- Add a global cache state for a device to indicate scopes of submitted
AQL packets
- Remove scopes for TS marker if hipEventReleaseToDevice is passed. Set
env ROC_EVENT_NO_FLUSH=1 to use NOP AQL for event records.
It would flush caches by default with system scope release.
- Calling finish() should ensure if caches are flushed, if not queue a
marker
Change-Id: Ibbbdbb1cd7ac61cb35649169212142545be159e0
Remove assert for kernel arg size, because COv5 reports a value
bigger than the actual usage in the most of cases
Change-Id: I8e15bc45a9e21b58a5894f9977511ca84408ce61
- Fix a crash with AMD_CPU_AFFINITY=1 as numa_bitmask_alloc isnt the
right api to allocate bitmask
- Do not set affinity for ROCr thread. It worsens performance rather
than any improvement.
- Fix regression from my previous change for event handler.
Change-Id: I3ea75adc2a6333f29752283eddd5b555e9b58cc5
- Queue handler for hipEventRecord(aka marker_ts_) only if there is a
callback associated with it.
Change-Id: I8a9877ae0e342556053abbaacc9510744a8e772a
Pass active queue for transfers in the cache coherency layer.
That will allow to use device transfer queue only for
cases when active queue isn't available, because using device
transfer queue from another active queue may cause a deadlock
Change-Id: Ifbe7e0303b77dbf6eeda3939ffbc25a3df7472de
If GlobalMemCacheLine reported is 0, runtime may run into an
infinite loop as the KernelSegmentAlignment is chosen as size of the
cache line.
Change-Id: Ide547940cc0407f16fab10ee210b4fd3ae4eaafc
Metadata in Codeobject version 5 is the extension of CO3 and CO4.
Add the detection of the new fields and program them in
the setup of the kernel arguments.
Change-Id: I27e58df77320ad00f4f16d35912668db803826af
Add a state indicator to retain ExternalSignals when needed.
Co-operative group launch uses external signals to indicate a dependency
to the next command.
Change-Id: I6d0daa006e2377c3bbf4aeca0fd5b63c7ac8fbbb
Crash was due to the fact that external signal structure was stale even
after destroyign the command. That is because we skipped wait due to a
missing check.
Detect external signals and dispatch a barrier in ReleaseGpuMemoryFence.
Also clear external_signals_ at ProfilingBegin.
Change-Id: I991387edcfe928b511bf5e780988ee131321ed5a
Add a threshold for ROCR/SDMA P2P transfers. ROCR copy path
requires extra barriers in compute for synchronization. That costs
extra performance with tiny transfers.
Reduce active wait time to 10us. Tensorflow uses extra thread
per GPU with constant hipEventQuery() calls. Longer active waits
in ROCr affect CPU performance.
Change-Id: I9020358438615fa2d4617f862f00a562f0a588e7
Stall in the host thread could occur earlier than the app expects.
Make sure rutnime can grow the signals to the queue size without
any stall. Also adding a new signal to the end of the pool could
break the dependency chain on signal reuse. The new logic will
insert the new signal after current to keep the chain intact.
Change-Id: I9c90b98515907db8b677528263c3e88cd9581a14
The optimization is controlled with ROCR_SKIP_KERNEL_ARG_COPY.
This is initial check-in for experiments. Extra changes are
necessary for full support:
- handle graph capture with the original sysmem alloc
- avoid memobject references, otherwise there is a race condition with
reusage of the arg buffer
- Remove arg setup from hip
Change-Id: Ib0af710f93e79834711fa4049a7c66093711e68b
The kernel arg pool will be divided into 8 chunks to avoid long stalls,
when the pool will be reused.
Change-Id: I228e6ca1c09e428c1775f1e5b685220a9a5d71af
The original logic left only one slot for HW processing in the queue.
For some reason there is a race condition on CPU overwrite of the slot
before the current active. The workaround is to avoid the previous to
the current active slot for possible unfinished HW processing.
Change-Id: I565495a8feeaedffc9fc8a505edbee5ff5816975
Pass the device agent specified by the user to the ROCr api instead of passing the device agent attached to the specified stream
Change-Id: I86c98935b9dc404eaa6d47ccdd082a8c3678fb36
There is a possible race condition when signal reuse can have
access to a destroyed Timestamp object, because the callback
was running asynchronously. Use reference counter and lock
to allow asynchronous timestamp update
Change-Id: I6224f7c62cb0a03a7466fcc512e5e5afb06736fa
Add ref counting to ProfilingSignal class to track the last release.
If a signal was used in the marker, then don't reuse it,
but create a new one for internal usage.
Don't rely on HSA callback for the command status update if there
are no pending dispatches.
Change-Id: I19f14ed9d80acfe79993b343b2187635f8428a20
Runtime can't assign internal HSA signals for HIP events, because
HIP application can destroy the HIP stream or signal reuse may
occur internally. Switch to global HSA signals for HIP events.
Change-Id: Ieaea2d6b039e492b2e7c5112782a8f4e601e50a1
Add lock protection for signal processing
If signal is reused, then disable reference to it from HIP
Increase the pool signal size to 32
Change-Id: I7d529b35910f83ce577c9eca6d3386759611ccc0
- Create an env var ROC_ACTIVE_WAIT_TIMEOUT to set active wait timeout
- Record profiling informaion if marker_ts_ property is valid.
Change-Id: If0d8aec8d9b0715027cf0f7c3dc8a4c722a6bae6
With HIP API callback runtime has to stall the queue until the
callback is done. Rocclr will introduce SW blocking HSA signal,
which will be released after the callback is done.
Change-Id: I6411f3efab31b468e3b87ebb5c8d155e116b613d
German Andryeyev
Saleel Kudchadker
Maxime Chambonnet
kjayapra-amd
Satyanvesh Dittakavi
Sarbojit Sarkar
Sourabh
Jason Tang
agunashe