The following builtins from the CUDA spec are implemented:
- __all_sync, __any_sync, __ballot_sync and __activemask
- __match_any_sync and __match_all_sync
- __shfl_sync, __shfl_up_sync, __shfl_down_sync, and __shfl_xor_sync
The following builtins are NOT implemented, pending support in the compiler:
- __reduce_add_sync, __reduce_min_sync, __reduce_max_sync
- __reduce_and_sync, __reduce_or_sync, __reduce_xor_sync
Change-Id: I07dedbbfe5449f4b5c9b040bed59f5603ccec8c3
- Use isFenceDirty() check in hipStreamQuery.
In case if there is a stopEvent for hipExtLaunch* API we avoid system
scope if there are consecutive hipExtLaunch* and replace them with agent
scope, but then if one uses hipStreamQuery it should detect we need a
barrier with system scope.
Change-Id: I539810f659d4e1c49104e3a0da6ff2803c6b3dee
This reverts commit 15b8cf911a.
Reason for revert: multiple tickets filed. i will re-instate this change after fixing those.
Change-Id: I66152e3643c73ec1d240d0fcddf2119f8f4848a8
This issue is due to the fact that we set fence_dirty_ flag for every
kernel launch. Whenever any next HIP API is caused, the stream logic
assumes fence is dirty and queues a marker, when its not needed.
Change-Id: I10a49ee951daa92221aa6be75e2334849579a45d
Recently some unused compiler options for HSAIL path were removed,
but it affected blit kernels compilation. Hence, remove those options.
Also delete assert for device to device copy in SDMA path for now.
Change-Id: Ib5d7f063af2ab4a3fc5d73d426e39c391b1011ac
Single OCL context with multiple devices should use direct map
even with persistent memory, because cache coherency layer.
Change-Id: I02a48830440203a1dea176acd4dff4452b70f133
- Make sure persistent memory from resource cache is properly adjusted
in free memory calculation.
Change-Id: I74ef68975ccde4694fb1cb904617c418e85dfc9f
hipIpcOpenMemHandle can be used to open memory handle and fetch
a device pointer on a different device than the actual device
where the memory was allocated. The device Id must be set to
the current device in such cases and not the original device.
Change-Id: Ie1c7eada928d02124a41125876876f96015552e5
This is an initial change before we refactor the build/link paths for
kernel launches for HIP. This current change is needed as compiler was
setting some dump file which needed fs access which has slowdowns for
NFS mounted file systems
Change-Id: I828f9bb04d789b4f8c05c1ed08767f325efeb47c
Add GPU_DEBUG_ENABLE to control ttpm behavior. If enabled,
then HW will collect more debug info at some perf cost
Change-Id: Icee0686b903a7b1bd483710b9d611877cd43c6aa
- For new AccumulateCommand we enqueue nop barrier packet. So during stream sync we need system flush.
Change-Id: I0b97b626bcdae582ef95e7c95030d78df1fa5a54
Runtime has to release extra memory, held by the pools,
in synchronization points for event, stream or device.
Change-Id: Id533a5e1d137812aa72bdfe101b4b333c6a43d66
When an offset is applied to the source or destination pointers plus the kind is set to
hipMemcpyDefault and the source or destination is allocated with hipMallocManaged
hipMemCpy2D erroneously fails with hipErrorInvalidValue.
Change-Id: I0db4c17514f743652d8f9a2691da6601a2abb2a1
Fix sporatic segment fault in texture test
via retaining image in texture object which
references the image.
The image will be released when the texture
object is destroyed.
Change-Id: Ic3fefa2d5dda6afebd1acd4d41ad310b138af6dd
Recent changes disabled system memory allocation
in the abstraciton layer. That requires memory
allocation/destruction in ROCR. Add destruction logic.
Change-Id: I68fe6b0a620ca743fe5850052ea0efa8bb7931c2
Extra CPU read back will be performed before every submission to make sure
previous writes over PCIE reached GPU. HDP flush is done by CP.
Change-Id: I402d28ca26c8cee4a3920feb3599af8c285d0889
Jatin Chaudhary
Sameer Sahasrabuddhe
Saleel Kudchadker
jujiang
Konstantin Zhuravlyov
German
kjayapra-amd
sdashmiz
David Salinas
Rahul Garg
ksankisa
Satyanvesh Dittakavi
Ioannis Assiouras
Konstantin Zhuravlyov
Jaydeep Patel
Anusha GodavarthySurya
Alex Xie
taosang2
jiabaxie