Merge branch 'metascroy-patch-7' into mlx-delegate

Author: metascroy

kernels/optimized/CMakeLists.txt

@@ -21,6 +21,36 @@ if(NOT EXECUTORCH_ROOT)
   set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../..)
 endif()
 
+# Apply pocketfft patch for ASAN-compatible aligned_alloc on POSIX systems. The
+# portable aligned_alloc in pocketfft stores metadata at ptr[-1], which
+# conflicts with ASAN's heap redzone and causes intermittent bus errors. This
+# patch uses posix_memalign instead, which is ASAN-compatible.
+set(POCKETFFT_DIR "${EXECUTORCH_ROOT}/third-party/pocketfft")
+set(POCKETFFT_PATCH
+    "${CMAKE_CURRENT_SOURCE_DIR}/patches/pocketfft_aligned_alloc.patch"
+)
+if(EXISTS "${POCKETFFT_PATCH}" AND EXISTS "${POCKETFFT_DIR}")
+  execute_process(
+    COMMAND git apply --check ${POCKETFFT_PATCH}
+    WORKING_DIRECTORY ${POCKETFFT_DIR}
+    RESULT_VARIABLE PATCH_CHECK_RESULT
+    ERROR_QUIET
+  )
+  if(PATCH_CHECK_RESULT EQUAL 0)
+    message(STATUS "Applying pocketfft aligned_alloc patch...")
+    execute_process(
+      COMMAND git apply ${POCKETFFT_PATCH}
+      WORKING_DIRECTORY ${POCKETFFT_DIR}
+      RESULT_VARIABLE PATCH_RESULT
+    )
+    if(NOT PATCH_RESULT EQUAL 0)
+      message(WARNING "Failed to apply pocketfft patch")
+    endif()
+  else()
+    message(STATUS "pocketfft patch already applied or not applicable")
+  endif()
+endif()
+
 set(_common_compile_options
     $<$<CXX_COMPILER_ID:MSVC>:/wd4996>
     $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-Wno-deprecated-declarations>

kernels/optimized/patches/pocketfft_aligned_alloc.patch

@@ -0,0 +1,36 @@
+diff --git a/pocketfft_hdronly.h b/pocketfft_hdronly.h
+index 57db7d7..e831ac0 100644
+--- a/pocketfft_hdronly.h
++++ b/pocketfft_hdronly.h
+@@ -162,8 +162,20 @@ template<> struct VLEN<double> { static constexpr size_t val=2; };
+ // std::aligned_alloc is a bit cursed ... it doesn't exist on MacOS < 10.15
+ // and in musl, and other OSes seem to have even more peculiarities.
+ // Let's unconditionally work around it for now.
+-# if 0
+-//#if (__cplusplus >= 201703L) && (!defined(__MINGW32__)) && (!defined(_MSC_VER)) && (__MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_15)
++#if defined(__APPLE__) || defined(__unix__)
++// Use posix_memalign on POSIX systems - it is ASAN-compatible.
++// The portable aligned_alloc below stores metadata at ptr[-1], which conflicts
++// with ASAN's heap redzone and causes intermittent bus errors.
++inline void *aligned_alloc(size_t align, size_t size)
++  {
++  void *ptr = nullptr;
++  if (posix_memalign(&ptr, align, size) != 0)
++    throw std::bad_alloc();
++  return ptr;
++  }
++inline void aligned_dealloc(void *ptr)
++    { free(ptr); }
++#elif (__cplusplus >= 201703L) && (!defined(__MINGW32__)) && (!defined(_MSC_VER))
+ inline void *aligned_alloc(size_t align, size_t size)
+   {
+   // aligned_alloc() requires that the requested size is a multiple of "align"
+@@ -173,7 +185,7 @@ inline void *aligned_alloc(size_t align, size_t size)
+   }
+ inline void aligned_dealloc(void *ptr)
+     { free(ptr); }
+-#else // portable emulation
++#else // portable emulation (NOT ASAN-compatible - stores metadata at ptr[-1])
+ inline void *aligned_alloc(size_t align, size_t size)
+   {
+   align = std::max(align, alignof(max_align_t));