Merge remote-tracking branch 'gerrit/release-4-6'

[alexxy/gromacs.git] / src / gromacs / mdlib / forcerec.c

diff --git a/src/gromacs/mdlib/forcerec.c b/src/gromacs/mdlib/forcerec.c
index bf53578cc6f77fdb8edb36fa5844d9df95b17a6a..4543deb383d7bc411dddc213c95d7bedbd3de6cb 100644 (file)
--- a/src/gromacs/mdlib/forcerec.c
+++ b/src/gromacs/mdlib/forcerec.c
@@ -61,7 +61,7 @@
 #include "qmmm.h"
 #include "copyrite.h"
 #include "mtop_util.h"
-
+#include "gmx_detectcpu.h"
 
 #ifdef _MSC_VER
 /* MSVC definition for __cpuid() */
@@ -1240,56 +1240,6 @@ gmx_bool can_use_allvsall(const t_inputrec *ir, const gmx_mtop_t *mtop,
 }
 
 
-/* Return 1 if SSE2 support is present, otherwise 0. */
-static int 
-forcerec_check_sse2()
-{
-#if ( defined(GMX_IA32_SSE2) || defined(GMX_X86_64_SSE2) || defined(GMX_IA32_SSE) || defined(GMX_X86_64_SSE)|| defined(GMX_SSE2) )
-       unsigned int level;
-       unsigned int _eax,_ebx,_ecx,_edx;
-       int status;
-       int CPUInfo[4];
-       
-       level = 1;
-#ifdef _MSC_VER
-       __cpuid(CPUInfo,1);
-       
-       _eax=CPUInfo[0];
-       _ebx=CPUInfo[1];
-       _ecx=CPUInfo[2];
-       _edx=CPUInfo[3];
-       
-#elif defined(__x86_64__)
-       /* GCC 64-bit inline asm */
-       __asm__ ("push %%rbx\n\tcpuid\n\tpop %%rbx\n"                 \
-                        : "=a" (_eax), "=S" (_ebx), "=c" (_ecx), "=d" (_edx) \
-                        : "0" (level));
-#elif defined(__i386__)
-       __asm__ ("push %%ebx\n\tcpuid\n\tpop %%ebx\n"                 \
-                        : "=a" (_eax), "=S" (_ebx), "=c" (_ecx), "=d" (_edx) \
-                        : "0" (level));
-#else
-       _eax=_ebx=_ecx=_edx=0;
-#endif
-    
-       /* Features:                                                                                                       
-        *                                                                                                                 
-        * SSE      Bit 25 of edx should be set                                                                            
-        * SSE2     Bit 26 of edx should be set                                                                            
-        * SSE3     Bit  0 of ecx should be set                                                                            
-        * SSE4.1   Bit 19 of ecx should be set                                                                            
-        */
-       status =  (_edx & (1 << 26)) != 0;
-    
-#else
-        int status = 0;
-#endif
-       /* Return SSE2 status */
-       return status;
-}
-
-
-
 
 void init_forcerec(FILE *fp,
                    const output_env_t oenv,
@@ -1318,6 +1268,16 @@ void init_forcerec(FILE *fp,
     t_nblists *nbl;
     int     *nm_ind,egp_flags;
     
+    gmx_detectcpu(&fr->cpu_information);
+    if(MASTER(cr))
+    {
+        /* Only print warnings from master */
+        gmx_detectcpu_check_acceleration(fr->cpu_information,fp);
+    }
+
+    /* By default we turn acceleration on, but it might be turned off further down... */
+    fr->use_acceleration = TRUE;
+
     fr->bDomDec = DOMAINDECOMP(cr);
 
     natoms = mtop->natoms;
@@ -1425,21 +1385,14 @@ void init_forcerec(FILE *fp,
         bNoSolvOpt         = TRUE;
     }
     
-    fr->UseOptimizedKernels = (getenv("GMX_NOOPTIMIZEDKERNELS") == NULL);
-    if(fp && fr->UseOptimizedKernels==FALSE)
+    if( (getenv("GMX_DISABLE_ACCELERATION") != NULL) || (getenv("GMX_NOOPTIMIZEDKERNELS") != NULL) )
     {
+        fr->use_acceleration = FALSE;
         fprintf(fp,
-                "\nFound environment variable GMX_NOOPTIMIZEDKERNELS.\n"
-                "Disabling SSE/SSE2/Altivec/ia64/Power6/Bluegene specific kernels.\n\n");
-    }    
-
-#if ( defined(GMX_IA32_SSE2) || defined(GMX_X86_64_SSE2) || defined(GMX_IA32_SSE) || defined(GMX_X86_64_SSE)|| defined(GMX_SSE2) )
-    if( forcerec_check_sse2() == 0 )
-    {
-        fr->UseOptimizedKernels = FALSE;
+                "\nFound environment variable GMX_DISABLE_ACCELERATION.\n"
+                "Disabling all architecture-specific (e.g. SSE2/SSE4/AVX) routines.\n\n");
     }
-#endif
-    
+
     /* Check if we can/should do all-vs-all kernels */
     fr->bAllvsAll       = can_use_allvsall(ir,mtop,FALSE,NULL,NULL);
     fr->AllvsAll_work   = NULL;
@@ -1842,7 +1795,7 @@ void init_forcerec(FILE *fp,
     init_ns(fp,cr,&fr->ns,fr,mtop,box);
     
     if (cr->duty & DUTY_PP){
-        gmx_setup_kernels(fp,bGenericKernelOnly);
+        gmx_setup_kernels(fp,fr,bGenericKernelOnly);
         if (ir->bAdress)
             gmx_setup_adress_kernels(fp,bGenericKernelOnly);
     }