vectorized_code.rst

Writing Vectorized Code

Assume that we have a simple function that computes the mean of two vectors, something like:

.. literalinclude:: ../../test/doc/writing_vectorized_code.cpp

How can we use xsimd to take advantage of vectorization?

Explicit Use of an Instruction Set

xsimd provides the template class :cpp:class:`xsimd::batch` parametrized by T and A types where T is the type of the values involved in SIMD instructions and A is the target architecture. If you know which instruction set is available on your machine, you can directly use the corresponding specialization of batch. For instance, assuming the AVX instruction set is available, the previous code can be vectorized the following way:

.. literalinclude:: ../../test/doc/explicit_use_of_an_instruction_set_mean.cpp

Note that the code is written in a form that's independent from the actual vector register width.

However, if you want to write code that is portable, you cannot rely on the use of batch<double, xsimd::avx>. Indeed this won't compile on a CPU where only SSE2 instruction set is available for instance. Fortunately, if you don't set the second template parameter, xsimd picks the best architecture among the one available, based on the compiler flag you use.

Aligned vs Unaligned Memory

In the previous example, you may have noticed the :cpp:func:`xsimd::batch::load_unaligned` and :cpp:func:`xsimd::batch::store_unaligned` functions. These are meant for loading values from contiguous dynamically allocated memory into SIMD registers and reciprocally. When dealing with memory transfer operations, some instructions sets required the memory to be aligned by a given amount, others can handle both aligned and unaligned modes. In that latter case, operating on aligned memory is generally faster than operating on unaligned memory.

xsimd provides an aligned memory allocator, namely :cpp:class:`xsimd::aligned_allocator` which follows the standard requirements, so it can be used with STL containers. Let's change the previous code so it can take advantage of this allocator:

.. literalinclude:: ../../test/doc/explicit_use_of_an_instruction_set_mean_aligned.cpp

Memory Alignment and Tag Dispatching

You may need to write code that can operate on any type of vectors or arrays, not only the STL ones. In that case, you cannot make assumption on the memory alignment of the container. xsimd provides a tag dispatching mechanism that allows you to easily write such a generic code:

.. literalinclude:: ../../test/doc/explicit_use_of_an_instruction_set_mean_tag_dispatch.cpp

Here, the Tag template parameter can be :cpp:class:`xsimd::aligned_mode` or :cpp:class:`xsimd::unaligned_mode`. Assuming the existence of a get_alignment_tag meta-function in the code, the previous code can be invoked this way:

mean(a, b, res, get_alignment_tag<decltype(a)>());

Writing Arch-Independent Code

If your code may target either SSE2, AVX2 or AVX512 instruction set, xsimd make it possible to make your code even more generic by using the architecture as a template parameter:

.. literalinclude:: ../../test/doc/explicit_use_of_an_instruction_set_mean_arch_independent.cpp

Then you just need to #include that file, force instantiation for a specific architecture and pass the appropriate flag to the compiler. For instance:

.. literalinclude:: ../../test/doc/sum_sse2.cpp

This can be useful to implement runtime dispatching, based on the instruction set detected at runtime. xsimd provides a generic machinery :cpp:func:`xsimd::dispatch()` to implement this pattern. Based on the above example, instead of calling mean{}(arch, a, b, res, tag), one can use xsimd::dispatch(mean{})(a, b, res, tag). More about this can be found in the :ref:`Arch Dispatching` section.

Breaking out of xsimd

Sometimes xsimd may not give you the whole availability of the instruction set you are targeting. This is not specific to this library but to all libraries that abstract something.

xsimd give you the possibility to break out of its :cpp:class:`~xsimd::batch` class and getting the underlying intrinsic register type. This is useful when an instruction set expose some intrinsicts for applications (e.g. video processing, cryptography...) that are too specific to include in xsimd, or when some instructions are currently missing.

There are many ways a user could add a special cases in their arch-independent SIMD code. One that is simple and compiles on all platforms is using C++17 if constexpr.

template<typename Arch>
auto sign_i8(
    xsimd::batch<int8_t, Arch> const& x,
    xsimd::batch<int8_t, Arch> const& y
) -> xsimd::batch<uint8_t, Arch> {
     // Dedicated instruction dispatch at compile time
     if constexpr(std::is_same_v<Arch, xsimd::avx2>){
         // Automatic conversion back and forth between xsimd::batch and native types
         return _mm256_sign_epi8(x, y);
         // When compiler complains we can be more explicit
         return xsimd::batch<int8_t, Arch>(_mm256_sign_epi8(x.to_native(), y.to_native()));
     }

     // General xsimd implementation
     auto const zero = xsimd::batch<uint8_t, Arch>(0);
     auto const c = xsimd::select(b < 0, -a, a);
     return xsimd::select(b == zero, zero, c);
}