Next Meeting : Aug 5th
- declare target pointer cannot be accessed in target region
Ye Luo said the number of transfers increased. Alex : this is due to the addition of extra pointer member.
The test case had 3 array data transfer, resulting in 6 more memory transfers.
Alex to investigate optimizing the extra pointer transfer for local variables and not globals.
- present clause
Joel said implementation Is done.
Currently present check is done on exit for both target data and target region.
Joel has a fix to not check on exit from target data region.
Need clarity from OpenMP lang committee if on exit present is false but on entry it is true is a valid program.
Present implementation for update broke some tests due to some memory corruption and issue has been identified.
- Issue with reference being passed to is_device_ptr instead of value.
attribute((aligned(64)))float (*ptr) = omp_target_alloc(100,0);
float (*ptr) = omp_target_alloc(100,0);
#pragma omp target device(0) map(i) is_device_ptr(ptr)
i = ptr +1;;
When aligned attribute is used on variable ptr, a reference to prt is used in is_device_ptr.
Need to file this as an issue.
AMD has lots of stuff working using a variant of gdb debugger.
Would like OMPD promoted to llvm.
Would like developers of OMPD to present in a future meeting. Johannes will follow up on this.
AMD have about 15 test fails in their version (pulldown around June or early July) in sollveC.
Test failures could be due to some task dependence clause change, need further investigation.
Status of failures when switching OpenMP version to 50
Default is 5.0.
Current test run for 4.5, 5.0 and default.
Alexey said this increases the run time significantly.
Default and 5.0 is redundant.
Decision is to run 4.5 and default version.
Heterogeneous LLVM-IR Modules
Initial proposal is here https://reviews.llvm.org/D84728
Not fully flushed. So far positive feedback.
Need more feedback.
Roll Call :