The MemoryAllocAll test in kfdtests exercises the new KFD memory
availability API by trying to allocate a single buffer object that
exactly fills all of vram. Desired object size is determined using the
memory availility KFD ioctl via libhsakmt, then an object is allocated
slightly larger than that size. If the allocation attempt fails then
the test tries to allocate a slightly smaller object, and continues
trying with smaller sizes until the allocation succeeds. The test
succeeds if the successfully allocated object is within some specified
tolerance of the available memory reported.
There are a number of known issues that can cause the successfully
allocated object to be significantly smaller than reported availability.
Until these issues are addressed, we should not fail the test, but just
log the actual divergence between the size of the object we thought we
could allocate, and what was actually possible.
Signed-off-by: Daniel Phillips <daniel.phillips@amd.com>
Change-Id: I165a30865ffbb2353286dcc896ad8e24af124615
Since KFD counts svm allocation as system memory usage,
KFDSVMEvictTest will fail on the case of small system
memory, adding check is to skip test.
Signed-off-by: Eric Huang <jinhuieric.Huang@amd.com>
Change-Id: I040f16f2dd0d4092d069a632cfba9c28293f781b
Updates:
* Fixed rsmi_dev_compute_partition_get
& rsmi_dev_nps_mode_get to properly check
for invalid arguments
* Updated compute partition & NPS mode tests
- Now properly confirms the invalid
argument is seen
- Spacing for multiple devices is added
to better see distinction between
separate device's tests (for verbose output)
- Changed expect to assert calls, so errors
are observed faster for test failures
- Fixed multiple device testing where a
variable should have been unset, but
having multiple devices caused it to
set
- Updated multiple device testing to iterate
accross all devices (previously returned,
instead of continuing checking support
after RSMI_STATUS_NOT_SUPPORTED detected)
- Fixed a few spelling errors & verbose output
Change-Id: Ieba9e5b46763c6cd880fbf27fcdf58be8ecbc683
Signed-off-by: Charis Poag <Charis.Poag@amd.com>
Implement hsaKmtExportDMABufHandle, which can be used for a new
upstreamable RDMA solution. It exports a DMABuf handle for an arbitrary
virtual address along with the offset of the address within the
allocation. It also checks that the size of the intended export does
not exceed the allocation.
This uses the new AMDKFD_IOC_EXPORT_DMABUF, which requires KFD ioctl
API version 1.12.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: Ie5fdb1f73ab3c7fa36c315ce326b1fb89eacc8b6
* Address and thread sanitizer fixes
- Fix compilation with clang
- Tweak perfetto copy to build tree
- Added suppression files to scripts
- fix LD_PRELOAD support in omnitrace-causal and omnitrace-sample
- use spin_mutex and spin_lock from timemory instead of atomic_mutex and atomic_lock
- state uses atomic
- fix some memory leaks
- tweak testing
- mpi tests do not use preload
- increase timeout when using sanitizers
- add env LD_PRELOAD when using sanitizers
* Tweak perfetto build
* Update timemory submodule
* Update version to 1.8.1
* Update omnitrace-leak.supp
* Update timemory submodule
- fixed spin_mutex implementation
* Remove previously added addr_space->allowTraps(instr_traps)
- this appears to cause errors during binary rewrite
* causal testing updates
- relaxed causal validation on CI systems (to account for hyperthreading decreasing prediction)
- improved impact calculation
- other general improvements to validate-causal-json.py
* Improve fork handling for perfetto
- numerous updates changing perfetto:: to ::perfetto::
- added perfetto_fwd.hpp
* Updated fork example
- user API for validation that stopping/starting perfetto is valid
* Misc fixes to perfetto + fork support
- tweak regions in fork example
- handle disabling tmp files
- get rid of stop/start with perfetto before/after fork
- fixed sampling support during fork
- tweak env of fork test
* Fix find_package in build-tree
* Fix buildtree export
* Fix buildtree export
* Restructured ConfigInstall before adding examples
* Guard against creating tmp file in sampling when disabled
* Fix buildtree package
* formatting
* exit handlers on child processes
- quick exit to avoid perfetto cleanup
* Further tweaking of causal tests for reliability
- enable PROCESSOR_AFFINITY
- decrease to 5 iterations
* Further tweaking of causal tests for reliability
- disable PROCESSOR_AFFINITY for fast func e2e tests
- enabling affinity results in (valid) speedup predictions greater than zero
* Fixes to fork handling
- use pthread_atfork for redundancy if fork_gotcha fails
* cmake formatting
* Fix fork init settings + install components
- remove dl from PROJECT_BUILD_TARGETS
* Testing tweaks
- fix mpi-binary-rewrite-run regex when OMNITRACE_VERBOSE set > 1 in env
- increase causal e2e iterations to 8
* Fix "Test User API"
- test-find-package.sh included dl component
* Further tweaks to causal validation
- further considerations of variance
Updates:
* Added RSMI_STATUS_SETTING_UNAVAILABLE for
rsmi_dev_compute_partition_set - gives users
better error output when attempting to set
compute partition to values not listed in
available_compute_partition SYSFS
* Updated python --setcomputepartition to
provide better output when receiving
RSMI_STATUS_SETTING_UNAVAILABLE
* Updated all test & example files to check for
RSMI_STATUS_SETTING_UNAVAILABLE when doing
rsmi_dev_compute_partition_set
Change-Id: Ida5d54880d9b9b6e4a0468cdb962fdc0c18d6257
Signed-off-by: Charis Poag <Charis.Poag@amd.com>
Remove BLACKLIST_GFX10_NV2X from GFX11 blacklists, update
BLACKLIST_GFX11 as needed.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I84bd91ba20a5d3df27478fb4c97afa12f8a3e76a
Use mwaitx instructions when busy waiting for signals to reduce CPU
energy usage.
This can be disabled by setting HSA_ENABLE_MWAITX=0
Change-Id: Ic207895a491b2bf6dacba47ef0921df3faad5b5a
Updates:
* Added rsmi_dev_compute_partition_reset & rsmi_dev_nps_mode_reset
* Added --resetcomputepartition and --resetnpsmode python smi calls
* Added temp data files rocmsmi_boot_compute_partition_<device num>
& rocmsmi_boot_nps_mode_partition_<device num>, writes UNKNOWN
if data cannot be read or device does not support
* Cleaned up NPS & compute API documentation
* Added creation and reading of API temp files (used in reset
functionality)
* Cleaned up output of rocm_smi_example
* Updated rocm_smi_example to check if running with sudo permission
before executing write API calls (cleans up erroneous output)
* Added template specialization for storing temp data, requires
specific rsmi_type_t enums (restrics what data can be stored)
* Added storage of temp data, if temp files do not exist
* Updated google tests for NPS & compute to include reset API calls
Change-Id: I69895a466b97107617e6dbb355737b84499a76c9
Signed-off-by: Charis Poag <Charis.Poag@amd.com>
Copying memory from device to host with a CPU agent
would cause a poor performance due to the reading of
uncahced device memory by CPU.
Fix it by using a GPU agent.
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Change-Id: Ia3b562758fe73ef9efaa284f47e67bf569cc7b7b
Daniel Phillips
Nirmal Unnikrishnan
Eric Huang
Charis Poag
coleramos425
Yifan Zhang
Hao Zhou
Bill(Shuzhou) Liu
Felix Kuehling
AravindanC
Jonathan R. Madsen
kent.russell@amd.com
David Yat Sin
Graham Sider
Karl W. Schulz
Lang Yu
Cole Ramos