AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-497759 - Fix memObj offset computation for hipHostRegister on Windows mgpu

Browse Source 
On Windows, hipHostRegister may add a single object in the MemObjMap
that maps to memory that is allocated on different devices.
This change ensures that the offset that is returned from
getMemoryObject() is computed relative to the memory that is allocated
on the current device.

Change-Id: I5fd3af200bf6f4926fdeaea12dcb9d0154d3a843


[ROCm/clr commit: e80442fdbf]
This commit is contained in:
Ioannis Assiouras
2024-12-04 22:47:44 +00:00
parent 12dc02b4f8
commit dc3ca8aab7
1 changed files with 14 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/src/hip_memory.cpp
+14
View File
@@ -40,6 +40,20 @@ amd::Memory* getMemoryObject(const void* ptr, size_t& offset, size_t size) {
memObj = (hip::getCurrentDevice()->asContext()->svmDevices()[0])->GetArenaMemObj(
ptr, offset, size);
}
// On Windows, when using hipHostRegister, the map may contain a single memory object for
// multiple devices. This is because device addresses can overlap.
// The offset needs to be calculated relative to the memory of the current device.
if (IS_WINDOWS && (memObj != nullptr) && (memObj->getMemFlags() & CL_MEM_USE_HOST_PTR)) {
device::Memory* currentDevMem =
memObj->getDeviceMemory(*hip::getCurrentDevice()->devices()[0], false);
if (currentDevMem != nullptr) {
size_t currentDevOffset = reinterpret_cast<uint64_t>(ptr) - currentDevMem->virtualAddress();
if (currentDevOffset < memObj->getSize()) {
offset = currentDevOffset;
}
}
}
return memObj;
}