- Add chicken bits to use host-side dependency management.
- Add optional PinInPlace path for unpinned copies
- Synchronize before pinned memcpy path.
- Add mutex to protect two threads launching to same stream.
On HIP path property obtaining done through hsa_iterate_agents and counting the devices of HSA_DEVICE_TYPE_GPU type.
P.S.
On multi-boards systems it might be problems with detection what board a GPU plugged into (not tested).
- Add device-side signal waits when transitioning between command classes
(Kernel, H2D copy, D2H copy).
- Support waiting in staged memory copies as well.
- Add several chicken bits to control implementation:
- HIP_DISABLE_ENQ_BARRIER
- HIP_DISABLE_BIDIR_MEMCPY
- HIP_ONESHOT_COPY_DEP
- Refactor signal pool to support efficient deallocation based on
signsequnm.
- Deallocate copy signals on eventSynchronize.
- Improve copy tests, add pingpong.
2. passed initial tests on different sets of
HIP_VISIBLE_DEVICES: (0),(1),(0,1),(1,2),(2,3),(1,2,3),(2,3,4),(1,5,2,3)
and achieved expected choice of GPU devices at the runtime.
3. Passed HIP test suite.
- add API to add / remove user-pointers from the tracker.
- test for thread-safety with MultiThreadtest_2 - rapid
insertions/removal.
- add mutex to provide thread-safety.
- rename tracker interface to "memtracker_..." for consistency.
- add am_memtracker_reset, connect to hipDeviceReset.
-
Tracks device where memory is allocated, pinned-host or device, and
more.
Uses memory-range-based lookups - so pointers that exist anywhere in
the range of hostPtr + size will find the associated AmPointerInfo.
The insertions and lookups use a self-balancing binary tree and
should support O(logN) lookup speed.
Device property MaxSharedMemoryPerMultiprocessor set equal to totalGlobalMem (HIP path).
Reason: MaxSharedMemoryPerMultiprocessor should be as the same as group memory size. Group memory will not be paged out, so, the physical memory size = total shared memory size = group region size. NVCC path remains untouched: CUDA's device property MaxSharedMemoryPerMultiprocessor is reported.
hipify is updated as well.
Reason: maxThreadsPerMultiProcessor should be as the same as group memory size. Group memory will not be paged out, so, the physical memory size = total shared memory size = group region size.
NVCC path remains untouched: CUDA's device property maxThreadsPerMultiProcessor is reported.
pensun
Aditya Avinash Atluri
Ben Sander
Evgeny Mankov
gargrahul