Closed stamatak closed 8 years ago
Hi Alexi, thanks for the comments. The implementation in its current status should work for arbitrary data types. Vectorizing it and keeping its functionality will definitely require some modifications. My very first thought (for vectorization) was to add padding to each CLV entry to get the number of entries to its the next multiple of 2 for SSE, or 4 for AVX -- point 3 in your suggestions.
My impression was that this could be straightforward to implement, but I haven't thought of it thoroughly.
arbitrary-state data types now work, with vectorization.
Just a few comments on the vectorization if you want to make it generic: