KOKKOS_RESTRICT...

[kyungjoo-kim]

Why do we KOKKOS_RESTRICT is defined for Intel architectures ? NVCC suppports the __restrict__ keyword, too.

[kyungjoo-kim]

@crtrott Can we add the #define KOKKOS_RESTRICT __restrict__ keyword if a compiler does not complain about it ?

@lucbv , put your hands up if u like this !!

[mhoemmen]

We would need to add a CMake test for this, but it's not hard:

INCLUDE(CheckCXXSourceCompiles)

FUNCTION(CHECK_RESTRICT_COMPILES VARNAME)
  SET(SOURCE
  "
#define KOKKOS_RESTRICT_LOCAL __restrict__

double test_restrict (const double* KOKKOS_RESTRICT_LOCAL x, const int N) {
  double sum = 0.0;
  for (int i = 0; i < N; ++i) {
    sum += x[i];
  }
  return sum;
}

int main() {
  constexpr int N = 10;
  const double arr[] = { 1.0, 2.0, 3.0, 4.0, 5.0 };
  const double sum = test_restrict (arr, N);
  return sum == 15.0 ? 0 : -1; 
}
  "
  )
  
  # This is a local variable name.  ${VARNAME} is the output variable.
  CHECK_CXX_SOURCE_COMPILES("${SOURCE}" KOKKOS_ENABLE_RESTRICT_LOCAL)

  IF(KOKKOS_ENABLE_RESTRICT_LOCAL)
    GLOBAL_SET(${VARNAME} TRUE)
    MESSAGE(STATUS "C++ compiler supports __restrict__")
  ELSE()
    GLOBAL_SET(${VARNAME} FALSE)
    MESSAGE(STATUS "C++ compiler does NOT support __restrict__")
  ENDIF()
ENDFUNCTION()

[lucbv]

@kyungjoo-kim thanks for adding this issue, I am in favor of allowing restrict in kokkos/kokkos-kernels. It is often useful for vectorization.
Also it would be nice to have this type of macro being maintained for GNU compiler, not only intel...

[mhoemmen]

@lucbv The above CMake test should work fine for any compiler.

[nmhamster]

You need to check which restrict is supported and if flags are needed to enable it. If memory serves, the compilers don’t all agree on how this is written.

[mhoemmen]

@nmhamster I don't know how to do that in Kokkos' raw Makefiles-based build system, but the above CMake test could be modified easily enough to cover variations on __restrict__.

[kyungjoo-kim]

When I have this kind of problem (cmake can test certain compiler options while makefile cannot do so), I just give a overriding option to users in the makefile configuration. Mostly, when people use makefile, we generally assume that those users know what they are doing.

[ibaned]

This is a duplicate of #1922, I'm going to close this one since it came later but we do want to start defining KOKKOS_RESTRICT for non-Intel compilers. Exactly how we do this may still be up for debate.