CULA R15, CUDA 5 RC: which libraries to use?

[v7kubkzs]

Hi, I'm having a difficulty that I haven't been able to find documented. I'm on a brand new MacBook Pro with Retina display (NVIDIA GeForce GT 650M), and my CULA codes are crashing out on culaInitialize(). I tried running some tests from the shipped CULA examples, with the same symptoms; abort() on culaInitialize(). The NVIDIA SDK examples all run fine. It seems that the problem is that I have to use the CUDA 5 RC libraries in place of the CULA R15 (CUDA 4.2) CUDA libraries. This makes sense, but I wanted to make sure that it was correct. And when I do this, I get "Insufficient memory" errors.

I'm using GCC 4.2.1 from XCode, the current CUDA 5 release candidate (V0.2.1221) with driver 5.0.24, CULA R15, OS X 10.8.1.

Backtrace from the debugger gives:
Program received signal SIGABRT, Aborted.
0x00007fff86addd46 in __kill ()
(gdb) bt
#0 0x00007fff86addd46 in __kill ()
#1 0x00007fff88cb4eec in __abort ()
#2 0x00007fff88cb5d43 in __stack_chk_fail ()
#3 0x00000001001dd1d2 in cudalib::GetDeviceCount ()
#4 0x00000001001f4828 in culaInitialize ()
#5 0x0000000100000e4d in main ()

It seems to be grabbing the CULA 4.2 cuda runtime and blas:
% otool -L gesvd
gesvd:
libcula_core.dylib (compatibility version 0.0.0, current version 0.0.0)
libcula_lapack.dylib (compatibility version 0.0.0, current version 0.0.0)
@rpath/libcublas.dylib (compatibility version 1.1.0, current version 4.2.0)
@rpath/libcudart.dylib (compatibility version 1.1.0, current version 4.2.0)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 169.3.0)

Recompiling with a pointer to the NVIDIA libraries:
% gcc -m64 -o gesvd gesvd.c -DNDEBUG -O3 -I/usr/local/cula/include -L/usr/local/cuda/lib -L/usr/local/cula/lib64 -lcula_core -lcula_lapack -lcublas -lcudart -pthread
yields the following:
% otool -L gesvd
gesvd:
libcula_core.dylib (compatibility version 0.0.0, current version 0.0.0)
libcula_lapack.dylib (compatibility version 0.0.0, current version 0.0.0)
@rpath/libcublas.dylib (compatibility version 1.1.0, current version 5.0.0)
@rpath/libcudart.dylib (compatibility version 1.1.0, current version 5.0.0)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 169.3.0)
% ./gesvd
--------------------------------------------------------------------------------
This example demonstrates using CULA to implement an image compression
algorithm. Two images will be generated:

Full-fidelity: image_original.bmp
Reduced: image_reduced.bmp

--------------------------------------------------------------------------------
Generating image ... done.
Performing singular value decomposition using CULA ... Insufficient memory to complete this operation

So, no crash, but I don't know if the failure is indicative of another problem. The same happens with the CULA "systemSolve" example:
% otool -L systemSolve
systemSolve:
libcula_core.dylib (compatibility version 0.0.0, current version 0.0.0)
libcula_lapack.dylib (compatibility version 0.0.0, current version 0.0.0)
@rpath/libcublas.dylib (compatibility version 1.1.0, current version 5.0.0)
@rpath/libcudart.dylib (compatibility version 1.1.0, current version 5.0.0)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 169.3.0)
% ./systemSolve
-------------------
SGESV
-------------------
Allocating Matrices
Initializing CULA
Calling culaSgesv
Insufficient memory to complete this operation


Any help is appreciated.

Thanks,
Chris






Proof of system details:
% which gcc
/usr/bin/gcc
% gcc --version
i686-apple-darwin11-llvm-gcc-4.2 (GCC) 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.11.00)
% which nvcc
/usr/local/cuda/bin/nvcc
% nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2012 NVIDIA Corporation
Built on Fri_Aug__3_17:12:53_PDT_2012
Cuda compilation tools, release 5.0, V0.2.1221
% env | grep CULA
CULA_ROOT=/usr/local/cula
CULA_INC_PATH=/usr/local/cula/include
CULA_LIB_PATH_32=/usr/local/cula/lib
CULA_LIB_PATH_64=/usr/local/cula/lib64
% pwd
/usr/local/cula/examples/imageCompressionSVD
% make build64
sh ../checkenvironment.sh
gcc -m64 -o gesvd gesvd.c -DNDEBUG -O3 -I/usr/local/cula/include -L/usr/local/cula/lib64 -lcula_core -lcula_lapack -lcublas -lcudart -pthread
% ./gesvd
--------------------------------------------------------------------------------
This example demonstrates using CULA to implement an image compression
algorithm. Two images will be generated:

Full-fidelity: image_original.bmp
Reduced: image_reduced.bmp

--------------------------------------------------------------------------------
Generating image ... done.
Abort


Device info:
NVIDIA GeForce GT 650M:
Chipset Model: NVIDIA GeForce GT 650M
Type: GPU
Bus: PCIe
PCIe Lane Width: x8
VRAM (Total): 1024 MB
Vendor: NVIDIA (0x10de)
Device ID: 0x0fd5
Revision ID: 0x00a2
ROM Revision: 3688
gMux Version: 3.2.19 [3.2.8]
CUDA Driver Version: 5.0.24
GPU Driver Version: 8.0.51 295.30.00f01

[john]

You definitely want to be sticking to the CUDA 4.2 runtime for this, and preferably the one that came with CULA. The driver can be 4.2 or 5.0, in theory, but the 4.2 driver would be the one closer to our test configuration.

This could also be a Mountain Lion compatibility issue. Mountain Lion is not on our supported OS list at this time.

[v7kubkzs]

Thanks. I'll give 4.2 a shot.

I'm also having this problem with a K20 on Chester (Titan testbed) at ORNL.