Andy Grove
Andy Grove
The main source of OOM errors for regex kernels comes from a stack allocation of 11 bytes per instruction per string at kernel launch. “Small” for 1-10 instructions, “medium” for...
@revans2 @jlowe Would this feature still be required if cuDF global-memory use can be made to perform near the speed of stack memory? The memory would come from RMM instead...
Fixed in https://github.com/NVIDIA/spark-rapids/pull/6006
I created a scala test with the example and get the same results and the debugger shows that the write is happening on GPU (there is no reason for it...
Write on GPU and column has type `Timestamp(Microsecond)`: ``` +----------------------------------------------------------+ | v | +----------------------------------------------------------+ | [4456-03-21 19:12:30.001070, 2015-01-01 02:11:05.123456] | +----------------------------------------------------------+ ``` Write on CPU and column has type `Timestamp(Millisecond)`:...
The changes look good, but I think we should also add integration tests so that we test this on all supported platforms.
Sorry, I probably won't have time to look at this until the weekend, but I will review this.
Code needs cleaning up but maybe this approach can work?
> This was the approach I was _crawling_ towards. Hopefully I'm not being too obvious, but the `filtering_during_aggregation` test case fails: Ugh, this was passing at one point. I will...
I will resume work on this tomorrow