- UUID needs to be specified in the format GPU-<body>, <body> encodes UUID as a 16 chars
- Convert set UUID in HIP_VISIBLE_DEVICES to device index internally
- Then use existing device index logic for HIP_VISIBLE_DEVICES
Change-Id: I654f492a49cd4d7a9b7339360ab558165240caa5
Avoid a deadlock on the host call buffer creation. Since the buffer will be
allocated in the queue thread, then use direct device memory allocation
skipping the global context lock.
Change-Id: I09b55ee03bb42ab5d320c152b52a8c842c5fdcc1
- UUID is Ascii string with a maximum of 21 chars which uniquely identifies a GPU
- Convert set UUID in HIP_VISIBLE_DEVICES to device index internally
- Then use existing device index logic for HIP_VISIBLE_DEVICES
Change-Id: I8cab4fe42459f8209b97f909300789e6e687b9ac
Refactor PrintfDbg::outputArgument() to remove potential risk.
Fix half vector printf issue on all devices.
Fix FEAT-56794 as well.
Change-Id: Iae39359d2128588def2e43d77fe58e868b8e71ff
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
alwaysResident setting doesn't require per queue residency tracking.
Thus, the logic can be skipped to avoid the lock of queues.
Change-Id: Ib5cff5b79d3ecb8c2f2eb2565cf069f9a69438b0
- Create a vector to allow multiple TS to be stored in Command.
- This would mean we dont wait for entire batch in Accumulate command
to finish when we exhaust signals.
- Reduce the number of signals created at init to 64. This min value
may still need to be tuned but the KFD allows max of 4094 interrupt
signals per device.
- Store kernel names whenever they are available and not just when
profiling. If we dynamically enable profiling like for Torch, a crash
can happen if hipGraphInstantiate wasnt included in Torch profile scope
beacuse we previously entered kernel names only when profiler is
attached.
Change-Id: I34e7881a25bbc763f82fdeb3408a8ea58e1ec006
- With https://gerrit-git.amd.com/c/lightning/ec/llvm-project/+/1002628 applied, at -O0 Kernel::dynamicParallelism() returns true but virtual queue isn't created
- This causes segfault inside VirtualGPU::submitKernelInternal() when getVQVirtualAddress() is called
Change-Id: Ia7af042adad2329e870c142caaac3e8fa886f8b8
Under ROCr physical allocations don't have initial VA and require extra
flag in ROCclr. Add an option to have a mempool of physical allocations.
Change-Id: I4d062fe0dd8113d4eaf6e8b51749ed56d8701d1e
PAL optimized the logic for the barriers, which caused failures with CP DMA on Navi4x.
Change barrier's code to match the most recent PAL optimizations.
Change-Id: I55eeab20f51eb8e920bcbb4b55fbe3c7f77fd3fa
Previously, we used the following approach and Comgr actions
for device lib linking:
AMD_COMGR_COMPILE_SOURCE_TO_BC (compile with clang driver)
AMD_COMGR_ADD_DEVICE_LIBRARIES (link in device libs with
llvm-link API)
However, the clang driver can link in device libraries as part
of compilation, assuming a --rocm-path is set. In this context,
this is accomplished by using the following Comgr action instead:
AMD_COMGR_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC (compile and
link in device libs with clang driver)
Change-Id: Ie0bbee7d9a12672536b6d751056a941128ed58be
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
Saleel Kudchadker
Jaydeep Patel
German Andryeyev
kjayapra-amd
Ioannis Assiouras
Konstantin Zhuravlyov
Rakesh Roy
pghafari
taosang2
Sourabh Betigeri
cadolphe
Anusha GodavarthySurya
Matt Arsenault
Jacob Lambert
Satyanvesh Dittakavi