Allocate aligned memory

Description

For runs on the CPU, the memory for the batches of the wavefunctions in the states is now allocated in an aligned way. The alignment is chosen according to the vectorization recognized at configure time. This improves vectorized instructions, especially for the finite difference kernels where vector instuction intrinsics are used.

This MR depends on MR !669 (merged).

News snippet

Improve vectorization by allocating aligned memory.

Checklist

• I have checked that my code follows the Octopus coding standards

src/grid/allocate_special.c

@@ -19,6 +19,8 @@
 */
 #include <stdlib.h>
 #include <config.h>
+#include <stdio.h>
+#include "vectors.h"
 
 #ifdef HAVE_CUDA
 #include <cuda_runtime.h>
@@ -42,15 +44,23 @@ void *my_allocate(int size_bytes) {
 #ifdef DEBUG_ALLOC
   printf("Allocating %d bytes, unpinned.\n", (unsigned int)size_bytes);
 #endif
-  return malloc(size_bytes);
-#else
+  void *aligned;
+  int status;
+  // align on vector size to improve vectorization
+  status = posix_memalign(&aligned, (size_t)sizeof(double)*VEC_SIZE, (unsigned int)size_bytes);
+  if (status != 0) {
+    printf("Error allocating aligned memory!\n");
+    return NULL;
+  }
+  return aligned;
+#else // HAVE_CUDA
 #ifdef DEBUG_ALLOC
   printf("Allocating %d bytes, pinned.\n", (unsigned int)size_bytes);
 #endif
   void *pinned;
   checkCuda(cudaMallocHost(&pinned, (unsigned int)size_bytes));
   return pinned;
-#endif
+#endif // HAVE_CUDA
 }
 
 void *dallocate_special(int size) {