Fix for some compiler warnings in parquet/page_decode.cuh #15518
Conversation
davidwendt
added
bug
|
/ok to test |
|
/ok to test |
| @@ -924,7 +924,7 @@ inline __device__ uint32_t InitLevelSection(page_state_s* s, | |||
|
|
|||
| auto start = cur; | |||
|
|
|||
| auto init_rle = [s, lvl, end, level_bits](uint8_t const* cur, uint8_t const* end) { | |||
| auto init_rle = [s, lvl, level_bits](uint8_t const* cur, uint8_t const* end) { | |||
There was a problem hiding this comment.
The behavior when a captured variable and a parameter have the same name is interesting. Thanks to godbolt, I found that all options are possible, with different compilers - captured variable value is used, or parameter value is used, or compilation fails. Fun!
There was a problem hiding this comment.
Wow that's interesting. We were in the same situation before, and the code runs differently on different architectures (due to compiled into different archs?).
There was a problem hiding this comment.
Thank you clangd! 😄
There was a problem hiding this comment.
I think nvcc stayed consistent in prioritizing the parameter over the capture. But I found online that other compilers disagree.
There was a problem hiding this comment.
We found that the capture overshadows the previous variable on H100+V100 while on other GPUs it doesn't.
|
/merge |
Description
Clangd generates several warnings/errors in cpp/src/io/parquet/page_decode.cuh. One is in regards to a lambda argument shadowing a captured value. The others involve the use of
thrust::optional::value()in device code...unlike pretty much every other member function,value()lacks the__device__decorator. This PR replaces two usages ofvalue()withoperator*()which does have the__device__decorator.Checklist