Releases: cornelisnetworks/opa-psm2
IFS_RELEASE_10_10_2_0_44
Syncs the open source version of PSM2 with the code released in IFS 10.10.2.0.44. It also corresponds with PSM2 tag PSM2_11.2.156.
IFS_RELEASE_10_10_1_0_36
Syncs the open source release of PSM2 with the version provided with IFS 10.10.1 and corresponds with tag PSM2_11.2.89.
IFS_10.10.0.0.437
Syncs the open source release of PSM2 with the version provided with IFS 10.10 and corresponds with tag PSM2_11.2.86.
IFS_10.9.0.0.210
Sync with PSM2 master up to remove unnecessary long waiting timeout for HFI driver
This code surface corresponds to version 11.2.77 of PSM2.
IFS_10.8.0.0.204
Update PSM2 to include HAL support and start 11.x series
This code surface corresponds to PSM2 release 11.2.23
In addition to HAL (Hardware Abstraction Layer), the change also includes:
- Modify PSM send side protocol selection logic to select correct flow at
start of send function - Update LMC disable thresholds
- Fix for Issue #30: Compilation error due to snprintf truncation warning
- Convert CUDA runtime API calls to CUDA Driver API calls
- Prevent SHM and IPS from closing other protocol connections
- Change CUDA_THRESH_RNDV and move CUDA init to separate function
- Fix selection logic for multi-HFI within NUMA node
- Refresh JKey to support SELinux
- Fix Issue #24: Uninitialized AM message
- Fix data data corruption on GPU buffers when when using am ptl
- Force EPID remap in shm if remote PID changed
IFS_10.7.0.0.145
Updating PSM2 to include GDR Copy feature.
This code surface corresponds to PSM2 release 10.3.35.
Some other code changes include-
- Addition of static version of libpsm2
- Fix SLES 12.3 libpsm2-2 Provides line
- Properly release memory mappings before closing context
- Avoid typecasting unknown addresses to aligned data types
- Cleanup cuda symbol loading procedure
- Fix GPU Direct Recv threshold default
- Fix uniformity of process mapping to HFIs
IFS_RELEASE_10_6_0_0_134
Major release starting v10.3 series of PSM2.
This code surface corresponds to libpsm2 10.3.8.
Changes also include:
- SLES 12.3 rpm name change
- V2 EPID
- Initialize CUDA support only if env variable is switched on
- Add PSM_PERF functionality (for measuring number of instructions
retired between two points in code) - Bug Fixes
IFS_RELEASE_10_5_0_0_155
Major release to enable Multi-Endpoints feature of libpsm2.
This code surface corresponds to libpsm2 10.2.235.
This change also includes:
- Optimized tunings for Intel Knights Mill Xeon Phi CPU's
- Set TidFlow starting sequence number to 0 to avoid wraparounds
- Generalize affinity logic to extend to non-CUDA workloads as well
- Cleanup build artifacts. "make" output will now be in $PWD/build_release by
default - Bug Fixes
- Mask non-DW msg size assert
- Fix hang associated with re-transmit of last fragment of non-DW aligned
SDMA eager message - Keep track of only last unsent ACK message
- Retry context operns to fix interrupted syscall
- Fix segfault in disconnect code path
IFS_RELEASE_10_4_2_0_7
Update PSM2 code to include CUDA support and psm2_ep_disconnect API
This code surface corresponds to 10.2.175 of PSM2. The changes include the following-
- Add psm2_ep_disconnect API
- Add GPU-Direct and support for CUDA buffers in PSM2
- Implement new HFI Selection algorithm for CUDA version of PSM2
- Related to this, Add new PSM2_MULTIRAIL option to allow multirail only
within a socket.
- Related to this, Add new PSM2_MULTIRAIL option to allow multirail only
- Bug Fixes and Driver interface changes
- Sync with driver interface changes for GPU-Direct feature support
- Fix hang in PSM2 during context open
- Fix bug in psmi_memalign()
- Fix data corruption due to unaligned bytes in msg
IFS_RELEASE_10_3_1_0_22
Sync with wfr-psm 10.3.1 This commit corresponds to 10.3.1 PSM2 codebase. Changes include fixes to segfault due to 4MB window size and optimization of PSM2 to for message size <= 32b (for SDMA transfers). Signed-off-by: Aravind Gopalakrishnan <[email protected]>