Since LAMMPS uses the low-level driver API of CUDA, it needs to ensure
that it is in the correct context when invoking such functions. At the
moment it creates and switches to its own context inside `UCL_Device::set`
but then assumes that the driver is still in that context for subsequent
calls into CUDA; if another part of the program uses a different context
(such as the CUDA runtime using the "primary" context) this will cause
failures inside LAMMPS.
This patch changes the context creation to instead use the primary
context for the requested device. While it's not perfect, in that it
still doesn't ensure that it's in the correct context before making
driver API calls, it at least allows it to work with libraries that use
the runtime API.
The following changes have been applied to src and lib folders:
regex replace: ([^"_])NULL ⇒ \1nullptr (8968 chgs in src, 1153 in lib)
Manually find/change: (void \*) nullptr ⇒ nullptr (1 case)
regex find: ".*?nullptr.*?"
Manually ~14 cases back to "NULL" in src, ~2 in lib
regex finds a few false positive where nullptr appears between two
strings in a function call
- document previously undocumented OpenCL tune settings
- implement OpenCL platform selection through prefixing the device type with the platform id separated by a colon
- allow passing custom tune parameters though postfixing the device type with the 13 tuneable parameters separated by commas
- remove an extra clear() that would delete device properties structs an cause LAMMPS to output garbage strings
The GPU package uses OpenCL events for measuring time. These have to be
released to free up memory. I removed the clReleaseEvent() calls in the
clear() method because in some cases they don't exist yet and I couldn't
find a way to check for a valid event (clRetainEvent didn't work). This
at least fixes the massive leak during simulations.
See issue #1006
