[mlir] [lldb] [libcxxabi] [clang] [compiler-rt] [llvm] [libcxx] Adding Separate OpenMP Offloading Backend to `libcxx/include/__algorithm/pstl_backends` (PR #66968)

[Anton Rydahl via cfe-commits]

================
@@ -466,6 +466,114 @@ Unpoisoning may not be an option, if (for example) you are not maintaining the a
 * You are using allocator, which does not call destructor during deallocation.
 * You are aware that memory allocated with an allocator may be accessed, even when unused by container.
 
+Offloading C++ Parallel Algorithms to GPUs
+------------------------------------------
+
+Experimental support for GPU offloading has been added to ``libc++``. The
+implementation uses OpenMP target offloading to leverage GPU compute resources.
+The OpenMP PSTL backend can target both NVIDIA and AMD GPUs.
+However, the implementation only supports contiguous iterators, such as 
+iterators for ``std::vector`` or ``std::array``.
+To enable the OpenMP offloading backend it must be selected with
+``LIBCXX_PSTL_BACKEND=openmp`` when installing ``libc++``. Further, when
+compiling a program, the user must specify the command line options
+``-fopenmp -fexperimental-library -stdlib=libc++``. To install LLVM with OpenMP
+offloading enabled, please read
+`the LLVM OpenMP FAQ. <https://openmp.llvm.org/SupportAndFAQ.html>`_ 
+You may also want to to visit
+`the OpenMP offloading command-line argument reference. <https://openmp.llvm.org/CommandLineArgumentReference.html#offload-command-line-arguments>`_ 
+
+Example
+~~~~~~~
+
+The following is an example of offloading vector addition to a GPU using our
+standard library extension.
+
+.. code-block:: cpp
+
+  #include <algorithm>
+  #include <execution>
+
+  template<typename T1, typename T2, typename T3>
+  void axpy(const T1 a,std::vector<T2>& x, std::vector<T3>& y)
+  {
+    std::transform(std::execution::par_unseq,x.begin(),x.end(), y.begin(), y.begin(),
+                  [=](T2 xi, T3 yi){ return a*xi + yi; });
+  }
+
+The execution policy ``std::execution::par_unseq`` states that the algorithm's
+execution may be parallelized, vectorized, and migrated across threads. This is
+the only execution mode that is safe to offload to GPUs, and for all other
+execution modes the algorithms will execute on the CPU.
+Special attention must be paid to the lambda captures when enabling GPU
+offloading. If the lambda captures by reference, the user must manually map the
+variables to the device. If capturing by reference, the above example could
+be implemented in the following way.
+
+.. code-block:: cpp
+
+  template<typename T1, typename T2, typename T3>
+  void axpy(const T1 a,std::vector<T2>& x, std::vector<T3>& y)
+  {
+  # pragma omp target data map(to:a)
+    std::transform(std::execution::par_unseq,x.begin(),x.end(), y.begin(), y.begin(),
----------------
AntonRydahl wrote:

I think they are correct, but I will const qualify the x vector. It is reading from x and y and writing to y, just like the classical BLAS AXPY routine. That corresponds to 4) from https://en.cppreference.com/w/cpp/algorithm/transform.

https://github.com/llvm/llvm-project/pull/66968