The new set tracks only the queues that have a command
submitted to them. This allows for fast iteration
in waitActiveStreams.
Change-Id: I2c832eefa01280d9a87a5f57874d36d2e9441de7
The variable is already set as cache, so that user can override.
But the hard coded setting is preventing override. Removed the same
Change-Id: I2aecc18ce4f1d1b523ba267ef1c8ef4ea1168d9c
1) currently cpu wait is set to true, which makes the host wait for last
command in queue to finish even if the kernel execution has already
finished causing delay in device sync call.
2) device sync only needs to await completion when hw event
is not ready.
Change-Id: I91e3e89d39a1193ae06abac822cea8ae651493a5
1) Since g_devices is not initialized when stream_per_thread constructor
is called on windows, m_streams is empty when hipDeviceReset is called.
2) clear_spt tries to access empty vector causing segfaults in
hipDeviceReset call.
3) on linux ROCCLR_INIT_PRIORITY makes sure that g_devices is initialized
first before tls constructor creates stream_per_thread object.
Change-Id: Ib2ba643d1278d820287ea3b242ed0878d7529165
The amdgpu-arch tool is not supported for static build.
This commit adds changes to detect the build type during
cmake config and use the rocm_agent_enumerator for static build.
Change-Id: I8a295e01f54075507390ef540f16b28bb20237a9
This change adds a new HIP API `hipExtHostAlloc` which preserves
the functionality of `hipHostMalloc`.
Change-Id: I13504c6fc13465ddd7aed329795bb4f2fef1baff
Integration into pytorch pointed out some issues, value narrowing, to
fix this we are now using unions. Also removed check for -munsafe*
compiler flag. The check is now just on builtin detection.
Change-Id: I49364503fa429bd862952f9b29879072afa6d553
SPT is destroyed with hipDeviceReset(). If a
stream is created right after reset, the same
object id could be reused. Later SPT destructor
incorrectly verifies that the stream is valid
referring to the reused object id causing the
corruption.
Change-Id: I3b1f7ffdf8bab874dca7b8fde22318162997b8f6
This change adds fixes in optimized multistream path for childGraph uses cases.
1) For childgraph nodes, rely on runNodes() only to process
the childgraph and skip calls to createCommand and enqueueCommands.
This ensures that the start/end markers are enqueued correctly
with respect to the childGraph commands.
In addition, the runNodes() for the childgraph should be called after
the dependency walkthrough to make sure that the subgraph is executed once.
2) Nodes with no outgoing edges should be marked
as a leafs regardless of which stream they are assigned to.
This is to ensure that marker dependencies from nodes
that run on non-zero stream to subgraph leafs that run on zero stream
are still set up correctly.
Change-Id: I4a5f4f3b0e0d01e515cdcb045b46c2798f291255
When source or destination pitch is set to zero in hip_Memcpy2D struct
it should default to WidthInBytes + [src/dst]XInBytes
Change-Id: Id57b53cab40ba72ced231258da9356554c4868c3
- Fixes -0.0 and +0.0 comparison. For atomicMax if the value on
address is -0.0 and on val is +0.0, gfx90a's unsafe atomics will swap
them. This behavior should be consistent with cas loop as well.
- _system variants of atomicMax and atomicMin are resulting in
incorrect output. Updated these to use the similar implementation as
atomicMax and atomicMin.
Change-Id: I20df36ee29ae0434a6b564f2ba71193fe41cfa59
Releasing graph exec after wait completes and before delete hip::stream obj
during stream destroy.
Change-Id: I1d68aa8d844f7d3af330c6d09c44af07f8553551
- Added the optimized multi stream path in graph execution. It uses a fixed number of async streams in the execution
- Optimize the launch latency, where commands
creation and execution is done at the same time
- Optimize the scheduling to use less barriers and waiting signals if
the same queue can be detected
- The new path is controlled by DEBUG_HIP_FORCE_GRAPH_QUEUES
environment variable, where 0 will use the original path and any other
value will force the number of asynchronous queues for execution
- DEBUG_HIP_FORCE_ASYNC_QUEUE can force single queue async
execution in graphs(applicable for Navi families only)
Change-Id: I7eb40bc15c45f508d6911868a6f6d4c3598d380e
=> Added support to capture multiple AQL Packets.
=> Added Interface to callback to hip runtime from rocclr to allocate
kernel args from the graph kernel arg pool.
=> Enabled Support to capture memset node.
Change-Id: I7e1c2ba06927459e024653058af142bd82192c43
hipDeviceSynchronize called from __hipUnregisterFatBinary
accesses static maps and monitors. This change ensures these ojects
are not destroyed before __hipUnregisterFatBinary is called.
Additionally it disables the teardown process for static build.
Change-Id: I46b58641d60efcf6637a8e99cdd786ffe9e2c77d
This issue was happening because of incorrect usage of getStream call,
if we get the null stream first and then typecast it, and call on
getStream again, we lose the advantage of simply passing "nullptr" to
indicate NULL stream. Thus we enter the waitActiveStream call and add
barriers to sync across streams.
Change-Id: I94dc4e3ec927295b9e1ab6dee4b37d7d3e00b0cc
Jaydeep Patel
Ioannis Assiouras
Ranjith Ramakrishnan
Rahul Manocha
Ajay
kjayapra-amd
Julia Jiang
victzhan
Marko Arandjelovic
Anusha GodavarthySurya
Tao Sang
Jatin Chaudhary
Vladana Stojiljkovic
Sameer Sahasrabuddhe
Sourabh Betigeri
Satyanvesh Dittakavi
German Andryeyev
Saleel Kudchadker