| Rev |
Log message |
Author |
Age |
Path |
| 3047 |
Fix the wrong affinity map generated in libopenmp on Magny-Cours RHEL 5.4
system, where the physical id does not start from 0. |
dcoakley |
1119d 04h |
/branches/x86_open64-4.2.3/ |
| 3046 |
Update version number and release notes for x86 Open64 4.2.3.2 release. |
dcoakley |
1119d 06h |
/branches/x86_open64-4.2.3/ |
| 3045 |
- Fixed bug 15146 (Shared libraries .rodata contained
dynamic relocations). Given the current implementation
for switch jump tables, the labels need to be put into the
.data.rel.ro.local section instead of .rodata.
- Applied fix from PSC compiler.
- Note that there is more work to be done see:
https://bugs.open64.net/show_bug.cgi?id=565
Internal bug 15739 points to this report. |
dcoakley |
1119d 08h |
/branches/x86_open64-4.2.3/ |
| 3044 |
- Provide a cleaner fix for bug 15607 (a bad processor
affinity map was being generated with newer Linux kernels
running on AMD Magny-Cours systems).
- When O64_OMP_VERBOSE is set, debugging trace information
is written to stderr when errors are detected during
the map generation process. Also when this environment
variable is set, in the situation where no errors, the
affinity map is written to stderr. |
dcoakley |
1119d 08h |
/branches/x86_open64-4.2.3/ |
| 3043 |
Fix in affinity mapping generation: do not mingle cores on different sockets.
For example, a machine's topology is (0-8 2-10 4-12 6-14) (1-9 3-11 5-13 7-15)
according to /proc/cpuinfo. 0-8 means proc 0 and proc 8 are two threads
in a core and they have the same socket id and core id. And cores
0-8,2-10,4-12,6-14 are on the same sockets. The affinity mapping should be
"0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15" instead of
"0 2 4 6 1 3 5 7 8 10 12 14 9 11 13 15". The later has problem when there
is one socket. |
dcoakley |
1119d 08h |
/branches/x86_open64-4.2.3/ |
| 3042 |
Fix thread-cpu binding with hyper threading (HT). When HT turns on,
there are 2 threads in a core. In /proc/cpuinfo, these 2 threads have
the same physical id and core id. Thus, a thread count is added to
distinguish them to avoid binding 2 OMP threads to the same cpu.
Also add better error detection in getting information from /proc/cpuinfo.
This fixed the cactusADM issue in SLES11 (bug 15607), where in /proc/cpuinfo,
"siblings" value is 6 while 6 "core id"s are numbered as 0, 1, 2, 0, 1, 2. |
dcoakley |
1119d 08h |
/branches/x86_open64-4.2.3/ |
| 2793 |
Update version number and release notes for x86 Open64 4.2.3.1 release. |
dcoakley |
1207d 04h |
/branches/x86_open64-4.2.3/ |
| 2771 |
Update version number for patch release candidate. |
dcoakley |
1220d 06h |
/branches/x86_open64-4.2.3/ |
| 2770 |
Update README for x86 Open64. |
dcoakley |
1221d 02h |
/branches/x86_open64-4.2.3/ |
| 2769 |
Added MTYPE_F4 WHIRL trees for reassociation. |
dcoakley |
1221d 02h |
/branches/x86_open64-4.2.3/ |
| 2713 |
Update version number and release notes for x86 Open64 4.2.3 release. |
dcoakley |
1261d 04h |
/branches/x86_open64-4.2.3/ |
| 2712 |
Merge r2684-r2711 from open64-booster. |
dcoakley |
1261d 06h |
/branches/x86_open64-4.2.3/ |
| 2685 |
Update documentation for x86 Open64 4.2.3 beta. |
dcoakley |
1279d 04h |
/branches/x86_open64-4.2.3/ |
| 2684 |
Merge r2673-r2683 from open64-booster. |
dcoakley |
1279d 05h |
/branches/x86_open64-4.2.3/ |
| 2674 |
Prepare for x86 Open64 4.2.3 beta release. |
dcoakley |
1284d 05h |
/branches/x86_open64-4.2.3/ |
| 2673 |
Create x86 Open64 4.2.3 release branch. |
dcoakley |
1284d 10h |
/branches/x86_open64-4.2.3/ |
| 2672 |
Fix invalid svn:mergeinfo property value. |
dcoakley |
1284d 10h |
/branches/open64-booster/ |
| 2671 |
Fix merge errors in updates r2405-2669. |
dcoakley |
1284d 11h |
/branches/open64-booster/ |
| 2669 |
Fix for 15520. Use without def from source program causes seg fault |
dcoakley |
1285d 07h |
/branches/open64-booster/ |
| 2668 |
Corrected a bug in induction variable incrementing in cost vector
calculation. |
dcoakley |
1285d 07h |
/branches/open64-booster/ |