BACtaki commented on PR #1650:
URL: https://github.com/apache/systemds/pull/1650#issuecomment-1201640792
Thanks for your review @phaniarnab! Comments inline:
> Are the optimizations only improve the naive cases where the input
dimension is less than 1024 for the given direction (row/col)?
The cache-conscious iteration optimization only applies to the naive
counting case where `nnz < minimumSize
= 1024`. Aside from this, I also refactored `MatrixSketch` and `KMVSketch`
to replace the `getScalarValue() -> Integer` method with the `getValue() ->
MatrixBlock` method to unify the handling of RowCol, Row, and Col cases (as per
@Baunsgaard 's suggestion from an earlier PR).
> I see that you are now iterating the dense and sparse inputs in a more
cache-conscious manner (reducing CPU cache misses). Are there any other
optimizations you are employing (e.g. reducing the number of intermediates)?
Really good point re intermediates. We are actually creating an intermediate
`SmallestPriorityQueue` of max size `k = 64` per row/col. The way we are
constructing the intermediates places an upper bound on additional memory:
1. RowCol: max 64 double values (64 bits each)
2. Row: `SmallestPriorityQueue spq` is reset between each iteration, so the
intermediate is at most 64 double values
3. Col: identical to Row, but the iteration occurs in column-major order and
`spq` is reset between iterations of consecutive columns
There are no further optimizations beyond cache-conscious iteration in CP
case.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: dev-unsubscr...@systemds.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
