The maxSharedMemoryPerMultiProcessor attribute is meant to describe
the number of bytes of shared memory (LDS space in AMD terminology)
in each SM (CU in AMD terminology). For instance, on AMD GPUs this
is often 64KB per CU, and some Nvidia GPUs it's 96KB per SM.
This shared memory is a different address space from the normal
global memory. However, the current HIP-HCC properties fill this
in with a size that matches the totalGlboalMem property. This gives
a drastically too-high calculation for the amount of LDS space that
each CU has -- tens of GBs vs. 10s of KBs.
This patch fixes this by pulling the maxSharedMemoryPerMultiProcessor
property from the HSA pool that describes how much workgroup-local
space is available on each CU. The HSA runtime eventually pulls
this from the topology information about LDSSizeInKB, defined as
"Size of Local Data Store in Kilobytes per SIMD".
Previously, this HSA query was used to fill in the value of the
sharedMemPerBlock property. On today's AMD GPUs, we know that
the amount of LDS avaialble to the workgroup is identical to the
amount of LDS space in the CU. However, in the future this may
differ. As such, this patch changes around the order and fills
in the "PerMultiProcessor" property from the HSA query (since
what's what the query is defined to return), and then separately
fills in the "PerBlock" property as we know it.
[ROCm/clr commit: 44b16df84f]
- Enable -O3 for hipDispatchLatency.
- Use nearly-null kernel to prevent it from being optimized away.
- Formatting for hipDispatchLatency.
- Formatting for hipInfo.
[ROCm/clr commit: e8beacbfb1]
- set USE_PEER_TO_PEER=3 (requires HCC "am_memtracker_update_peers")
- when enabling peer, turn it on for previously allocated memory.
- hipDeviceCanAccessPeer is no longer self-ware (self does not qualify
as a peer)
- device peerlist always includes self, so when we call allow_access
we never remove self access.
- hipDeviceReset() removes old peer mappings.
[ROCm/clr commit: 9e7efd7c65]
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).
[ROCm/clr commit: 7bb0f17656]
Joseph Greathouse
Sarbojit2019
chrispaquot
Evgeny Mankov
Maneesh Gupta
Aditya Atluri
Ben Sander
Aditya Atluri