SVE backend

CMakeLists.txt

@@ -287,13 +287,29 @@ set(arbor_supported_components)
 # Target microarchitecture for building arbor libraries, tests and examples
 #---------------------------------------------------------------------------
 
+set(ARB_SVE_WIDTH "auto" CACHE STRING "Default SVE vector length in bits. Default: auto (detection during configure time).")
+mark_as_advanced(ARB_SVE_WIDTH)
+if (ARB_SVE_WIDTH STREQUAL "auto")
+    get_sve_length(ARB_HAS_SVE ARB_SVE_BITS)
+    if (ARB_HAS_SVE)
+        message(STATUS "SVE detected with vector size = ${ARB_SVE_BITS} bits")
+        set(ARB_CXX_SVE_FLAGS " -msve-vector-bits=${ARB_SVE_BITS}")
+    else()
+        message(STATUS "NO SVE detected")
+        set(ARB_CXX_SVE_FLAGS "")
+    endif()
+else()
+    set(ARB_SVE_BITS ${ARB_SVE_WIDTH})
+    set(ARB_CXX_SVE_FLAGS " -msve-vector-bits=${ARB_SVE_BITS}")
+endif()
+
 # Set the full set of target flags in ARB_CXX_FLAGS_TARGET_FULL, which
 # will include target-specific -march flags if ARB_ARCH is not "none".
 if(ARB_ARCH STREQUAL "none")
-    set(ARB_CXX_FLAGS_TARGET_FULL ${ARB_CXX_FLAGS_TARGET})
+    set(ARB_CXX_FLAGS_TARGET_FULL ${ARB_CXX_FLAGS_TARGET} ${ARB_CXX_SVE_FLAGS})
 else()
     set_arch_target(ARB_CXXOPT_ARCH ${ARB_ARCH})
-    set(ARB_CXX_FLAGS_TARGET_FULL ${ARB_CXX_FLAGS_TARGET} ${ARB_CXXOPT_ARCH})
+    set(ARB_CXX_FLAGS_TARGET_FULL ${ARB_CXX_FLAGS_TARGET} ${ARB_CXXOPT_ARCH} ${ARB_CXX_SVE_FLAGS})
 endif()
 
 # Compile with `-fvisibility=hidden` to ensure that the symbols of the generated

arbor/CMakeLists.txt

@@ -147,4 +147,20 @@ export_visibility(arbor)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/include/arbor/export.hpp
     DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/arbor)
 
+configure_file(
+    ${PROJECT_SOURCE_DIR}/cmake/vls_sve_bits.hpp.in
+    ${CMAKE_CURRENT_BINARY_DIR}/include/arbor/simd/vls_sve_bits.hpp
+    @ONLY)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/include/arbor/simd/vls_sve_bits.hpp
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/arbor/simd)
+
+configure_file(
+    ${PROJECT_SOURCE_DIR}/cmake/sve_bits.hpp.in
+    ${CMAKE_CURRENT_BINARY_DIR}/include/arbor/simd/sve_bits.hpp
+    @ONLY)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/include/arbor/simd/sve_bits.hpp
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/arbor/simd)
+
 install(TARGETS arbor EXPORT arbor-targets ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR})

arbor/include/arbor/simd/native.hpp

@@ -68,8 +68,9 @@ ARB_DEF_NATIVE_SIMD_(double, 8, avx512)
 #if defined(__ARM_FEATURE_SVE)
 
 #include "sve.hpp"
-ARB_DEF_NATIVE_SIMD_(int, 0, sve)
-ARB_DEF_NATIVE_SIMD_(double, 0, sve)
+#include "vls_sve.hpp"
+ARB_DEF_NATIVE_SIMD_(int, detail::vls_sve_width, vls_sve)
+ARB_DEF_NATIVE_SIMD_(double, detail::vls_sve_width, vls_sve)
 
 #elif defined(__ARM_NEON)
 

arbor/include/arbor/simd/simd.hpp

@@ -395,7 +395,7 @@ namespace detail {
         template <typename V>
         friend class indirect_expression;
 
-        simd_impl() = default;
+        constexpr simd_impl() = default;
 
         // Construct by filling with scalar value.
         simd_impl(const scalar_type& x) {
@@ -751,7 +751,7 @@ namespace detail {
         using base::width;
         using base::value_;
 
-        simd_mask_impl() = default;
+        constexpr simd_mask_impl() = default;
 
         // Construct by filling with scalar value.
         simd_mask_impl(bool b) {

arbor/include/arbor/simd/sve.hpp

@@ -9,6 +9,7 @@
 #include <cstdint>
 #include <iostream>
 
+#include <arbor/simd/sve_bits.hpp>
 #include <arbor/util/pp_util.hpp>
 
 #include "approx.hpp"
@@ -17,6 +18,9 @@ namespace arb {
 namespace simd {
 namespace detail {
 
+// number of elements in a vector
+static constexpr unsigned max_sve_width = sve_bits/64;
+
 struct sve_double;
 struct sve_int;
 struct sve_mask;
@@ -33,7 +37,7 @@ template<> struct is_sve<svbool_t>    : std::true_type {};
 
 template <>
 struct simd_traits<sve_mask> {
-    static constexpr unsigned width = 8;
+    static constexpr unsigned width = max_sve_width;
     using scalar_type = bool;
     using vector_type = svbool_t;
     using mask_impl = sve_mask;
@@ -43,7 +47,7 @@ struct simd_traits<sve_mask> {
 
 template <>
 struct simd_traits<sve_double> {
-    static constexpr unsigned width = 8;
+    static constexpr unsigned width = max_sve_width;
     using scalar_type = double;
     using vector_type = svfloat64_t;
     using mask_impl = sve_mask;
@@ -53,7 +57,7 @@ struct simd_traits<sve_double> {
 
 template <>
 struct simd_traits<sve_int> {
-    static constexpr unsigned width = 8;
+    static constexpr unsigned width = max_sve_width;
     using scalar_type = int32_t;
     using vector_type = svint64_t;
     using mask_impl = sve_mask;
@@ -453,7 +457,7 @@ struct sve_double {
 
         // Compute n and g.
 
-        auto n = svrintz_f64_z(svptrue_b64(), add(mul(broadcast(ln2inv), x), broadcast(0.5)));
+        auto n = svrintm_f64_z(svptrue_b64(), add(mul(broadcast(ln2inv), x), broadcast(0.5)));
 
         auto g = fma(n, broadcast(-ln2C1), x);
         g = fma(n, broadcast(-ln2C2), g);