| Re: Any interest in READ/WRITE functions for sparse (CSR) matrices? [message #82257] |
Mon, 26 November 2012 05:49 |
tom.grydeland
Messages: 51
Registered: September 2012
|
Member |
|
|
On Monday, November 26, 2012 10:14:23 AM UTC+1, tom.gr...@gmail.com wrote:
> I tried searching for a bit, but couldn't find an established (binary) format for sparse matrices. Hence, I've done what one should never do -- dreamt up my own file format for saving such data.
I should perhaps point out that IDL's SPRSIN and friends would not work for us for a number of reasons, since we use the sparse format as a data container on a subset of a regular grid, and with no particular eye towards doing linear algebra on the resulting matrices (so I should probably not call them matrices at all, but only arrays). Therefore, being restricted to square matrices makes little sense. Also, the mask often applies to a number of facets or properties, which is often useful to save together in one file. It is also useful to us to have a pure index file, which can be used for saving other data (e.g. after subsequent processing) in a congruent sparse format.
A quick comparison wrt file size: a 500x500 single-precision float array randomly filled at about 10% resulted in a CSR file of ~150k and an IDL AS file of about 300k. The values themselves take up ~100k, so the overhead is about 50% for CSR, about 200% for AS in this simple case. The ratio of 1/4 for the overhead is as expected. IDL AS files save a single ULONG64 index (8 bytes) per element, while CSR saves one UINT x index (2 bytes) per element, and also one UINT x count per line. (If there are more than 2^16-1 elements per line, the indices will automatically be ULONG -- 4 bytes -- instead).
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
