Support avx10.1-512 variant

It seems there's a new variant of AVX 512. Clang has a flag to enable
it, so we may as well try it.

Author: mhucka

CMakeLists.txt

@@ -68,6 +68,7 @@ if(WIN32)
 else()
     check_cxx_compiler_flag("-mavx2" HAVE_AVX2)
     check_cxx_compiler_flag("-mavx512f" HAVE_AVX512)
+    check_cxx_compiler_flag("-mavx10.1-512" HAVE_AVX512_10)
 endif()
 
 
@@ -107,7 +108,7 @@ if(NOT CMAKE_APPLE_SILICON_PROCESSOR)
       add_subdirectory(pybind_interface/avx2)
     endif()
 
-    if(HAVE_AVX512)
+    if(HAVE_AVX512 OR HAVE_AVX512_10)
       add_subdirectory(pybind_interface/avx512)
     endif()
 endif()
@@ -133,6 +134,7 @@ endif()
 message(STATUS "${MSG_PREFIX} host has SSE2 = ${HAVE_SSE2}")
 message(STATUS "${MSG_PREFIX} host has AVX2 = ${HAVE_AVX2}")
 message(STATUS "${MSG_PREFIX} host has AVX512 = ${HAVE_AVX512}")
+message(STATUS "${MSG_PREFIX} host has AVX10.1-512 = ${HAVE_AVX512_10}")
 
 message(STATUS "${MSG_PREFIX} shell $PATH = $ENV{PATH}")
 message(STATUS "${MSG_PREFIX} shell $CUQUANTUM_ROOT = $ENV{CUQUANTUM_ROOT}")

pybind_interface/avx512/CMakeLists.txt

@@ -17,6 +17,8 @@ project(qsim)
 
 IF (WIN32)
     add_compile_options(/arch:AVX512)
+ELSEIF(HAVE_AVX512_10)
+    add_compile_options(-mavx10.1-512 -mbmi2)
 ELSE()
     add_compile_options(-mavx512f -mbmi2)
 ENDIF()

pybind_interface/basic/CMakeLists.txt

@@ -15,6 +15,10 @@
 cmake_minimum_required(VERSION 3.31)
 project(qsim)
 
+if(NOT WIN32)
+    add_compile_options(-march=native)
+endif()
+
 INCLUDE(../GetPybind11.cmake)
 pybind11_add_module(qsim_basic pybind_main_basic.cpp)