Move to vectorclass v2

CMakeLists.txt

@@ -63,23 +63,38 @@ else()
 endif()
 
 # vectorclass
-set(vectorclass_VERSION "1.30")
-set(vectorclass_SRC_URI "https://www.agner.org/optimize/vectorclass.zip")
+set(vectorclass_VERSION "2.00.01")
+set(vectorclass_SRC_URI "https://github.com/vectorclass/version2/archive/v${vectorclass_VERSION}.tar.gz")
 set(vectorclass_DESTDIR "${CMAKE_BINARY_DIR}/include")
 set(vectorclass_ROOTDIR "${vectorclass_DESTDIR}/vectorclass")
 
 ExternalProject_Add(vectorclass
   URL ${vectorclass_SRC_URI}
-  URL_HASH SHA256=f9cb70a3e865dd019b58f449d11f90147ce8ba5f2c60410389ec0ead92944b97
+  URL_HASH SHA256=73b239876ff3453abbc644a367e7cf2af6738099c60bd4bd32b6985f7ff23aef
   SOURCE_DIR vectorclass
   INSTALL_DIR ${vectorclass_ROOTDIR}
   LOG_DOWNLOAD 1 LOG_BUILD 1 LOG_INSTALL 1
   CONFIGURE_COMMAND ""
-  BUILD_COMMAND unzip -d <SOURCE_DIR> -o <SOURCE_DIR>/special.zip
+  BUILD_COMMAND ""
   INSTALL_COMMAND ${CMAKE_COMMAND} -E copy_directory <SOURCE_DIR> <INSTALL_DIR>
   STEP_TARGETS install
 )
 
+set(vectorclass_addon_TARBALL "20415ff928ef0586f1574d8598bf4a52aed8d706.tar.gz")
+ExternalProject_Add(vectorclass_addon
+  URL "https://github.com/vectorclass/add-on/archive/${vectorclass_addon_TARBALL}"
+  URL_HASH SHA256=007e9e99b4561d511bf8a77ea3f350b251d7af00a2fb92fafa6fc402b7e5bf06
+  DOWNLOAD_NAME "vectorclass_addon.tar.gz"
+  DOWNLOAD_NO_EXTRACT TRUE
+  INSTALL_DIR ${vectorclass_ROOTDIR}
+  LOG_DOWNLOAD 1 LOG_BUILD 1 LOG_INSTALL 1
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND ""
+  INSTALL_COMMAND tar -C <INSTALL_DIR> -xvf <DOWNLOADED_FILE> --strip-components 1
+  STEP_TARGETS install
+)
+ExternalProject_Add_StepDependencies(vectorclass_addon configure vectorclass-install)
+
 # FIXME: This is a workaround to let ROOT find the headers at runtime if
 # they are in the build directory. This is necessary until we decide how to
 # treat externals with headers used by ROOT
@@ -92,8 +107,6 @@ if(NOT EXISTS ${CMAKE_BINARY_DIR}/include/Vc)
 endif()
 # end of workaround
 
-# install(DIRECTORY ${Vc_DESTDIR}/ DESTINATION ".")
-
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake")
 
 set(CMAKE_CXX_STANDARD 17)
@@ -138,7 +151,7 @@ if (USE_AVX2)
 target_link_libraries(generate PUBLIC ${Vc_LIBRARIES})
 target_include_directories(generate
     PUBLIC ${CMAKE_BINARY_DIR}/include/vectorclass)
-add_dependencies(generate vectorclass-install)
+add_dependencies(generate vectorclass_addon-install)
 target_compile_definitions(generate PUBLIC "-DVc_IMPL=AVX2")
 if (Vc_BUILTIN)
   add_dependencies(generate VC-install)

src/generate/generate_avx2.cpp

@@ -18,9 +18,8 @@
 #include <functional>
 
 #include <Vc/Vc>
-#include <instrset.h>
-#include <vectori256.h>
-#include <ranvec1.h>
+#include <vectorclass.h>
+#include <random/ranvec1.h>
 
 #include <storage.h>
 #include <generate_common.h>
@@ -45,17 +44,17 @@ float GenAVX2::dot_product(const Vc::SimdArray<float, 3>& left, const Vc::SimdAr
 
 inline int_v scan_AVX(int_v x) {
   // first shift then add
-  auto t0 = permute8i<3, 0, 1, 2, 7, 4, 5, 6>(x.data());
-  auto t1 = permute8i<-1, -1, -1, -1, 0, 1, 2, 3>(t0);
+  auto t0 = permute8<3, 0, 1, 2, 7, 4, 5, 6>(Vec8i{x.data()});
+  auto t1 = permute8<-1, -1, -1, -1, 0, 1, 2, 3>(t0);
   x += _mm256_blend_epi32(t0, t1, 0x11);
 
   // second shift then add
-  t0 = permute8i<2, 3, 0, 1, 6, 7, 4, 5>(x.data());
-  t1 = permute8i<-1, -1, -1, -1, 0, 1, 2, 3>(t0);
+  t0 = permute8<2, 3, 0, 1, 6, 7, 4, 5>(Vec8i{x.data()});
+  t1 = permute8<-1, -1, -1, -1, 0, 1, 2, 3>(t0);
   x += _mm256_blend_epi32(t0, t1, 0x33);
 
   // final shift and add
-  x += int_v{permute8i<-1, -1, -1, -1, 0, 1, 2, 3>(x.data())};
+  x += int_v{permute8<-1, -1, -1, -1, 0, 1, 2, 3>(Vec8i{x.data()})};
   return x;
 }
 
@@ -148,7 +147,7 @@ std::tuple<storage_t, storage_t> generate_avx2(const long time_start, const long
           last = second[long_v::size() - 1];
 
           //broadcast last element
-          offset.data() = permute4q<3, 3, 3, 3>(second.data());
+          offset.data() = permute4<3, 3, 3, 3>(Vec4q{second.data()});
 
           // Generate ToT as a gauss and pmt flat.
           // Only do it every other pass to make use of the double

src/generate/vectorclass_ranvec1.cpp

-#include <ranvec1.cpp>
+#include <random/ranvec1.cpp>