Use non-temporal instructions also on ARM processors

configure.ac

@@ -101,6 +101,18 @@ fi
 AC_CHECK_DECLS([_mm_stream_si32], [have_mm_stream_si32=yes], [have_mm_stream_si32=no], [[#include <x86intrin.h>]])
 AM_CONDITIONAL(HAVE_MM_STREAM_SI32, test x$have_mm_stream_si32 = xyes)
 
+AC_CHECK_DECLS([vst1q_lane_s32, vdupq_n_s32], [have_arm_nt_store=yes], [have_arm_nt_store=no], [[#include <arm_neon.h>]])
+AM_CONDITIONAL(HAVE_ARM_NT_STORE, test x$have_arm_nt_store = xyes)
+if test x$have_arm_nt_store = xyes; then
+    AC_DEFINE([HAVE_ARM_NT_STORE], [1], [Define to 1 if you have non-temporal store instructions on ARM architecture])
+fi
+
+have_non_temporal_store=no
+if test x$have_mm_stream_si32 = xyes -o x$have_arm_nt_store = xyes; then
+    AC_DEFINE([HAVE_NON_TEMPORAL_STORE], [1], [Define to 1 if you have non-temporal store instructions])
+    have_non_temporal_store=yes
+fi
+
 
 # Compiler
 AC_PROG_CC
@@ -165,11 +177,12 @@ AC_OUTPUT
 AC_MSG_NOTICE([
 
 Configuration summary:
-    CPU:    $host_cpu
-    MPI:    $have_mpi ($mpicc_path)
-    likwid: $have_likwid
-    StarPU: $have_starpu
-    MKL:    $have_mkl
-    AVX:    $available_avx
-    Python: $install_python_package
+    CPU:      $host_cpu
+    MPI:      $have_mpi ($mpicc_path)
+    likwid:   $have_likwid
+    StarPU:   $have_starpu
+    MKL:      $have_mkl
+    AVX:      $available_avx
+    Python:   $install_python_package
+    nt store: $have_non_temporal_store
 ])

src/openmp_memset.c

@@ -6,8 +6,12 @@
 
 #include "config.h"
 
-#if defined(HAVE_DECL__MM_STREAM_SI32)
+#if HAVE_NON_TEMPORAL_STORE
+#if HAVE_DECL__MM_STREAM_SI32
 #include <x86intrin.h>
+#elif HAVE_ARM_NT_STORE
+#include <arm_neon.h>
+#endif
 #endif
 
 #include "cli.h"
@@ -37,6 +41,18 @@ static double* memset_perfs_per_thread_comm;
 volatile int* memset_comm_bench_ended = NULL;
 
 
+#if HAVE_NON_TEMPORAL_STORE
+static inline __attribute__((always_inline)) void non_temporal_store(int *p, int a)
+{
+#if HAVE_DECL__MM_STREAM_SI32
+    _mm_stream_si32(p, a);
+#elif HAVE_ARM_NT_STORE
+    vst1q_lane_s32((int32_t *) p, vdupq_n_s32(a), 0);
+#endif
+}
+#endif
+
+
 static void memset_set_comm_bench_ended_ptr(volatile int* _comm_bench_ended)
 {
     memset_comm_bench_ended = _comm_bench_ended;
@@ -201,14 +217,14 @@ static double memset_run_kernel(enum bench_type bench_type)
 	    double last_iter_duration = 0;
 	    do
 	    {
-#if defined(HAVE_DECL__MM_STREAM_SI32)
+#if HAVE_NON_TEMPORAL_STORE
 		if (use_non_temporal)
 		{
 		    PUK_GET_TICK(start_iter_time);
 		    int value = scalar * (nb_iter_per_thread[i]+1);
 		    for (int k = 0; k < array_size; k++)
 		    {
-			_mm_stream_si32(a[i]+k, value);
+			non_temporal_store(a[i]+k, value);
 		    }
 		    PUK_GET_TICK(end_iter_time);
 		}
@@ -236,7 +252,7 @@ static double memset_run_kernel(enum bench_type bench_type)
 	#pragma omp parallel for
 	for (int i = 0; i < nb_threads; i++)
 	{
-#if defined(HAVE_DECL__MM_STREAM_SI32)
+#if HAVE_NON_TEMPORAL_STORE
 	    if (use_non_temporal)
 	    {
 		PUK_GET_TICK(thread_start_times[i]);
@@ -245,7 +261,7 @@ static double memset_run_kernel(enum bench_type bench_type)
 		    int value = scalar * j;
 		    for (int k = 0; k < array_size; k++)
 		    {
-			_mm_stream_si32(a[i]+k, value);
+			non_temporal_store(a[i]+k, value);
 		    }
 		}
 		PUK_GET_TICK(thread_end_times[i]);
@@ -403,7 +419,7 @@ static void memset_release()
 static void memset_man()
 {
     printf("Memset-related options:\n");
-#if defined(HAVE_DECL__MM_STREAM_SI32)
+#if HAVE_NON_TEMPORAL_STORE
     printf("\t--nt\tuse non-temporal stores to bypass the LLC\n");
 #endif
     printf("\t--throughput\tmeasure memory throughput\n");
@@ -417,7 +433,7 @@ static void memset_print_params()
 
 static int memset_parse_arg(char* arg)
 {
-#if defined(HAVE_DECL__MM_STREAM_SI32)
+#if HAVE_NON_TEMPORAL_STORE
     if (strcmp(arg, "--nt") == 0)
     {
 	use_non_temporal = 1;