Make qsimcirq Apple Silicon (M1, M2, ARM64) compatible and improve build scripts

.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:
@@ -32,7 +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"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --overwrite [email protected] && 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"

.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"
+      CIBW_BEFORE_BUILD_MACOS: "brew install libomp llvm && brew link --overwrite [email protected] && 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"

CMakeLists.txt

@@ -1,24 +1,51 @@
+set(CMAKE_CXX_STANDARD 11)
 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)
+# Set APPLE_ARM to TRUE if running on Apple Silicon
+if(APPLE)
+    execute_process(COMMAND uname -m OUTPUT_VARIABLE OSX_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if (OSX_ARCH STREQUAL "arm64")
+        set(APPLE_ARM TRUE)
+    else()  # x86_64
+        set(APPLE_ARM FALSE)
     endif()
 else()
-    project(qsim LANGUAGES CXX CUDA)
-    ADD_SUBDIRECTORY(pybind_interface/cuda)
-    if(DEFINED ENV{CUQUANTUM_ROOT})
-        ADD_SUBDIRECTORY(pybind_interface/custatevec)
+    set(APPLE_ARM FALSE)
+endif(APPLE)
+
+# Set the project name and language
+if(APPLE)
+    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()
     endif()
 endif()
 
-ADD_SUBDIRECTORY(pybind_interface/sse)
-ADD_SUBDIRECTORY(pybind_interface/avx512)
-ADD_SUBDIRECTORY(pybind_interface/avx2)
+find_package(OpenMP REQUIRED)
+
+# Add subdirectories based on the architecture or available compilers
 ADD_SUBDIRECTORY(pybind_interface/basic)
 ADD_SUBDIRECTORY(pybind_interface/decide)
+if(NOT APPLE_ARM)
+    if(has_nvcc)
+        ADD_SUBDIRECTORY(pybind_interface/cuda)
+        if(DEFINED ENV{CUQUANTUM_ROOT})
+            ADD_SUBDIRECTORY(pybind_interface/custatevec)
+        endif()
+    elseif(has_hipcc)
+        ADD_SUBDIRECTORY(pybind_interface/hip)
+    endif()
+
+    ADD_SUBDIRECTORY(pybind_interface/sse)
+    ADD_SUBDIRECTORY(pybind_interface/avx512)
+    ADD_SUBDIRECTORY(pybind_interface/avx2)
+endif()

pybind_interface/GetPybind11.cmake

@@ -8,6 +8,21 @@ FetchContent_Declare(
 )
 FetchContent_GetProperties(pybind11)
 find_package(pybind11 "${MIN_PYBIND_VERSION}" CONFIG)
+
+if (pybind11_FOUND)
+  message(STATUS "Found pybind11 v${pybind11_VERSION}: ${pybind11_INCLUDE_DIRS}")
+  # The pybind11_add_module doesn't correctly set the CXX_INCLUDES properly if a system pybind11 is found.
+  # Using `include_directories(${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
+  # 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()
+
 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})

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)
@@ -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)

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)
@@ -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)

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)
@@ -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)

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)
@@ -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)

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)
@@ -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)

pybind_interface/decide/CMakeLists.txt

@@ -1,66 +1,48 @@
 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()
+if(WIN32)
+    set(CMAKE_CXX_FLAGS "/O2 /openmp")
 else()
-  project(qsim LANGUAGES CXX CUDA)
+    set(CMAKE_CXX_FLAGS "-O3")
 endif()
 
-IF (WIN32)
-    set(CMAKE_CXX_FLAGS "/O2 /openmp")
-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")
 endif()
 
-INCLUDE(../GetPybind11.cmake)
-
-if(has_nvcc STREQUAL "")
-  if(has_hipcc STREQUAL "")
-    pybind11_add_module(qsim_decide decide.cpp)
-  else()
+include(../GetPybind11.cmake)
+
+# Configure based on the detected platform
+if(has_nvcc)
+    find_package(CUDA REQUIRED)
+    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 "${PYTHON_MODULE_EXTENSION}"
+    )
+    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE CUDA)
+elseif(has_hipcc)
     list(APPEND CMAKE_MODULE_PATH "/opt/rocm/lib/cmake/hip")
     find_package(HIP REQUIRED)
-    find_package(PythonLibs 3.7 REQUIRED)
-
-    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
-
     hip_add_library(qsim_decide MODULE decide.cpp)
-
+    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
+    find_package(Python3 3.7 REQUIRED COMPONENTS Interpreter Development)
+    include_directories(${PYTHON_INCLUDE_DIRS} ${pybind11_SOURCE_DIR}/include)
     set_target_properties(qsim_decide PROPERTIES
-        PREFIX "${PYTHON_MODULE_PREFIX}"
-        SUFFIX "${PYTHON_MODULE_EXTENSION}"
+            PREFIX "${PYTHON_MODULE_PREFIX}"
+            SUFFIX "${PYTHON_MODULE_EXTENSION}"
     )
-    set_source_files_properties(decide.cpp PROPERTIES LANGUAGE HIP)
-  endif()
 else()
-  find_package(PythonLibs 3.7 REQUIRED)
-  find_package(CUDA REQUIRED)
-
-  include_directories(${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()
-
-  set_target_properties(qsim_decide PROPERTIES
-        PREFIX "${PYTHON_MODULE_PREFIX}"
-        SUFFIX "${PYTHON_MODULE_EXTENSION}"
-  )
-  set_source_files_properties(decide.cpp PROPERTIES LANGUAGE CUDA)
+    pybind11_add_module(qsim_decide decide.cpp)
 endif()
+
+target_link_libraries(qsim_decide PUBLIC OpenMP::OpenMP_CXX)

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]);
@@ -34,10 +34,11 @@ enum Instructions { AVX512F = 0, AVX2 = 1, SSE4_1 = 2, BASIC = 3};
 
 int detect_instructions() {
   Instructions instr = BASIC;
-  int info[4];
 
+  #if !defined(__aarch64__) || !defined(__APPLE__)
+  // 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);
@@ -47,14 +48,14 @@ 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);
 }

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)
@@ -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)