noraabiakar authored 6 years ago and Benjamin Cumming committed 6 years ago

Remove redundant profiler calls that caused crashes when using event generators.

Make the execution context presented to users an opaque handle, moving all implementation of the gpu, thread and distributed contexts into the back end.

* move `execution_context` and `distributed_context` definitions to the back end
* create `execution_context` handle called `context` in the public API
* provide `make_context` helper functions that build different context configurations (default, user-specified local resources, with MPI)
* update documentation for all parts of the public API that touch contexts
* move `distributed_context` docs to the developer documentation (from the public API docs)

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Fixes #575.

* Guard CPU architecture option for nvcc with generator expression.

Benjamin Cumming authored 6 years ago and Sam Yates committed 6 years ago

* Add new ring benchmark to examples.
* Refactored common functionality for reading miniapp parameters from a json file to `aux` (used by both bench and ring).

Fixes #516.

Move implementation of `gpu_context` from header to `cpp` file, so that `ARB_WITH_CUDA` doesn't leak from library implementation.

noraabiakar authored 6 years ago and Sam Yates committed 6 years ago

* Add gpu_context as part of execution context containing information about GPU availability, managed_memory synchronization, and atomic double availability.
* Choose between ON and OFF for ARB_GPU in CMake. If ON compile for K20, K80, and P100

Note that we still need compile time information about the GPU in cuda_atomic.hpp for atomicAdd(double*, double*). This is because the function is only defined when the program is compiled  for sm_60 or more.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Fixes #568.

Benjamin Cumming authored 6 years ago and Sam Yates committed 6 years ago

Fixes #564

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Use a compat::fma wrapper for std::fma to avoid a bug in the tree optimizer in GCC version < 8.2.

See: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87046
Fixes #568.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Global temperature for mechanisms.

* Make 'celsius' magic in modcc: now an indexed variable.
* Add a new temperature data source for indexed variables.
* Add support to printers for indexed variables that reference a scalar.
* Check that indexed variables aren't used in PROCEDURE blocks (this is a problem not just for 'celsius').
* Modify built-in mod files to pass celsius as a parameter to rates() procedures.
* Add global temperature to shared_state classes, and initialize through backend mechanism superclasses.
* Add some infrastructure for unit-test only mechanisms.
* Set modcc flags globally in top level CMakeLists.txt.
* Add test mechanism/module for checking celsius setting.
* Add unit test for multicore and gpu mechanism celsius setting.
* Make common mechanism private field data access helper for unit tests.
* Use helper in temperature, synapses tests.
* Fix warning in `distribued_context.hpp` about errant semicolon.
* Fix global scal...

Two MacPorts/gcc7 issues:

std::uint64_t is unsigned long long on OS X, breaking an assumption about size_t in the distributed_context interface.
Problems with missing errno defines in the standard library headers.
With MacPorts gcc7, the installed c++config.h defines _GLIBCXX_HAVE_EOWNERDEAD and _GLIBCXX_HAVE_ENOTRECOVERABLE, but the corresponding errno defines are not provided by
sys/errno.h unless __DARWIN_C_SOURCE, which takes its value from _POSIX_C_SOURCE if defined, is greater than or equal to 200809L. Technically a MacPorts configuration bug? but easily worked around.

Use basic integral types for communication collectives interfaces.
Define _POSIX_C_SOURCE to be 200809L for glob.cpp.
Fixes #562.

* Remove dependency on memory library and range utils from `multi_event_stream.cu` source.

Fixes #545

noraabiakar authored 6 years ago and Sam Yates committed 6 years ago

* Replace distributed_contest with shared_ptr<distributed_context> in execution_context and pass around the shared pointer instead of a raw pointer.
* Fix construction of mpi_context
* Remove num_threads() from arb and arb::threading. Modify mpi_context so it also returns a shared_ptr. proc_allocation is initialized from execution context to determine available resources.
* Rename threading backend files. Delete useless files.
* Pass execution_context by const reference or value.
* Remove code duplication in thread_system constructors.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

* Update the install docs for architecture build options
* Update to reflect new install target

Reduce redundant functionality across event_generator, time_seq
and schedule by providing a low-heap overhead interface to schedule
and using that for time sequences in event_generator and specialized
cell groups.

* Have schedule return pair of pointers as view to generated times.
* Fix missing DEBUG/TRACE functionality.
* Use rate instead of mean_dt for Poisson schedule.
* Move merge_events() functionality to simulator.cpp.
* Migrate event_generator to event span interface.
* Migrate tourney_tree to event span interface.
* Only invoke tourney_tree merge if generators have events in the epoch.
* Use schedule for times in event_generator implementations.
* Replace seq_generator with explicit_generator that keeps a copy of events.
* Replace vector_backed_generator and poisson_generator with schedule_generator.
* Replace time_seq uses with schedule.
* Add default empty schedule.
* Move rounding error test for regular time sequence into sch...

noraabiakar authored 6 years ago and Benjamin Cumming committed 6 years ago

Use correct profiling and mpi flags.
Fix profiler doc

This small refactor simplifies the interface and implementation of the `tree` type.
* use `std::vector` instead of `memory::array` for internal storage in `arb::tree`
* return a `util::range` intstead of a view for `tree::children(int)`
* remove unused functionality for changing the root of a tree.

noraabiakar authored 6 years ago and Benjamin Cumming committed 6 years ago

Add initialize method to the profiler to set up the needed threading parameters given a simulation's task system.

noraabiakar authored 6 years ago and Benjamin Cumming committed 6 years ago

- Task system is no longer a single system private to the implementation of the threading backend and used everywhere. A separate task_system can be used (with a specified number of threads) for every simulation.
- arb::execution_context is the interface to task_system  and the previously defined distributed_context
- TBB and serial support has been removed. Cthreads is the only threading backend available. 

Fixes #139.

* Split colours and `pprintf(...)` into `io/pprintf.hpp` header.
* Remove generic `to_string()` function, replacing its very occasional usage with `pprintf`.
* Move block pretty printing into own .cpp file; this is the only place that the vector ostream printer was used.
* Remove `enum_hash`, as not needed with C++14.
* Move `is_in` utility function to `util.hpp`.
* Remove old SIMD printer backend code.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

* Perform device-to-device copy when device_reference is assigned a device_reference.

Better error message on missing file.

* Don't call a missing file an 'internal compiler error'.

noraabiakar authored 6 years ago and Sam Yates committed 6 years ago

Cthreads classes:
- Notification queue : Manages tasks: tries or forces popping and pushing tasks.  
- Task system : manages the notifications queues; controls which queue to pop from/push to; controls spinning on queues if necessary; manages creating/joining threads. Is a singleton.
- Task group : manages synchronization on a group of tasks. 

Operation: 
- Each thread has an associated queue
- Task system _tries to_ push tasks in one of the available queues. If it is unable to acquire a lock on a queue, it tries the next in a round robin fashion. After it loops all queues if it still hasn't successfully pushed the task, it spins on a single queue until lock is acquired and task is pushed. 
- Task system _tries to_ pop a task from the calling thread's queue. If it is unable to acquire the lock, it tries to steal the task from another thread's queue, in a round robin loop. If it is still unable to pop a task, it spins on its the c...

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

All example code and validation tests no longer require access to private include directories. This provides the minimal requirement for an installable target

Note that it is still not possible to separately build mechanisms from NMODL with just the public includes, and there is not yet any package configuration file creation for use with CMake or pkg-config.

* Replace `hw::node_info` with `proc_allocation`, describing local resources for the purposes of domain decomposition.
* Group processor counting and gpu counting implementation under `node_info.cpp`.
* Remove `domain_decomposition` dependency from `cell_group_factory.hpp` so we can use the latter to test for backend support for a cell kind.
* Add `arb::cell_kind_implementation()` which performs the mapping from cell kind and backend kind to a `cell_group_ptr`-producing function (this will then become the site for custom cell group kind mapping support in future work).
* Move headers for aux library ...

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Fixes #526.

Fixes issue #524.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Fixes #182.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Migrate source/build to c++14                                                                                                                    (#522)

* Update `CMakeLists.txt` for C++14 option.
* Update to gcc 6 minimum.
* Update travis CI from gcc-5 to gcc-6
* Use `std::..._t` style type traits, replacing `util::` aliases.
* Use `std::cbegin`, `std::cend`, and `std::make_unique`, replacing `util::` versions.
* Remove `DEDUCED_RETURN_TYPE` macros.
* Remove redundant return type specifications.
* Use correct ADL for `begin` and `end` in (almost all) the range utilities.
* Remove redundant `mechinfo` ctor (aggregate initialization suffices).
* Use lambda capture initializers where appropriate.
* Use generic `std::equal_to`.
* Use variable templates for `math::infinity` and `math::pi`.
* Remove `enum_hash` workaround.
* Use `""s` string literals where we were using our own `""_s` construction.
* Use generic lambda for recursive lambda instead of `std::function` wrapper.
* Use generic lambda for generic arithmetic tests.

Fixes #358.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Who broke the build? Sam did!

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Fixes issue #517.

Deprecate the IBM xlC compiler.
xlC generates code that is an order of a magnitude slower than gcc, while generating spurious warnings, and requiring hacks and workarounds to pass all tests.
Supporting it makes no sense.

* Add test and fatal error for xlC detection in CheckCompilerXLC.cmake.
* Move xlC 13 misdetection work around to CheckCompilerXLC.cmake.
* Remove xlC-specific compatibility workarounds from code.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

This time we're moving `recipe.hpp` and `simulation.hpp`, plus the requirements they bring.

Code changes:
* Pimplize `simulation`.
* Consolidate arbor exceptions: all non-cell kind specific exceptions that might be expected to reach user code now have consistent messages and fit in an exception hierarchy based at `arb::arbor_exception`. Internal errors throw an `arb::arbor_internal_error` exception.
* Renamed `postsynaptic_spike_event` to `spike_event`. (Note: `pse_vector` name is unchanged.)
* Repurposed `pprintf` and moved it into `strprintf.h` — further consolidation is a TODO.
* Made a generic `util::to_string` to avoid redundancy of `operator<<` overloads and other `to_string` definitions. Defaults to ADL `to_string`, `std::to_string`, and finally tries using `operator<<`.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

Further work to public install target.

* Move SIMD classes, cell description classes, simple sampler to public include.
* Rename `cell` to `mc_cell`, `segment` to `mc_segment`, and remove `_description` from cell description class names and includes.
* Move `compartment_model` out of `mc_cell` interface and use only in `fvm_layout.cpp`.
* (Provisionally) remove area/volume methods on `mc_cell` and `mc_segment`.

Sam Yates authored 6 years ago and Benjamin Cumming committed 6 years ago

CMake and build refactoring

*   Use CUDA as first-class language (leading to CMake 3.9 minimum version requirement).

*   Use 'modern CMake' interface libraries for compiler options, include file and library dependency tracking. Interface library targets:
    * `arbor-deps`: compiler options and library requirements for the `libarbor.a` static library, as governed by configure-time options and environment.
    * `arbor-private-headers`: include path for non-installed headers, as required by unit tests and arbor itself.
    * `arbor-aux`: helper classes and utilities used across tests and examples.
    * `ext-json`, `ext-tclap`, `ext-tbb`, `ext-benchmark`, `ext-sphinx_rtd_theme`: externally maintained software that we include (directly or via submodule) in the `ext/` subdirectory.
 
*   Single static library `libarbor.a` includes all built-in modules and CUDA objects.

*   Simply configuration options:
    *  `ARB_WITH_TRACE`, `ARB_AUTORUN_MODCC_ON_CHA...

Add a new cell type, and corresponding cell_group implementation, for benchmarking the simulator library architecture.

Add an benchmark_cell_group, where each cell in the group

generates a spike train prescribed by a time_seq
takes a prescribed time interval per cell to perform the cell_group::advance method.
With this cell type, one can easily build arbitrary networks with prescribed spiking and cell update overheads.
A miniapp that uses this cell type to build a benchmark model is implemented in example/bench.

Fixes #493
Fixes #501

The built in profiler generates timings for state and current for individual multicore mechanisms.

Modcc generates and PE(advance_integrate_{state,current}_X) profiler calls (along with corresponding PL() for calls to multicore mechanism nrn_state and nrn_current API calls.

No timings are made for the gpu back end, which is not properly supported by the current profiling tools.

noraabiakar authored 6 years ago and Benjamin Cumming committed 6 years ago

Changes have been made to the simd implementation of mechansim functions: 

- The node_index array (array of indices that specifies for each mechanism the CVs where it is present), is now partitioned into 4 arrays according to the constraint on each simd_vector in node_index:
    1. contiguous array: contains the indices of all simd_vectors in node_index where the elements in simd_vector are contiguous
    2. constant array: contains the indices of all simd_vectors in node_index where the elements in simd_vector are identical
    3. independent array: contains the indices of all simd_vectors in node_index where the elements in simd_vector are independent (no repetitions) but not contiguous 
    4. none array: contains the indices of all simd_vectors in node_index where the none of the above constraints apply

    When mechanism functions are executed, they loop over each of the 4 arrays separately. This allows for optimizations in every category. 

- The modcc...

Changes to libarbor
-------------------------

Time sequences were added in `src/time_sequence.hpp`:
- added new `time_seq` type that implements a type-erasure interface for the
  concept of a time sequence generator.
- added poisson, regular and vector-backed implementations of the time sequence
  concept.

Event generators:
- The poisson, regular and vector-backed implementations of the event generator
  concept were refactored to use the.

Cell groups:
- Removed the `dss_cell_group` and `rss_cell_group` and associated types.
- Added a generic spike source cell  that generates a sequence of spikes
  at time points specified by a `time_seq`. Using this approach, an
  additional `cell_group` specialization is not required for each type of
  sequence, and user-defined sequences can be used with minimal overhead.

Unit tests
------------

- Added unit tests for `time_seq`.
- Simplified `event_generator` unit tests, because much of the testing
  of the sequences was moved to the `time_seq` tests.
- Added unit tests for `spike_source_cell_group`.

Changes to miniapp
-------------------------

- simplified the miniapp by removing the command line options for using an input spike chain from file.
- updated the miniapp recipe to use `spike_source` cell group instead of `dss_cell_group`.

Benjamin Cumming authored 6 years ago and Sam Yates committed 6 years ago

Move from choosing the distributed communication model from a compile time choice (the old `arb::communication::communication_policy` type) to a run time decision.

* Add `arb::distributed_context` class that provides the required interface for distributed communication implementations, using type-erasure to provide value semantics.
* Add two implementations for the distributed context: `arb::mpi_context` and `arb::local_context`.
* Allow distribution over a user-supplied MPI communicator by providing it as an argument to `arb::mpi_context`.
* Add `mpi_error` exception type to wrap MPI errors.
* Move contents of the `arb::communication` namespace to the `arb` namespace.
* Add preprocessor for-each utility `ARB_PP_FOREACH`.
* Rewrite all examples and tests to use the new distributed context interface.
* Add documentation for distributed context class and semantics, and update documentation for load balancer and simulation classes accordingly.

Fixes #472