.morphs.MJCF(file='Genesis/genesis/assets/xml/franka_emika_panda/panda.xml'),
)
scene.build()

thanks a lot, just double confirm with: box_box_detection=True with Genesis 0.4.5 + Quadrants 0.5.2 and it works. looks this bug is out-of-date, thanks for confirm again.

[duburcqa]

This bug is not fixed. Let's keep this issue open.

[v01dXYZ]

Done with LLVM 20.1.0, using the archive provided by quadrants CI.

Below are the faulty module and its reduced version:

• fail_llvm_ll.txt
• reduced_ll.txt

The reduced version is:

target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-p9:192:256:256:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8:9"
target triple = "amdgcn-amd-amdhsa"

define amdgpu_kernel void @func_narrow_phase_convex_specializations_c424_0_kernel_0_serial() {
entry:
  %0 = alloca [9 x float], i32 0, align 4
  store float 0.000000e+00, ptr %0, align 4
  ret void
}

[v01dXYZ]

I think it is related to LLVM 20 not rejecting alloca without addrspace(5). I ve got to play with LLVM 20 and LLVM 22 before being sure about that.

[hughperkins]

I think it is related to LLVM 20 not rejecting alloca without addrspace(5). I ve got to play with LLVM 20 and LLVM 22 before being sure about that.

Note that we don't use LLVM 20 any more.