Merge branch 'add_monopole_work_aggregation' of github.com:STEllAR-GR…

…OUP/octotiger into add_monopole_work_aggregation

CMakeLists.txt

@@ -85,6 +85,7 @@ set(OCTOTIGER_CUDA_ARCH "sm_50"
 
 set(OCTOTIGER_KOKKOS_SIMD_LIBRARY "KOKKOS" CACHE STRING "Choose either KOKKOS (for kokkos simd types) or STD (for std::experimental::simd.") 
 set(OCTOTIGER_KOKKOS_SIMD_EXTENSION "DISCOVER" CACHE STRING "Choose either AVX512 AVX2 VSX NEON SVE SCALAR or DISCOVER (automatic discovery)") 
+set(OCTOTIGER_SVE_LEN "512" CACHE STRING "Choose number of bits for SVE (must be 128, 256 or 512)") 
 
 # silence warnings for deprecated HPX includes
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-cpp")
@@ -166,6 +167,7 @@ FetchContent_Declare(
   GIT_TAG        c845b7d9d8da461457cc5aa88fdda5923773a4bf
 )
 FetchContent_MakeAvailable(kokkossimd)
+set(SVE_LENGTH ${OCTOTIGER_SVE_LEN})
 FetchContent_Declare(
   svesimd
   GIT_REPOSITORY https://github.com/srinivasyadav18/sve.git
@@ -825,10 +827,14 @@ elseif(${OCTOTIGER_KOKKOS_SIMD_EXTENSION} STREQUAL SVE)
     if(OCTOTIGER_WITH_CXX17)
       message(FATAL_ERROR "SVE builds with std::experimental::simd require C++20. Please set OCTOTIGER_WITH_CXX20=ON")
     endif()
-    target_compile_definitions(octotiger PUBLIC SVE_LEN=512)
-    target_compile_definitions(optionslib PUBLIC SVE_LEN=512)
-    target_compile_definitions(octolib PUBLIC SVE_LEN=512)
-    target_compile_definitions(hydrolib PUBLIC SVE_LEN=512)
+    target_compile_definitions(octotiger PUBLIC SVE_LEN=${OCTOTIGER_SVE_LEN})
+    target_compile_definitions(optionslib PUBLIC SVE_LEN=${OCTOTIGER_SVE_LEN})
+    target_compile_definitions(octolib PUBLIC SVE_LEN=${OCTOTIGER_SVE_LEN})
+    target_compile_definitions(hydrolib PUBLIC SVE_LEN=${OCTOTIGER_SVE_LEN})
+    target_compile_options(hydrolib PUBLIC "-msve-vector-bits=${OCTOTIGER_SVE_LEN}")
+    target_compile_options(octolib PUBLIC "-msve-vector-bits=${OCTOTIGER_SVE_LEN}")
+    target_compile_options(optionslib PUBLIC "-msve-vector-bits=${OCTOTIGER_SVE_LEN}")
+    target_compile_options(octotiger PUBLIC "-msve-vector-bits=${OCTOTIGER_SVE_LEN}")
   endif()
 elseif(${OCTOTIGER_KOKKOS_SIMD_EXTENSION} STREQUAL SCALAR)
   message(STATUS "KOKKOS SIMD extension manually set to SCALAR (SIMD is OFF)")

octotiger/common_kernel/std_simd.hpp

@@ -51,8 +51,9 @@ static_assert(host_simd_mask_t::size() == 2);
 namespace SIMD_NAMESPACE = sve::experimental::parallelism_v2;
 using host_simd_t = sve::experimental::native_simd<double>;
 using host_simd_mask_t = sve::experimental::native_simd_mask<double>;
-static_assert(host_simd_t::size() == 8);
-static_assert(host_simd_mask_t::size() == 8);
+static_assert(host_simd_t::size() == 8 || host_simd_t::size() == 4 || host_simd_t::size() == 2);
+static_assert(host_simd_mask_t::size() == 8 || host_simd_mask_t::size() == 4 ||
+    host_simd_mask_t::size() == 2);
 #elif defined(OCTOTIGER_KOKKOS_SIMD_VSX)
 #if !(defined(__VSX__)
 #error "VSX Kokkos kernels are specified explicitly but build is without VSX support."