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-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]
- Use getBuffer/releaseBuffer in BlitManager
- Cleanup XferBuffer as we use ManagedBuffer for both reads/writes
Change-Id: I2661b85dd012763b17a38a743fec1b1d79125f67
[ROCm/clr commit: 37d606d193]
- Fix regression for D2H pinned copies which adds systemscope release.
- Skip cpu wait for D2H unpinned copies as we can pass the signal of the
barrier to rocr copy.
- Fix an old bug in sdmaEngineRetainCount_ logic
- Improve logging
Change-Id: If074bddb05564b15949b0d5f9bf12acd3692174e
[ROCm/clr commit: 4c95ee5e1e]
BatchMemop should be positioned before the image support kernels
because the total number of kernels is determined by BlitLinearTotal,
when there is no image support on the device.
Change-Id: I8e53caf744ba54259ac04bad1762eef21806f3f2
[ROCm/clr commit: 3e01da3dac]
This reverts commit 0830d95f6d.
Reason for revert: There needs to be memcpy size change
Change-Id: If4f51769731e54743ac705b19b4f81b2d5925d5a
[ROCm/clr commit: 446ed661a0]
We are passing this arg as an address, and memcpy complains about
overreading (8 bytes instead of 4).
Change-Id: Ica9207f6c5f6056a4bfc968280c76e779ded13ae
[ROCm/clr commit: a6f2a2c2af]
- Refactor blit code and clean ASAN instrumentation
- Use unified function for rocr copy
- Enable shader copy path for unpinned writeBuffer/readBuffer paths
- Set GPU_FORCE_BLIT_COPY_SIZE=16 which means we will use BLIT copy for
pinned copies or unpinned H2D/D2H copies < 16KB
Change-Id: I42045cca79234b340dbf53dafb93044199736ae4
[ROCm/clr commit: 7863eb92dc]
Although unpinned copies require synchronizations
in HIP, runtime can avoid syncs for H2D copies with
a staging buffer
Change-Id: If2203c6bc0cbd89742823688dc8e89e9acd873b2
[ROCm/clr commit: 29cc678d8d]
- Gfx12 TCC cacheline size is 256B, Increase to have alignment
compatible. Eventually this needs to be replaced with what the query
returns.
Change-Id: I545929446c4faa3f26872a6290b3a89657888596
[ROCm/clr commit: bb01b4c3b4]
If we are using the mask returned by getLastUsedSdmaEngine() then we
need to apply the SDMA Read/Write mask to it before using with HSA
copy_on_engine API.
Change-Id: I6e5dc6c187eeb3c61ee159e9d2a0fa7b4737c06e
[ROCm/clr commit: 3f0bcf7834]
The copies can get blocked if the last SDMA engine is used by another
copy and this can lead to perf drop in some of the tests like Gromacs.
Resetting the last engine by checking the engine status and fetching the
new mask after few copies can avoid this.
Change-Id: I8fe8ea678db508d291c6242f3741fa9215e99921
[ROCm/clr commit: 1b25484f0f]
Add hipMemcpyDeviceToDeviceNoCU to force a non blit copy path. This
helps in cases where an app may determine that CU may be busy and copies
with SDMA may be quicker.
Change-Id: I59b415dd8f6022c244e8d75f265464d5c635df1e
[ROCm/clr commit: f316a30e5d]
- Remove large bar memcpy path. Since we end up waiting for a barrier,
its defeating the true intent of the copy, Also memcpy over PCIE\XGMI is
introducing variability in perf for HPC apps like GROMACS
Change-Id: I3b5c9d9ce93333959c39023bf4f703e2ccb6e3af
[ROCm/clr commit: 0a3d4bd4d4]
- Maintain a map of SDMA engine# to stream allocated following a greedy
approach
- Anything past that will query SDMA engine status always and go with a
SDMA or Blit copy path
Change-Id: Ibfaed7f951ab84d80cb0430596a4d11b5aec9202
[ROCm/clr commit: 5865c642d4]
- Leverage managed buffer that would use chunks for fill pattern. Use a
different chunk for the next fill to avoid wait
Change-Id: I254483c867e112f66564ffd8f55e0a605d8896c9
[ROCm/clr commit: 175ad024d3]
Implementation to use a blit kernel to perform
a hipStreamWait/write instead of an AQL packet.
Change-Id: I462671ed5cec37144dfe97ff66439249196117c1
[ROCm/clr commit: cbb8d82bdb]
For the fillBuffer shader, if there are two 32bit writes to a MMIO
register, it can get dropped. It has to be a single 64bit write.
Add optimization to fillBuffer to write 64bit and 16bit writes.
Change-Id: I3aa78e027898f8ae01e9c8f09004615673720c2b
[ROCm/clr commit: 21ba34d0fe]
- The logic will trace compute, sdma read/write operations and
apply signals when necessary
- ROC_CPU_WAIT_FOR_SIGNAL, ROC_SYSTEM_SCOPE_SIGNAL
and ROC_SKIP_COPY_SYNC were added to control the tracking
Change-Id: I9e8e6174c63bf7784f7ab00964e2918c8667d364
[ROCm/clr commit: dbc7abaecf]
Implement the global class for signals tracking per device queue.
Switch to the new tracking mechanism.
Change-Id: I3c4dda04b34e6d18d6a95510d84102909633b415
[ROCm/clr commit: 8698aeef0d]
On behalf of Tony Tye. The bootleg passes CQE Baffin/Ellesmere test.
Change-Id: I4c21d21b3aaba360682ef15b8a4dda239f8af276
[ROCm/clr commit: d4316141b7]
When HIP_ENABLE_DEFERRED_LOADING=0, many global variables will be
referenced but they are not initialized in that early time. The patch
will use constexpr to initialze global constant varables in compile
time.
Change-Id: I9d538b7abc6a0ce700ec3332b97fc144db5fc1ef
[ROCm/clr commit: fdef6f722f]
Optimization for the fence release removed a sync for mem fill.
Add simple const buffer management forr the filled pattern to avoid
pattern overwriting with the async fills.
Change-Id: I63773ac09ceec31d5396d24570e4647ff096326b
[ROCm/clr commit: 2ce6bbebc4]
Apply the optimization to change for OpenCL too.
Clean up some unnecessary checks.
Change-Id: I840261fe35baeeadeba7388e86779d482f509aad
[ROCm/clr commit: 6c5a42b33c]
[ROCm][TCT][HIP] cooperative stream test case is failing.
Make sure lockXfer() in the blit manager returns a valid value.
Port the latest PAL backend logic into the ROCr backend.
This change doesn't fix the issue, reported in the ticket.
Change-Id: I54101a824f49a2dcfbbf5414cb5b3af41745306d
[ROCm/clr commit: 89133a7301]
SaleelK
Danylo Lytovchenko
Andryeyev, German
Rahul Manocha
kjayapra-amd
Ioannis Assiouras
Jimbo Xie
Jatin Jaikishan Chaudhary
Sourabh Betigeri
Satyanvesh Dittakavi
Anusha GodavarthySurya
Ajay
Ravi C Akkenapally
Jason Tang
Tao Sang
Michael LIAO
Alex Xie
jujiang