Page MenuHome

CUDA fails to reinitialize
Open, NormalPublic

Description

System Information
Operating system:W10
Graphics card:980Ti

Blender Version
Broken: 2.80 19-04-02
Worked: 2.79

Short description of error
When CUDA worker crashes for any reason (for example you keep modding an object with rendered view enabled the memory is not being freed until the computing is correctly done once, so memory stacks and boom out of memory with a 100mb scene) it cannot be reinitialized correctly.
In 2.79 if your CUDA crashed you only needed to go out of rendered view to solid view, come back to rendered and it was working correctly.
Now once it crashed - it can't render at all. Not even with F12.
You have to restart blender.
Here's console:

CUDA error: Invalid value in cuCtxDestroy(cuContext), line 318
CUDA error at cuCtxCreate: Launch failed

Refer to the Cycles GPU rendering documentation for possible solutions:
https://docs.blender.org/manual/en/dev/render/cycles/gpu_rendering.html

CUDA error: Invalid value in cuCtxPushCurrent(device->cuContext)
CUDA error: Invalid context in mem_alloc_result, line 850
CUDA error: Invalid context in cuCtxPopCurrent(NULL)
CUDA error: Invalid value in cuCtxPushCurrent(device->cuContext)
CUDA error: Invalid context in mem_alloc_result, line 850
CUDA error: Invalid context in cuCtxPopCurrent(NULL)

Exact steps for others to reproduce the error
Just get CUDA to crash in any way. Most times when it happens I just move the view with rendered view enabled and go too close to mesh, or I keep modifying an object on rendered view and memory is not correctly flushed after each modification.

Details

Type
Bug