Previously, the initiators were implicitly coupled to the responses of
the memory, calculating each new initiator clock relatively based on the
time of the BEGIN_RESP phase. This lead to an implicit coupling as
same rounding error of the initiator clock to the memory clock was
applied each time again.
Now, initiators are in itself self-clocked and only send requests based
on backpressure.
Previously, the subdirectories in which the sub-json files were searched
in were hardcoded. Now, DRAMSys simply searches in the directory of the
base config, making this approach more flexible.
The concept of a resource directory was confusing, error-prone and
was only used to specify the directory of the base config json anyway.
Therefore, remove the concept of the resource directory and use the
parent directory of the base config directly.
When the controller accepts requests in the same clock cycle as it
handles them, undeterministic simulations can occur as the outcome
depends on if the new request is accepted before the controllerMethod
is called or not.
Therefore, a minimum delay of one clock cylce should be used to always
handle request only in the next clock cycle, removing the disambiguity.
- Remove nested minimum required to supress warnings.
- Declare SystemC as system library to supress warnings in headers.
- Add a BUILD_SHARED_LIBS option
- Remove hardcoded STATIC in various add_library calls to honor the
BUILD_SHARED_LIBS option
- Remove _deps/ directory from the build directory in GitLab pipeline
- Remove *.tdb files after test stage in pipeline
- Set Ninja as the default generator for the dev preset and re-enable
colored diagnostics
Remove DRAMSysRecordable.h/cpp as the functionality has been incorporated into
DRAMSys.h/cpp. The databaseRecording config is now completely handled by
DRAMSys itself without needing the user of the library to instanciate DRAMSys
or DRAMSysRecordable depending on this config.
When the generator clock did not match the memory clock,
the generator always created a constant delay to the
next transaction.
This is not correct as due to rounding, the delay should be
one cycle more or less depending on the current simulation time.
