FFTE backend

[af-ayala]

Currently, we need to address:

• Modify CMake for correct compilation, using GCC 7 or higher. In case PGI is needed, we need to put that option accordingly.
• Once compiling, use the test3D_CPU_C2C.c and test3D_GPU_C2C.c , for verifying if routines from FFTE are being correctly called. The Error checker can be deactivated at this point (comment that section out).
• If possible, verify correctness by printing out the output of the forward FFTE and compare to the one from FFTW or heFFTe backends.

[af-ayala]

Do we have any update on FFTE backend? @luszczek

[luszczek]

I only had to change this for OpenMPI 3:

diff --git a/benchmarks/CMakeLists.txt b/benchmarks/CMakeLists.txt
index 7aa1d7e..504a9aa 100644
--- a/benchmarks/CMakeLists.txt
+++ b/benchmarks/CMakeLists.txt
@@ -24,7 +24,7 @@ macro(fiber_add_benchmark fiber_benchmark)
         target_link_libraries(${fiber_benchmark} p3dfft.3 stdc++ fftw3 fftw3f)
     endif()
     if (FIBER_ENABLE_FFTE)
-        target_link_libraries(${fiber_benchmark} ffte)
+        target_link_libraries(${fiber_benchmark} fftempi ffte gfortran mpi_mpifh)
     endif()
     if (FIBER_ENABLE_SWFFT)
         target_link_libraries(${fiber_benchmark} swfft)

Then I used this command to configure:
env CC=mpicc LDFLAGS=-L/path/to/ffte cmake -D CMAKE_INSTALL_PREFIX=/path/to/fiber/install -D FIBER_ENABLE_FFTE=ON /path/to/fiber

Then simple make worked as expected.

PGI is only needed when you use old CUDA Fortran. Now you can just use NVIDIA's NVHPC SDK that has CUDA Fortran bundled with other tools.