Using cuSparse library for Tridiagonal solver on a CUDA compatible GPU with compute capability 1.1, has decreased the performance drastically, up to 50 times slower when compared with a traditional serial solver on a Core 2 Duo CPU.
Is it because of the low compute compatibility or is my implementation wrong?
Think you don't choose proper format of sparse matrix(CSR,COO,CSC).
http://docs.nvidia.com/cuda/pdf/CUDA_CUSPARSE_Users_Guide.pdf
First You can try CUSP or cula sparse(it's free for 32-bit float) and measure performance not to take care of that:
http://cusplibrary.github.io/
Cusp has diagnal matrix storage format:
https://github.com/cusplibrary/cusplibrary/blob/master/examples/MatrixFormats/dia.cu
Boston University lections:
http://www.bu.edu/pasi/files/2011/01/NathanBell1-10-1000.pdf
http://www.bu.edu/pasi/files/2011/01/NathanBell2-11-1130.pdf
http://www.bu.edu/pasi/files/2011/01/NathanBell3-12-1000.pdf
http://www.bu.edu/pasi/files/2011/07/Lecture6.pdf
http://www.bu.edu/pasi/files/2011/07/Lecture7.pdf
Maybe it worth to use it with Thrust - i dont know, both have high level C++ interface.
Thank you very much for the links..
I am getting different values when variables are stored in host and device memory....Is there any significance?. and some values are "nan" when variables are in device memory .... what might be wrong?
Good morning Suhas,
As far as you programming CUDA you must do cudaMemcpy(devA,A,numberBytes, cudaMemcpyDeviceToHost) before you call kernel function and cudaMemcpy(A,devA,numberBytes, cudaMemcpyHostToDevice) after. If you not want to store it and work with them on CPU, you might use MKL(blas already implemented, don't know about sparse) or C/C++(Fortran) with OpenMP(MPI if you have distibuted network).You should also learn about shared memory , if you care about performance of your program. You are not to worry about this, if you use cusp/thrust, because they use C++ vector type, similar to boost/std.
Best regards, Ivan.
Sorry, I didn't read your code.
Of course you have three elements when i=0(-1 offset) and i=2(1 offset), you must store 0,1,1,1 and 1,1,1,0 instead of 1,1,1,1. 1,1,1,1 must be only when i=0(0 offset). And try to cusp :: print.Or want you formatted output?
I don't know about interoperability cusp and thrust, whether it works or not thought.
Think if you try to use thrust :: copy with diagonal format stored matrix it maybe will not work.
Hi Ivan,
It doesnt matter if it is 1,1,1,1 or 0,1,1,1, because that value is outside the matrix and its not considered, hence the result remains same. I just gave 1,1,1,1 so that its easier to write inside the "for" loop.
As you mentioned earlier, that the values must be copied to device memory using this command
"cudaMemcpy(devA,A,numberBytes, cudaMemcpyHostToDevice) "
But I think devA and A here are the address of the arrays and not the array itself. So how to initialize devA and A here?
Thanks..
Hi Suhas,
Don't know, it works well for me:
sparse matrix with 10 entries
0 0 1
0 1 1
1 0 1
1 1 1
1 2 1
2 1 1
2 2 1
2 3 1
3 2 1
3 3 1
array1d
1
1
1
1
u[0]=1.000000
u[1]=-0.000000
u[2]=-0.000000
u[3]=1.000000
u[4]=0.000000
Check the installation.
I have CUDA 4.1(CUSP 0.3 not working with lesser) and latest the github repository clone.
nvcc dia.cu -I/opt/cusplibrary-master
command compiles it.
Hi Ivan,
In that file "dia.cu" variables are stored in host memory. This works for me too. But when you change the variables to device memory, it doesn't work.
Thank you very much.. cuda-memcheck helped me a lot and of course the tridiagonal matrix solution using CUSP is running much faster than using CUSPARSE.
- Badges
- Science topic