Create cuNDArray

[Christoph Kolbitsch]

Hello,

I tried to create a cuNDArray from an hoNDArray which I loaded from file:

boost::shared_ptr< hoNDArray<float> > dcf_ho = read_nd_array<float>("/data/ck09/data.dcf");
boost::shared_ptr< cuNDArray<float> > dcf_cu(new cuNDArray<float>(dcf_ho.get()));

When I try to run this code I get a segmentation fault for the second line

Does somebody know what I am doing wrong here?

[Thomas Sangild Sørensen]

What is this .dcf format? 'read_nd_array' only supports the simple header+data format described in the manual/wiki. What are the dimensions of the host array after creation?

/ Thomas

[Christoph Kolbitsch]

I created the file "data.dcf" with the write_nd_array function. Nevertheless, I have reduced the problem to something a bit more reproducible:

std::vector< size_t > dcf_dim(2); 
dcf_dim[0] = 500; dcf_dim[1] = 10;
hoNDArray<float>* dcf_ho;
dcf_ho = new hoNDArray<float>();
dcf_ho->create(&dcf_dim);
GADGET_DEBUG2("hoNDArray [%d %d]\n", dcf_ho->get_size(0), dcf_ho->get_size(1));

boost::shared_ptr< cuNDArray<float> > dcf_cu(new cuNDArray<float>(dcf_ho));

Here I create a hoNDArray with the dimensions 500 x 10 and also get_size(0) returns 500 and get_size(1) returns 10 as expected. Nevertheless when I execute the next line to create the cuNDArray I get:

[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 105] Gadget (RpeAccumulateGadget) Close Called with flags = 0
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 117] Gadget (RpeAccumulateGadget) thread finished
Segmentation fault (core dumped)

[David Hansen]

This is a bit of a strange error. Most of the things that could go wrong when creating a cuNDArray would result in a cuda_error being thrown, rather than a segmentation fault.

Do you also get the error if you create the cuNDArray<float> directly, and not in a shared_ptr or via new?</float>

[Christoph Kolbitsch]

I definitely agree that this is strange.

I tried the following:

1.Run the following code inside Gadget B (gpuRpeGridGadget):

GADGET_DEBUG1("Start creating cuNDArray...\n");
std::vector< size_t > dcf_dim(2); 
dcf_dim[0] = 500; dcf_dim[1] = 10;
cuNDArray<float>* dcf_cu;
dcf_cu = new cuNDArray<float>();
dcf_cu->create(&dcf_dim);
GADGET_DEBUG2("cuNDArray [%d %d]\n", dcf_cu->get_size(0), dcf_cu->get_size(1));

This leads to the following error message similar to the one posted above:

[file /data/ck09/projects_ck09/GadgetronRecon/gpuRpeGridGadget.cpp, line 265] Start creating cuNDArray...
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 105] Gadget (RpeAccumulateGadget) Close Called with flags = 0
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 117] Gadget (RpeAccumulateGadget) thread finished
Segmentation fault (core dumped)

2.Run the above code inside Gadget B (gpuRpeGridGadget) and in another Gadget A (RpeAccumulateGadget) which is placed BEFORE Gadget B:

This leads to the following messages:
[file /data/ck09/projects_ck09/GadgetronRecon/RpeAccumulateGadget.cpp, line 342] Start creating cuNDArray...
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/GadgetStreamController.cpp, line 80] Received close signal from client. Closing stream...
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 105] Gadget (RpeAccumulateGadget) Close Called with flags = 1
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 115] Gadget (RpeAccumulateGadget) waiting for thread to finish
[file /data/ck09/projects_ck09/GadgetronRecon/RpeAccumulateGadget.cpp, line 348] cuNDArray [500 10]

and then a bit later

[file /data/ck09/projects_ck09/GadgetronRecon/gpuRpeGridGadget.cpp, line 265] Start creating cuNDArray...
[file /data/ck09/projects_ck09/GadgetronRecon/gpuRpeGridGadget.cpp, line 271] cuNDArray [500 10]

This means to me that there is a problem when creating the cuNDArray in Gadget A but it manages to do it and this then means that the cuNDArray can be created successfully in the following Gadget B.

Does this make sense to anyone? I am grateful for any ideas or suggestions where I should be looking for the mistake.

[Michael Hansen]

Could you run this in a debugger and figure out which line causes the segfault?

---

Sent from a mobile device - please excuse brevity and/or typos.

On Jun 18, 2014, at 8:11 AM, "Christoph Kolbitsch" ck09@users.sf.net wrote:

I definitely agree that this is strange.

I tried the following:

1.Run the following code inside Gadget B (gpuRpeGridGadget):
GADGET_DEBUG1("Start creating cuNDArray...\n");
std::vector< size_t > dcf_dim(2);
dcf_dim[0] = 500; dcf_dim[1] = 10;
cuNDArray* dcf_cu;
dcf_cu = new cuNDArray();
dcf_cu->create(&dcf_dim);
GADGET_DEBUG2("cuNDArray [%d %d]\n", dcf_cu->get_size(0), dcf_cu->get_size(1));

This leads to the following error message similar to the one posted above:

[file /data/ck09/projects_ck09/GadgetronRecon/gpuRpeGridGadget.cpp, line 265] Start creating cuNDArray...
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 105] Gadget (RpeAccumulateGadget) Close Called with flags = 0
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 117] Gadget (RpeAccumulateGadget) thread finished
Segmentation fault (core dumped)

2.Run the above code inside Gadget B (gpuRpeGridGadget) and in another Gadget A (RpeAccumulateGadget) which is placed BEFORE Gadget B:

This leads to the following messages:
[file /data/ck09/projects_ck09/GadgetronRecon/RpeAccumulateGadget.cpp, line 342] Start creating cuNDArray...
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/GadgetStreamController.cpp, line 80] Received close signal from client. Closing stream...
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 105] Gadget (RpeAccumulateGadget) Close Called with flags = 1
[file /data/ck09/GADGETRON/gadgetron/apps/gadgetron/Gadget.h, line 115] Gadget (RpeAccumulateGadget) waiting for thread to finish
[file /data/ck09/projects_ck09/GadgetronRecon/RpeAccumulateGadget.cpp, line 348] cuNDArray [500 10]

and then a bit later

[file /data/ck09/projects_ck09/GadgetronRecon/gpuRpeGridGadget.cpp, line 265] Start creating cuNDArray...
[file /data/ck09/projects_ck09/GadgetronRecon/gpuRpeGridGadget.cpp, line 271] cuNDArray [500 10]

This means to me that there is a problem when creating the cuNDArray in Gadget A but it manages to do it and this then means that the cuNDArray can be created successfully in the following Gadget B.

Does this make sense to anyone? I am grateful for any ideas or suggestions where I should be looking for the mistake.

Create cuNDArray

Sent from sourceforge.net because you indicated interest in https://sourceforge.net/p/gadgetron/discussion/general/

To unsubscribe from further messages, please visit https://sourceforge.net/auth/subscriptions/

[Thomas Sangild Sørensen]

Also, can you confirm that the Cuda samples (from the Cuda SDK) run? - e.g. deviceQuery. If you have Cuda installed with the wrong driver you might be able to compile Gadgetron but encounter errors at runtime.

/ Thomas

[Christoph Kolbitsch]

I finally found time to look into this problem again.

When I looked at the segmentation fault with a debugger I found out that the error occurs when the destructor of an boost:shared_ptr to an hoNDArray is called.

I am not really sure how this is linked to my code fragment of creating a cuNDArray but it seems this hoNDArray is linked via a boost:shared_ptr to a global variable in one of my gadgets. In the destructor of this gadget I delete this global variable. Nevertheless, this does seem to cause problems with the boost:shared_ptr destructor.

I changed this variable from being global to a local variable and now it seems to be working fine.

Thank you for your suggestions.

Regards,
Christoph