*
* Copyright (c) 1991-2000, University of Groningen, The Netherlands.
* Copyright (c) 2001-2004, The GROMACS development team.
- * Copyright (c) 2013, by the GROMACS development team, led by
+ * Copyright (c) 2013,2014, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
* To help us fund GROMACS development, we humbly ask that you cite
* the research papers on the package. Check out http://www.gromacs.org.
*/
-#ifdef HAVE_CONFIG_H
-#include <config.h>
-#endif
+#include "gmxpre.h"
+
+#include "gromacs/legacyheaders/nrnb.h"
+#include <stdlib.h>
#include <string.h>
-#include "types/commrec.h"
-#include "sysstuff.h"
-#include "gmx_fatal.h"
-#include "names.h"
-#include "macros.h"
-#include "nrnb.h"
-#include "main.h"
-#include "smalloc.h"
+
+#include "gromacs/legacyheaders/macros.h"
+#include "gromacs/legacyheaders/names.h"
+#include "gromacs/legacyheaders/types/commrec.h"
+#include "gromacs/utility/smalloc.h"
typedef struct {
const char *name;
* - GPU always does exclusions, which requires 2-4 flops, but as invsqrt
* is always counted as 6 flops, this roughly compensates.
*/
- { "NxN RF Elec. + VdW [F]", 38 }, /* nbnxn kernel LJ+RF, no ener */
- { "NxN RF Elec. + VdW [V&F]", 54 },
- { "NxN QSTab Elec. + VdW [F]", 41 }, /* nbnxn kernel LJ+tab, no en */
- { "NxN QSTab Elec. + VdW [V&F]", 59 },
- { "NxN Ewald Elec. + VdW [F]", 66 }, /* nbnxn kernel LJ+Ewald, no en */
- { "NxN Ewald Elec. + VdW [V&F]", 107 },
- { "NxN VdW [F]", 33 }, /* nbnxn kernel LJ, no ener */
- { "NxN VdW [V&F]", 43 },
+ { "NxN RF Elec. + LJ [F]", 38 }, /* nbnxn kernel LJ+RF, no ener */
+ { "NxN RF Elec. + LJ [V&F]", 54 },
+ { "NxN QSTab Elec. + LJ [F]", 41 }, /* nbnxn kernel LJ+tab, no en */
+ { "NxN QSTab Elec. + LJ [V&F]", 59 },
+ { "NxN Ewald Elec. + LJ [F]", 66 }, /* nbnxn kernel LJ+Ewald, no en */
+ { "NxN Ewald Elec. + LJ [V&F]", 107 },
+ { "NxN LJ [F]", 33 }, /* nbnxn kernel LJ, no ener */
+ { "NxN LJ [V&F]", 43 },
{ "NxN RF Electrostatics [F]", 31 }, /* nbnxn kernel RF, no ener */
{ "NxN RF Electrostatics [V&F]", 36 },
{ "NxN QSTab Elec. [F]", 34 }, /* nbnxn kernel tab, no ener */
{ "NxN QSTab Elec. [V&F]", 41 },
{ "NxN Ewald Elec. [F]", 61 }, /* nbnxn kernel Ewald, no ener */
{ "NxN Ewald Elec. [V&F]", 84 },
+ /* The switch function flops should be added to the LJ kernels above */
+ { "NxN LJ add F-switch [F]", 12 }, /* extra cost for LJ F-switch */
+ { "NxN LJ add F-switch [V&F]", 22 },
+ { "NxN LJ add P-switch [F]", 27 }, /* extra cost for LJ P-switch */
+ { "NxN LJ add P-switch [V&F]", 20 },
+ { "NxN LJ add LJ Ewald [F]", 36 }, /* extra cost for LJ Ewald */
+ { "NxN LJ add LJ Ewald [V&F]", 33 },
{ "1,4 nonbonded interactions", 90 },
{ "Born radii (Still)", 47 },
{ "Born radii (HCT/OBC)", 183 },
#endif
}
+/* Returns in enr is the index of a full nbnxn VdW kernel */
+static gmx_bool nrnb_is_nbnxn_vdw_kernel(int enr)
+{
+ return (enr >= eNR_NBNXN_LJ_RF && enr <= eNR_NBNXN_LJ_E);
+}
+
+/* Returns in enr is the index of an nbnxn kernel addition (LJ modification) */
+static gmx_bool nrnb_is_nbnxn_kernel_addition(int enr)
+{
+ return (enr >= eNR_NBNXN_ADD_LJ_FSW && enr <= eNR_NBNXN_ADD_LJ_EWALD_E);
+}
+
void print_flop(FILE *out, t_nrnb *nrnb, double *nbfs, double *mflop)
{
- int i;
+ int i, j;
double mni, frac, tfrac, tflop;
const char *myline = "-----------------------------------------------------------------------------";
for (i = 0; (i < eNRNB); i++)
{
mni = 1e-6*nrnb->n[i];
- *mflop += mni*nbdata[i].flop;
- frac = 100.0*mni*nbdata[i].flop/tflop;
- tfrac += frac;
- if (out && mni != 0)
+ /* Skip empty entries and nbnxn additional flops,
+ * which have been added to the kernel entry.
+ */
+ if (mni > 0 && !nrnb_is_nbnxn_kernel_addition(i))
{
- fprintf(out, " %-32s %16.6f %15.3f %6.1f\n",
- nbdata[i].name, mni, mni*nbdata[i].flop, frac);
+ int flop;
+
+ flop = nbdata[i].flop;
+ if (nrnb_is_nbnxn_vdw_kernel(i))
+ {
+ /* Possibly add the cost of an LJ switch/Ewald function */
+ for (j = eNR_NBNXN_ADD_LJ_FSW; j <= eNR_NBNXN_ADD_LJ_EWALD; j += 2)
+ {
+ int e_kernel_add;
+
+ /* Select the force or energy flop count */
+ e_kernel_add = j + ((i - eNR_NBNXN_LJ_RF) % 2);
+
+ if (nrnb->n[e_kernel_add] > 0)
+ {
+ flop += nbdata[e_kernel_add].flop;
+ }
+ }
+ }
+ *mflop += mni*flop;
+ frac = 100.0*mni*flop/tflop;
+ tfrac += frac;
+ if (out != NULL)
+ {
+ fprintf(out, " %-32s %16.6f %15.3f %6.1f\n",
+ nbdata[i].name, mni, mni*flop, frac);
+ }
}
}
if (out)
fprintf(out, " %-32s %16s %15.3f %6.1f\n",
"Total", "", *mflop, tfrac);
fprintf(out, "%s\n\n", myline);
+
+ if (nrnb->n[eNR_NBKERNEL_GENERIC] > 0)
+ {
+ fprintf(out,
+ "WARNING: Using the slow generic C kernel. This is fine if you are\n"
+ "comparing different implementations or MD software. Routine\n"
+ "simulations should use a different non-bonded setup for much better\n"
+ "performance.\n\n");
+ }
}
}
fprintf(log, "\nDetailed load balancing info in percentage of average\n");
- fprintf(log, " Type NODE:");
+ fprintf(log, " Type RANK:");
for (i = 0; (i < cr->nnodes); i++)
{
fprintf(log, "%3d ", i);
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff