Fix potential CNMEM_NOT_INITIALIZED errors #26
Conversation
| @@ -229,6 +229,9 @@ class MemoryHandlerActivator { | |||
| using_pool_ = true; | |||
| MemoryHandler::usePool(); | |||
| MemoryHandler::setGPUs(gpus); | |||
| void* temp; | |||
| MemoryHandler::mallocGPU(&temp, 4); | |||
| MemoryHandler::freeGPU(temp); | |||
There was a problem hiding this comment.
This probably warrants an explanatory comment.
There was a problem hiding this comment.
This needs to be wrapped in a #ifndef CPU_ONLY (or #ifdef USE_CNMEM) for the CPU-only build.
|
This fixed my HDF5Data layer runtime error - thanks! But now the tests are failing hard (https://travis-ci.org/NVIDIA/caffe/builds/77675932), and they even fail for me when I build with cuDNN and CNMeM (with 6,000+ lines of stderr output!). Can you verify this? It's not clear to me why such a small change would affect the tests so dramatically. |
|
It looks like the TravisCI builds are failing to build Caffe, whereas on my machine I'm failing to build the tests. |
|
The tests are completely unchanged -- they don't use the memory pool at all. Only time I've seen something similar it ended up being a bug in gcc -- moving to something newer (in my case from gcc 4.6 -> 4.9) fixed the issue for me |
Well that seems like a giant test hole then! Looking into the gcc version now ... |
|
I have gcc version 4.8.4 - I don't think it's a gcc problem. But it does seem to be a CMake problem (make works fine), and I'm seeing it with For this pull reqeuest, can you add the |
dummy alloc / free in MemoryHandlerActivator to ensure that the memory pool has been set up before any potential operations
|
Added the #ifndef guard and squashed commits |
|
Ok great, that fixed the Travis tests. Thanks! |
Fix potential CNMEM_NOT_INITIALIZED errors
|
Tagged as |
|
Ok, sounds good. |
dummy alloc / free in MemoryHandlerActivator to ensure that
the memory pool has been set up before any potential operations