Question about seq_exec compiler optimizations

[yencal]

Greetings, could someone please explain what sort of compiler optimization happens behind the scenes when one uses seq_exec?
Because for a simple 3d finite difference nested loop, I see that the RAJA version is about four times faster than the C code.

Here is the C code. It does not show the body of the loop for brevity, but it is the same as the RAJA version:

  for (int k = 0; k < nz; ++k ) {
    for (int j = 0; j < ny; ++j ) {
      for (int i = 0; i < nx; ++i ) {
         A[i + nx * (j + ny * k)] = ...

and RAJA version of the same loop as such

  using EXEC_POLICY_3D =
    RAJA::KernelPolicy<
      RAJA::statement::For<2, RAJA::seq_exec,      // k
        RAJA::statement::For<1, RAJA::seq_exec,    // j
          RAJA::statement::For<0, RAJA::seq_exec,  // i
            RAJA::statement::Lambda<0>
          >
        >
      >
    >;
  RAJA::kernel<EXEC_POLICY_3D>(
    RAJA::make_tuple( RAJA::TypedRangeSegment<int>(0, nz),
                      RAJA::TypedRangeSegment<int>(0, ny),
                      RAJA::TypedRangeSegment<int>(0, nx) ),

    [=] RAJA_DEVICE ( int k, int j, int i) {
        A[i + nx * (j + ny * k)] = ...

Note that I use the same compiler flags for both codes:

set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17 -Ofast -march=native")

And I am using M3 MacBook.
Thanks

[rhornung67]

I believe that the compiler sees essentially the same source code whether written as a C-style for-loop or using a RAJA kernel exec method with a seq_exec policy. https://github.com/LLNL/RAJA/blob/develop/include/RAJA/policy/sequential/forall.hpp#L65

That is, there are no pragmas or other annotations applied in RAJA internals. That said, we often observe cases where RAJA code runs faster than native C-style code, but it is not clear why. However, 4x faster seems extraordinary. Have you compared the assembly code for the two versions?

[yencal]

Ok, I will check the assembly code and ensure the cmake flags are propagated accordingly. Thanks