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
- Make sure barrier with dependent signals issues before queue
index reservation
- Don't issue extra barrier if it's already a barrier command
with dependent signals
Change-Id: I179a8b7adac79eed698f4a4d9eca2606d8e913aa
This addresses the rocprof hang seen with direct dispatch. The
workaround queues the handler back if any of the signal value in the batch
is not decremented. To rememmber the last position in the list, we save
the parsed command in the current timestamp struct.
Change-Id: I02959e463cfe3cee83c54808ffd6e6f48f43b4e8
MT doesn't use GPU waits, but CPU for sync between engines.
Change the threshold values for CPU waits for direct dispatch.
That will bring behavior closer to MT.
Change-Id: Ia41c3cb812614962aff2746b6cf858f1bf77dda2
The check has to be performed inside the signal loop, because
active signals need to be processed to avoid a stale timestamp
class.
Change-Id: I26af8287aae18eb19c096d9358cd0b86cfd1c561
- With direct disaptch profiling state is enabled to trigger the
callback on HSA signal. However ROCr has very low peformance on
the first call to get the profiling info. That impacts some tiny
performance tests.
Change-Id: Idacd1b10a473fcfb5feef3074b7191d35743f769
The copy image workaround could be recursively used by ROCclr blit kernel.
Avoid such situation by using stack variable.
Change-Id: Iadaa8cad9216220194760dd461a9533bb236aea0
HIP tests require HIP callbacks to be processed in another thread.
This change will use a thread from HSA signal callbacks to make
sure a HIP callback was done asynchronously.
Also process the callback before changing the status of command
Change-Id: Icef85d0e0f808663882cf6881ff1be3e5eca29ac
- Don't notify if the batch is empty, because that means
the current command was processed already.
- Disable pinning optimization to avoid a race condition on stall.
- TS marker submition requires extra AQL barrier
to track the status.
Change-Id: I17eff4ad12ac66cfe1bb44048bebb1891805279d
GPU waits have noticeable overheads on compute with extra
AQL barrier packet and on SDMA with power saving features. This
change introduces a wait on CPU for 30 us in case the app has tiny
operations.
Change-Id: I761ba3af595f3f48544980058a9077dda15aa5f9
This change unifies the hostcall implementation for all the backends,
by pushing the common logic to the device layer. This is done by
replacing the use of hsa_signal_t with device::Signal (a light wrapper
around it).
Change-Id: I7b6fca7930b5a0b199da5d85e2e048354cc04e7b
- Avoid GPU wait on the marker submission and update the command
batch after HSA signal callback upon HSA barrier completion.
Change-Id: I5c1c97212aefc2ae4b99aa9e2a81627ee9a38c1c
In pitch workaround, we need to copy the image to copy buffer
when application wants to read image to buffer. After this
patch, we unconditionally copy the image data to the copy buffer.
Change-Id: I71b0d19459542dfbb3ca51a2c8a3a81367fa2fb5
The existing workgroup calculation logic for GWS initialization is
incorrect. It tries to add together workgroups across dimensions,
leading to major under-count in 2D and 3D kernels. An (x,y,z) kernel
uses x * y * z blocks, not x + y + z.
In addition, the previous logic was incorrect for the case of launching
a single-threaded kernel. It calculated 0 workgroups, leading to
initializing GWS to -1.
Change-Id: I1bb20a0d5b6e0cc10ac55901c28d8f93aac61c09
German Andryeyev
kjayapra-amd
Sarbojit Sarkar
Sourabh
Jason Tang
Satyanvesh Dittakavi
Saleel Kudchadker
agunashe
Alex Xie
Vladislav Sytchenko
Ravi C Akkenapally
Payam
Joseph Greathouse