Opened 15 years ago
Closed 15 years ago
#230 closed defect (fixed)
LHAPDF bug fixed, but extensive testing required
| Reported by: | Juergen Reuter | Owned by: | Christian Speckner |
|---|---|---|---|
| Priority: | P1 | Milestone: | v1.95 |
| Component: | core | Version: | 1.94 |
| Severity: | blocker | Keywords: | |
| Cc: |
Description
Something has gone nuts.
Change History (8)
comment:1 Changed 15 years ago by
| Milestone: | → v1.95 |
|---|---|
| Version: | 1.93 → 1.94 |
comment:2 Changed 15 years ago by
| Owner: | changed from jr_reuter, cnspeckn to Christian Speckner |
|---|---|
| Status: | new → assigned |
comment:3 Changed 15 years ago by
Should add that the commit was r1904; also, spin 3/2 is already affected, so this might not be ideal.
comment:4 Changed 15 years ago by
We have to decide quickly what to do,and put 1.95 on the web. I guess, wrt. W2 we can afford losing spin 3/2 and spin 2 in W1??? Can't we?
comment:5 Changed 15 years ago by
I guess so, since spin >= 3/2 must have been broken from the beginning. However, I would feel more at rest if someone reviewed the patch; in addition, some more tests should be run in order to make sure this doesn't introduce any more regressions (I can do so tomorrow, but not today anymore).
comment:6 Changed 15 years ago by
| Summary: | LHAPDF does only give zero cross sections for WHIZARD 1.94 → LHAPDF bug fixed, but extensive testing required |
|---|
It is as I though; there are many places where the spin 1 limit was hardwired, and I am indeed surprised that the implementation of spin 3/2 / 2 did not try to bite the user earlier. I have reverted r1903, skimmed the code for all places which had spin <= 1 hardcoded and tried to fix the spin <= 2 implementation. The result can be found in r1907, and LHAPDF still seems to be working.
However, this is a bit "choose your poison": the changes are numerous, and I am far from confident that no new regressions have been introduced. So, if we want to keep this fix, we have to do some extensive testing, otherwise, we should simply drop gravitons and gravitinos. In particular, we have to make sure CIRCE is not affected. I am going to perform some tests regarding without structure functions and with LHAPDF, but I would be glad if someone else could give CIRCE a try.
comment:7 Changed 15 years ago by
OK, I've got agreement between 1.93 and svn 1.xx at the 3\sigma level for numerous 2->2 processes both with and without LHAPDF structure functions, and no new memory corruption. If someone can confirm that CIRCE is working consistently too, this bug can be closed and a new release prepared.
comment:8 Changed 15 years ago by
| Resolution: | → fixed |
|---|---|
| Status: | assigned → closed |
After extensive testing, even with CIRCE and ISR, this has been fixed in r1918. Thanks, CS! Closing the ticket.
Seems to have been introduced in r1165 while fixing an array overflow I had discovered. I suspect that the code makes implicit assumptions about the shape of rho_hel elsewhere, causing this nasty issue with LHAPDF. I have reverted the change and introduced another fix for the bounds violation (intentionally breaking particles with spin > 3/2 on the way), please check if this is acceptable.