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]
The profiler can retrieve this correlation ID to attribute waves to
specific dispatch locations.
Change-Id: I700e8a91219d612f6a2028c0dda0c92753f3526a
[ROCm/clr commit: b043b4f5a2]
HIP can't rely on the resource tracking, used in OCL and requires different explicit sync.
Make sure ROCCLR syncs compute only when SDMA is used and vise versa.
The new logic will allow to enable CPDMA without unnecessary waits.
Change-Id: Ib9d1788cfd5afa5ea2fec4c96a37d8b9c4d0059d
[ROCm/clr commit: ff6b4db70b]
Blender creates and destroys big allocations during the benchmark.
That causes big delays, because vidmm has to page-in/page-out memory.
Change-Id: I2baf4545807127406e3d2870a7581ff9ae7bcdb5
[ROCm/clr commit: dc4ad8c99c]
- Make sure SQQT trace is captured for RGP server if the queue is destroyed before normal capture is done.
- Remove prepare queue from the logic. It's not really used for any HW capture and can cause RGP server abort if destroyed before capture is even started(delayed capture)
Change-Id: I6eb19963190a5769c6477a5496c1b831a6d59b89
[ROCm/clr commit: c1c5127875]
In X86 on Windows, sizeof(size_t)=4, but size=8, for
amd::KernelParameterDescriptor::HiddenGlobalOffsetX/Y/Z items.
Loose the condition to prevent crash.
Change-Id: I2216f71f4d4fd6dd3766023b1c821cb3d35d7848
[ROCm/clr commit: 3d281114fb]
The ROCclr assigns zero-based IDs to GPUs in the order they are
discovered. That zero-based ID is what is used to identify the GPU
on which the HIP_OPS activity took place.
When multiple ranks are used, each rank's first logical device always
has GPU ID 0, regardless of which physical device is selected with
CUDA_VISIBLE_DEVICES. Because of this, when merging trace files from
multiple ranks, GPU IDs from different processes may overlap.
The long term solution is to use the KFD's gpu_id which is stable
across APIs and processes. Unfortunately the gpu_id is not yet exposed
by the ROCr, so for now use the driver's node id.
Change-Id: Ib78854527d600d175bb76e2df0747c33f898c615
[ROCm/clr commit: 9a82118c85]
- Use a dirty flag to determine fence optimization
- If fence is dirty submit a marker at top level to sync.
Change-Id: I53fb19b5bb05b7c7b37c41637a6c7aaf870b639a
[ROCm/clr commit: 6405b6cdba]
- 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]
Remove the activity_prof::CallbacksTable. The table was redundant with
the information already stored in the roctracer library. Instead use a
single callback into the roctracer library to query whether the activity
is enabled, and to report it.
Change-Id: I2e05b0881bb4a1953c14361d00ea310d02eb6e0c
[ROCm/clr commit: 52eb28930a]
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]
The activity_ is only instantiated if profiling is enabled.
Remove the HIP private global record ID. Instead, use the correlation ID
stored in the hip_api_data_t by the profiler while the last HIP function
is in scope.
For NDRange and Copy commands, store the kernel name and byte size
(respectively) in the record.
General cleanups to improve the code's readability.
Change-Id: I01907484b0d9611eb9440c3a7c4865479dc42289
[ROCm/clr commit: 4fbae91468]
Chauncey Hui
kjayapra-amd
German
Tao Sang
Laurent Morichetti
jatang
Saleel Kudchadker
Ajay
Alex (Bin) Xie
Sourabh Betigeri
Julia Jiang
Christophe Paquot
Jeremy Newton