In this pull request, I changed the following things.
Entirely remove CodeBuilder. Originally, the CodeBuilder module contains two parts, which are the imperative DSL and some properties belonging to a stage such as var_dict and axis_list. Now all the imperative DSL is merged with dsl.py. In addition, a new module Stage is created for those properties. Module Stage can be found in schedule.py.
For all compute APIs (e.g., compute, update, etc.), it either returns a new tensor or None. Each compute API itself is a stage built by the Stage module, which creates an ExternOp object. From now on, it should be very clear that a Tensor is different from a Stage.
Users should use hcl.init() before each algorithm definition. To be more specific, users do not need to use hcl.init() again if they are building different schedules for the same algorithm. However, if they want to run several algorithms sequentially, they need to call hcl.init(). With this new command, now HeteroCL automatically detects the final stage. Namely, users do not need to specify that during create_schedule and related APIs.
Simplify tvm.Tensor, which becomes tvm._Tensor after this pull request. Now it is just a module for C/Python interface. The real computations happen in heterocl.Tensor.
In this pull request, I changed the following things.
Entirely remove
CodeBuilder. Originally, theCodeBuildermodule contains two parts, which are the imperative DSL and some properties belonging to a stage such asvar_dictandaxis_list. Now all the imperative DSL is merged withdsl.py. In addition, a new moduleStageis created for those properties. ModuleStagecan be found inschedule.py.For all compute APIs (e.g.,
compute,update, etc.), it either returns a new tensor orNone. Each compute API itself is a stage built by theStagemodule, which creates anExternOpobject. From now on, it should be very clear that aTensoris different from aStage.Users should use
hcl.init()before each algorithm definition. To be more specific, users do not need to usehcl.init()again if they are building different schedules for the same algorithm. However, if they want to run several algorithms sequentially, they need to callhcl.init(). With this new command, now HeteroCL automatically detects the final stage. Namely, users do not need to specify that duringcreate_scheduleand related APIs.Simplify
tvm.Tensor, which becomestvm._Tensorafter this pull request. Now it is just a module for C/Python interface. The real computations happen inheterocl.Tensor.