remove posix_memalign

Author: JacobSzwejbka

backends/cuda/runtime/cuda_backend.cpp

@@ -146,7 +146,7 @@ class ET_EXPERIMENTAL CudaBackend final
     outfile.close();
 
     // Load the lib
-    Result<void*> lib_handle_res = load_library(so_path);
+    Result<void*> lib_handle_res = load_library(so_path.c_str());
     if (!lib_handle_res.ok()) {
       return lib_handle_res.error();
     }

backends/cuda/runtime/platform/platform.h

@@ -12,10 +12,12 @@
 #include <string>
 
 #ifdef _WIN32
+#include <malloc.h>
 #include <windows.h>
 #else // Posix
 #include <dlfcn.h>
 #include <unistd.h>
+#include <cstdlib>
 #endif
 
 namespace executorch {
@@ -26,11 +28,7 @@ executorch::runtime::Result<void*> load_library(const char* path) {
 #ifdef _WIN32
   auto lib_handle = LoadLibrary(path);
   if (lib_handle == NULL) {
-    ET_LOG(
-        Error,
-        "Failed to load %s with error: %lu",
-        path,
-        GetLastError());
+    ET_LOG(Error, "Failed to load %s with error: %lu", path, GetLastError());
     return executorch::runtime::Error::AccessFailed;
   }
 
@@ -96,6 +94,22 @@ int32_t get_process_id() {
 #endif
 }
 
+void* aligned_alloc(size_t alignment, size_t size) {
+#ifdef _WIN32
+  return _aligned_malloc(size, alignment);
+#else
+  return std::aligned_alloc(alignment, size);
+#endif
+}
+
+void aligned_free(void* ptr) {
+#ifdef _WIN32
+  _aligned_free(ptr);
+#else
+  std::free(ptr);
+#endif
+}
+
 } // namespace cuda
 } // namespace backends
 } // namespace executorch

backends/cuda/runtime/shims/memory.cpp

@@ -13,7 +13,6 @@
 #include <executorch/backends/cuda/runtime/utils.h>
 #include <executorch/runtime/platform/log.h>
 #include <cstdint>
-#include <cstdlib> // For posix_memalign
 #include <memory>
 #include <unordered_map>
 #include <unordered_set>
@@ -230,15 +229,11 @@ AOTITorchError aoti_torch_empty_strided(
         cudaMallocAsync(&ptr, static_cast<size_t>(nbytes), cudaStreamDefault));
   } else if (device_type == static_cast<int32_t>(SupportedDevices::CPU)) {
     // Ensure 16-byte alignment for CPU memory to match CUDA requirements
-    int result = posix_memalign(&ptr, 16, nbytes);
-    ET_CHECK_OR_RETURN_ERROR(
-        result == 0,
-        MemoryAllocationFailed,
-        "Failed to allocate aligned CPU memory");
+    &ptr = aligned_alloc(16, nbytes);
     ET_CHECK_OR_RETURN_ERROR(
         ptr != nullptr,
         MemoryAllocationFailed,
-        "Failed to call posix_memalign");
+        "Failed to allocate aligned CPU memory");
   } else {
     ET_CHECK_OR_RETURN_ERROR(
         false,
@@ -339,7 +334,7 @@ AOTITorchError aoti_torch_delete_tensor_object(Tensor* tensor) {
                 Internal,
                 "Expected host memory but got managed!")
             // This is CPU memory - free immediately
-            free(data_ptr);
+            aligned_free(data_ptr);
             data_ptr = nullptr;
           }