temp

.github/workflows/prerelease.yml

@@ -323,7 +323,7 @@ jobs:
         run: |
           python -m pip install --upgrade pip
           pip install pytest pytest-repeat numpy pyarrow
-          python -m pip install .
+          python -m pip install . -vv
       - name: Test Python
         run: pytest scripts/test.py -s -x
 

include/stringzilla/stringzilla.h

@@ -1106,50 +1106,104 @@ SZ_PUBLIC void sz_sort_intro(sz_sequence_t *sequence, sz_sequence_comparator_t l
  *  All of those can be controlled by the user.
  */
 #ifndef SZ_USE_X86_AVX512
-#ifdef __AVX512BW__
+#if defined(__AVX512BW__) //&& SZ_TARGET_X86
 #define SZ_USE_X86_AVX512 1
+#pragma message("SZ_USE_X86_AVX512")
 #else
 #define SZ_USE_X86_AVX512 0
+#pragma message("NOT SZ_USE_X86_AVX512")
 #endif
 #endif
 
 #ifndef SZ_USE_X86_AVX2
-#ifdef __AVX2__
+#if defined(__AVX2__) //&& SZ_TARGET_X86
 #define SZ_USE_X86_AVX2 1
+#pragma message("SZ_USE_X86_AVX2")
 #else
 #define SZ_USE_X86_AVX2 0
+#pragma message("NOT SZ_USE_X86_AVX2")
 #endif
 #endif
 
 #ifndef SZ_USE_ARM_NEON
-#if defined(__ARM_NEON) || defined(_M_ARM64)
+#if (defined(__ARM_NEON) || defined(_M_ARM64))// && SZ_TARGET_ARM
 #define SZ_USE_ARM_NEON 1
+#pragma message("SZ_USE_ARM_NEON")
 #else
 #define SZ_USE_ARM_NEON 0
+#pragma message("NOT SZ_USE_ARM_NEON")
 #endif
 #endif
 
 #ifndef SZ_USE_ARM_SVE
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE)// && SZ_TARGET_ARM
 #define SZ_USE_ARM_SVE 1
+#pragma message("SZ_USE_ARM_SVE")
 #else
 #define SZ_USE_ARM_SVE 0
+#pragma message("NOT SZ_USE_ARM_SVE")
+#endif
+#endif
+
+#if 0
+#ifndef SZ_TARGET_ARM
+#if defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || defined(_M_ARM64)
+#define SZ_TARGET_ARM 1
+#pragma message("SZ_TARGET_ARM")
+#else
+#define SZ_TARGET_ARM 0
+#pragma message("NOT SZ_TARGET_ARM")
+#endif
+#endif
+
+#ifndef SZ_TARGET_X86
+#if defined(__i386__) || defined(__amd64__) || defined(_M_IX86) || defined(_M_AMD64)
+#define SZ_TARGET_X86 1
+#pragma message("SZ_TARGET_X86")
+#else
+#define SZ_TARGET_X86 0
+#pragma message("NOT SZ_TARGET_X86")
+#endif
+#endif
+#endif
+
+#ifndef SZ_TARGET_X86
+#if !(defined(__i386__) || defined(__amd64__) || defined(_M_IX86) || defined(_M_AMD64))
+#undef SZ_USE_X86_AVX2
+#define SZ_USE_X86_AVX2 0
+#undef SZ_USE_X86_AVX512
+#define SZ_USE_X86_AVX512 0
+#pragma message("NOT SZ_TARGET_X86")
+#else
+#pragma message("SZ_TARGET_X86")
+#endif
+#endif
+
+#ifndef SZ_TARGET_ARM
+#if !(defined(__arm__) || defined(__aarch64__) || defined(_M_ARM) || defined(_M_ARM64))
+#undef SZ_USE_ARM_NEON
+#define SZ_USE_ARM_NEON 0
+#undef SZ_USE_ARM_SVE
+#define SZ_USE_ARM_SVE 0
+#pragma message("NOT SZ_TARGET_ARM")
+#else
+#pragma message("SZ_TARGET_ARM")
 #endif
 #endif
 
 /*
  *  Include hardware-specific headers.
  */
-#if SZ_USE_X86_AVX512 || SZ_USE_X86_AVX2
+#if (SZ_USE_X86_AVX512 || SZ_USE_X86_AVX2)// && SZ_TARGET_X86
 #include <immintrin.h>
 #endif // SZ_USE_X86...
-#if SZ_USE_ARM_NEON
+#if SZ_USE_ARM_NEON// && SZ_TARGET_ARM
 #if !defined(_MSC_VER)
 #include <arm_acle.h>
 #endif
 #include <arm_neon.h>
 #endif // SZ_USE_ARM_NEON
-#if SZ_USE_ARM_SVE
+#if SZ_USE_ARM_SVE// && SZ_TARGET_ARM
 #if !defined(_MSC_VER)
 #include <arm_sve.h>
 #endif

pyproject.toml

@@ -116,3 +116,9 @@ environment.SZ_ARM64="1"
 select = "*-macos*_arm64"
 inherit.environment = "append"
 environment.SZ_ARM64="1"
+
+[[tool.cibuildwheel.overrides]]
+select = "*-macos*_universal2"
+inherit.environment = "append"
+environment.SZ_X86_64="1"
+environment.SZ_ARM64="1"

setup.py

@@ -80,16 +80,19 @@ def darwin_settings() -> Tuple[List[str], List[str], List[Tuple[str]]]:
         "-fPIC",  # to enable dynamic dispatch
     ]
 
+    print("system", sysconfig.get_platform())
+
     # Apple Clang doesn't support the `-march=native` argument,
     # so we must pre-set the CPU generation. Technically the last Intel-based Apple
     # product was the 2021 MacBook Pro, which had the "Coffee Lake" architecture.
     # During Universal builds, however, even AVX header cause compilation errors.
-    can_use_avx2 = is_64bit_x86() and sysconfig.get_platform().startswith("universal")
+    is_building_x86 = is_64bit_x86() and "universal" in sysconfig.get_platform()
+    is_building_arm = is_64bit_arm() and "universal" in sysconfig.get_platform()
     macros_args = [
         ("SZ_USE_X86_AVX512", "0"),
-        ("SZ_USE_X86_AVX2", "1" if can_use_avx2 else "0"),
+        ("SZ_USE_X86_AVX2", "1" if is_building_x86 else "0"),
         ("SZ_USE_ARM_SVE", "0"),
-        ("SZ_USE_ARM_NEON", "1" if is_64bit_arm() else "0"),
+        ("SZ_USE_ARM_NEON", "1" if is_building_arm else "0"),
     ]
 
     return compile_args, link_args, macros_args