One executable for all Generation

[mrprajesh]

Typically, our make should result in one executable (say StarPlat compiler) which can generate files for all the architectures. i.e., should output cuda, omp, or mpi files based on the cmd line args. Currently, it is

• make cuda for cuda,
• make omp, similarly to others.

We may have to use namespace and modify our makefile to achieve it. A small prototype here

[mrprajesh]

Let's follow these as our namespace conventions. Prefix sp meaning StarPlat.

• cuda spcuda
• omp spomp
• acc spacc
• mpi spmpi

// header
namespace X
{
  class Foo
  {
  public:
    void TheFunc();
  };
}

// cpp
namespace X
{
  void Foo::TheFunc()
  {
    return;
  }
}

We follow option 1 in Source

More notes

• we should never miss out #ifndef macros in headers.
• To add new backend.
  1. Always create a branch from WorkingBranch
  2. create a new folder as src/backends/backend_<new> folder with .cpp and .h file
  3. create a folder named graphcode/generated_<new>
  4. Append line 4 in Makefile where bin/<new>_dsl_cpp_generator.o replace new with two-letter short code of the new backend.
  5. Add lines the below lines.

bin/<new>_dsl_cpp_generator.o: backends/backend_<new>/dsl_cpp_generator.cpp backends/dslCodePad.h
    $(CC) -c backends/backend_<new>/dsl_cpp_generator.cpp -o bin/<new>_dsl_cpp_generator.o

[rupesh0508]

we should never miss out #ifndef macros in headers.

Or use #pragma once

[mrprajesh]

Thanks, Rupesh. #pragma one has one tiny drawback.
This feature is completed in working branch as of eb34823
except for MPI.

• CUDA, OMP, ACC
• Testing out generation part.
  • OMP
  • CUDA // segfaulting. getUserVar returning variable with no type. v in forall
  • MPI //awaiting TC
  • ACC //asserton fail.
• MPI // We are good to merge once MPI code is frozen.

[rupesh0508]

With the heterogeneous backend Nibedita is working on, we will need all the three backends under the same binary.