[FEA] Performance Updates for row/column conversions

[hyperbolic2346]

The row/column conversion code isn't as performance as desired. This is a list of things that should be tried to attempt to improve the performance.

• Different tile sizes. Right now the tile is square in size, but this may not be the best option. Tall and wide tiles should be investigated.
• Uncoalesced memory access is called out in nsight traces. Investigation into the source of these accesses is needed.
• It was suggested in the review of Rewriting row/column conversions for Spark <-> cudf data conversions #8444 that it might make more sense for a block to not consume the entire shared memory available in an attempt to allow more blocks to run concurrently in order to hide the memory latency.
• Investigate tile ordering
• Investigate thread stride changes to improve throughput

[hyperbolic2346]

Tile sizes of 1 column, 1 row, 32 columns, and 32 rows were tried with little to no performance gain. Most saw a loss of performance.

[hyperbolic2346]

Changed shared memory sizes to 1/2, 1/4, and 1/16 of the available memory and performance was the same for column to row and degraded for row to column.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.

[hyperbolic2346]

Still relevant