Fix ORC writer output corruption with string columns #7565
Conversation
Codecov Report
@@ Coverage Diff @@
## branch-0.19 #7565 +/- ##
===============================================
+ Coverage 81.86% 82.38% +0.51%
===============================================
Files 101 101
Lines 16884 17339 +455
===============================================
+ Hits 13822 14284 +462
+ Misses 3062 3055 -7
Continue to review full report at Codecov.
|
vuule
added
bug
vuule
requested review from
trxcllnt,
jrhemstad,
rgsl888prabhu,
nvdbaranec and
devavret
| } | ||
| s->scratch_red[t] = v; | ||
| s->scratch_red[t] = valid; |
There was a problem hiding this comment.
This whole block looks to be over optimized. It's pre-loading an offseted mask word into scratch_red only to have individual threads treating this shared memory as the new mask, and then doing an exclusive sum over the validity values.
This could be achieved using an iterator for validity, directly accessing the mask itself. and that iterator could be given an appropriate offset, which I believe would be the column offset + this dictionary chunk's start row.
There was a problem hiding this comment.
Agreed. I also had a few ideas on what can be simplified around this code, but wanted to keep such changes separate from the actual fix.
|
@gpucibot merge |
Closes: rapidsai#7346 Fixes an issue in ORC writer where null counting would not read the mask for every row. The issue occurs when the column offset is not divisible by 32 so that two words are always read to get 32bits of mask (each read is effectively offset by the columns offset, so when reading the mask for 32 rows, we need to get two words to account for the offset). Namely, the second word is not read when the row is closer than 32 to the end of the chunk. This condition is incorrect for most column offsets, as the current row is not really the first bit of the mask word. The fix is to adjust the condition when the second mask word is read (assuming that mask in padded to multiple of 32). Authors: - Vukasin Milovanovic (@vuule) Approvers: - @nvdbaranec - Mike Wilson (@hyperbolic2346) - Devavret Makkar (@devavret) URL: rapidsai#7565
Closes: #7346
Fixes an issue in ORC writer where null counting would not read the mask for every row.
The issue occurs when the column offset is not divisible by 32 so that two words are always read to get 32bits of mask (each read is effectively offset by the columns offset, so when reading the mask for 32 rows, we need to get two words to account for the offset). Namely, the second word is not read when the row is closer than 32 to the end of the chunk. This condition is incorrect for most column offsets, as the current row is not really the first bit of the mask word.
The fix is to adjust the condition when the second mask word is read (assuming that mask in padded to multiple of 32).