Segfault with command line matcher tools #7

Activity

1. ahojnnes commented on May 1, 2016

   

   ahojnnes

   on May 1, 2016

   Contributor

   More actions

   Hi Andrew,

   Thanks for reporting!

   Unfortunately, I upgraded to the latest Macbook Pro series some time ago which ships without an Nvidia card, so I cannot check CUDA support on OSX anymore. I will try to find a suitable machine over the next days to fix this and will follow up with you here.

   Thanks again.

   Cheers,
   Johannes
2. ahojnnes commented on May 21, 2016

   

   ahojnnes

   on May 21, 2016

   Contributor

   More actions

   I could not reproduce this issue here. Can you share your specific Hardware?
3. QuantumEntangledAndy commented on May 21, 2016

   

   QuantumEntangledAndy

   on May 21, 2016

   Last edited by QuantumEntangledAndy

   Author

   More actions

   Hmm that's interesting, an un-reproducable one.
   Anyways here are my details not sure which part you would be interested in

   • OS

   osX El-Capitain  10.11.4
   

   • Model

     Model Name:   MacBook Pro
     Model Identifier: MacBookPro9,1
   

   • CPU

     Processor Name:   Intel Core i7
     Processor Speed:  2.3 GHz
     Number of Processors: 1
     Total Number of Cores:    4
     L2 Cache (per Core):  256 KB
     L3 Cache: 6 MB
   

   • RAM

     Memory:   16 GB
     Size: 2  x 8 GB
     Type: DDR3
     Speed:    1600 MHz
   

   • GPU

     Chipset Model:    NVIDIA GeForce GT 650M
     Type: GPU
     Bus:  PCIe
     PCIe Lane Width:  x8
     VRAM (Total): 512 MB
     Vendor:   NVIDIA (0x10de)
   

   I will recompile with the latest code and see if its been squashed since then by another commit.
   Hope all goes well, let me know if you'd like anything else from me.
4. QuantumEntangledAndy commented on May 21, 2016

   

   QuantumEntangledAndy

   on May 21, 2016

   Author

   More actions

   Ok here the the retest results:

   Summary

   • Doesn't work with CUDA enabled resulting in a segfault
   • Does work with CUDA disabled

   Details

   I will layout all of my steps so this may be a long one

   With CUDA enabled

   git clone https://github.com/colmap/colmap/ 01_Test
   cd  01_Test
   

   set up LocalConfig.cmake to include

   set(Qt5_CMAKE_DIR "/usr/local/opt/qt5/lib/cmake")
   set(Qt5Core_DIR ${Qt5_CMAKE_DIR}/Qt5Core)
   set(Qt5OpenGL_DIR ${Qt5_CMAKE_DIR}/Qt5OpenGL)
   

   mkdir build && cd build
   cmake ..
   make
   

   Build succeeds and cmake detects CUDA, cu files are built with nvcc

   Now to test:
   images come from http://zhuoliang.me/gargoyle.zip which is from this page http://zhuoliang.me/meshrecon.html

   mkdir test
   src/exe/database_creator --General.database_path "test/colmap.db"
   src/exe/feature_extractor --General.database_path "test/colmap.db" --General.image_path "gargoyle"
   src/exe/sequential_matcher" --General.database_path "test/colmap.db"
   

   The following segfault occurs when running sequential_matcher

   *** Aborted at 1463834627 (unix time) try "date -d @1463834627" if you are using GNU date ***
   PC: @     0x7fff95051c30 glGetString
   *** SIGSEGV (@0x3b0) received by PID 16919 (TID 0x700000081000) stack trace: ***
       @     0x7fff8b7ba52a _sigtramp
       @        0x1082e5010 glewIsSupported
       @        0x1082b79b1 glewInit
       @        0x1076295fd GlobalUtil::InitGLParam()
       @        0x107645af9 SiftMatchGL::InitSiftMatch()
       @        0x1076477e2 SiftMatchGPU::_VerifyContextGL()
       @        0x10734a73f colmap::FeatureMatcher::SetupWorkers()
       @        0x10734a0d6 colmap::FeatureMatcher::run()
       @        0x10916757a QThreadPrivate::start()
       @     0x7fff8af8a99d _pthread_body
       @     0x7fff8af8a91a _pthread_start
       @     0x7fff8af88351 thread_start
   

   With CUDA disabled

   Same set up as above except now cloned to a fresh directory:

   git clone https://github.com/colmap/colmap/ 02_Test_NOCUDA
   

   and the cmake command is

   cmake -DCUDA_ENABLED=0 ..
   

   Testing:
   Same images etc

   mkdir test
   src/exe/database_creator --General.database_path "test/colmap.db"
   src/exe/feature_extractor --General.database_path "test/colmap.db" --General.image_path "gargoyle"
   src/exe/sequential_matcher" --General.database_path "test/colmap.db"
   

   Succeeds printing out

   ==============================================================================
   Sequential feature matching
   ==============================================================================
   
   Matching image [1/23] in 0.000s
   Matching image [2/23] in 0.789s
   Matching image [3/23] in 0.599s
   Matching image [4/23] in 0.692s
   Matching image [5/23] in 0.875s
   Matching image [6/23] in 1.447s
   Matching image [7/23] in 1.444s
   

   Debuging

   To debug further I put the code (with CUDA enabled) into Xcode using

   cd  01_Test
   mkdir xcode && cd xcode
   cmake -G Xcode ..
   

   Opening in xcode and running with the debugger attached catches the segfault at:

   01_Test/src/ext/SiftGPU/GlobalUtil.cpp: Line 384
   which reads:

   //first time in this function
   glewInit();
   

   throwing a EXC_BAD_ACCESS
   I believe this is occurring because there is no active GLContex at this point and the Mac CUDA version must be relying on it. Which is hinted at in the siftGPU manual, albeit for all GPU implementations not just mac, http://www.cs.unc.edu/~ccwu/siftgpu/manual.pdf on page 6.

   When you tested this on a mac with CUDA, was CUDA successfully detected by cmake i.e. is CUDA in the PATH and DYLD_LIBRARY_PATH, like this:

   export PATH=/Developer/NVIDIA/CUDA-7.5/bin:$PATH
   export DYLD_LIBRARY_PATH=/Developer/NVIDIA/CUDA-7.5/lib:$DYLD_LIBRARY_PATH
   

   Sorry I just find it surprising that you didn't observe this, and thought I should ask just in case.

   A fix that works for me

   Again if I explicitly initialise the GLContext then the program works fine, which I tested by making this alteration to call SiftGPU::CreateContextGL. in FeatureMatching.cc line on 155

   sift_gpu_->CreateContextGL();
   sift_match_gpu_->CreateContextGL();
   

   If you'd like I could create a pull request with the changes I made that got the program working for me. I just don't have another build machine to test the changes on and see if it breaks other OS's.

   Anyway hope all this helps, sorry if it's a bit too much info
   Best Regards
   Andy
5. ahojnnes commented on May 23, 2016

   

   ahojnnes

   on May 23, 2016

   Contributor

   More actions

   Thanks for the detailed traceback!

   Can you try to change L52 in src/ext/SiftGPU/CMakeLists.txt from:

   if(CUDA_ENABLED AND NOT APPLE)
   

   to

   if(CUDA_ENABLED)
   

   and check if that solves the problem?
6. QuantumEntangledAndy commented on May 25, 2016

   

   QuantumEntangledAndy

   on May 25, 2016

   Author

   More actions

   I pulled your latest version, with the changed

   if(CUDA_ENABLED)
   

   After which the program successfully compiled and ran without any further changes.
   I will let you know if I find anything else but you can probably close :).
7. ahojnnes commented on May 25, 2016

   

   ahojnnes

   on May 25, 2016

   Contributor

   More actions

   Thanks Andrew for your efforts!
8. ahojnnes

   closed this as completedon May 25, 2016
9. Zador-Pataki

   added 2 commits that reference this issue on Apr 27, 2026

   review fixes: docstrings, fail-loud emplace, dead-code notes

   ...

   fd77b95

   Retire glomap_ra GP+RA; route through colmap4-native (#5)

   ...

   d48d727