JSON tree algorithm code reorg #16836
Conversation
karthikeyann
added
3 - Ready for Review
| bitmask_type* validity; | ||
| }; | ||
|
|
||
| std::pair<cudf::detail::host_vector<uint8_t>, |
There was a problem hiding this comment.
Can we change ignore_vals returned here to a bool vector? We can also move this change to #16545
EDIT: Similar suggestion for is_pruned and is_str_column_all_nulls
There was a problem hiding this comment.
IIRC, I've seen concerns raised in the past regarding the use of bool vectors, and recommendations to use int8_t/uint8_t instead. I didn't think host_vector<bool> would be plagued with the same problems as std::vector<bool>, but I could be wrong.
Watching this conversation for @karthikeyann's opinion.
There was a problem hiding this comment.
@mythrocks That's right. 💯
std::vector<bool> stores each element as a bit (to save memory). so, iteration and writing to data pointer causes issues.
cudf::detail::host_vector is based on thrust::host_vector which are simple vectors. (without any specialization for bool). So, that should be suitable.
There was a problem hiding this comment.
Yes, thrust::host_vector<bool> will use byte-packing (not bits).
There was a problem hiding this comment.
Another reason why I used uint8_t is because using 0 and 1 is easier to read. (😄)
- Updating
ignore_valsto host_vector is fine since it needs to be copied to device, so pinned memory will help. I will update it in next PR [WIP] JSON host tree algorithms #16545 is_prunedandis_str_column_all_nullsare accessed only in host, and not copied to device. So, no benefit in using pinned memory, also allocation is slower. (not sure if CPU read time or write time is slower still).
The only real drawbacks to pinning memory are the reduction in available physical ram to the host demand-paging system, and the time it takes to pin memory (which is significantly longer than an ordinary malloc of the same size).
https://forums.developer.nvidia.com/t/advantages-disadvantages-of-using-pinned-memory/34422/2
https://forums.developer.nvidia.com/t/pinned-memory-slower-than-pageable-memory/18821
Note: These posts are old.
Description
This PR moves JSON host tree algorithms to separate file.
This code movement will help #16545 review easier.
The code is moved to new file and reorganized for code reuse.
Very long function
make_device_json_columnis split intoreduce_to_column_treecallbuild_treescatter_offsetsNo new functionality is added in this PR.
Checklist