This now ensures that finalizeReq is indeed called from mrntt,
since exception-experiencing threads will post an exceptionInd
to mrntt, which will then call finalizeReq.
We added a new centralized OpenCL Compute manager. This can later
be extended to support CUDA, SyCL, etc. SMO can be configured at
build time to choose which API it will use for compute.
Moreover, the ComputeMgr allows us to register buffers which are
available to all cl_contexts.
This symbol is defined as a static member object inside of a
boost detail header. When boost headers are used in a project
that uses Boost in both the main binary as well as dlopen()'d
shlibs, the top_ symbol gets duplicated and the metadata gets
partitioned.
We use the Boost shlib to unify both the main binary and the
shlibs to use the same memory address for top_.
This involves marking the templated object call_stack::top_ as
"extern" and then declaring to Boost that we intend to use the
shlibs.
Async: Use new [Non]PostedAsyncCont and callOriginalCb
This new hierarchy of classes gives us a central mechanism for
managing both reply-posting and lockSpec unlocking.
* callOriginalCb: Now uses a modern C++ variadic template design
enabling it to handle both direct calling and std::bind()
re-binding of an arbitrary number of arguments from the caller.
This enables us to mostly eliminate the repeated, bespoke
definitions of callOriginalCb littered throughout the codebase.
We've also propagated these changes throughout the codebase in
this patch.
This leverages the new clean dynamic allocation of the globalMind
object to make the mrntt::main and SMO's initialization and
shutdown much cleaner. We no longer concern ourselves with
shutting down the Mind threads inside of mrntt::main, but rather
we leave that state machine to the Mind class and Mrntt component.
This makes the initialization sequence much cleaner and conceptually
well encapsulated.
We also now dynamically allocate the Mind objects. They're allocated
dynamically by Mrntt inside of initializeReq. This means that we no
longer have to worry about jolting and cleaning up the running threads
of global mind object even when we never explicitly called
Mind.initializeReq.
Along with other conceptual improvements to our abstractions, this
patch also gets us to a real "end of program initialization" point
for the first time.
We now allocate globalMind locally inside of marionetteMain. Why?
Before now, we had an asymmetric threading situation where the
globalMind's threads were initialized at during global constructor
invocation and not on demand. This meant that we had to shut down
those threads even if we had never got to the point of calling
Mind::initializeReq.
This significantly complicated our shutdown sequence since we had
to factor in the lifetime of the std::thread objects inside of the
ComponentThreads which were inside of the globalMind object.
Now, if we hadn't called Mind::initializeReq, we don't have to
perform any Mind::finalizeReq or adjacent operations. Shutdown is
symmetrically mirrored against the operations we actually performed
during execution.
We introduced some complexity by splitting ComponentThreads into
two derivative types (MindThread and MarionetteThread) but I think
in the long term we'll be able to massage this split into a much
cleaner situation overall.
In Mrntt, we now initialize Mind:: object threads before calling
initializeSalmanoffReq().
We've also propagated the spinscale async pattern into the Mind
class.
This is the culmination of a lot of changes over the last week. We're
making SMO basically fully async in many areas, and then preparing to
implement the spinqueueing mechanism for locking.
Now we have modularized the Mind class to contain all of its
ComponentThreads. This enables us to run multiple mind instances
within the same SMO process, at least in theory.
We probably won't actually do this, but we want to ensure that the
design is clean enough to enable it.
We don't post a message to the event loop telling it to execute
initializeSalmanoff anymore. We now execute it in the main control
flow.
Also, we've unified the logic to call finalizeReq() in response to
exceptions in the outermost try block.
This is generally frowned upon but it makes this code 10x cleaner.
We handle commandLine usage msg printing by using exceptions for
control flow. This allows us to centralize the logic for killing
the Mind threads in one place. At least with respect to printing
the usage msg.
We moved the instance of smo::Mind to global scope. I suppose we'll
only support one instance of Mind per SMO process at least for now.
We now track the state of Mind threads' JOLT-waiting. This allows us
to centralize the Mind thread shutdown logic. Mind::finalizeReq()
now takes care of all Mind thread shutdown state logic by tracking
whether Mind threads need to be JOLTed first before being told to
exit.
We moved initializeSalmanoff and shutdownSalmanoff into
salmanoff.cpp. Now we also invoke shutdownSalmanoff when exiting
to destroy subsystems and components gracefully.
This fixes the segfault that was thrown on every program exit
when xcbWindow had captured a window.
hayodea