rocFFT and hipFFT examples (part II)

[Beanavil]

This pull request contains the second batch of the new rocFFT and hipFFT examples. Added samples:

• rocFFT
  • complex_complex
  • complex_real
  • real_complex
• hipFFT
  • multi_gpu

[evetsso]

I think this would be easier to understand and use if it was either renamed to row_major (since the deffault looks like column-major), or if this was replaced by an enum that spelled out exactly what the caller wanted.

[Beanavil]

Hm yes, this was added to support printing data with either layout. IINM the default is row-major (np[0] is assumed to be the size of the outermost dimension), and when reversing the ordering of the dimension sizes we assume a column-major layout. Changed this parameter to column_major.

[malcolmroberts]

I'm having some trouble getting the example to work with the cuda backend; cmake issues show up.

[malcolmroberts]

cuda compilation doesn't seem to work. CI coverage is probably needed here.

[Beanavil]

cuda compilation doesn't seem to work. CI coverage is probably needed here.

@malcolmroberts what errors are you getting? On our end (internal CI, and I also built it locally to double-check) it builds without problems

[evetsso]

data should be a const reference?

[Beanavil]

Could be, yes. Updated this!

[Beanavil]

@malcolmroberts @evetsso what's the status of the review for this? I think the only issues left to tackle were the ones mentioned in this comment about the CUDA build, but this is not showing up on our end so without further details I cannot move forward on that

[malcolmroberts]

Sorry for the delay - jumping back into this again.

When I configure with
cmake -DCMAKE_CXX_COMPILER=hipcc -DGPU_RUNTIME=CUDA ..
I get errors like

[ 16%] Building CUDA object multi_gpu/CMakeFiles/hipfft_multi_gpu.dir/main.cpp.o
In file included from /home/AMD/marobert/repo/rocm-examples/Libraries/hipFFT/multi_gpu/main.cpp:24:
/home/AMD/marobert/repo/rocm-examples/Libraries/hipFFT/multi_gpu/../../../Common/example_utils.hpp:50:10: fatal error: hip/hip_runtime.h: No such file or directory
   50 | #include <hip/hip_runtime.h>
      |          ^~~~~~~~~~~~~~~~~~~
```

[Beanavil]

@malcolmroberts Oh I see, but IINM that error should be happening for other samples because it seems the HIP headers are not found. I can reproduce this on my side if for instance I rename the /opt/rocm/include/hip folder to something else. Could you verify if you have the hip headers present on your machine?

[malcolmroberts]

Hey, @Beanavil ; I do have /opt/rocm/include/hip populated on my test machine.

However, if there isn't a find_package( hip REQUIRED ) then I wouldn't expect it to include or link the hip runtime. The files are .cpp, so the cmake setup isn't detecting the hip programming language, and I don't think that the runtime is going to be automatically included. I'm configuring/compiling from the Libraries/hipFFT subdirectory.

[evetsso]

@Beanavil Just to be totally clear here, are you also building these examples under CUDA?

[Beanavil]

@evetsso @malcolmroberts sorry for the delayed response, I got totally caught up on other tasks. Yes, we build the examples for CUDA and HIP backend. In fact, we use the dockerfiles provided by this repo for this, so I think if you use a container built from the cuda image the build should be successful.

I'm trying to figure out what could be triggering the failures in your end, I don't think a find_package( hip REQUIRED ) is what's missing here because IINM when we link to hipFFT we are already including the dependency on HIP. On my end, for instance, I can see that when I invoke cmake it does add the compiler option for including the hip headers (-isystem=/opt/rocm/include):

developer@2c1b9a595f95:~/rocm-examples/Libraries/hipFFT/build$ cmake --build . -v
...
[ 16%] Building CUDA object multi_gpu/CMakeFiles/hipfft_multi_gpu.dir/main.cpp.o
cd /home/developer/rocm-examples/Libraries/hipFFT/build/multi_gpu && /usr/local/cuda/bin/nvcc 
-forward-unknown-to-host-compiler  -I/home/developer/rocm-examples/Libraries/hipFFT/multi_gpu/../../../Common 
+ -isystem=/opt/rocm/include --generate-code=arch=compute_52,code=[compute_52,sm_52] -std=c++17 -MD -MT 
multi_gpu/CMakeFiles/hipfft_multi_gpu.dir/main.cpp.o -MF CMakeFiles/hipfft_multi_gpu.dir/main.cpp.o.d -x cu 
-c /home/developer/rocm-examples/Libraries/hipFFT/multi_gpu/main.cpp -o CMakeFiles/hipfft_multi_gpu.dir/main.cpp.o
...

Could you also show how is the compiler being run on your end? This is done with the -v option that I show in the previous code.