Maniphest T39526

CUDA kernel compilation fails with PTX assembler internal failure
Closed, Duplicate
Actions

Assigned To
Brecht Van Lommel (brecht)
Authored By
Benjamin (Benjamin-L)
Mar 31 2014, 5:39 AM
Tags
  • BF Blender
  • Cycles
Subscribers
Bastien Montagne (mont29)
Benjamin (Benjamin-L)
Brecht Van Lommel (brecht)
Daniel Salazar (zanqdo)
Martijn Berger (juicyfruit)
Thomas Dinges (dingto)

Description

System Information
OS: Gentoo Linux (kernel version: 3.0), Graphics Card: Nvidia GeForce GT 640, Cuda version: 5.5

Blender Version
Broken: 2.70, c24c343
Worked: 2.70, unknown hash (downloaded as source code from blender website for 2.70 release rather than using git)

Short description of error
When rendering using Cuda for the first time it needed to compile the kernels. It failed at this step with the following error message:

CUDA version 5.5 detected, build may succeed but only CUDA 5.0 is officially supported.
Compiling CUDA kernel ...
"nvcc" -arch=sm_30 -m32 --cubin "/root/blender-versions/mycopy/build/bin/2.70/scripts/addons/cycles/kernel/kernel.cu" -o "/root/.config/blender/2.70/cache/cycles_kernel_sm30_95088DA6C7DFF4F1EA0DD66BE94D2583.cubin" --ptxas-options="-v" --maxrregcount=32 --use_fast_math -I"/root/blender-versions/mycopy/build/bin/2.70/scripts/addons/cycles/kernel" -DNVCC -DKERNEL_CUDA_VERSION=55
nvcc error : 'ptxas' died due to signal 11 (Invalid memory reference)
CUDA kernel compilation failed, see console for details.

Refer to the Cycles GPU rendering documentation for possible solutions:
ht

I do not think that this is actually a problem with Cuda 5.5 because it worked on the other build using the same Cuda version. I noticed two unusual problems here: the error message is not a compiler error but an assembler error, and it is actually an internal error in ptxas itself. I found a single reference to this problem but it was on the Nvidia forums (which have been down for multiple years and force a redirection loop with Midori) :http://devtalk.nvidia.com/default/topic/395921/cuda-programming-and-performance/-ptxas-died/

Exact steps for others to reproduce the error
Press f12.

default.blend442 KBDownload

Related Objects

Event Timeline

Benjamin (Benjamin-L) created this task.Mar 31 2014, 5:39 AM
Benjamin (Benjamin-L) raised the priority of this task from to 90.
Benjamin (Benjamin-L) updated the task description.
Benjamin (Benjamin-L) added a project: Cycles.
Benjamin (Benjamin-L) edited a custom field.
Benjamin (Benjamin-L) added a subscriber: Benjamin (Benjamin-L).
Benjamin (Benjamin-L) added a comment.EditedMar 31 2014, 6:24 AM
This comment has been deleted.
Benjamin (Benjamin-L) added subscribers: Martijn Berger (juicyfruit), Thomas Dinges (dingto), Daniel Salazar (zanqdo), Brecht Van Lommel (brecht).Mar 31 2014, 6:25 AM

◀ Merged tasks: T39513.

Bastien Montagne (mont29) assigned this task to Brecht Van Lommel (brecht).Mar 31 2014, 8:22 AM
Bastien Montagne (mont29) added a subscriber: Bastien Montagne (mont29).

Next time, please do not preform "management" tasks like merge & co, this is for devs to decide! And Cuda 5.0 and 5.5 are different things, as stated by your error message, 5.5 is not officially supported, while 5.0 is.

And even though the error happens in ptxas, it does not mean it’s not a CUDA issue! Also note that our kernel changes over time, so a hash might build on 5.5, and the next one fail.

We may close this report, since 5.5 is not supported, but will let Brecht decide.

Bastien Montagne (mont29) added a project: BF Blender.Mar 31 2014, 8:52 AM
Martijn Berger (juicyfruit) renamed this task from PTX assembler internal failure to CUDA kernel compilation fails with PTX assembler internal failure.Mar 31 2014, 10:51 AM
Martijn Berger (juicyfruit) changed the task status from Unknown Status to Duplicate.
Martijn Berger (juicyfruit) lowered the priority of this task from 90 to 50.

✘ Merged into T39513.