Timeline for Speed up sparse Hermitian matrix-vector product
Current License: CC BY-SA 4.0
13 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 13 at 20:48 | history | edited | user64494 | CC BY-SA 4.0 |
edited title
|
| Mar 13 at 20:17 | answer | added | Henrik Schumacher | timeline score: 2 | |
| Mar 13 at 0:17 | comment | added | Henrik Schumacher | Yeah, the problem with GPU computing like this is 1.) data has to be transported back an forth from main memory to GPU and 2.) sparse matrix arithmetic is not really well suited for this kind of processors... | |
| Mar 12 at 21:45 | comment | added | flinty |
There is a matrixMul.cu example for CUDALink that works on real dense matrices. However, I tried this and found it slower than just a CPU A.B in Mathematica.
|
|
| Mar 12 at 21:12 | comment | added | Felipe | My experience with these special array types is that they are not faster than PackedArrays, and I believe that you can use packed arrays in this case. Also, take a look at the code here, maybe you can copy and paste it to your problem, with some change for matrix-vector multiplication. | |
| Mar 12 at 18:52 | history | edited | march |
edited tags
|
|
| Mar 12 at 17:17 | history | edited | creidhne | CC BY-SA 4.0 |
fix typo in title
|
| Mar 12 at 17:12 | answer | added | Henrik Schumacher | timeline score: 6 | |
| Mar 12 at 14:05 | comment | added | Henrik Schumacher | Do you happen to have to multiply many vectors at once? | |
| Mar 12 at 12:44 | history | edited | Matteo | CC BY-SA 4.0 |
added 721 characters in body
|
| Mar 12 at 12:36 | comment | added | Roman | "...can I create a compiled function with the list of indexes and values of 𝐻 and 𝑣 as inputs and do better?" – This is precisely what a sparse-sparse multiplication does and I doubt you can do better. However, if you can calculate/enumerate the sparse matrix elements faster than you can retrieve them from a pre-computed list, then you could write a function $\vec{v}\mapsto H\cdot \vec{v}$ that is faster than a sparse-sparse matrix multiplication. This becomes particularly relevant for extremely large vectors, or if you use ARPACK directly. | |
| Mar 12 at 12:21 | comment | added | MarcoB | Please add a sample matrix and vector so people have something to play with as they experiment. | |
| Mar 12 at 11:59 | history | asked | Matteo | CC BY-SA 4.0 |