Oh, no! Missed noticing my first link from HN for a few days. Hopefully I can clarify some of this.
First, this is in no way faster than cuSPARSE or cusp. I wrote this for an originally for a school assignment (hence older cuda version) and was hoping to convey what I had learned my first time using cuda.
The size of the shared memory is so the reduction has no buffer overflow without using conditionals. However I am using more than needed, it should be set to 32+16. I don't expect this to affect performance as the kernel already reaches 100% theoretical occupancy.
Could you explain the desire for more threads/blocks per row? I can't immediately think of one, having more additions per thread sounds good so long as all are running.
Thanks very much for the read and the reply!
1) The implementation may not be the most effecient for a larger matrix or an even more dense matrix. 1% of 52000 is 520. That divided by 32 is 8-9 additions per thread. As that number increases, increasing the number of threads (and eventually using more blocks per row) would be a good idea.
2) He is allocating twice as much shared memory than required. I genuinely hope that was an artefact from before. If not, that is a killer for performance. Using more shared memory per block reduces the number of concurrent blocks.
Note: Not sure why he is still using cuda 3.2. cuda has had csr multiplication for a few months now, and has even gone through a revision to make it even faster.