Merge branch 'release-5-0'

[alexxy/gromacs.git] / src / programs / mdrun / runner.cpp

diff --git a/src/programs/mdrun/runner.cpp b/src/programs/mdrun/runner.cpp
index 10c6cc47fa1712367db2e76beb33c0c930a9a5c2..2e122a897d016f9b7e1c57cb15bd98414944fb7f 100644 (file)
--- a/src/programs/mdrun/runner.cpp
+++ b/src/programs/mdrun/runner.cpp
@@ -309,20 +309,22 @@ static int get_tmpi_omp_thread_division(const gmx_hw_info_t *hwinfo,
          * two CPUs with HT, so we need a limit<16; thus we use 12.
          * A reasonable limit for Intel Sandy and Ivy bridge,
          * not knowing the topology, is 16 threads.
+         * Below we check for Intel and AVX, which for now includes
+         * Sandy/Ivy Bridge, Has/Broadwell. By checking for AVX instead of
+         * model numbers we ensure also future Intel CPUs are covered.
          */
         const int nthreads_omp_always_faster             =  4;
         const int nthreads_omp_always_faster_Nehalem     = 12;
-        const int nthreads_omp_always_faster_SandyBridge = 16;
-        gmx_bool  bIntel_Family6;
+        const int nthreads_omp_always_faster_Intel_AVX   = 16;
+        gmx_bool  bIntelAVX;
 
-        bIntel_Family6 =
+        bIntelAVX =
             (gmx_cpuid_vendor(hwinfo->cpuid_info) == GMX_CPUID_VENDOR_INTEL &&
-             gmx_cpuid_family(hwinfo->cpuid_info) == 6);
+             gmx_cpuid_feature(hwinfo->cpuid_info, GMX_CPUID_FEATURE_X86_AVX));
 
         if (nthreads_tot <= nthreads_omp_always_faster ||
-            (bIntel_Family6 &&
-             ((gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_Nehalem && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
-              (gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_SandyBridge && nthreads_tot <= nthreads_omp_always_faster_SandyBridge))))
+            ((gmx_cpuid_is_intel_nehalem(hwinfo->cpuid_info) && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
+             (bIntelAVX && nthreads_tot <= nthreads_omp_always_faster_Intel_AVX)))
         {
             /* Use pure OpenMP parallelization */
             nthreads_tmpi = 1;