avoid mdrun crash when rdtscp is not supported

When using rdtscp, mdrun now detects at runtime whether the CPU supports
this instruction and if this is not the case, it issues a fatal error
and instructs the user to recompile mdrun for the compute host. Note
that this will happen rarely, only when cross-compiling from a newer
host for a rather old one.

Additionally, when the user manually picks AVX, we also turn on RDTSCP
as all AVX-capable CPUs support it.

Also made CMake advanced cache option for GMX_USE_RDTSCP. This replaces
the previously hidden GMX_DISTRIBUTABLE_BUILD option.

Fixes #1428

Change-Id: I8bc884ef9ea8ea4661626b60490182ae2b302648

diff --git a/CMakeLists.txt b/CMakeLists.txt
index b95dab8befb68a310dbd108feac06bc985d044b0..535ec6d3a750075f01cb01dbfab42f9a41a55f78 100644 (file)
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -610,10 +610,26 @@ gmx_test_inline_asm_gcc_x86(GMX_X86_GCC_INLINE_ASM)
 
 include(gmxSetBuildInformation)
 gmx_set_build_information()
-if(BUILD_CPU_FEATURES MATCHES "rdtscp" AND NOT GMX_DISTRIBUTABLE_BUILD)
+# Turn on RDTSCP if:
+# - the build system's CPU supports it
+# - the acceleration is set to AVX as all AVX-capable CPUs support AVX (which
+#   at this point means that the user set it).
+# Note: it's better to not use the later set value of GMX_CPU_ACCELERATION because
+# it reflects the system's capability of both compiling and running AVX code.
+# TODO: After merge with 5.0 one could implement a cache variable dependency
+# such that GMX_USE_RDTSCP can change if GMX_CPU_ACCELERATION is changed to AVX
+# after the first cmake pass.
+if (BUILD_CPU_FEATURES MATCHES "rdtscp" OR GMX_CPU_ACCELERATION MATCHES "AVX")
+    set(GMX_USE_RDTSCP_DEFAULT_VALUE ON)
+else()
+    set(GMX_USE_RDTSCP_DEFAULT_VALUE OFF)
+endif()
+option(GMX_USE_RDTSCP "Use RDTSCP for better CPU-based timers (available on recent x86 CPUs; might need to be off when compiling for heterogeneous environments)" ${GMX_USE_RDTSCP_DEFAULT_VALUE})
+mark_as_advanced(GMX_USE_RDTSCP)
+if(GMX_USE_RDTSCP)
     # The timestep counter headers do not include config.h
     add_definitions(-DHAVE_RDTSCP)
-endif(BUILD_CPU_FEATURES MATCHES "rdtscp" AND NOT GMX_DISTRIBUTABLE_BUILD)
+endif()
 
 include(gmxTestFloatFormat)
 gmx_test_float_format(GMX_FLOAT_FORMAT_IEEE754 

diff --git a/src/gmxlib/gmx_detect_hardware.c b/src/gmxlib/gmx_detect_hardware.c
index a1c38df268da84ef4507a5764747d906b106d267..a61585739ba5586e3607e5bb9a5df8a029ae1e94 100644 (file)
--- a/src/gmxlib/gmx_detect_hardware.c
+++ b/src/gmxlib/gmx_detect_hardware.c
@@ -45,6 +45,8 @@
 #include "types/commrec.h"
 #include "gmx_fatal.h"
 #include "gmx_fatal_collective.h"
+#include "md_logging.h"
+#include "gmx_cpuid.h"
 #include "smalloc.h"
 #include "gpu_utils.h"
 #include "statutil.h"
@@ -210,6 +212,40 @@ static void parse_gpu_id_csv_string(const char *idstr, int *nid, int *idlist)
     gmx_incons("Not implemented yet");
 }
 
+/* Give a suitable fatal error or warning if the build configuration
+   and runtime CPU do not match. */
+static void
+check_use_of_rdtscp_on_this_cpu(FILE                *fplog,
+                                const t_commrec     *cr,
+                                const gmx_hw_info_t *hwinfo)
+{
+    gmx_bool bCpuHasRdtscp, bBinaryUsesRdtscp;
+#ifdef HAVE_RDTSCP
+    bBinaryUsesRdtscp = TRUE;
+#else
+    bBinaryUsesRdtscp = FALSE;
+#endif
+
+    bCpuHasRdtscp = gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_RDTSCP);
+
+    if (!bCpuHasRdtscp && bBinaryUsesRdtscp)
+    {
+        gmx_fatal(FARGS, "The %s executable was compiled to use the rdtscp CPU instruction. "
+                  "However, this is not supported by the current hardware and continuing would lead to a crash. "
+                  "Please rebuild GROMACS with the GMX_USE_RDTSCP=OFF CMake option.",
+                  ShortProgram());
+    }
+
+    if (bCpuHasRdtscp && !bBinaryUsesRdtscp)
+    {
+        md_print_warn(cr, fplog, "The current CPU can measure timings more accurately than the code in\n"
+                      "%s was configured to use. This might affect your simulation\n"
+                      "speed as accurate timings are needed for load-balancing.\n"
+                      "Please consider rebuilding %s with the GMX_USE_RDTSCP=OFF CMake option.\n",
+                      ShortProgram(), ShortProgram());
+    }
+}
+
 void gmx_check_hw_runconf_consistency(FILE *fplog,
                                       const gmx_hw_info_t *hwinfo,
                                       const t_commrec *cr,
@@ -252,6 +288,8 @@ void gmx_check_hw_runconf_consistency(FILE *fplog,
              the case! Might not hurt to add an extra check over MPI. */
     gmx_cpuid_acceleration_check(hwinfo->cpuid_info, fplog, SIMMASTER(cr));
 
+    check_use_of_rdtscp_on_this_cpu(fplog, cr, hwinfo);
+
     /* NOTE: this print is only for and on one physical node */
     print_gpu_detection_stats(fplog, &hwinfo->gpu_info, cr);