- 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
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
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
Add a env var ROC_USE_FGS_KERNARG to toggle kernel arg placement
By default its in Fine Grain Kernel arg segment for supported asics.
Change-Id: I3d57ed69a1a4db2b392b0438ead499f3ddca4716
This patch allows to substitute binary for the opencl program. It supposed to be used as:
1. Run the opencl program with -save-temps.
2. Open the cl temp and find the following text in the program header:
Hash to override:
Source: 0xd66bcfa20e69e605
Source + clang options: 0x656a9dd8aedcbfb6
3. Create config file (ascii text) with a pair(s):
<hash> <path_to_binary_to_substitute>
where hash is the hex value from step 2 (without leading 0x), you can use either hash
depending on what you're going to match:
only the source text of the program or along with it's clang options.
4. Set the env variable AMD_OCL_SUBST_OBJFILE to the path of your config file.
5. Rerun the opencl program.
Change-Id: I977c80fe529ea14458194918c6ddfbe2de6a8857
- 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
Setting AMD_CPU_AFFINITY = 1 will make runtime honor core affinity that
the process may set. This is disabled by default as it can prevent
worker thread or any thread that runtime creates from getting scheduled
thus affecting performance.
Change-Id: Ibe4cc95e7b99caee5ce750b7bf66e09e999cc9a3
- 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
The hack dosn't really track the commands status. It may be not
necessary for HIP, but will cause early resource release.
Change-Id: I791ad36dd8abd3b6b3d2c9b16a210a555c08ca64
This is helpfull to do when debugging issues on lowend asics. Navi14 can be emulated as Navi10. So can Navi22 be emulated as Navi21.
Change-Id: I693ffd45a5b03657822afdc872781901bc69b65c
The change reuses HSA signals for dispatches as a wait signal.
Skipping the barrier requires to disable L2 cache for sysmem
allocations and extra tracking for HDP access with the large bar.
ROC_BARRIER_SYNC=0 activates the new logic. Barrier sync is
still used by default.
ROC_ACTIVE_WAIT=1 enables unconditional active wait in ROCr.
The change also consolidated ROCr wait logic under single function.
Change-Id: I6bd1be30aa88258da1b1f9de319ef5a45852afd8
This reverts commit 5f055d227d.
This change broke the legacy-complib build in p4. It seems that we can't use any flags in debug.cpp.
Change-Id: I17bb83651b85d6f415d9074634b479658fd4c3f9
- Expose ROCclr interfaces for HIP usage
- ROCr interfaces aren't available in staging, thus control the
build with AMD_HMM_SUPPORT define
Change-Id: Iadc2bcc230e78d3b0dc22b235189c8cc80843446
This reverts commit a3b730b595.
Reason for revert: HIP_ENABLE_LAZY_KERNEL_LOADING is needed before the runtime is initialized, so this utility cannot be used
Change-Id: I49f8ddb98c9a85b9a77b8fd4b236d06b6b2b0f32
This workaround is to avoid performance penalty of SDMA engine
taking a while to clock up from a lower DPM state. Add env var
GPU_FORCE_BLIT_COPY_SIZE (1024 by default for HIP in KB). Forcing
Src and Dst agent to be amdgpu makes ROCr take blit copy path for
what otherwise should have been SDMA copy
Change-Id: I222f687155f86000d17d66d25182e490b6710463
[hipclang-vdi-rocm][perf]~45% to 50% of Performance drop on
rocBLAS_int8 test
- Enable AMD_OPT_FLUSH optimization by default to match HCC
- Disable CPU writes to GPU memory on boards with large bar,
because it requires HDP flush tracking.
- Enable L2 cache on kernel arguments, because L2 will be
invalidated on memory reuse .
Change-Id: I124cf250bdd4d19c523ce542c163813828f8fbdc
German Andryeyev
Saleel Kudchadker
Satyanvesh Dittakavi
Sourabh
Jason Tang
vpykhtin
agunashe
Anusha GodavarthySurya
Vladislav Sytchenko
Julia Jiang
Sarbojit Sarkar
Aryan Salmanpour
Payam Ghafari
Aakash Sudhanwa
Laurent Morichetti