Require BML_USE_POSIX_MEMALIGN to use posix_memalign()

Using the HAVE_POSIX_MEMALIGN code path results in ~1.7x slowdown of an MD code which frequently allocates arrays.

Require a new macro BML_USE_POSIX_MEMALIGN to be defined, to use the code path. Define by default

CMakeLists.txt

@@ -671,9 +671,14 @@ endif()
 
 add_definitions(-D_POSIX_C_SOURCE=200112L)
 
-check_function_exists(posix_memalign HAVE_POSIX_MEMALIGN)
-if(HAVE_POSIX_MEMALIGN)
-  add_definitions(-DHAVE_POSIX_MEMALIGN)
+option(BML_POSIX_MEMALIGN "Use posix_memalign() for allocations if available" ON)
+
+if(BML_POSIX_MEMALIGN)
+  check_function_exists(posix_memalign HAVE_POSIX_MEMALIGN)
+  if(HAVE_POSIX_MEMALIGN)
+    add_definitions(-DBML_USE_POSIX_MEMALIGN)
+    message(STATUS "Will use posix_memalign() for allocations")
+  endif()
 endif()
 
 set(INTEL_OPT FALSE

build.sh

@@ -83,6 +83,7 @@ EOF
     echo "EXTRA_FFLAGS               Extra fortran flags         (default is '${EXTRA_FFLAGS}')"
     echo "EXTRA_LINK_FLAGS           Add extra link flags        (default is '${EXTRA_LINK_FLAGS}')"
     echo "BML_OMP_OFFLOAD            {yes,no}                    (default is ${BML_OMP_OFFLOAD})"
+    echo "BML_POSIX_MEMALIGN         Use if available {yes,no}   (default is ${BML_POSIX_MEMALIGN})"
     echo "BML_OFFLOAD_ARCH           {NVIDIA, AMD}               (default is ${BML_OFFLOAD_ARCH})"
     echo "GPU_ARCH                   GPU architecture            (default is '${GPU_ARCH}')"
     echo "BML_CUDA                   Build with CUDA             (default is ${BML_CUDA})"
@@ -139,6 +140,7 @@ set_defaults() {
     : ${FORTRAN_FLAGS:=}
     : ${EXTRA_LINK_FLAGS:=}
     : ${BML_OMP_OFFLOAD:=no}
+    : ${BML_POSIX_MEMALIGN:=yes}
     : ${BML_OFFLOAD_ARCH:=NVIDIA}
     : ${GPU_ARCH:=}
     : ${BML_CUDA:=no}
@@ -232,6 +234,7 @@ configure() {
         -DSCALAPACK_LIBRARIES="${SCALAPACK_LIBRARIES}" \
         -DBML_ELPA="${BML_ELPA}" \
         -DBML_OPENMP="${BML_OPENMP}" \
+        -DBML_POSIX_MEMALIGN="${BML_POSIX_MEMALIGN}" \
         -DMKL_GPU="${MKL_GPU}" \
         -DBML_MPI="${BML_MPI}" \
         -DBML_MPIEXEC_EXECUTABLE="${BML_MPIEXEC_EXECUTABLE}" \

src/C-interface/bml_allocate.c

@@ -53,7 +53,7 @@ bml_allocate_memory(
         __assume_aligned(ptr, MALLOC_ALIGNMENT);
         ptr[i] = 0;
     }
-#elif defined(HAVE_POSIX_MEMALIGN)
+#elif defined(BML_USE_POSIX_MEMALIGN)
     char *ptr;
     posix_memalign((void **) &ptr, MALLOC_ALIGNMENT, size);
 #pragma omp simd
@@ -86,7 +86,7 @@ bml_noinit_allocate_memory(
 {
 #if defined(INTEL_OPT)
     void *ptr = _mm_malloc(size, MALLOC_ALIGNMENT);
-#elif defined(HAVE_POSIX_MEMALIGN)
+#elif defined(BML_USE_POSIX_MEMALIGN)
     void *ptr;
     posix_memalign(&ptr, MALLOC_ALIGNMENT, size);
 #else