tree 36734188fe4e9afd4bc63da0a27920bdadfff035
parent 472d778b4c0713b6df6b3a10bcfd7475ef167cc8
author Matthias Boehm <mboehm7@gmail.com> 1603106483 +0200
committer Matthias Boehm <mboehm7@gmail.com> 1603106498 +0200

[SYSTEMDS-2688] Performance diag V2M operations (shallow data copy)

For diag V2M, we already output the sparse result in a space-efficient
CSR representation. However, for the special case of input vectors v
with nrow(v)==nnz(v) -- which often happens in practice -- we now use a
shallow copy of the data array because the CSR values array is
equivalent. Similarly, we now reuse a integer single seq(0,nrow(v)) for
both the pointer and column index arrays (by setting the available
logical size of the column index array to the nnz, this shared array is
logically truncated by the last entry).

For a scenario of 1000 diag V2M operations over a (? x 1) input vectors,
we got the following improvements:
* 1000 times diag(1M x 1): 9.1s -> 3.6s
* 1000 times diag(10M x 1): 81.9s -> 22.5s