* clr: Implement dynamic stream to HW queue assignment
This change implements dynamic stream to hardware queue (HWq) mapping
with the following features:
* Queue depth heuristics with weights for optimal HWq assignment
* Make last used queue sticky for better locality
* Use pipe HWq to pipe mapping - gfx9 follows a round-robin queue to
pipe mapping based on creation order (single process per device only,
as pipe ID is statically assigned by runtime)
* More aggressive heuristic usage for better queue distribution
* Extend dynamic queues support for all stream priorities
Environment variables:
* DEBUG_HIP_DYNAMIC_QUEUE: 0 - disabled, 1 - Depth heuristics 2 -
Depth+Pipe heuristics
* DEBUG_HIP_IGNORE_STREAM_PRIORITY=1: ignore priority stream creation
* clr: Clean up last_used_queue_
Problem:
The existing SDMA engine selection logic had several issues:
1. Same VirtualGPU/stream could use different SDMA engines for consecutive
async copies since copy_engine_status may report engines as busy
2. Busy and Preferred engine check for every copy
3. No global tracking of which VirtualGPU uses which engine, leading to
suboptimal resource allocation
Solution:
Implemented a global SDMA engine allocator with per-stream affinity:
- Added Device::SdmaEngineAllocator to manage VirtualGPU → engine assignments
* Maintains global map of active assignments
* Enforces exclusivity: different streams use different engines (except
inter-GPU copies where preferred engines are prioritized for optimal
hardware paths like XGMI links)
* Thread-safe allocation/release with Monitor lock
- Modified VirtualGPU to cache assigned engine locally (assigned_sdma_engine_)
for fast lookup without map access on hot path
- Refactored rocrCopyBuffer() to:
1. Check local cached engine first → use if assigned
2. Call AllocateSdmaEngine() if not assigned → cache result
- Moved HSA API queries (memory_copy_engine_status, memory_get_preferred_copy_engine)
into AllocateEngine() for cleaner separation of concerns
- Engine release on HostQueue::finish() instead of only VirtualGPU destruction
* Improves engine utilization by releasing earlier
* Added virtual ReleaseSdmaEngines() method to device::VirtualDevice
- Added future path for simple round-robin allocation (kUseSimpleRR) for
next-gen GPUs with uniform SDMA bandwidth (disabled by default)
Cleanup:
- Removed selectSdmaEngine() helper (logic moved to allocator)
- Removed getSdmaRWMasks() (allocator accesses maxSdmaReadMask_/WriteMask_ directly)
- Removed unused sdmaEngineReadMask_/WriteMask_ member variables from DmaBlitManager
Benefits:
- Ensures consistent per-stream SDMA engine usage
- Prevents cross-stream contention and engine thrashing
- Prioritizes hardware-optimal paths for inter-GPU transfers
- Better resource utilization through earlier release
- Cleaner, more maintainable code structure
* SWDEV-555347 - Remove lock contention in async events loop
* SWDEV-555347 - Introduce Pool of AsyncEventItems
* create generic mempool for AsyncEventItem
* Use BaseShared allocate and free for async event pool
---------
Co-authored-by: Rahul Manocha <rmanocha@amd.com>
* SWDEV-465041 - Add support for user events with DD
User events can be replaced with HSA signals. Add the interface
to allocate HSA signal for user events and update the status on
CL_COMPLETE.
Force pinned path with DD to avoid blocking calls. Pinned memory
can be released only when the command is complete.
Simplify device enqueue path to use generic kernel arg buffer and
signals
* Fix notifyCmdQueue() logic for OCL
* Avoid blocking calls in OCL with DD
* Add event destruciton in a case of the failure.
[ROCm/clr commit: 2305f8ae56]
- When doing device/stream sync, we can submit a handler which may
introduce some host side delays. Use DEBUG_CLR_BATCH_CPU_SYNC_SIZE to
batch commands for host wait. Default for HIP is 8 commands.
- Investigation is underway in ROCr but need to address this for now in
HIP runtime.
[ROCm/clr commit: 9b045922a8]
* SWDEV-520352 - Remove HostThread and legacy monitor
Remove HostThread, semaphore and legacy monitor.
Make original logics of thread and command queue stricker.
Add more comments to make logics clearer.
Some other minor improvement.
Also part of SWDEV-458943.
[ROCm/clr commit: 96cadbc9e9]
Fix data validation issue of rocFFT when dynamic queue on.
ReleaseHwQueue() can be called only when no command in HostQueue.
The checking condition need be protected by lock.
[ROCm/clr commit: 18d191fd1d]
- This change tries to save extra synchronization packets we may insert
as we didnt track the completion signals for every command. We track
the current enqueued command until it exits the enqueue stage. We also
record the exit scope to know if we flushed the caches
- Handle correct release scopes and store completion signal as HW events
- Use a new finishCommand implementation to only wait for the command
passed as the argument
Change-Id: Ie4350c5dd24f5d48dfa6ccbabd892f0544caadcc
[ROCm/clr commit: e03e4f3b5d]
There are 2 functional changes to this patch:
* Use GPU timing for internal markers for HIP.
* Measure CPU time closer to GPU timer, to reduce delta between GPU/CPU timestamp measurements.
There are some smaller non-functional updates:
* waifForFence -> waitForFence typo
* Remove unused drmProfiling
Change-Id: I4c5fa600a842ab60e454888779edcac8449a902a
[ROCm/clr commit: 179801a750]
- Resolve signal dependencies for barrier value packet if there are > 1
depenent signals. Barrier Value packet accounts for only 1 dep signal
- Better log
Change-Id: Ia506ad5d80b91d598f92e7b539f41756e9b4b64b
[ROCm/clr commit: 2d450e8b06]
Add an atomic counter to track the outstanding HSA handlers.
Wait on CPU for the callbacks if the number exceeds the value
in DEBUG_HIP_BLOCK_SYNC env variable.
Change-Id: I95dc8c4bf0258c7e59411b7504220709ed6898c5
[ROCm/clr commit: 403f624bf8]
Applications may submit commands withoout waits
for GPU. That causes a growth of SW unreleased commands.
Make sure runtime flushes SW queue, if it grows over some
threshold, controlled by DEBUG_CLR_MAX_BATCH_SIZE.
Change-Id: Ia4d85c24210ef91c394f638ab6b53b14323a0396
[ROCm/clr commit: 8657a77029]
- Don't generate callbacks for HIP events
- Don't process profiling info in the callback for HIP events
- Wait for CPU status update of the submitted commands
every 50 calls. That will allow to drain the commands and
destroy HSA signals.
Change-Id: Ib601a350e7e7c2b6c6209a172385389baccf73a9
[ROCm/clr commit: 364dfb0ed1]
The new set tracks only the queues that have a command
submitted to them. This allows for fast iteration
in waitActiveStreams.
Change-Id: I2c832eefa01280d9a87a5f57874d36d2e9441de7
[ROCm/clr commit: bcc545e6b8]
We must be in protected way to get last command when calling
awaitCompletion() where lastCommand will be released and
possibly destroyed.
This can solve scope lock(notify_lock_) crash in
Event::notifyCmdQueue() with AMD_DIRECT_DISPATCH = true.
Change-Id: I4297166f912a71112f4a8945d993160ba9afdc34
[ROCm/clr commit: 749385155a]
- No cpu wait is needed when profiler is attached, Doing this changes
the application profile when roctracer is attached.
Change-Id: I2b9cfc48d697cf5ed54bb6a240d8c12bdb079171
[ROCm/clr commit: 51e4368723]
It seems that due to removal of vdev()->isHandlerPending(),
Marker queued to ensure finish is not enqueued and that cause
hung at waiting event for kernel enqueue command.
Change-Id: I364abb2dcb4897b11a7eb61b5d85013b69292792
[ROCm/clr commit: eecbc2e436]
- Refactor code and cleanup logic for callback saving for event records
Change-Id: I5c56aa8e9c968a5bca70fb07ad1796da318e9e89
[ROCm/clr commit: 1338ff37e8]
Some pytorch tests use a tracer plugin and rely on profiling information
to be reported right after hipDeviceSynchronize()
Change-Id: Ib021a1e7b1a30b3c24de72627c471810f7f7878d
[ROCm/clr commit: 5438b6362e]
Allocate marker only if HW event doesn't exist for the last command.
Change-Id: I3e7284202365a9c75313fb5403f0c1908ab51d1e
[ROCm/clr commit: 596b496c16]
hipStreamSynchronize and hipDeviceSynchronize won't longer wait
for CPU commands in DD mode
Change-Id: I079c8bbfc34ddc6d3e2d74c92a34665877e512a5
[ROCm/clr commit: fbea58ba11]
Use large signal pool if profiler is connected or profiling forced
enabled. This is needed to mitigate signal creation overhead when
profiling as signals are attached to every packet and deeper batch may
show overhead of signal allocation.
Change-Id: I8034b8a20b55328b87d593bf044f59672f9653e8
[ROCm/clr commit: 1ec0ba3537]
- Enable CUs adjacent pairwise for WGP mode
- In HostQueue::terminate() do not segfault if virtual device hasn't been created
Change-Id: I94402ff333308af5824878086cc238b3993d534d
[ROCm/clr commit: 8c1232124e]
OS can terminate unfinished queue thread from default stream at any
time. Potentially leaving the queue lock in a bad state and causing a
deadlock if runtime destroys VirtualGPU later from the host thread.
Change-Id: I247f102ee84e6b4dba947504933395071945c85d
[ROCm/clr commit: 28daf98f1f]
Windows kills threads on exit without any notification. However,
runtime can still destroy VirtualGPU object from the host thread with
HostQueue destruction.
This change also forces RGP trace transfer on the last capture without
any delays.
Change-Id: I768e87e99e1d23a021e63c12f36e450817743759
[ROCm/clr commit: ad33a021cb]
- Store last fence scopes and use the last value to determine if we need a cache flush again. This helps cases where hipExtLaunchKernel API is
used.
- Purge code for ROC_EVENT_NO_FLUSH
Change-Id: I531cf9c9c60d5e2b3a9e265d0f52f79ed2fa8a8c
[ROCm/clr commit: 9b5cbd37a2]
The fix for SWDEV-329789 moved down the last use of the a
command object pointer in order to prevent a race condition.
However, the previous patch did not move down the release of
that command. By releasing the command early, another thread
could get a command with the same pointer. That second thread
could later submit work to the queue using that new command.
The first thread could then perform a comparison against the
queue's last command using its own now-stale pointer. This
could eventually allow the second thread to skip synchornizing
on the queue. This would result in host synchronizations
completing before their device work was actually complete.
Change-Id: I292b7b369743251ceafe453a4c5cae14a6d01046
[ROCm/clr commit: 6b956f7627]
To support both hip and ocl. HIP_FORCE_QUEUE_PROFILING will be replaced with this later on.
Change-Id: I6d3514b1568ff049584ed9fd74bbdb3e4f4bf0c3
[ROCm/clr commit: d92b3a2d90]
SaleelK
Ioannis Assiouras
German Andryeyev
Rahul Manocha
Godavarthy Surya, Anusha
Danylo Lytovchenko
Manocha, Rahul
Stojiljkovic, Vladana
Andryeyev, German
Patel, Jaydeepkumar
Jayaprakash, Karthik
Sang, Tao
Arandjelovic, Marko
Aidan Belton-Schure
Anusha GodavarthySurya
Ioannis Assiouras
Rakesh Roy
Ajay
Joseph Greathouse
Jason Tang