"unspecified launch failure" in init_random()

[GoogleCodeExporter]

What steps will reproduce the problem?
install latest version.
import cudamat as cm
cm.cublas_init()
mat=cm.CUDAMatrix.init_random(1)

What is the expected output? What do you see instead?
"cudamat.CUDAMatException: CUDA error: no error"
Traces back to an unspecified launch failure in cudamat.cu

What version of the product are you using? On what operating system?
cudamat version 0.3
python 2.6.6
OS: CentOS release 5.3 (Final)
x86_64 GNU/Linux
2.6.18-128.2.1.el5 
NVIDIA UNIX x86_64 Kernel Module  185.18.14
cublas 2.2

Please provide any additional information below.

Changing alloc and copy code in cudamat.cu to: 

   //cublasAlloc(NUM_RND_STREAMS, sizeof(unsigned int), (void**)&rnd_state->dev_mults);
    //cublasAlloc(NUM_RND_STREAMS, sizeof(unsigned int), (void**)&rnd_state->dev_words);
    //cublasSetVector(NUM_RND_STREAMS, sizeof(unsigned int), host_mults, 1, rnd_state->dev_mults, 1);
    cudaMalloc((void **)&rnd_state->dev_mults, NUM_RND_STREAMS * sizeof(unsigned int));
    cudaMalloc((void **)&rnd_state->dev_words, NUM_RND_STREAMS * sizeof(unsigned long long));
    cudaMemcpy(rnd_state->dev_mults, host_mults, NUM_RND_STREAMS * sizeof(unsigned int), cudaMemcpyHostToDevice);

fixes the problem.

Original issue reported on code.google.com by seamus.c...@gmail.com on 6 Nov 2010 at 12:10

[GoogleCodeExporter]

This issue refers to the version in the repository, not the the code in  
cudamat-01-15-2010.tar.gz

Original comment by seamus.c...@gmail.com on 6 Nov 2010 at 12:36