- Implement workaround to ensure HDP writes are done by writing and
reading the HDP MMIO register.
- Implement the same workaround for graphs, we no longer need sentinel
write/readback
Change-Id: I0d3027b46a1f61131ec62e3c8c669ff5184fa6b2
[ROCm/clr commit: f138e0d113]
Add GPU_DEBUG_ENABLE to control ttpm behavior. If enabled,
then HW will collect more debug info at some perf cost
Change-Id: Icee0686b903a7b1bd483710b9d611877cd43c6aa
[ROCm/clr commit: 7d661bc7df]
- Add the new fillBuffer kernel, which allows to launch a limited
number of workgroups for memory fill operation
- Switch fill memory to 16 bytes write by default
- Allow to limit the workgroups with DEBUG_CLR_LIMIT_BLIT_WG
Change-Id: Ibad1822f2d42b2fc71bcfc1917c31409c0623e8e
[ROCm/clr commit: f1dc81f427]
- Rename HIP_USE_SDMA_QUERY to DEBUG_CLR_USE_SDMA_QUERY as this is
supposed to be a temporary env var for debug purposes only.
Change-Id: If6ebd52ab87624375a3df24ceccdcc05c60a65af
[ROCm/clr commit: 770b2a4711]
- Use HIP_USE_SDMA_QUERY env var toggle for new API use. Env var is 0 by
default
Change-Id: If725a0c41e15f78a1a6c3f47942954fe9240b4db
[ROCm/clr commit: 8d193c32bb]
With recent upstream changes (D145770), we can now use the
Comgr unbundler without requiring an env field in the supplied
targetID. For users, this is consistent with previous legacy
unbundler behavior.
Change-Id: I5f085b0fa1ad352bbbb282b75367c206b75f279f
[ROCm/clr commit: 443f912c7f]
HIP_FORCE_DEV_KERNARG=1 will create a device allocation for kernel arg
segment. Flag is 0 by default.
Change-Id: Iaaf5a149f3be8596568878d5d272268baf067c60
[ROCm/clr commit: 5436d362b1]
The change enables VM support in graphs on Windows. That allows
to avoid caching of all allocations at the cost of map/unmap
overhead during memory create/destroy.
Change-Id: I792be00fba099e5e5d3cd44a963e1dfd6976a86d
[ROCm/clr commit: 04b696abee]
This reverts commit dfa7790030.
Reason for revert: Deferred to a future release.
Change-Id: Ia66c37f0ab9734dee73c930d10d7469d5fd57254
[ROCm/clr commit: 5dc104b3ea]
Pinned copy can cause big performance drops, because slow pinning under Windows.
Use up to 128MB for staging transfers. Change staging buffer size to 4MB.
Linux path should still have the old defaults.
Change-Id: I954edceb3ec89e8e670be116aa2d0a9564c8b11c
[ROCm/clr commit: 79d12df147]
Sync between compute and SDMA engines can be very expensive under Windows.
Use CP DMA for tiny transfers (< 1KiB) to avoid syncs and improve performance.
Change-Id: I9db39a2199f7b9e337ed08fd36d9cbc150502f1f
[ROCm/clr commit: 473621c008]
- 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]
Profiling should be enabled for any command reporting activities as the
activity record captures the profilingInfo's start and end timestamps.
Since IS_PROFILER_ON is only used to determine whether API tracing is
enabled, there is no need to expose it globally, it should be a property
of the activity_prof::CallbacksTable.
Change-Id: I44a0d19ed2862606cfbc9a98c1a07a336ab7e26c
[ROCm/clr commit: e713b5c7d0]
HIP_FORCE_QUEUE_PROFILING has been replaced by GPU_FORCE_QUEUE_PROFILING.
Change-Id: Ic32ecdf829a2725ace84e76abab8a81c8790e13f
[ROCm/clr commit: dd49cf0fa0]
To support both hip and ocl. HIP_FORCE_QUEUE_PROFILING will be replaced with this later on.
Change-Id: I6d3514b1568ff049584ed9fd74bbdb3e4f4bf0c3
[ROCm/clr commit: d92b3a2d90]
- 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
[ROCm/clr commit: 8eeaa998c0]
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
[ROCm/clr commit: 008133cf41]
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
[ROCm/clr commit: 7e12cf6318]
ROC_AQL_QUEUE_SIZE will control the size of AQL queue.
The current sefault value is 4096.
Change-Id: Icd2a4ee3ba554c06aa05b08defd922d2c63e43fd
[ROCm/clr commit: 7fe696b6ef]
Implementation to use a blit kernel to perform
a hipStreamWait/write instead of an AQL packet.
Change-Id: I462671ed5cec37144dfe97ff66439249196117c1
[ROCm/clr commit: cbb8d82bdb]
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
[ROCm/clr commit: e29b9c00ee]
Saleel Kudchadker
Anusha GodavarthySurya
German
kjayapra-amd
Ioannis Assiouras
jiabaxie
Todd tiantuo Li
Jacob Lambert
Alex Voicu
Maneesh Gupta
Laurent Morichetti
Jason Tang
Satyanvesh Dittakavi
Sourabh