Crash using get_devices_for_type() to choose cycles device with python #81877

New Issue

Cedar Myers · 2020-10-20T03:34:03+02:00

Cedar Myers commented

2020-10-20 03:34:03 +02:00

System Information
Operating system: Ubuntu 20.04.1 LTS
Graphics card: GTX 1070 + GTX 1070 + GTX 1050 Ti, NVIDIA 440.100

Blender Version
Broken: 2.91.0 nightly, Commit date: 2020-10-19 22:48, Hash ea4d28aea0
Worked: 2.83.7

Short description of error
This is a follow-up of #81520. Commit 1001adb500 fixes get_devices() but get_devices_for_type() remains broken. As previously something about the combination of OPTIX and OPENCL messes up the CUDA devices. Also I have since discovered that by changing the order getting OPTIX devices can mess with OPENCL devices.

This may be an issue specific to multi-gpu systems. I am willing to reconfigure hardware to help test if needed.

Exact steps for others to reproduce the error

blender --factory-startup --python-expr "import bpy; \
  cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \
  optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \
  opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \
  cuda_devices[0].use = True"

or

blender --factory-startup --python-expr "import bpy; \
  cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \
  opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \
  optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \
  cuda_devices[0].use = True"

More information:
It seems like something is overwriting something else. Are these device locations supposed to be inter-related across types? Perhaps this can help. Consider the different CUDA and OPENCL devices for the CPU at 0x7fa3700c8088 and 0x7fa370412188 respectively in the first example. In the second example below the single device at 0x7f72950c8088 is both the CUDA and OPENCL device. Which behavior is correct? (And does any of this potentially indicate other issues?)

My output (formatted)

blender --factory-startup --python-expr "import bpy; \
  cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \
  print('CUDA devices before getting anything else', cuda_devices); \
  optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \
  print('CUDA devices before getting OPENCL devices', cuda_devices); \
  print('OPTIX devices before getting OPENCL devices', optix_devices); \
  opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \
  print('CUDA devices after getting OPENCL devices', cuda_devices); \
  print('OPTIX devices after getting OPENCL devices', optix_devices); \
  print('OPENCL devices', opencl_devices); \
  cuda_devices[0].use = True"

CUDA devices before getting anything else [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f08>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f88>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa3700c8008>,
  <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7fa3700c8088>
]
CUDA devices before getting OPENCL devices [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f08>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f88>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa3700c8008>,
  <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7fa3700c8088>
]
OPTIX devices before getting OPENCL devices [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412208>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412288>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa370412308>
]
CUDA devices after getting OPENCL devices [
  <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c7f08>,
  <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c7f88>,
  <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c8008>,
  <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c8088>
]
OPTIX devices after getting OPENCL devices [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412208>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412288>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa370412308>
]
OPENCL devices [
  <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7fa370412188>
]
Writing: /tmp/blender.crash.txt
Segmentation fault (core dumped)

and

blender --factory-startup --python-expr "import bpy; \
  cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \
  print('CUDA devices before getting anything else', cuda_devices); \
  opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \
  print('CUDA devices before getting OPTIX devices', cuda_devices); \
  print('OPENCL devices before getting OPTIX devices', opencl_devices); \
  optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \
  print('CUDA devices after getting OPTIX devices', cuda_devices); \
  print('OPENCL devices after getting OPTIX devices', opencl_devices); \
  print('OPTIX devices', optix_devices); \
  cuda_devices[0].use = True"

CUDA devices before getting anything else [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f08>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f88>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7f72950c8008>,
  <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7f72950c8088>
]
CUDA devices before getting OPTIX devices [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f08>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f88>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7f72950c8008>,
  <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7f72950c8088>
]
OPENCL devices before getting OPTIX devices [
  <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7f72950c8088>
]
CUDA devices after getting OPTIX devices [
  <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c7f08>,
  <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c7f88>,
  <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c8008>,
  <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c8088>
]
OPENCL devices after getting OPTIX devices [
  <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c8088>
]
OPTIX devices [
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f7295420108>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f7295420188>,
  <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7f7295420208>
]
Writing: /tmp/blender.crash.txt
Segmentation fault (core dumped)

**System Information** Operating system: Ubuntu 20.04.1 LTS Graphics card: GTX 1070 + GTX 1070 + GTX 1050 Ti, NVIDIA 440.100 **Blender Version** Broken: 2.91.0 nightly, Commit date: 2020-10-19 22:48, Hash ea4d28aea034 Worked: 2.83.7 **Short description of error** This is a follow-up of #81520. Commit 1001adb500 fixes get_devices() but get_devices_for_type() remains broken. As previously something about the combination of OPTIX and OPENCL messes up the CUDA devices. Also I have since discovered that by changing the order getting OPTIX devices can mess with OPENCL devices. This may be an issue specific to multi-gpu systems. I am willing to reconfigure hardware to help test if needed. **Exact steps for others to reproduce the error** ``` blender --factory-startup --python-expr "import bpy; \ cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \ optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \ opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \ cuda_devices[0].use = True" ``` or ``` blender --factory-startup --python-expr "import bpy; \ cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \ opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \ optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \ cuda_devices[0].use = True" ``` **More information:** It seems like something is overwriting something else. Are these device locations supposed to be inter-related across types? Perhaps this can help. Consider the _different_ CUDA and OPENCL devices for the CPU at 0x7fa3700c8088 and 0x7fa370412188 respectively in the first example. In the second example below the single device at 0x7f72950c8088 is both the **CUDA and OPENCL** device. Which behavior is correct? (And does any of this potentially indicate other issues?) My output (formatted) ``` blender --factory-startup --python-expr "import bpy; \ cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \ print('CUDA devices before getting anything else', cuda_devices); \ optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \ print('CUDA devices before getting OPENCL devices', cuda_devices); \ print('OPTIX devices before getting OPENCL devices', optix_devices); \ opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \ print('CUDA devices after getting OPENCL devices', cuda_devices); \ print('OPTIX devices after getting OPENCL devices', optix_devices); \ print('OPENCL devices', opencl_devices); \ cuda_devices[0].use = True" CUDA devices before getting anything else [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f08>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f88>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa3700c8008>, <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7fa3700c8088> ] CUDA devices before getting OPENCL devices [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f08>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa3700c7f88>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa3700c8008>, <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7fa3700c8088> ] OPTIX devices before getting OPENCL devices [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412208>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412288>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa370412308> ] CUDA devices after getting OPENCL devices [ <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c7f08>, <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c7f88>, <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c8008>, <bpy_struct, CyclesDeviceSettings("") at 0x7fa3700c8088> ] OPTIX devices after getting OPENCL devices [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412208>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7fa370412288>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7fa370412308> ] OPENCL devices [ <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7fa370412188> ] Writing: /tmp/blender.crash.txt Segmentation fault (core dumped) ``` and ``` blender --factory-startup --python-expr "import bpy; \ cuda_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('CUDA'); \ print('CUDA devices before getting anything else', cuda_devices); \ opencl_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPENCL'); \ print('CUDA devices before getting OPTIX devices', cuda_devices); \ print('OPENCL devices before getting OPTIX devices', opencl_devices); \ optix_devices = bpy.context.preferences.addons['cycles'].preferences.get_devices_for_type('OPTIX'); \ print('CUDA devices after getting OPTIX devices', cuda_devices); \ print('OPENCL devices after getting OPTIX devices', opencl_devices); \ print('OPTIX devices', optix_devices); \ cuda_devices[0].use = True" CUDA devices before getting anything else [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f08>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f88>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7f72950c8008>, <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7f72950c8088> ] CUDA devices before getting OPTIX devices [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f08>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f72950c7f88>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7f72950c8008>, <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7f72950c8088> ] OPENCL devices before getting OPTIX devices [ <bpy_struct, CyclesDeviceSettings("AMD Ryzen 9 3900X 12-Core Processor") at 0x7f72950c8088> ] CUDA devices after getting OPTIX devices [ <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c7f08>, <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c7f88>, <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c8008>, <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c8088> ] OPENCL devices after getting OPTIX devices [ <bpy_struct, CyclesDeviceSettings("") at 0x7f72950c8088> ] OPTIX devices [ <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f7295420108>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1070") at 0x7f7295420188>, <bpy_struct, CyclesDeviceSettings("GeForce GTX 1050 Ti") at 0x7f7295420208> ] Writing: /tmp/blender.crash.txt Segmentation fault (core dumped) ```

Cedar Myers commented

2020-10-20 03:34:03 +02:00

Added subscriber: @Cedar-Myers

Campbell Barton commented

2020-10-20 04:42:05 +02:00

Added subscriber: @ideasman42

Campbell Barton commented

2020-10-20 04:42:05 +02:00

This is caused by re-allocation each time get_devices_for_type is called, suggest either...

Initialize devices for all types (once) so re-allocation doesn't occur.

Assume this would cause problems if devices are added at run-time (is that even supported?)

Return copies from get_devices_for_type (named-tuples for example).

Is it ever necessary to write to the returned values?

We could make the current get_devices_for_type into a private function, which can be used for drawing, then have a public version that copies the values to avoid these allocation issues.

This is caused by re-allocation each time `get_devices_for_type` is called, suggest either... - Initialize devices for all types (once) so re-allocation doesn't occur. ``` Assume this would cause problems if devices are added at run-time (is that even supported?) ``` - Return copies from `get_devices_for_type` (named-tuples for example). ``` Is it ever necessary to write to the returned values? ``` ---- We could make the current `get_devices_for_type` into a private function, which can be used for drawing, then have a public version that copies the values to avoid these allocation issues.

Raimund Klink commented

2020-10-20 09:15:30 +02:00

Added subscriber: @Raimund58

Philipp Oeser commented

2020-10-20 09:19:05 +02:00

Changed status from 'Needs Triage' to: 'Confirmed'

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Crash using get_devices_for_type() to choose cycles device with python #81877