Fix: aligned\_alloc not declared when building PyTorch on Termux (Android, clang)

[Manamama]

Related: #5.

I have a deja vu that I had solved it some month ago but I forgot how, maybe via some cmake switches.

Symptom:

... /data/data/com.termux/files/home/downloads/pytorch/torch/csrc/inductor/aoti_torch/generated/c_shim_cuda.cpp                      [ 79%] Building CXX object caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/native/mkl/SpectralOps.cpp.o                            In file included from /data/data/com.termux/files/home/downloads/pytorch/aten/src/ATen/native/mkl/SpectralOps.cpp:207:              /data/data/com.termux/files/home/downloads/pytorch/third_party/pocketfft/pocketfft_hdronly.h:159:15: error: no member named 'aligned_alloc' in the global namespace; did you mean simply 'aligned_alloc'?                                                                 159 |   void *ptr = ::aligned_alloc(align,(size+align-1)&(~(align-1)));                                                                 |               ^~~~~~~~~~~~~~~
      |               aligned_alloc                               /data/data/com.termux/files/home/downloads/pytorch/third_party/pocketfft/pocketfft_hdronly.h:156:14: note: 'aligned_alloc' declared here
  156 | inline void *aligned_alloc(size_t align, size_t size)           |              ^
1 error generated.                                                 

Ran into the same issue as in the older ticket most likely, but while building PyTorch on Termux (Android + clang/gcc), where ::aligned_alloc is technically C++17 but not provided by Android libc.

I vaguely remember that a version of this worked: sed -i 's/::aligned_alloc/aligned_alloc/g' third_party/pocketfft/pocketfft_hdronly.h

but not now: circular definition warning, etc.

More context: appears on pure Termux when compiling PyTorch, via cmake .. \ -DCMAKE_BUILD_TYPE=Release \ -DINTERN_BUILD_MOBILE=ON \ -DUSE_CUDA=OFF \ -DUSE_ROCM=OFF \ -DUSE_VULKAN=OFF \ -DUSE_XPU=OFF \ -DBUILD_PYTHON=OFF \ -DBUILD_TEST=OFF \ -DBUILD_BINARY=OFF \ -DBUILD_SHARED_LIBS=ON \ -DUSE_FBGEMM=OFF \ -DUSE_OPENMP=ON \ -DUSE_MKLDNN=OFF \ -DUSE_MKL=OFF \ -DUSE_NUMPY=ON

Ver. 1.4

[Manamama]

Update: After reading: numpy/numpy#25940, etc, I tried:
sed -i 's/#if (__cplusplus >= 201703L) && (!defined(__MINGW32__)) && (!defined(_MSC_VER))/#if (__cplusplus >= 201703L) && (!defined(__MINGW32__)) && (!defined(_MSC_VER)) && (!defined(__ANDROID__))/g' /data/data/com.termux/files/home/downloads/pytorch/third_party/pocketfft/pocketfft_hdronly.h
Replacement: Adds && (!defined(ANDROID)) to the condition, for Termux.

[mreineck]

Thanks for the report! I'll get this as soon as possible.
(In my other codes I have now completely switched to the homegrown aligned allocator, since aligned_alloc is just missing on too many platforms...)

[Manamama]

Thanks! FYI, this sed worked to overcome it: the fix is banal : "Add also Android" (Termux mostly).

[mreineck]

I just looked through the current HEAD again, and there is no occurrence of ::aligned_alloc in pocketfft sources (except in a comment and a location disabled by #if 0). I had actually unconditionally disabled it.
Are you perhaps using an older version of the sources?

[mreineck]

Ah, it looks like Pytorch is using an outdated version of pocketfft ... you may have to prod them to update.

[Manamama]

Hm. A quick sanity check:

~/.../third_party/test1 $ git clone https://github.com/mreineck/pocketfft/
Cloning into 'pocketfft'...
remote: Enumerating objects: 595, done.
remote: Counting objects: 100% (205/205), done.
remote: Compressing objects: 100% (45/45), done.
remote: Total 595 (delta 176), reused 171 (delta 160), pack-reused 390 (from 1)
Receiving objects: 100% (595/595), 230.97 KiB | 209.00 KiB/s, done.
Resolving deltas: 100% (379/379), done.
~/.../third_party/test1 $ ls
 pocketfft
~/.../third_party/test1 $ cd pocketfft/
~/.../test1/pocketfft $ ls
 LICENSE.md   pocketfft_demo.cc   pocketfft_hdronly.h   README.md
~/.../test1/pocketfft $ grep aligned_alloc *                          pocketfft_hdronly.h:// std::aligned_alloc is a bit cursed ... it doesn't exist on MacOS < 10.15
pocketfft_hdronly.h:inline void *aligned_alloc(size_t align, size_t size)
pocketfft_hdronly.h:  // aligned_alloc() requires that the requested size is a multiple of "align"
pocketfft_hdronly.h:  void *ptr = ::aligned_alloc(align,(size+align-1)&(~(align-1)));
pocketfft_hdronly.h:inline void *aligned_alloc(size_t align, size_t size)
pocketfft_hdronly.h:      void *ptr = aligned_alloc(64, num*sizeof(T));
pocketfft_hdronly.h:template <typename T> struct aligned_allocator
pocketfft_hdronly.h:  aligned_allocator(const aligned_allocator<U>&) {}
pocketfft_hdronly.h:  aligned_allocator() = default;
pocketfft_hdronly.h:    void* mem = aligned_alloc(alignof(T), n*sizeof(T));
pocketfft_hdronly.h:    std::vector<worker, aligned_allocator<worker>> workers_;
~/.../test1/pocketfft $ time clang++ -o pocketfft_demo pocketfft_demo.cc -I.

real    0m9.609s
user    0m8.480s
sys     0m0.563s
~/.../test1/pocketfft $

I see ::aligned_alloc in code. But then it compiles and runs somehow, if on its own:

~/.../test1/pocketfft $ ./pocketfft_demo
0
3.21314e-08
2.47097e-08
4.66234e-08                                                           3.86054e-08
6.61671e-08                                                           4.30934e-08
6.67378e-08
9.60117e-08
4.47281e-08
6.50549e-08                                                           5.08413e-08                                                           6.84954e-08
6.77912e-08
6.74436e-08
6.63298e-08                                                           8.86252e-08
8.44595e-08
7.88568e-08
8.57887e-08
8.00278e-08
6.27916e-08
7.45258e-08
6.83147e-08
7.86977e-08

FYI:

clang version 20.1.5
Target: aarch64-unknown-linux-android24
Thread model: posix
InstalledDir: /data/data/com.termux/files/usr/bin
Found candidate GCC installation: /data/data/com.termux/files/usr/lib/gcc/aarch64-linux-android/12.3.0
Found candidate GCC installation: /data/data/com.termux/files/usr/lib/gcc/aarch64-linux-android/6.5.0
Found candidate GCC installation: /data/data/com.termux/files/usr/lib/gcc/aarch64-linux-android/8.3.0
Selected GCC installation: /data/data/com.termux/files/usr/lib/gcc/aarch64-linux-android/12.3.0
Candidate multilib: .;@m64
Selected multilib: .;@m64

Mystery. Maybe PyTorch guys mess up smth adding e.g. wrong logic after DINTERN_BUILD_MOBILE=ON ? Me no expert, just thinking aloud...

[mreineck]

It's correct, there is still a call to aligned_alloc in the sources. But it is within an #if 0 ... #endif pair, so the compiler does not see it.
PyTorch has fixed "their" version of pocketft to commit 9d3ab05 (see https://github.com/pytorch/pytorch/tree/main/third_party), which still has the active ::aligned_alloc. They can change the particular commit they want, but it's not something I can do on my side.

[Manamama]

Yes, seems to have worked asap: pytorch/pytorch#154845 .