Arm Forge Release History
This page describes the changes between releases of Arm Forge.
What's new in Allinea Forge 6.1
Arm Forge 6.1 covers the following releases:
- Allinea Forge 6.1.0
- Allinea Forge 6.1.1
- Allinea Forge 6.1.2
Release Note for Release History Allinea Forge 6.1
Changes for Allinea Forge 6.1.2:
- Fix missing metrics when using MAP via remote launch or reverse connect.
- Fix crash when another profile is loaded after zooming in the metrics view.
- Fix the current line(s) view not working when a file is closed and reopened in DDT.
- Fix detection of overwrites of reserved memory in the memory debugger.
- Fix CUDA thread selection when stepping through warp.
Changes for Allinea Forge 6.1.1:
- Fix a case where standard input would not be forwarded during non-interactive sessions when the source was a file on a Lustre filesystem.
- Add support for profiling programs where all MPI communication takes place on a single non-main thread.
- Fix an issue where CUDA runtime and driver version mismatches were sometimes silently ignored.
- Fix an issue with the memory debugger deadlocking on some platforms when one thread frees memory from dlclose while another thread is allocating or freeing memory.
- Fix an issue where a deallocation trace would sometimes be shown for a newly allocated object when freed memory was reused.
- Fix an issue which prevented expressions from being used in the MDA viewer if a language had been set.
- Fix an issue with the Ctrl+K stopping working when switching tools.
- Fix an issue that prevented the graphical installer working on Suse 12.
- Remove the temporary .build directories for the precompiled MPI wrappers.
- Address a debugger start up problem affecting Intel MPI, MPICH, MVAPICH and generic MPI users.
- Fix crash on startup for Cray systems with an old version of wlm_detect.
- Fix multiple compilers being used to compile MPI wrappers.
Changes for Allinea Forge 6.1:
- Debugging and profiling support for Knights Landing (KNL) Xeon Phi product family.
- Support for CUDA 8.0.
- Enable CUDA support on POWER.
- Lustre I/O metrics added to MAP.
- Preloading the Allinea profiler libraries is now supported for dynamically linked programs on Cray X-Series / Cray MPT systems.
- Preloading the Allinea profiler libraries and DDT memory debugging library when using Slurm versions older than 14.11 is now supported.
- Precompiled Allinea MPI wrappers now provided for OpenMPI (1.6 and up), Intel MPI (5.0 and up), MPICH2, MPICH3, Bullx MPI, MVAPICH2 and Cray MPI.
- In DDT's offline log, display the local variables and stack arguments for a configurable number of stack frames under every stack.
- For dangling pointers, display the stack at the point of deallocation in the Pointer Details window.
- Highlight dangling pointers in the Cross-Process Comparison and Cross-Thread Comparison windows.
- The command-line arguments for DDT offline logs are now consistent with those for creating MAP profile files. To start DDT in offline mode, use the --offline flag. DDT will generate an HTML offline log with an automatically chosen filename. Optionally, you can add "--output=<filename>" to direct the output to a particular text or HTML file. Warning: Existing batch scripts that use the "--offline log.html" syntax must be changed to use the new "--offline --output=log.html" syntax.
- Add support in DDT's disassembly view for 64-bit POWER specific instructions.
- Include stat/lstat/fstat calls in MAP's I/O metrics.
- MAP supports functions export in profiling mode.
- Mac OS X client installer is now signed.
- Licences are now tied to architectures. x86_64 and i686 are taken as defaults if no architectures are specified in the licence file.
- Fix CUDA 7.5 not working on some Cray systems.
- Fix missing OpenMP thread IDs in Cross-Thread Comparison window.
- Forge will no longer show spurious TopologyFile::readEntry warnings on some systems.
- Programs that dynamically load OpenMP at runtime will no longer crash.
- Fix array viewer modifying lower bounds whilst user is still typing maximum bounds.
- Fix an issue that would cause memory debugging to hang on ARMv8 (aarch64) systems.
- Add support for Platform MPI to Express Launch mode.
- Programs using Platform MPI will now have their working directory set correctly.
- Add support for statically linking the profiler libraries on Cray systems that are using binutils >= 2.25.
- Add automatic SLURM detection on Cray.
- Add support for UPC and CAF programs on Cray systems using native SLURM.
- Memory allocations are aligned correctly for SSE data types when guard pages are set to "after".
- Fix incorrectly reported number of nodes in summary and program details.
- Fix "--nodes" command line argument handling.
- Fix an issue that could cause unnecessary sampling overhead when profiled program makes redundant calls to dlopen or dlclose.
- Prevent spurious "Maximum backtrace size in sampler exceeded, stack too deep." messages appearing when running MAP.
- Updates to example QTF files.