From b4bc7bfd8187382bff1903fb5c2bd08e352c1570 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 08:28:37 -0800
Subject: [PATCH 01/21] Compile for M1

---
 CMakeLists.txt                         | 48 ++++++++-----
 pybind_interface/decide/CMakeLists.txt | 94 ++++++++++++++++----------
 pybind_interface/decide/decide.cpp     | 14 ++--
 3 files changed, 98 insertions(+), 58 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index f902a421..d4574b73 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,24 +1,38 @@
 cmake_minimum_required(VERSION 3.11)
 
-execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
-if(has_nvcc STREQUAL "")
-    execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
-    if(has_hipcc STREQUAL "")
-        project(qsim)
-    else()
-        project(qsim LANGUAGES CXX HIP)
-        ADD_SUBDIRECTORY(pybind_interface/hip)
-    endif()
+# Check if we are on an Apple system and if the architecture is ARM
+if(APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
+    # Set a variable to indicate we are on Apple ARM
+    set(APPLE_ARM TRUE)
 else()
-    project(qsim LANGUAGES CXX CUDA)
-    ADD_SUBDIRECTORY(pybind_interface/cuda)
-    if(DEFINED ENV{CUQUANTUM_ROOT})
-        ADD_SUBDIRECTORY(pybind_interface/custatevec)
-    endif()
+    set(APPLE_ARM FALSE)
 endif()
 
-ADD_SUBDIRECTORY(pybind_interface/sse)
-ADD_SUBDIRECTORY(pybind_interface/avx512)
-ADD_SUBDIRECTORY(pybind_interface/avx2)
 ADD_SUBDIRECTORY(pybind_interface/basic)
 ADD_SUBDIRECTORY(pybind_interface/decide)
+
+# You can then use this variable to control conditional compilation
+if(APPLE_ARM)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8-a")
+else()
+    execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
+    if(has_nvcc STREQUAL "")
+        execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
+        if(has_hipcc STREQUAL "")
+            project(qsim)
+        else()
+            project(qsim LANGUAGES CXX HIP)
+            ADD_SUBDIRECTORY(pybind_interface/hip)
+        endif()
+    else()
+        project(qsim LANGUAGES CXX CUDA)
+        ADD_SUBDIRECTORY(pybind_interface/cuda)
+        if(DEFINED ENV{CUQUANTUM_ROOT})
+            ADD_SUBDIRECTORY(pybind_interface/custatevec)
+        endif()
+    endif()
+
+    ADD_SUBDIRECTORY(pybind_interface/sse)
+    ADD_SUBDIRECTORY(pybind_interface/avx512)
+    ADD_SUBDIRECTORY(pybind_interface/avx2)
+endif()
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
index b94b0ade..182e10f7 100644
--- a/pybind_interface/decide/CMakeLists.txt
+++ b/pybind_interface/decide/CMakeLists.txt
@@ -1,15 +1,28 @@
 cmake_minimum_required(VERSION 3.11)
 
-execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
-if(has_nvcc STREQUAL "")
-  execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
-  if(has_hipcc STREQUAL "")
-    project(qsim)
-  else()
-    project(qsim LANGUAGES CXX HIP)
-  endif()
+# Check if we are on an Apple system and if the architecture is ARM
+if(APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
+    # Set a variable to indicate we are on Apple ARM
+    set(APPLE_ARM TRUE)
 else()
-  project(qsim LANGUAGES CXX CUDA)
+    set(APPLE_ARM FALSE)
+endif()
+
+# Set the project based on the detected platform
+if(APPLE_ARM)
+    project(qsim LANGUAGES CXX)
+else()
+    execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
+    if(has_nvcc STREQUAL "")
+        execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
+        if(has_hipcc STREQUAL "")
+            project(qsim)
+        else()
+            project(qsim LANGUAGES CXX HIP)
+        endif()
+    else()
+        project(qsim LANGUAGES CXX CUDA)
+    endif()
 endif()
 
 IF (WIN32)
@@ -26,41 +39,48 @@ endif()
 
 INCLUDE(../GetPybind11.cmake)
 
-if(has_nvcc STREQUAL "")
-  if(has_hipcc STREQUAL "")
+# Configure based on the detected platform
+if(APPLE_ARM)
+    # Configuration for Apple ARM
+    # Set additional flags if necessary
+    # Compile the module using standard C++
     pybind11_add_module(qsim_decide decide.cpp)
-  else()
-    list(APPEND CMAKE_MODULE_PATH "/opt/rocm/lib/cmake/hip")
-    find_package(HIP REQUIRED)
-    find_package(PythonLibs 3.7 REQUIRED)
 
-    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
+elif(has_nvcc STREQUAL "")
+    if(has_hipcc STREQUAL "")
+        pybind11_add_module(qsim_decide decide.cpp)
+    else()
+        list(APPEND CMAKE_MODULE_PATH "/opt/rocm/lib/cmake/hip")
+        find_package(HIP REQUIRED)
+        find_package(PythonLibs 3.7 REQUIRED)
 
-    hip_add_library(qsim_decide MODULE decide.cpp)
+        include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
 
-    set_target_properties(qsim_decide PROPERTIES
-        PREFIX "${PYTHON_MODULE_PREFIX}"
-        SUFFIX "${PYTHON_MODULE_EXTENSION}"
-    )
-    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
-  endif()
+        hip_add_library(qsim_decide MODULE decide.cpp)
+
+        set_target_properties(qsim_decide PROPERTIES
+            PREFIX "${PYTHON_MODULE_PREFIX}"
+            SUFFIX "${PYTHON_MODULE_EXTENSION}"
+        )
+        set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
+    endif()
 else()
-  find_package(PythonLibs 3.7 REQUIRED)
-  find_package(CUDA REQUIRED)
+    find_package(PythonLibs 3.7 REQUIRED)
+    find_package(CUDA REQUIRED)
 
-  include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
+    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
 
-  cuda_add_library(qsim_decide MODULE decide.cpp)
+    cuda_add_library(qsim_decide MODULE decide.cpp)
 
-  if(DEFINED ENV{CUQUANTUM_ROOT})
-      target_compile_options(qsim_decide PRIVATE
-          $<$<COMPILE_LANGUAGE:CUDA>:-D__CUSTATEVEC__>
-      )
-  endif()
+    if(DEFINED ENV{CUQUANTUM_ROOT})
+        target_compile_options(qsim_decide PRIVATE
+            $<$<COMPILE_LANGUAGE:CUDA>:-D__CUSTATEVEC__>
+        )
+    endif()
 
-  set_target_properties(qsim_decide PROPERTIES
-        PREFIX "${PYTHON_MODULE_PREFIX}"
-        SUFFIX "${PYTHON_MODULE_EXTENSION}"
-  )
-  set_source_files_properties(decide.cpp PROPERTIES LANGUAGE CUDA)
+    set_target_properties(qsim_decide PROPERTIES
+          PREFIX "${PYTHON_MODULE_PREFIX}"
+          SUFFIX "${PYTHON_MODULE_EXTENSION}"
+    )
+    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE CUDA)
 endif()
diff --git a/pybind_interface/decide/decide.cpp b/pybind_interface/decide/decide.cpp
index 5e9a63f1..c611ac39 100644
--- a/pybind_interface/decide/decide.cpp
+++ b/pybind_interface/decide/decide.cpp
@@ -34,10 +34,15 @@ enum Instructions { AVX512F = 0, AVX2 = 1, SSE4_1 = 2, BASIC = 3};
 
 int detect_instructions() {
   Instructions instr = BASIC;
-  int info[4];
 
+  // Check for ARM architecture, specifically for Apple M1 or similar
+  #if defined(__aarch64__) && defined(__APPLE__)
+  // On Apple ARM systems, always use BASIC
+  instr = BASIC;
+  #else
+  // Existing x86/x86_64 specific instruction set detection logic
+  int info[4];
   cpuid(info, 0);
-
   int nIds = info[0];
   if (nIds >= 1) {
     cpuid(info, 1);
@@ -47,18 +52,19 @@ int detect_instructions() {
   }
   if (nIds >= 7) {
     cpuid(info, 7);
-    if ((info[1] & (1 <<  5))!= 0) {
+    if ((info[1] & (1 << 5)) != 0) {
       instr = AVX2;
     }
     if ((info[1] & (1 << 16)) != 0) {
       instr = AVX512F;
     }
-
   }
+  #endif
 
   return static_cast<int>(instr);
 }
 
+
 enum GPUCapabilities {
     CUDA = 0, CUSTATEVEC = 1, HIP = 2, NO_GPU = 10, NO_CUSTATEVEC = 11 };
 

From 7e2a8f9681dd94a3f209b23922f613f1a42f7598 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 09:53:10 -0800
Subject: [PATCH 02/21] Make qsimcirq Apple Silicon (M1, M2, ARM64) compatible

---
 CMakeLists.txt                         | 22 ++++++++++++++++------
 Makefile                               |  4 ++--
 pybind_interface/GetPybind11.cmake     |  1 +
 pybind_interface/decide/CMakeLists.txt | 24 ++++++++++++++++--------
 pybind_interface/decide/decide.cpp     |  4 ++--
 setup.py                               |  4 ++--
 6 files changed, 39 insertions(+), 20 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index d4574b73..98aae939 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,25 +1,34 @@
+set(CMAKE_CXX_STANDARD 11)
+set(OpenMP_INCLUDE_DIR "/Users/bas.nijholt/micromamba/lib/clang/17.0.4/include")
+include_directories(${OpenMP_INCLUDE_DIR})
 cmake_minimum_required(VERSION 3.11)
 
-# Check if we are on an Apple system and if the architecture is ARM
-if(APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
-    # Set a variable to indicate we are on Apple ARM
-    set(APPLE_ARM TRUE)
+if(APPLE)
+    # Returns x86_64 or arm64
+    execute_process(COMMAND uname -m OUTPUT_VARIABLE OSX_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if (OSX_ARCH STREQUAL "arm64")
+        set(APPLE_ARM TRUE)
+    else()
+        set(APPLE_ARM FALSE)
+    endif()
 else()
     set(APPLE_ARM FALSE)
-endif()
+endif(APPLE)
+message(STATUS "APPLE_ARM: ${APPLE_ARM}")
 
 ADD_SUBDIRECTORY(pybind_interface/basic)
 ADD_SUBDIRECTORY(pybind_interface/decide)
 
 # You can then use this variable to control conditional compilation
 if(APPLE_ARM)
+    project(qsim LANGUAGES CXX)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8-a")
 else()
     execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
     if(has_nvcc STREQUAL "")
         execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
         if(has_hipcc STREQUAL "")
-            project(qsim)
+            project(qsim LANGUAGES CXX)
         else()
             project(qsim LANGUAGES CXX HIP)
             ADD_SUBDIRECTORY(pybind_interface/hip)
@@ -36,3 +45,4 @@ else()
     ADD_SUBDIRECTORY(pybind_interface/avx512)
     ADD_SUBDIRECTORY(pybind_interface/avx2)
 endif()
+find_package(OpenMP REQUIRED)
diff --git a/Makefile b/Makefile
index 227ca603..7ebbaaba 100644
--- a/Makefile
+++ b/Makefile
@@ -4,11 +4,11 @@ EIGEN_URL = "https://gitlab.com/libeigen/eigen/-/archive/"
 TARGETS = qsim
 TESTS = run-cxx-tests
 
-CXX=g++
+CXX=x86_64-apple-darwin13.4.0-clang++
 NVCC=nvcc
 HIPCC=hipcc
 
-CXXFLAGS = -O3 -fopenmp
+CXXFLAGS = -O3 -fopenmp -I
 ARCHFLAGS = -march=native
 NVCCFLAGS = -O3
 HIPCCFLAGS = -O3
diff --git a/pybind_interface/GetPybind11.cmake b/pybind_interface/GetPybind11.cmake
index e8516a0b..56223801 100644
--- a/pybind_interface/GetPybind11.cmake
+++ b/pybind_interface/GetPybind11.cmake
@@ -1,3 +1,4 @@
+add_compile_options(-I/Users/bas.nijholt/micromamba/envs/py311/include/python3.11 -I/Users/bas.nijholt/micromamba/envs/py311/lib/python3.11/site-packages/pybind11/include)
 include(FetchContent)
 
 set(MIN_PYBIND_VERSION "2.10.1")
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
index 182e10f7..b3b24ca4 100644
--- a/pybind_interface/decide/CMakeLists.txt
+++ b/pybind_interface/decide/CMakeLists.txt
@@ -1,12 +1,19 @@
 cmake_minimum_required(VERSION 3.11)
+set(CMAKE_SYSTEM_PROCESSOR arm64)
 
 # Check if we are on an Apple system and if the architecture is ARM
-if(APPLE AND CMAKE_SYSTEM_PROCESSOR MATCHES "arm64")
-    # Set a variable to indicate we are on Apple ARM
-    set(APPLE_ARM TRUE)
+if(APPLE)
+    # Returns x86_64 or arm64
+    execute_process(COMMAND uname -m OUTPUT_VARIABLE CMAKE_OSX_ARCHITECTURES OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if (OSX_ARCH STREQUAL "arm64")
+        set(APPLE_ARM TRUE)
+    else()
+        set(APPLE_ARM FALSE)
+    endif()
 else()
     set(APPLE_ARM FALSE)
-endif()
+endif(APPLE)
+message(STATUS "APPLE_ARM: ${APPLE_ARM}")
 
 # Set the project based on the detected platform
 if(APPLE_ARM)
@@ -45,14 +52,15 @@ if(APPLE_ARM)
     # Set additional flags if necessary
     # Compile the module using standard C++
     pybind11_add_module(qsim_decide decide.cpp)
-
-elif(has_nvcc STREQUAL "")
+    # Print a message to indicate the module is being compiled using standard C++
+    message(STATUS "Compiling qsim_decide using standard C++ on Apple ARM")
+elseif(has_nvcc STREQUAL "")
     if(has_hipcc STREQUAL "")
         pybind11_add_module(qsim_decide decide.cpp)
     else()
         list(APPEND CMAKE_MODULE_PATH "/opt/rocm/lib/cmake/hip")
         find_package(HIP REQUIRED)
-        find_package(PythonLibs 3.7 REQUIRED)
+        find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
 
         include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
 
@@ -65,7 +73,7 @@ elif(has_nvcc STREQUAL "")
         set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
     endif()
 else()
-    find_package(PythonLibs 3.7 REQUIRED)
+    find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
     find_package(CUDA REQUIRED)
 
     include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
diff --git a/pybind_interface/decide/decide.cpp b/pybind_interface/decide/decide.cpp
index c611ac39..3641833c 100644
--- a/pybind_interface/decide/decide.cpp
+++ b/pybind_interface/decide/decide.cpp
@@ -21,8 +21,8 @@ namespace py = pybind11;
 #include <intrin.h>
 #define cpuid(info, x)    __cpuidex(info, x, 0)
 
-#else
-//  GCC Intrinsics
+#elif defined(__x86_64__) || defined(__i386__)
+// GCC Intrinsics for x86/x86_64
 #include <cpuid.h>
 void cpuid(int info[4], int infoType){
     __cpuid_count(infoType, 0, info[0], info[1], info[2], info[3]);
diff --git a/setup.py b/setup.py
index 593620a6..06b7ba16 100644
--- a/setup.py
+++ b/setup.py
@@ -60,8 +60,8 @@ def build_extension(self, ext):
 
         if platform.system() == "Darwin":
             cmake_args += [
-                "-DCMAKE_C_COMPILER=/usr/local/opt/llvm/bin/clang",
-                "-DCMAKE_CXX_COMPILER=/usr/local/opt/llvm/bin/clang++",
+                "-DCMAKE_C_COMPILER=/Users/bas.nijholt/micromamba/envs/py311/bin/clang",
+                "-DCMAKE_CXX_COMPILER=/Users/bas.nijholt/micromamba/envs/py311/bin/clang++",
             ]
 
         if shutil.which("hipcc") is not None:

From 66949709c6879730481d50c19c7cb8c3d6a5aac5 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 10:01:37 -0800
Subject: [PATCH 03/21] make work

---
 Makefile                               |  4 ++--
 pybind_interface/decide/CMakeLists.txt | 23 +----------------------
 pybind_interface/decide/decide.cpp     |  1 -
 3 files changed, 3 insertions(+), 25 deletions(-)

diff --git a/Makefile b/Makefile
index 7ebbaaba..227ca603 100644
--- a/Makefile
+++ b/Makefile
@@ -4,11 +4,11 @@ EIGEN_URL = "https://gitlab.com/libeigen/eigen/-/archive/"
 TARGETS = qsim
 TESTS = run-cxx-tests
 
-CXX=x86_64-apple-darwin13.4.0-clang++
+CXX=g++
 NVCC=nvcc
 HIPCC=hipcc
 
-CXXFLAGS = -O3 -fopenmp -I
+CXXFLAGS = -O3 -fopenmp
 ARCHFLAGS = -march=native
 NVCCFLAGS = -O3
 HIPCCFLAGS = -O3
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
index b3b24ca4..7724a3e9 100644
--- a/pybind_interface/decide/CMakeLists.txt
+++ b/pybind_interface/decide/CMakeLists.txt
@@ -1,21 +1,5 @@
 cmake_minimum_required(VERSION 3.11)
-set(CMAKE_SYSTEM_PROCESSOR arm64)
 
-# Check if we are on an Apple system and if the architecture is ARM
-if(APPLE)
-    # Returns x86_64 or arm64
-    execute_process(COMMAND uname -m OUTPUT_VARIABLE CMAKE_OSX_ARCHITECTURES OUTPUT_STRIP_TRAILING_WHITESPACE)
-    if (OSX_ARCH STREQUAL "arm64")
-        set(APPLE_ARM TRUE)
-    else()
-        set(APPLE_ARM FALSE)
-    endif()
-else()
-    set(APPLE_ARM FALSE)
-endif(APPLE)
-message(STATUS "APPLE_ARM: ${APPLE_ARM}")
-
-# Set the project based on the detected platform
 if(APPLE_ARM)
     project(qsim LANGUAGES CXX)
 else()
@@ -38,7 +22,7 @@ ELSE()
     set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
 ENDIF()
 
-if(APPLE)
+if(APPLE AND NOT APPLE_ARM)
     set(CMAKE_CXX_STANDARD 14)
     include_directories("/usr/local/include" "/usr/local/opt/llvm/include")
     link_directories("/usr/local/lib" "/usr/local/opt/llvm/lib")
@@ -48,12 +32,7 @@ INCLUDE(../GetPybind11.cmake)
 
 # Configure based on the detected platform
 if(APPLE_ARM)
-    # Configuration for Apple ARM
-    # Set additional flags if necessary
-    # Compile the module using standard C++
     pybind11_add_module(qsim_decide decide.cpp)
-    # Print a message to indicate the module is being compiled using standard C++
-    message(STATUS "Compiling qsim_decide using standard C++ on Apple ARM")
 elseif(has_nvcc STREQUAL "")
     if(has_hipcc STREQUAL "")
         pybind11_add_module(qsim_decide decide.cpp)
diff --git a/pybind_interface/decide/decide.cpp b/pybind_interface/decide/decide.cpp
index 3641833c..0da7b0a8 100644
--- a/pybind_interface/decide/decide.cpp
+++ b/pybind_interface/decide/decide.cpp
@@ -64,7 +64,6 @@ int detect_instructions() {
   return static_cast<int>(instr);
 }
 
-
 enum GPUCapabilities {
     CUDA = 0, CUSTATEVEC = 1, HIP = 2, NO_GPU = 10, NO_CUSTATEVEC = 11 };
 

From 44ed30523d590a042131cc855581b3eb24d528c6 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 10:10:39 -0800
Subject: [PATCH 04/21] Support Apple x86 and ARM in setup.py with homebrew and
 current PATH

---
 setup.py | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)

diff --git a/setup.py b/setup.py
index 06b7ba16..f460b2b5 100644
--- a/setup.py
+++ b/setup.py
@@ -59,10 +59,23 @@ def build_extension(self, ext):
             build_args += ["--", "-j2"]
 
         if platform.system() == "Darwin":
-            cmake_args += [
-                "-DCMAKE_C_COMPILER=/Users/bas.nijholt/micromamba/envs/py311/bin/clang",
-                "-DCMAKE_CXX_COMPILER=/Users/bas.nijholt/micromamba/envs/py311/bin/clang++",
-            ]
+            homebrew_x86 = "/usr/local/opt/llvm/bin"
+            homebrew_arm = "/opt/homebrew/opt/llvm/bin"
+            # Add clang
+            if shutil.which("clang") is not None:
+                cmake_args.append("-DCMAKE_C_COMPILER=clang")
+            elif os.path.exists(f"{homebrew_x86}/clang"):
+                cmake_args.append(f"-DCMAKE_C_COMPILER={homebrew_x86}/clang")
+            elif os.path.exists(f"{homebrew_arm}/clang"):
+                cmake_args.append(f"-DCMAKE_C_COMPILER={homebrew_arm}/clang")
+
+            # Add clang++
+            if shutil.which("clang++") is not None:
+                cmake_args.append("-DCMAKE_CXX_COMPILER=clang++")
+            elif os.path.exists(f"{homebrew_x86}/clang++"):
+                cmake_args.append(f"-DCMAKE_CXX_COMPILER={homebrew_x86}/clang++")
+            elif os.path.exists(f"{homebrew_arm}/clang++"):
+                cmake_args.append(f"-DCMAKE_CXX_COMPILER={homebrew_arm}/clang++")
 
         if shutil.which("hipcc") is not None:
             cmake_args += [

From 31536d81adfaee14279b136368042ad1acd5e565 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 10:22:23 -0800
Subject: [PATCH 05/21] Style

---
 CMakeLists.txt | 15 +++++++++++----
 setup.py       |  4 ++--
 2 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 98aae939..ff9148f1 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -3,12 +3,12 @@ set(OpenMP_INCLUDE_DIR "/Users/bas.nijholt/micromamba/lib/clang/17.0.4/include")
 include_directories(${OpenMP_INCLUDE_DIR})
 cmake_minimum_required(VERSION 3.11)
 
+# Set APPLE_ARM to TRUE if running on Apple Silicon
 if(APPLE)
-    # Returns x86_64 or arm64
     execute_process(COMMAND uname -m OUTPUT_VARIABLE OSX_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE)
     if (OSX_ARCH STREQUAL "arm64")
         set(APPLE_ARM TRUE)
-    else()
+    else()  # x86_64
         set(APPLE_ARM FALSE)
     endif()
 else()
@@ -16,10 +16,14 @@ else()
 endif(APPLE)
 message(STATUS "APPLE_ARM: ${APPLE_ARM}")
 
+message(STATUS "CXX Flags: " ${CMAKE_CXX_FLAGS})
+message(STATUS "C Flags: " ${CMAKE_C_FLAGS})
+
+
+# basic and decide can be compiled with any compiler
 ADD_SUBDIRECTORY(pybind_interface/basic)
 ADD_SUBDIRECTORY(pybind_interface/decide)
 
-# You can then use this variable to control conditional compilation
 if(APPLE_ARM)
     project(qsim LANGUAGES CXX)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8-a")
@@ -45,4 +49,7 @@ else()
     ADD_SUBDIRECTORY(pybind_interface/avx512)
     ADD_SUBDIRECTORY(pybind_interface/avx2)
 endif()
-find_package(OpenMP REQUIRED)
+
+if(APPLE)
+    find_package(OpenMP REQUIRED)  # Automatically happens for Linux
+endif()
diff --git a/setup.py b/setup.py
index f460b2b5..ed4720c3 100644
--- a/setup.py
+++ b/setup.py
@@ -62,7 +62,7 @@ def build_extension(self, ext):
             homebrew_x86 = "/usr/local/opt/llvm/bin"
             homebrew_arm = "/opt/homebrew/opt/llvm/bin"
             # Add clang
-            if shutil.which("clang") is not None:
+            if shutil.which("clang") is not None:  # Always prefer from PATH
                 cmake_args.append("-DCMAKE_C_COMPILER=clang")
             elif os.path.exists(f"{homebrew_x86}/clang"):
                 cmake_args.append(f"-DCMAKE_C_COMPILER={homebrew_x86}/clang")
@@ -70,7 +70,7 @@ def build_extension(self, ext):
                 cmake_args.append(f"-DCMAKE_C_COMPILER={homebrew_arm}/clang")
 
             # Add clang++
-            if shutil.which("clang++") is not None:
+            if shutil.which("clang++") is not None:  # Always prefer from PATH
                 cmake_args.append("-DCMAKE_CXX_COMPILER=clang++")
             elif os.path.exists(f"{homebrew_x86}/clang++"):
                 cmake_args.append(f"-DCMAKE_CXX_COMPILER={homebrew_x86}/clang++")

From a10c693031a91f4adebf829a16013bc85dfe5bb1 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 10:23:29 -0800
Subject: [PATCH 06/21] remove debug messages in cmake

---
 CMakeLists.txt | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index ff9148f1..bf87799c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -14,11 +14,6 @@ if(APPLE)
 else()
     set(APPLE_ARM FALSE)
 endif(APPLE)
-message(STATUS "APPLE_ARM: ${APPLE_ARM}")
-
-message(STATUS "CXX Flags: " ${CMAKE_CXX_FLAGS})
-message(STATUS "C Flags: " ${CMAKE_C_FLAGS})
-
 
 # basic and decide can be compiled with any compiler
 ADD_SUBDIRECTORY(pybind_interface/basic)

From 0fbbc89e0c1a61289f71e6637c1a53d65ccd18ed Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 10:55:01 -0800
Subject: [PATCH 07/21] Generalize pybind11 finding

---
 pybind_interface/GetPybind11.cmake | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)

diff --git a/pybind_interface/GetPybind11.cmake b/pybind_interface/GetPybind11.cmake
index 56223801..9df238b9 100644
--- a/pybind_interface/GetPybind11.cmake
+++ b/pybind_interface/GetPybind11.cmake
@@ -1,4 +1,3 @@
-add_compile_options(-I/Users/bas.nijholt/micromamba/envs/py311/include/python3.11 -I/Users/bas.nijholt/micromamba/envs/py311/lib/python3.11/site-packages/pybind11/include)
 include(FetchContent)
 
 set(MIN_PYBIND_VERSION "2.10.1")
@@ -9,6 +8,24 @@ FetchContent_Declare(
 )
 FetchContent_GetProperties(pybind11)
 find_package(pybind11 "${MIN_PYBIND_VERSION}" CONFIG)
+
+if (pybind11_FOUND)
+  message(STATUS "Found pybind11 v${pybind11_VERSION}: ${pybind11_INCLUDE_DIRS}")
+  # The pybind11_add_module doesn't correctly set the CXX_INCLUDES properly if a system pybind11 is found.
+  # Using `include_directories(...)` doesn't result in anything in
+  # CXX_INCLUDES. e.g., `pybind_interface/basic/CMakeFiles/qsim_basic.dir/flags.make` would only
+  # have `CXX_INCLUDES = -isystem $PREFIX/include/python3.11` and would miss `$PREFIX/include`.
+  # This problem would result in `fatal error: pybind11/complex.h: No such file or directory`
+  # This is a hack to get around that by passing `-I/path/to/include` to CXX_FLAGS
+  execute_process(COMMAND python3 -m pybind11 --includes OUTPUT_VARIABLE PYBIND11_INCLUDES OUTPUT_STRIP_TRAILING_WHITESPACE)
+  message(STATUS "pybind11 includes: ${PYBIND11_INCLUDES}")
+  # Separate the includes into a list and add them as individual compile options
+  separate_arguments(PYBIND11_INCLUDES_LIST UNIX_COMMAND "${PYBIND11_INCLUDES}")
+  foreach(INCLUDE_DIR ${PYBIND11_INCLUDES_LIST})
+    add_compile_options("${INCLUDE_DIR}")
+  endforeach()
+endif()
+
 if((NOT pybind11_FOUND) AND (NOT pybind11_POPULATED)) # check first on system path, then attempt git fetch
   FetchContent_Populate(pybind11)
   add_subdirectory(${pybind11_SOURCE_DIR} ${pybind11_BINARY_DIR})

From 08ad3d74b22d3b758c1cc445b911e064c14ba615 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 10:55:28 -0800
Subject: [PATCH 08/21] Set project first

---
 CMakeLists.txt | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index bf87799c..59c45c28 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -15,13 +15,10 @@ else()
     set(APPLE_ARM FALSE)
 endif(APPLE)
 
-# basic and decide can be compiled with any compiler
-ADD_SUBDIRECTORY(pybind_interface/basic)
-ADD_SUBDIRECTORY(pybind_interface/decide)
-
 if(APPLE_ARM)
     project(qsim LANGUAGES CXX)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8-a")
+    ADD_SUBDIRECTORY(pybind_interface/basic)
+    ADD_SUBDIRECTORY(pybind_interface/decide)
 else()
     execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
     if(has_nvcc STREQUAL "")
@@ -43,6 +40,8 @@ else()
     ADD_SUBDIRECTORY(pybind_interface/sse)
     ADD_SUBDIRECTORY(pybind_interface/avx512)
     ADD_SUBDIRECTORY(pybind_interface/avx2)
+    ADD_SUBDIRECTORY(pybind_interface/basic)
+    ADD_SUBDIRECTORY(pybind_interface/decide)
 endif()
 
 if(APPLE)

From e8b6ed6816e4a65f269524f1eec2206c14864b94 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 11:00:22 -0800
Subject: [PATCH 09/21] Simplify

---
 pybind_interface/GetPybind11.cmake | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)

diff --git a/pybind_interface/GetPybind11.cmake b/pybind_interface/GetPybind11.cmake
index 9df238b9..9a55c904 100644
--- a/pybind_interface/GetPybind11.cmake
+++ b/pybind_interface/GetPybind11.cmake
@@ -12,17 +12,14 @@ find_package(pybind11 "${MIN_PYBIND_VERSION}" CONFIG)
 if (pybind11_FOUND)
   message(STATUS "Found pybind11 v${pybind11_VERSION}: ${pybind11_INCLUDE_DIRS}")
   # The pybind11_add_module doesn't correctly set the CXX_INCLUDES properly if a system pybind11 is found.
-  # Using `include_directories(...)` doesn't result in anything in
+  # Using `include_directories(${pybind11_INCLUDE_DIRS})` doesn't result in anything in
   # CXX_INCLUDES. e.g., `pybind_interface/basic/CMakeFiles/qsim_basic.dir/flags.make` would only
   # have `CXX_INCLUDES = -isystem $PREFIX/include/python3.11` and would miss `$PREFIX/include`.
   # This problem would result in `fatal error: pybind11/complex.h: No such file or directory`
   # This is a hack to get around that by passing `-I/path/to/include` to CXX_FLAGS
-  execute_process(COMMAND python3 -m pybind11 --includes OUTPUT_VARIABLE PYBIND11_INCLUDES OUTPUT_STRIP_TRAILING_WHITESPACE)
-  message(STATUS "pybind11 includes: ${PYBIND11_INCLUDES}")
-  # Separate the includes into a list and add them as individual compile options
-  separate_arguments(PYBIND11_INCLUDES_LIST UNIX_COMMAND "${PYBIND11_INCLUDES}")
-  foreach(INCLUDE_DIR ${PYBIND11_INCLUDES_LIST})
-    add_compile_options("${INCLUDE_DIR}")
+  # Iterate over each include directory and add it as a compile option
+  foreach(INCLUDE_DIR ${pybind11_INCLUDE_DIRS})
+    add_compile_options("-I${INCLUDE_DIR}")
   endforeach()
 endif()
 

From 423b8ed7d6b9f610b4a9217175c0263a46d395b5 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 11:10:40 -0800
Subject: [PATCH 10/21] Simplify logic in CMakeLists.txt

---
 CMakeLists.txt | 25 +++++++++++++------------
 1 file changed, 13 insertions(+), 12 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 59c45c28..736cd4a3 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,6 +1,4 @@
 set(CMAKE_CXX_STANDARD 11)
-set(OpenMP_INCLUDE_DIR "/Users/bas.nijholt/micromamba/lib/clang/17.0.4/include")
-include_directories(${OpenMP_INCLUDE_DIR})
 cmake_minimum_required(VERSION 3.11)
 
 # Set APPLE_ARM to TRUE if running on Apple Silicon
@@ -15,10 +13,10 @@ else()
     set(APPLE_ARM FALSE)
 endif(APPLE)
 
-if(APPLE_ARM)
+# Set the project name and language
+if(APPLE)
     project(qsim LANGUAGES CXX)
-    ADD_SUBDIRECTORY(pybind_interface/basic)
-    ADD_SUBDIRECTORY(pybind_interface/decide)
+    find_package(OpenMP REQUIRED)  # Automatically happens for Linux
 else()
     execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
     if(has_nvcc STREQUAL "")
@@ -27,23 +25,26 @@ else()
             project(qsim LANGUAGES CXX)
         else()
             project(qsim LANGUAGES CXX HIP)
-            ADD_SUBDIRECTORY(pybind_interface/hip)
         endif()
     else()
         project(qsim LANGUAGES CXX CUDA)
+    endif()
+endif()
+
+# Add subdirectories based on the architecture or available compilers
+ADD_SUBDIRECTORY(pybind_interface/basic)
+ADD_SUBDIRECTORY(pybind_interface/decide)
+if(NOT APPLE_ARM)
+    if(NOT has_nvcc STREQUAL "")
         ADD_SUBDIRECTORY(pybind_interface/cuda)
         if(DEFINED ENV{CUQUANTUM_ROOT})
             ADD_SUBDIRECTORY(pybind_interface/custatevec)
         endif()
+    elseif(NOT has_hipcc STREQUAL "")
+        ADD_SUBDIRECTORY(pybind_interface/hip)
     endif()
 
     ADD_SUBDIRECTORY(pybind_interface/sse)
     ADD_SUBDIRECTORY(pybind_interface/avx512)
     ADD_SUBDIRECTORY(pybind_interface/avx2)
-    ADD_SUBDIRECTORY(pybind_interface/basic)
-    ADD_SUBDIRECTORY(pybind_interface/decide)
-endif()
-
-if(APPLE)
-    find_package(OpenMP REQUIRED)  # Automatically happens for Linux
 endif()

From e333c2ec39e56c7396b39d7590175be3fb0850a6 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 20 Nov 2023 11:46:21 -0800
Subject: [PATCH 11/21] Simplify CMakeLists.txt even more

---
 CMakeLists.txt                         | 20 ++++----
 pybind_interface/decide/CMakeLists.txt | 71 ++++++++------------------
 2 files changed, 32 insertions(+), 59 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 736cd4a3..151ad6b8 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -18,16 +18,16 @@ if(APPLE)
     project(qsim LANGUAGES CXX)
     find_package(OpenMP REQUIRED)  # Automatically happens for Linux
 else()
-    execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
-    if(has_nvcc STREQUAL "")
-        execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
-        if(has_hipcc STREQUAL "")
-            project(qsim LANGUAGES CXX)
-        else()
+    execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if(has_nvcc)
+        project(qsim LANGUAGES CXX CUDA)
+    else()
+        execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc OUTPUT_STRIP_TRAILING_WHITESPACE)
+        if(has_hipcc)
             project(qsim LANGUAGES CXX HIP)
+        else()
+            project(qsim LANGUAGES CXX)
         endif()
-    else()
-        project(qsim LANGUAGES CXX CUDA)
     endif()
 endif()
 
@@ -35,12 +35,12 @@ endif()
 ADD_SUBDIRECTORY(pybind_interface/basic)
 ADD_SUBDIRECTORY(pybind_interface/decide)
 if(NOT APPLE_ARM)
-    if(NOT has_nvcc STREQUAL "")
+    if(has_nvcc)
         ADD_SUBDIRECTORY(pybind_interface/cuda)
         if(DEFINED ENV{CUQUANTUM_ROOT})
             ADD_SUBDIRECTORY(pybind_interface/custatevec)
         endif()
-    elseif(NOT has_hipcc STREQUAL "")
+    elseif(has_hipcc)
         ADD_SUBDIRECTORY(pybind_interface/hip)
     endif()
 
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
index 7724a3e9..159e8b1f 100644
--- a/pybind_interface/decide/CMakeLists.txt
+++ b/pybind_interface/decide/CMakeLists.txt
@@ -1,26 +1,10 @@
 cmake_minimum_required(VERSION 3.11)
 
-if(APPLE_ARM)
-    project(qsim LANGUAGES CXX)
-else()
-    execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc)
-    if(has_nvcc STREQUAL "")
-        execute_process(COMMAND which hipcc OUTPUT_VARIABLE has_hipcc)
-        if(has_hipcc STREQUAL "")
-            project(qsim)
-        else()
-            project(qsim LANGUAGES CXX HIP)
-        endif()
-    else()
-        project(qsim LANGUAGES CXX CUDA)
-    endif()
-endif()
-
-IF (WIN32)
+if(WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
-ELSE()
+else()
     set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
-ENDIF()
+endif()
 
 if(APPLE AND NOT APPLE_ARM)
     set(CMAKE_CXX_STANDARD 14)
@@ -28,46 +12,35 @@ if(APPLE AND NOT APPLE_ARM)
     link_directories("/usr/local/lib" "/usr/local/opt/llvm/lib")
 endif()
 
-INCLUDE(../GetPybind11.cmake)
+include(../GetPybind11.cmake)
 
 # Configure based on the detected platform
-if(APPLE_ARM)
-    pybind11_add_module(qsim_decide decide.cpp)
-elseif(has_nvcc STREQUAL "")
-    if(has_hipcc STREQUAL "")
-        pybind11_add_module(qsim_decide decide.cpp)
-    else()
-        list(APPEND CMAKE_MODULE_PATH "/opt/rocm/lib/cmake/hip")
-        find_package(HIP REQUIRED)
-        find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
-
-        include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
-
-        hip_add_library(qsim_decide MODULE decide.cpp)
-
-        set_target_properties(qsim_decide PROPERTIES
-            PREFIX "${PYTHON_MODULE_PREFIX}"
-            SUFFIX "${PYTHON_MODULE_EXTENSION}"
-        )
-        set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
-    endif()
-else()
-    find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
+if(has_nvcc)
     find_package(CUDA REQUIRED)
-
-    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
-
     cuda_add_library(qsim_decide MODULE decide.cpp)
-
     if(DEFINED ENV{CUQUANTUM_ROOT})
         target_compile_options(qsim_decide PRIVATE
             $<$<COMPILE_LANGUAGE:CUDA>:-D__CUSTATEVEC__>
         )
     endif()
-
+    find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
+    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
     set_target_properties(qsim_decide PROPERTIES
-          PREFIX "${PYTHON_MODULE_PREFIX}"
-          SUFFIX "${PYTHON_MODULE_EXTENSION}"
+            PREFIX "${PYTHON_MODULE_PREFIX}"
+            SUFFIX "${PYTHON_MODULE_EXTENSION}"
     )
     set_source_files_properties(decide.cpp PROPERTIES LANGUAGE CUDA)
+elseif(has_hipcc)
+    list(APPEND CMAKE_MODULE_PATH "/opt/rocm/lib/cmake/hip")
+    find_package(HIP REQUIRED)
+    hip_add_library(qsim_decide MODULE decide.cpp)
+    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
+    find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
+    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
+    set_target_properties(qsim_decide PROPERTIES
+            PREFIX "${PYTHON_MODULE_PREFIX}"
+            SUFFIX "${PYTHON_MODULE_EXTENSION}"
+    )
+else()
+    pybind11_add_module(qsim_decide decide.cpp)
 endif()

From fff819ee24b4d234e83b8fee573cebe7838fb39a Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 09:19:28 -0800
Subject: [PATCH 12/21] Implement suggestion by @sergeisakov

---
 pybind_interface/decide/decide.cpp | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/pybind_interface/decide/decide.cpp b/pybind_interface/decide/decide.cpp
index 0da7b0a8..b40f6975 100644
--- a/pybind_interface/decide/decide.cpp
+++ b/pybind_interface/decide/decide.cpp
@@ -35,11 +35,7 @@ enum Instructions { AVX512F = 0, AVX2 = 1, SSE4_1 = 2, BASIC = 3};
 int detect_instructions() {
   Instructions instr = BASIC;
 
-  // Check for ARM architecture, specifically for Apple M1 or similar
-  #if defined(__aarch64__) && defined(__APPLE__)
-  // On Apple ARM systems, always use BASIC
-  instr = BASIC;
-  #else
+  #if !defined(__aarch64__) || !defined(__APPLE__)
   // Existing x86/x86_64 specific instruction set detection logic
   int info[4];
   cpuid(info, 0);

From 1ca54779b37c8972fa42d9014c9fa64ee8c0879f Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 09:32:28 -0800
Subject: [PATCH 13/21] Set flags for OpenMP_CXX

---
 CMakeLists.txt | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 151ad6b8..75b609cb 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -17,6 +17,8 @@ endif(APPLE)
 if(APPLE)
     project(qsim LANGUAGES CXX)
     find_package(OpenMP REQUIRED)  # Automatically happens for Linux
+    set(OpenMP_CXX_FLAGS "-fopenmp=libomp")
+    set(OpenMP_CXX_LIB_NAMES "libomp")
 else()
     execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc OUTPUT_STRIP_TRAILING_WHITESPACE)
     if(has_nvcc)

From 6f074e939bc8ea50fcc22ca88addffefd53d53c7 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 11:06:17 -0800
Subject: [PATCH 14/21] Link libomp with homebrew

---
 .github/workflows/release_wheels.yml | 2 +-
 .github/workflows/testing_wheels.yml | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/release_wheels.yml b/.github/workflows/release_wheels.yml
index 292f4dc5..5b63bf29 100644
--- a/.github/workflows/release_wheels.yml
+++ b/.github/workflows/release_wheels.yml
@@ -32,7 +32,7 @@ jobs:
       CIBW_SKIP: "*musllinux*"
       CIBW_ARCHS: "${{ matrix.cibw.arch || 'auto' }}"
       CIBW_MANYLINUX_X86_64_IMAGE: "${{ matrix.cibw.manylinux_image }}"
-      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force libomp"
       CIBW_REPAIR_WHEEL_COMMAND_MACOS: "delocate-listdeps {wheel} && delocate-wheel --verbose --require-archs {delocate_archs} -w {dest_dir} {wheel}"
       # to install latest delocate package
       CIBW_DEPENDENCY_VERSIONS: "latest"
diff --git a/.github/workflows/testing_wheels.yml b/.github/workflows/testing_wheels.yml
index 4d3c1ca3..80233abd 100644
--- a/.github/workflows/testing_wheels.yml
+++ b/.github/workflows/testing_wheels.yml
@@ -37,7 +37,7 @@ jobs:
       CIBW_SKIP: "*musllinux*"
       CIBW_ARCHS: "${{ matrix.cibw.arch || 'auto' }}"
       CIBW_MANYLINUX_X86_64_IMAGE: "${{ matrix.cibw.manylinux_image }}"
-      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force libomp"
       CIBW_REPAIR_WHEEL_COMMAND_MACOS: "delocate-listdeps {wheel} && delocate-wheel --verbose --require-archs {delocate_archs} -w {dest_dir} {wheel}"
       # to install latest delocate package
       CIBW_DEPENDENCY_VERSIONS: "latest"

From b605cad39fb05752930beac9fe108d53de4b7947 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 11:07:19 -0800
Subject: [PATCH 15/21] Remove setting of CXX flags and lib

---
 CMakeLists.txt | 2 --
 1 file changed, 2 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 75b609cb..151ad6b8 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -17,8 +17,6 @@ endif(APPLE)
 if(APPLE)
     project(qsim LANGUAGES CXX)
     find_package(OpenMP REQUIRED)  # Automatically happens for Linux
-    set(OpenMP_CXX_FLAGS "-fopenmp=libomp")
-    set(OpenMP_CXX_LIB_NAMES "libomp")
 else()
     execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc OUTPUT_STRIP_TRAILING_WHITESPACE)
     if(has_nvcc)

From bb92b2e64e32666429bbad2f5ee5ff0da35c8f80 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 11:18:59 -0800
Subject: [PATCH 16/21] Rely on 'find_package(OpenMP REQUIRED)' to set -fopenmp

---
 CMakeLists.txt                             | 3 ++-
 pybind_interface/avx2/CMakeLists.txt       | 2 +-
 pybind_interface/avx512/CMakeLists.txt     | 2 +-
 pybind_interface/basic/CMakeLists.txt      | 8 ++++----
 pybind_interface/cuda/CMakeLists.txt       | 8 ++++----
 pybind_interface/custatevec/CMakeLists.txt | 8 ++++----
 pybind_interface/decide/CMakeLists.txt     | 2 +-
 pybind_interface/hip/CMakeLists.txt        | 8 ++++----
 pybind_interface/sse/CMakeLists.txt        | 2 +-
 9 files changed, 22 insertions(+), 21 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 151ad6b8..b98e1bbe 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -16,7 +16,6 @@ endif(APPLE)
 # Set the project name and language
 if(APPLE)
     project(qsim LANGUAGES CXX)
-    find_package(OpenMP REQUIRED)  # Automatically happens for Linux
 else()
     execute_process(COMMAND which nvcc OUTPUT_VARIABLE has_nvcc OUTPUT_STRIP_TRAILING_WHITESPACE)
     if(has_nvcc)
@@ -31,6 +30,8 @@ else()
     endif()
 endif()
 
+find_package(OpenMP REQUIRED)
+
 # Add subdirectories based on the architecture or available compilers
 ADD_SUBDIRECTORY(pybind_interface/basic)
 ADD_SUBDIRECTORY(pybind_interface/decide)
diff --git a/pybind_interface/avx2/CMakeLists.txt b/pybind_interface/avx2/CMakeLists.txt
index eebba584..18047a1e 100644
--- a/pybind_interface/avx2/CMakeLists.txt
+++ b/pybind_interface/avx2/CMakeLists.txt
@@ -4,7 +4,7 @@ project(qsim)
 IF (WIN32)
     set(CMAKE_CXX_FLAGS "/arch:AVX2 /O2 /openmp")
 ELSE()
-    set(CMAKE_CXX_FLAGS "-mavx2 -mfma -O3 -fopenmp")
+    set(CMAKE_CXX_FLAGS "-mavx2 -mfma -O3")
 ENDIF()
 
 if(APPLE)
diff --git a/pybind_interface/avx512/CMakeLists.txt b/pybind_interface/avx512/CMakeLists.txt
index 86cfdfa8..5875c08c 100644
--- a/pybind_interface/avx512/CMakeLists.txt
+++ b/pybind_interface/avx512/CMakeLists.txt
@@ -5,7 +5,7 @@ project(qsim)
 IF (WIN32)
     set(CMAKE_CXX_FLAGS "/arch:AVX512 /O2 /openmp")
 ELSE()
-    set(CMAKE_CXX_FLAGS "-mavx512f -mbmi2 -O3 -fopenmp")
+    set(CMAKE_CXX_FLAGS "-mavx512f -mbmi2 -O3")
 ENDIF()
 
 if(APPLE)
diff --git a/pybind_interface/basic/CMakeLists.txt b/pybind_interface/basic/CMakeLists.txt
index 35347211..36bdfc8d 100644
--- a/pybind_interface/basic/CMakeLists.txt
+++ b/pybind_interface/basic/CMakeLists.txt
@@ -1,11 +1,11 @@
 cmake_minimum_required(VERSION 3.11)
 project(qsim)
 
-IF (WIN32)
+if(WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
-ELSE()
-    set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
-ENDIF()
+else()
+    set(CMAKE_CXX_FLAGS "-O3")
+endif()
 
 
 if(APPLE)
diff --git a/pybind_interface/cuda/CMakeLists.txt b/pybind_interface/cuda/CMakeLists.txt
index 4e46eb32..29fd367c 100644
--- a/pybind_interface/cuda/CMakeLists.txt
+++ b/pybind_interface/cuda/CMakeLists.txt
@@ -1,11 +1,11 @@
 cmake_minimum_required(VERSION 3.11)
 project(qsim LANGUAGES CXX CUDA)
 
-IF (WIN32)
+if(WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
-ELSE()
-    set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
-ENDIF()
+else()
+    set(CMAKE_CXX_FLAGS "-O3")
+endif()
 
 
 if(APPLE)
diff --git a/pybind_interface/custatevec/CMakeLists.txt b/pybind_interface/custatevec/CMakeLists.txt
index 1ef63166..1c34a1d9 100644
--- a/pybind_interface/custatevec/CMakeLists.txt
+++ b/pybind_interface/custatevec/CMakeLists.txt
@@ -15,11 +15,11 @@
 cmake_minimum_required(VERSION 3.11)
 project(qsim LANGUAGES CXX CUDA)
 
-IF (WIN32)
+if(WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
-ELSE()
-    set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
-ENDIF()
+else()
+    set(CMAKE_CXX_FLAGS "-O3")
+endif()
 
 if(APPLE)
     set(CMAKE_CXX_STANDARD 14)
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
index 159e8b1f..b3dac71c 100644
--- a/pybind_interface/decide/CMakeLists.txt
+++ b/pybind_interface/decide/CMakeLists.txt
@@ -3,7 +3,7 @@ cmake_minimum_required(VERSION 3.11)
 if(WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
 else()
-    set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
+    set(CMAKE_CXX_FLAGS "-O3")
 endif()
 
 if(APPLE AND NOT APPLE_ARM)
diff --git a/pybind_interface/hip/CMakeLists.txt b/pybind_interface/hip/CMakeLists.txt
index fe4b1c54..b56687ab 100644
--- a/pybind_interface/hip/CMakeLists.txt
+++ b/pybind_interface/hip/CMakeLists.txt
@@ -1,11 +1,11 @@
 cmake_minimum_required(VERSION 3.18)
 project(qsim LANGUAGES CXX HIP)
 
-IF (WIN32)
+if(WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
-ELSE()
-    set(CMAKE_CXX_FLAGS "-O3 -fopenmp")
-ENDIF()
+else()
+    set(CMAKE_CXX_FLAGS "-O3")
+endif()
 
 INCLUDE(../GetPybind11.cmake)
 find_package(PythonLibs 3.7 REQUIRED)
diff --git a/pybind_interface/sse/CMakeLists.txt b/pybind_interface/sse/CMakeLists.txt
index fe9b218e..9be08400 100644
--- a/pybind_interface/sse/CMakeLists.txt
+++ b/pybind_interface/sse/CMakeLists.txt
@@ -4,7 +4,7 @@ project(qsim)
 IF (WIN32)
     set(CMAKE_CXX_FLAGS "/O2 /openmp")
 ELSE()
-    set(CMAKE_CXX_FLAGS "-msse4.1 -O3 -fopenmp")
+    set(CMAKE_CXX_FLAGS "-msse4.1 -O3")
 ENDIF()
 
 

From 67273764b32ca8b43b403d6bfa633793ace36013 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 11:42:42 -0800
Subject: [PATCH 17/21] Build wheels for Apple Silicon

---
 .github/workflows/release_wheels.yml | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/release_wheels.yml b/.github/workflows/release_wheels.yml
index 5b63bf29..7e1201b6 100644
--- a/.github/workflows/release_wheels.yml
+++ b/.github/workflows/release_wheels.yml
@@ -12,10 +12,14 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - os: macos-12
+          - os: macos-12 # x86_64
             name: mac
             cibw:
               build: "cp37* cp38* cp39* cp310* cp311*"
+          - os: macos-13-large # Apple Silicon
+            name: mac_arm64
+            cibw:
+              build: "cp37* cp38* cp39* cp310* cp311*"
           - os: ubuntu-20.04
             name: manylinux2014
             cibw:

From 0d324c0b8a14e8a6a75162f3e2c8081f4bf85624 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 11:56:02 -0800
Subject: [PATCH 18/21] Add --overwrite

---
 .github/workflows/release_wheels.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/release_wheels.yml b/.github/workflows/release_wheels.yml
index 7e1201b6..971bd286 100644
--- a/.github/workflows/release_wheels.yml
+++ b/.github/workflows/release_wheels.yml
@@ -36,7 +36,7 @@ jobs:
       CIBW_SKIP: "*musllinux*"
       CIBW_ARCHS: "${{ matrix.cibw.arch || 'auto' }}"
       CIBW_MANYLINUX_X86_64_IMAGE: "${{ matrix.cibw.manylinux_image }}"
-      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force libomp"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force --overwrite libomp"
       CIBW_REPAIR_WHEEL_COMMAND_MACOS: "delocate-listdeps {wheel} && delocate-wheel --verbose --require-archs {delocate_archs} -w {dest_dir} {wheel}"
       # to install latest delocate package
       CIBW_DEPENDENCY_VERSIONS: "latest"

From e7649ffb2af3d0d05f13972358065c37c4e845c4 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 12:03:41 -0800
Subject: [PATCH 19/21] Add macos-13-large to testing wheels too

---
 .github/workflows/testing_wheels.yml | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/testing_wheels.yml b/.github/workflows/testing_wheels.yml
index 80233abd..a73fa771 100644
--- a/.github/workflows/testing_wheels.yml
+++ b/.github/workflows/testing_wheels.yml
@@ -17,10 +17,14 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - os: macos-12
+          - os: macos-12 # x86_64
             name: mac
             cibw:
               build: "cp37* cp38* cp39* cp310* cp311*"
+          - os: macos-13-large # Apple Silicon
+            name: mac_arm64
+            cibw:
+              build: "cp37* cp38* cp39* cp310* cp311*"
           - os: ubuntu-20.04
             name: manylinux2014
             cibw:
@@ -37,7 +41,7 @@ jobs:
       CIBW_SKIP: "*musllinux*"
       CIBW_ARCHS: "${{ matrix.cibw.arch || 'auto' }}"
       CIBW_MANYLINUX_X86_64_IMAGE: "${{ matrix.cibw.manylinux_image }}"
-      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force libomp"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force --overwrite libomp"
       CIBW_REPAIR_WHEEL_COMMAND_MACOS: "delocate-listdeps {wheel} && delocate-wheel --verbose --require-archs {delocate_archs} -w {dest_dir} {wheel}"
       # to install latest delocate package
       CIBW_DEPENDENCY_VERSIONS: "latest"

From b06b395d31d2d3866341c484237a9227f4b4dee1 Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Mon, 27 Nov 2023 12:11:02 -0800
Subject: [PATCH 20/21] overwrite Python

---
 .github/workflows/release_wheels.yml | 2 +-
 .github/workflows/testing_wheels.yml | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/release_wheels.yml b/.github/workflows/release_wheels.yml
index 971bd286..d6161a47 100644
--- a/.github/workflows/release_wheels.yml
+++ b/.github/workflows/release_wheels.yml
@@ -36,7 +36,7 @@ jobs:
       CIBW_SKIP: "*musllinux*"
       CIBW_ARCHS: "${{ matrix.cibw.arch || 'auto' }}"
       CIBW_MANYLINUX_X86_64_IMAGE: "${{ matrix.cibw.manylinux_image }}"
-      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force --overwrite libomp"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --overwrite python@3.11 && brew link --force libomp"
       CIBW_REPAIR_WHEEL_COMMAND_MACOS: "delocate-listdeps {wheel} && delocate-wheel --verbose --require-archs {delocate_archs} -w {dest_dir} {wheel}"
       # to install latest delocate package
       CIBW_DEPENDENCY_VERSIONS: "latest"
diff --git a/.github/workflows/testing_wheels.yml b/.github/workflows/testing_wheels.yml
index a73fa771..f30921af 100644
--- a/.github/workflows/testing_wheels.yml
+++ b/.github/workflows/testing_wheels.yml
@@ -41,7 +41,7 @@ jobs:
       CIBW_SKIP: "*musllinux*"
       CIBW_ARCHS: "${{ matrix.cibw.arch || 'auto' }}"
       CIBW_MANYLINUX_X86_64_IMAGE: "${{ matrix.cibw.manylinux_image }}"
-      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --force --overwrite libomp"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --overwrite python@3.11 && brew link --force libomp"
       CIBW_REPAIR_WHEEL_COMMAND_MACOS: "delocate-listdeps {wheel} && delocate-wheel --verbose --require-archs {delocate_archs} -w {dest_dir} {wheel}"
       # to install latest delocate package
       CIBW_DEPENDENCY_VERSIONS: "latest"

From 34ca2a5fe57b78b04dab38da95776285232ef07b Mon Sep 17 00:00:00 2001
From: Bas Nijholt <bas@nijho.lt>
Date: Tue, 28 Nov 2023 12:55:43 -0500
Subject: [PATCH 21/21] Call target_link_libraries for OpenMP

---
 pybind_interface/avx2/CMakeLists.txt       | 2 ++
 pybind_interface/avx512/CMakeLists.txt     | 2 ++
 pybind_interface/basic/CMakeLists.txt      | 2 ++
 pybind_interface/cuda/CMakeLists.txt       | 2 ++
 pybind_interface/custatevec/CMakeLists.txt | 2 ++
 pybind_interface/decide/CMakeLists.txt     | 2 ++
 pybind_interface/hip/CMakeLists.txt        | 2 ++
 pybind_interface/sse/CMakeLists.txt        | 2 ++
 8 files changed, 16 insertions(+)

diff --git a/pybind_interface/avx2/CMakeLists.txt b/pybind_interface/avx2/CMakeLists.txt
index 18047a1e..ab7d6d83 100644
--- a/pybind_interface/avx2/CMakeLists.txt
+++ b/pybind_interface/avx2/CMakeLists.txt
@@ -14,3 +14,5 @@ if(APPLE)
 endif()
 INCLUDE(../GetPybind11.cmake)
 pybind11_add_module(qsim_avx2 pybind_main_avx2.cpp)
+
+target_link_libraries(qsim_avx2 PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/avx512/CMakeLists.txt b/pybind_interface/avx512/CMakeLists.txt
index 5875c08c..9d635dd0 100644
--- a/pybind_interface/avx512/CMakeLists.txt
+++ b/pybind_interface/avx512/CMakeLists.txt
@@ -16,3 +16,5 @@ endif()
 
 INCLUDE(../GetPybind11.cmake)
 pybind11_add_module(qsim_avx512 pybind_main_avx512.cpp)
+
+target_link_libraries(qsim_avx512 PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/basic/CMakeLists.txt b/pybind_interface/basic/CMakeLists.txt
index 36bdfc8d..ab9cd94b 100644
--- a/pybind_interface/basic/CMakeLists.txt
+++ b/pybind_interface/basic/CMakeLists.txt
@@ -16,3 +16,5 @@ endif()
 
 INCLUDE(../GetPybind11.cmake)
 pybind11_add_module(qsim_basic pybind_main_basic.cpp)
+
+target_link_libraries(qsim_basic PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/cuda/CMakeLists.txt b/pybind_interface/cuda/CMakeLists.txt
index 29fd367c..89df1948 100644
--- a/pybind_interface/cuda/CMakeLists.txt
+++ b/pybind_interface/cuda/CMakeLists.txt
@@ -26,3 +26,5 @@ set_target_properties(qsim_cuda PROPERTIES
        SUFFIX "${PYTHON_MODULE_EXTENSION}"
 )
 set_source_files_properties(pybind_main_cuda.cpp PROPERTIES LANGUAGE CUDA)
+
+target_link_libraries(qsim_cuda PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/custatevec/CMakeLists.txt b/pybind_interface/custatevec/CMakeLists.txt
index 1c34a1d9..77fe6631 100644
--- a/pybind_interface/custatevec/CMakeLists.txt
+++ b/pybind_interface/custatevec/CMakeLists.txt
@@ -44,3 +44,5 @@ set_target_properties(qsim_custatevec PROPERTIES
        SUFFIX "${PYTHON_MODULE_EXTENSION}"
 )
 set_source_files_properties(pybind_main_custatevec.cpp PROPERTIES LANGUAGE CUDA)
+
+target_link_libraries(qsim_custatevec PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/decide/CMakeLists.txt b/pybind_interface/decide/CMakeLists.txt
index b3dac71c..15d51707 100644
--- a/pybind_interface/decide/CMakeLists.txt
+++ b/pybind_interface/decide/CMakeLists.txt
@@ -44,3 +44,5 @@ elseif(has_hipcc)
 else()
     pybind11_add_module(qsim_decide decide.cpp)
 endif()
+
+target_link_libraries(qsim_decide PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/hip/CMakeLists.txt b/pybind_interface/hip/CMakeLists.txt
index b56687ab..8d540c90 100644
--- a/pybind_interface/hip/CMakeLists.txt
+++ b/pybind_interface/hip/CMakeLists.txt
@@ -22,3 +22,5 @@ set_target_properties(qsim_hip PROPERTIES
        SUFFIX "${PYTHON_MODULE_EXTENSION}"
 )
 set_source_files_properties(pybind_main_hip.cpp PROPERTIES LANGUAGE HIP)
+
+target_link_libraries(qsim_hip PUBLIC OpenMP::OpenMP_CXX)
diff --git a/pybind_interface/sse/CMakeLists.txt b/pybind_interface/sse/CMakeLists.txt
index 9be08400..a9123a5a 100644
--- a/pybind_interface/sse/CMakeLists.txt
+++ b/pybind_interface/sse/CMakeLists.txt
@@ -16,3 +16,5 @@ endif()
 
 INCLUDE(../GetPybind11.cmake)
 pybind11_add_module(qsim_sse pybind_main_sse.cpp)
+
+target_link_libraries(qsim_sse PUBLIC OpenMP::OpenMP_CXX)