-/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
+/*
+ * This file is part of the GROMACS molecular simulation package.
*
+ * Copyright (c) 2012,2013,2014, 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.
*
- * This source code is part of
- *
- * G R O M A C S
- *
- * GROningen MAchine for Chemical Simulations
- *
- * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
- * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
- * Copyright (c) 2001-2012, The GROMACS development team,
- * check out http://www.gromacs.org for more information.
- *
- * This program is free software; you can redistribute it and/or
- * modify it under the terms of the GNU General Public License
- * as published by the Free Software Foundation; either version 2
+ * GROMACS is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public License
+ * as published by the Free Software Foundation; either version 2.1
* of the License, or (at your option) any later version.
*
- * If you want to redistribute modifications, please consider that
- * scientific software is very special. Version control is crucial -
- * bugs must be traceable. We will be happy to consider code for
- * inclusion in the official distribution, but derived work must not
- * be called official GROMACS. Details are found in the README & COPYING
- * files - if they are missing, get the official version at www.gromacs.org.
+ * GROMACS is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
*
- * To help us fund GROMACS development, we humbly ask that you cite
- * the papers on the package - you can find them in the top README file.
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with GROMACS; if not, see
+ * http://www.gnu.org/licenses, or write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
*
- * For more info, check our website at http://www.gromacs.org
+ * If you want to redistribute modifications to GROMACS, please
+ * consider that scientific software is very special. Version
+ * control is crucial - bugs must be traceable. We will be happy to
+ * consider code for inclusion in the official distribution, but
+ * derived work must not be called official GROMACS. Details are found
+ * in the README & COPYING files - if they are missing, get the
+ * official version at http://www.gromacs.org.
*
- * And Hey:
- * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
+ * 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 "nbnxn_cuda_data_mgmt.h"
+
+#include "config.h"
-#include <stdlib.h>
-#include <stdio.h>
#include <assert.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
-#include "gmx_fatal.h"
-#include "smalloc.h"
-#include "tables.h"
-#include "typedefs.h"
-#include "types/nb_verlet.h"
-#include "types/interaction_const.h"
-#include "types/force_flags.h"
-#include "../nbnxn_consts.h"
+#include <cuda.h>
+
+#include "gromacs/gmxlib/cuda_tools/cudautils.cuh"
+#include "gromacs/legacyheaders/gmx_detect_hardware.h"
+#include "gromacs/legacyheaders/gpu_utils.h"
+#include "gromacs/legacyheaders/pmalloc_cuda.h"
+#include "gromacs/legacyheaders/tables.h"
+#include "gromacs/legacyheaders/typedefs.h"
+#include "gromacs/legacyheaders/types/enums.h"
+#include "gromacs/legacyheaders/types/force_flags.h"
+#include "gromacs/legacyheaders/types/interaction_const.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_consts.h"
+#include "gromacs/pbcutil/ishift.h"
+#include "gromacs/utility/common.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/fatalerror.h"
+#include "gromacs/utility/smalloc.h"
#include "nbnxn_cuda_types.h"
-#include "../../gmxlib/cuda_tools/cudautils.cuh"
-#include "nbnxn_cuda_data_mgmt.h"
-#include "pmalloc_cuda.h"
-#include "gpu_utils.h"
static bool bUseCudaEventBlockingSync = false; /* makes the CPU thread block */
/* Functions from nbnxn_cuda.cu */
extern void nbnxn_cuda_set_cacheconfig(cuda_dev_info_t *devinfo);
-extern const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_nbfp_texref();
-extern const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_coulomb_tab_texref();
+extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_texref();
+extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_comb_texref();
+extern const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_coulomb_tab_texref();
+
+/* We should actually be using md_print_warn in md_logging.c,
+ * but we can't include mpi.h in CUDA code.
+ */
+static void md_print_warn(FILE *fplog,
+ const char *fmt, ...)
+{
+ va_list ap;
+
+ if (fplog != NULL)
+ {
+ /* We should only print to stderr on the master node,
+ * in most cases fplog is only set on the master node, so this works.
+ */
+ va_start(ap, fmt);
+ fprintf(stderr, "\n");
+ vfprintf(stderr, fmt, ap);
+ fprintf(stderr, "\n");
+ va_end(ap);
+
+ va_start(ap, fmt);
+ fprintf(fplog, "\n");
+ vfprintf(fplog, fmt, ap);
+ fprintf(fplog, "\n");
+ va_end(ap);
+ }
+}
+
/* Fw. decl. */
static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb);
and the table GPU array. If called with an already allocated table,
it just re-uploads the table.
*/
-static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp)
+static void init_ewald_coulomb_force_table(cu_nbparam_t *nbp,
+ const cuda_dev_info_t *dev_info)
{
float *ftmp, *coul_tab;
- int tabsize;
- double tabscale;
- cudaError_t stat;
+ int tabsize;
+ double tabscale;
+ cudaError_t stat;
tabsize = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
/* Subtract 2 iso 1 to avoid access out of range due to rounding */
pmalloc((void**)&ftmp, tabsize*sizeof(*ftmp));
table_spline3_fill_ewald_lr(ftmp, NULL, NULL, tabsize,
- 1/tabscale, nbp->ewald_beta);
+ 1/tabscale, nbp->ewald_beta, v_q_ewald_lr);
/* If the table pointer == NULL the table is generated the first time =>
the array pointer will be saved to nbparam and the texture is bound.
nbp->coulomb_tab = coul_tab;
- cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
- stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
- coul_tab, &cd, tabsize*sizeof(*coul_tab));
- CU_RET_ERR(stat, "cudaBindTexture on coul_tab failed");
+#ifdef TEXOBJ_SUPPORTED
+ /* Only device CC >= 3.0 (Kepler and later) support texture objects */
+ if (dev_info->prop.major >= 3)
+ {
+ cudaResourceDesc rd;
+ memset(&rd, 0, sizeof(rd));
+ rd.resType = cudaResourceTypeLinear;
+ rd.res.linear.devPtr = nbp->coulomb_tab;
+ rd.res.linear.desc.f = cudaChannelFormatKindFloat;
+ rd.res.linear.desc.x = 32;
+ rd.res.linear.sizeInBytes = tabsize*sizeof(*coul_tab);
+
+ cudaTextureDesc td;
+ memset(&td, 0, sizeof(td));
+ td.readMode = cudaReadModeElementType;
+ stat = cudaCreateTextureObject(&nbp->coulomb_tab_texobj, &rd, &td, NULL);
+ CU_RET_ERR(stat, "cudaCreateTextureObject on coulomb_tab_texobj failed");
+ }
+ else
+#endif
+ {
+ GMX_UNUSED_VALUE(dev_info);
+ cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
+ stat = cudaBindTexture(NULL, &nbnxn_cuda_get_coulomb_tab_texref(),
+ coul_tab, &cd, tabsize*sizeof(*coul_tab));
+ CU_RET_ERR(stat, "cudaBindTexture on coulomb_tab_texref failed");
+ }
}
cu_copy_H2D(coul_tab, ftmp, tabsize*sizeof(*coul_tab));
cudaError_t stat;
ad->ntypes = ntypes;
- stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
+ stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
ad->bShiftVecUploaded = false;
ad->nalloc = -1;
}
+/*! Selects the Ewald kernel type, analytical on SM 3.0 and later, tabulated on
+ earlier GPUs, single or twin cut-off. */
+static int pick_ewald_kernel_type(bool bTwinCut,
+ const cuda_dev_info_t *dev_info)
+{
+ bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
+ int kernel_type;
+
+ /* Benchmarking/development environment variables to force the use of
+ analytical or tabulated Ewald kernel. */
+ bForceAnalyticalEwald = (getenv("GMX_CUDA_NB_ANA_EWALD") != NULL);
+ bForceTabulatedEwald = (getenv("GMX_CUDA_NB_TAB_EWALD") != NULL);
+
+ if (bForceAnalyticalEwald && bForceTabulatedEwald)
+ {
+ gmx_incons("Both analytical and tabulated Ewald CUDA non-bonded kernels "
+ "requested through environment variables.");
+ }
+
+ /* By default, on SM 3.0 and later use analytical Ewald, on earlier tabulated. */
+ if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
+ {
+ bUseAnalyticalEwald = true;
+
+ if (debug)
+ {
+ fprintf(debug, "Using analytical Ewald CUDA kernels\n");
+ }
+ }
+ else
+ {
+ bUseAnalyticalEwald = false;
+
+ if (debug)
+ {
+ fprintf(debug, "Using tabulated Ewald CUDA kernels\n");
+ }
+ }
+
+ /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
+ forces it (use it for debugging/benchmarking only). */
+ if (!bTwinCut && (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL))
+ {
+ kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA : eelCuEWALD_TAB;
+ }
+ else
+ {
+ kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA_TWIN : eelCuEWALD_TAB_TWIN;
+ }
+
+ return kernel_type;
+}
+
+/*! Copies all parameters related to the cut-off from ic to nbp */
+static void set_cutoff_parameters(cu_nbparam_t *nbp,
+ const interaction_const_t *ic)
+{
+ nbp->ewald_beta = ic->ewaldcoeff_q;
+ nbp->sh_ewald = ic->sh_ewald;
+ nbp->epsfac = ic->epsfac;
+ nbp->two_k_rf = 2.0 * ic->k_rf;
+ nbp->c_rf = ic->c_rf;
+ nbp->rvdw_sq = ic->rvdw * ic->rvdw;
+ nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
+ nbp->rlist_sq = ic->rlist * ic->rlist;
+
+ nbp->sh_lj_ewald = ic->sh_lj_ewald;
+ nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
+
+ nbp->rvdw_switch = ic->rvdw_switch;
+ nbp->dispersion_shift = ic->dispersion_shift;
+ nbp->repulsion_shift = ic->repulsion_shift;
+ nbp->vdw_switch = ic->vdw_switch;
+}
+
/*! Initializes the nonbonded parameter data structure. */
-static void init_nbparam(cu_nbparam_t *nbp,
+static void init_nbparam(cu_nbparam_t *nbp,
const interaction_const_t *ic,
- const nonbonded_verlet_t *nbv)
+ const nbnxn_atomdata_t *nbat,
+ const cuda_dev_info_t *dev_info)
{
cudaError_t stat;
- int ntypes, nnbfp;
+ int ntypes, nnbfp, nnbfp_comb;
+
+ ntypes = nbat->ntype;
- ntypes = nbv->grp[0].nbat->ntype;
+ set_cutoff_parameters(nbp, ic);
- nbp->ewald_beta = ic->ewaldcoeff;
- nbp->sh_ewald = ic->sh_ewald;
- nbp->epsfac = ic->epsfac;
- nbp->two_k_rf = 2.0 * ic->k_rf;
- nbp->c_rf = ic->c_rf;
- nbp->rvdw_sq = ic->rvdw * ic->rvdw;
- nbp->rcoulomb_sq= ic->rcoulomb * ic->rcoulomb;
- nbp->rlist_sq = ic->rlist * ic->rlist;
- nbp->sh_invrc6 = ic->sh_invrc6;
+ if (ic->vdwtype == evdwCUT)
+ {
+ switch (ic->vdw_modifier)
+ {
+ case eintmodNONE:
+ case eintmodPOTSHIFT:
+ nbp->vdwtype = evdwCuCUT;
+ break;
+ case eintmodFORCESWITCH:
+ nbp->vdwtype = evdwCuFSWITCH;
+ break;
+ case eintmodPOTSWITCH:
+ nbp->vdwtype = evdwCuPSWITCH;
+ break;
+ default:
+ gmx_incons("The requested VdW interaction modifier is not implemented in the CUDA GPU accelerated kernels!");
+ break;
+ }
+ }
+ else if (ic->vdwtype == evdwPME)
+ {
+ if (ic->ljpme_comb_rule == ljcrGEOM)
+ {
+ assert(nbat->comb_rule == ljcrGEOM);
+ nbp->vdwtype = evdwCuEWALDGEOM;
+ }
+ else
+ {
+ assert(nbat->comb_rule == ljcrLB);
+ nbp->vdwtype = evdwCuEWALDLB;
+ }
+ }
+ else
+ {
+ gmx_incons("The requested VdW type is not implemented in the CUDA GPU accelerated kernels!");
+ }
if (ic->eeltype == eelCUT)
{
{
nbp->eeltype = eelCuRF;
}
- else if ((EEL_PME(ic->eeltype) || ic->eeltype==eelEWALD))
+ else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
{
- /* Initially rcoulomb == rvdw, so it's surely not twin cut-off, unless
- forced by the env. var. (used only for benchmarking). */
- if (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL)
- {
- nbp->eeltype = eelCuEWALD;
- }
- else
- {
- nbp->eeltype = eelCuEWALD_TWIN;
- }
+ /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
+ nbp->eeltype = pick_ewald_kernel_type(false, dev_info);
}
else
{
}
/* generate table for PME */
- if (nbp->eeltype == eelCuEWALD)
+ nbp->coulomb_tab = NULL;
+ if (nbp->eeltype == eelCuEWALD_TAB || nbp->eeltype == eelCuEWALD_TAB_TWIN)
{
- nbp->coulomb_tab = NULL;
- init_ewald_coulomb_force_table(nbp);
+ init_ewald_coulomb_force_table(nbp, dev_info);
}
- nnbfp = 2*ntypes*ntypes;
- stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
- CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
- cu_copy_H2D(nbp->nbfp, nbv->grp[0].nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
+ nnbfp = 2*ntypes*ntypes;
+ nnbfp_comb = 2*ntypes;
- cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
- stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
- nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
- CU_RET_ERR(stat, "cudaBindTexture on nbfp failed");
-}
+ stat = cudaMalloc((void **)&nbp->nbfp, nnbfp*sizeof(*nbp->nbfp));
+ CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp");
+ cu_copy_H2D(nbp->nbfp, nbat->nbfp, nnbfp*sizeof(*nbp->nbfp));
-/*! Re-generate the GPU Ewald force table, resets rlist, and update the
- * electrostatic type switching to twin cut-off (or back) if needed. */
-void nbnxn_cuda_pme_loadbal_update_param(nbnxn_cuda_ptr_t cu_nb,
- const interaction_const_t *ic)
-{
- cu_nbparam_t *nbp = cu_nb->nbparam;
- nbp->rlist_sq = ic->rlist * ic->rlist;
- nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
- nbp->ewald_beta = ic->ewaldcoeff;
+ if (ic->vdwtype == evdwPME)
+ {
+ stat = cudaMalloc((void **)&nbp->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
+ CU_RET_ERR(stat, "cudaMalloc failed on nbp->nbfp_comb");
+ cu_copy_H2D(nbp->nbfp_comb, nbat->nbfp_comb, nnbfp_comb*sizeof(*nbp->nbfp_comb));
+ }
- /* When switching to/from twin cut-off, the electrostatics type needs updating.
- (The env. var. that forces twin cut-off is for benchmarking only!) */
- if (ic->rcoulomb == ic->rvdw &&
- getenv("GMX_CUDA_NB_EWALD_TWINCUT") == NULL)
+#ifdef TEXOBJ_SUPPORTED
+ /* Only device CC >= 3.0 (Kepler and later) support texture objects */
+ if (dev_info->prop.major >= 3)
{
- nbp->eeltype = eelCuEWALD;
+ cudaResourceDesc rd;
+ cudaTextureDesc td;
+
+ memset(&rd, 0, sizeof(rd));
+ rd.resType = cudaResourceTypeLinear;
+ rd.res.linear.devPtr = nbp->nbfp;
+ rd.res.linear.desc.f = cudaChannelFormatKindFloat;
+ rd.res.linear.desc.x = 32;
+ rd.res.linear.sizeInBytes = nnbfp*sizeof(*nbp->nbfp);
+
+ memset(&td, 0, sizeof(td));
+ td.readMode = cudaReadModeElementType;
+ stat = cudaCreateTextureObject(&nbp->nbfp_texobj, &rd, &td, NULL);
+ CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_texobj failed");
+
+ if (ic->vdwtype == evdwPME)
+ {
+ memset(&rd, 0, sizeof(rd));
+ rd.resType = cudaResourceTypeLinear;
+ rd.res.linear.devPtr = nbp->nbfp_comb;
+ rd.res.linear.desc.f = cudaChannelFormatKindFloat;
+ rd.res.linear.desc.x = 32;
+ rd.res.linear.sizeInBytes = nnbfp_comb*sizeof(*nbp->nbfp_comb);
+
+ memset(&td, 0, sizeof(td));
+ td.readMode = cudaReadModeElementType;
+ stat = cudaCreateTextureObject(&nbp->nbfp_comb_texobj, &rd, &td, NULL);
+ CU_RET_ERR(stat, "cudaCreateTextureObject on nbfp_comb_texobj failed");
+ }
}
else
+#endif
+ {
+ cudaChannelFormatDesc cd = cudaCreateChannelDesc<float>();
+ stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_texref(),
+ nbp->nbfp, &cd, nnbfp*sizeof(*nbp->nbfp));
+ CU_RET_ERR(stat, "cudaBindTexture on nbfp_texref failed");
+
+ if (ic->vdwtype == evdwPME)
+ {
+ stat = cudaBindTexture(NULL, &nbnxn_cuda_get_nbfp_comb_texref(),
+ nbp->nbfp_comb, &cd, nnbfp_comb*sizeof(*nbp->nbfp_comb));
+ CU_RET_ERR(stat, "cudaBindTexture on nbfp_comb_texref failed");
+ }
+ }
+}
+
+/*! Re-generate the GPU Ewald force table, resets rlist, and update the
+ * electrostatic type switching to twin cut-off (or back) if needed. */
+void nbnxn_cuda_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
+ const interaction_const_t *ic)
+{
+ if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_CUDA)
{
- nbp->eeltype = eelCuEWALD_TWIN;
+ return;
}
+ nbnxn_cuda_ptr_t cu_nb = nbv->cu_nbv;
+ cu_nbparam_t *nbp = cu_nb->nbparam;
- init_ewald_coulomb_force_table(cu_nb->nbparam);
+ set_cutoff_parameters(nbp, ic);
+
+ nbp->eeltype = pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw,
+ cu_nb->dev_info);
+
+ init_ewald_coulomb_force_table(cu_nb->nbparam, cu_nb->dev_info);
}
/*! Initializes the pair list data structure. */
static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
{
cudaError_t stat;
- int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync: cudaEventDefault );
+ int eventflags = ( bUseCudaEventBlockingSync ? cudaEventBlockingSync : cudaEventDefault );
stat = cudaEventCreateWithFlags(&(t->start_atdat), eventflags);
CU_RET_ERR(stat, "cudaEventCreate on start_atdat failed");
t->nb_h2d_t = 0.0;
t->nb_d2h_t = 0.0;
- t->nb_c = 0;
+ t->nb_c = 0;
t->pl_h2d_t = 0.0;
t->pl_h2d_c = 0;
for (i = 0; i < 2; i++)
{
- for(j = 0; j < 2; j++)
+ for (j = 0; j < 2; j++)
{
t->ktime[i][j].t = 0.0;
t->ktime[i][j].c = 0;
}
}
-/* Decide which kernel version to use (default or legacy) based on:
- * - CUDA version
- * - non-bonded kernel selector environment variables
- * - GPU SM version TODO ???
- */
-static int pick_nbnxn_kernel_version()
-{
- bool bLegacyKernel, bDefaultKernel, bCUDA40, bCUDA32;
- char sbuf[STRLEN];
- int kver;
-
- /* legacy kernel (former k2), kept for now for backward compatibility,
- faster than the default with CUDA 3.2/4.0 (TODO: on Kepler?). */
- bLegacyKernel = (getenv("GMX_CUDA_NB_LEGACY") != NULL);
- /* default kernel (former k3). */
- bDefaultKernel = (getenv("GMX_CUDA_NB_DEFAULT") != NULL);
-
- if ((unsigned)(bLegacyKernel + bDefaultKernel) > 1)
- {
- gmx_fatal(FARGS, "Multiple CUDA non-bonded kernels requested; to manually pick a kernel set only one \n"
- "of the following environment variables: \n"
- "GMX_CUDA_NB_DEFAULT, GMX_CUDA_NB_LEGACY");
- }
-
- bCUDA32 = bCUDA40 = false;
-#if CUDA_VERSION == 3200
- bCUDA32 = true;
- sprintf(sbuf, "3.2");
-#elif CUDA_VERSION == 4000
- bCUDA40 = true;
- sprintf(sbuf, "4.0");
-#endif
-
- /* default is default ;) */
- kver = eNbnxnCuKDefault;
-
- if (bCUDA32 || bCUDA40)
- {
- /* use legacy kernel unless something else is forced by an env. var */
- if (bDefaultKernel)
- {
- fprintf(stderr,
- "\nNOTE: CUDA %s compilation detected; with this compiler version the legacy\n"
- " non-bonded kernels perform best. However, the default kernels were\n"
- " selected by the GMX_CUDA_NB_DEFAULT environment variable.\n"
- " For best performance upgrade your CUDA toolkit.",
- sbuf);
- }
- else
- {
- kver = eNbnxnCuKLegacy;
- }
- }
- else
- {
- /* issue not if the non-default kernel is forced by an env. var */
- if (bLegacyKernel)
- {
- fprintf(stderr,
- "\nNOTE: Legacy non-bonded CUDA kernels were selected by the GMX_CUDA_NB_LEGACY\n"
- " env. var. Consider using using the default kernels which should be faster!\n");
-
- kver = eNbnxnCuKLegacy;
- }
- }
-
- return kver;
-}
-
-void nbnxn_cuda_init(FILE *fplog,
- nbnxn_cuda_ptr_t *p_cu_nb,
- gmx_gpu_info_t *gpu_info, int my_gpu_index,
- gmx_bool bLocalAndNonlocal)
+void nbnxn_cuda_init(FILE *fplog,
+ nbnxn_cuda_ptr_t *p_cu_nb,
+ const gmx_gpu_info_t *gpu_info,
+ const gmx_gpu_opt_t *gpu_opt,
+ int my_gpu_index,
+ gmx_bool bLocalAndNonlocal)
{
- cudaError_t stat;
+ cudaError_t stat;
nbnxn_cuda_ptr_t nb;
- char sbuf[STRLEN];
- bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86;
+ char sbuf[STRLEN];
+ bool bStreamSync, bNoStreamSync, bTMPIAtomics, bX86, bOldDriver;
+ int cuda_drv_ver;
assert(gpu_info);
- if (p_cu_nb == NULL) return;
+ if (p_cu_nb == NULL)
+ {
+ return;
+ }
snew(nb, 1);
snew(nb->atdat, 1);
init_plist(nb->plist[eintLocal]);
+ /* set device info, just point it to the right GPU among the detected ones */
+ nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, gpu_opt, my_gpu_index)];
+
/* local/non-local GPU streams */
stat = cudaStreamCreate(&nb->stream[eintLocal]);
CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
if (nb->bUseTwoStreams)
{
init_plist(nb->plist[eintNonlocal]);
+
+ /* CUDA stream priority available in the CUDA RT 5.5 API.
+ * Note that the device we're running on does not have to support
+ * priorities, because we are querying the priority range which in this
+ * case will be a single value.
+ */
+#if CUDA_VERSION >= 5500
+ {
+ int highest_priority;
+ stat = cudaDeviceGetStreamPriorityRange(NULL, &highest_priority);
+ CU_RET_ERR(stat, "cudaDeviceGetStreamPriorityRange failed");
+
+ stat = cudaStreamCreateWithPriority(&nb->stream[eintNonlocal],
+ cudaStreamDefault,
+ highest_priority);
+ CU_RET_ERR(stat, "cudaStreamCreateWithPriority on stream[eintNonlocal] failed");
+ }
+#else
stat = cudaStreamCreate(&nb->stream[eintNonlocal]);
CU_RET_ERR(stat, "cudaStreamCreate on stream[eintNonlocal] failed");
+#endif
}
/* init events for sychronization (timing disabled for performance reasons!) */
stat = cudaEventCreateWithFlags(&nb->misc_ops_done, cudaEventDisableTiming);
CU_RET_ERR(stat, "cudaEventCreate on misc_ops_one failed");
- /* set device info, just point it to the right GPU among the detected ones */
- nb->dev_info = &gpu_info->cuda_dev[get_gpu_device_id(gpu_info, my_gpu_index)];
-
/* On GPUs with ECC enabled, cudaStreamSynchronize shows a large overhead
* (which increases with shorter time/step) caused by a known CUDA driver bug.
* To work around the issue we'll use an (admittedly fragile) memory polling
* waiting to preserve performance. This requires support for atomic
* operations and only works on x86/x86_64.
* With polling wait event-timing also needs to be disabled.
+ *
+ * The overhead is greatly reduced in API v5.0 drivers and the improvement
+ * is independent of runtime version. Hence, with API v5.0 drivers and later
+ * we won't switch to polling.
+ *
+ * NOTE: Unfortunately, this is known to fail when GPUs are shared by (t)MPI,
+ * ranks so we will also disable it in that case.
*/
bStreamSync = getenv("GMX_CUDA_STREAMSYNC") != NULL;
bTMPIAtomics = false;
#endif
-#if defined(i386) || defined(__x86_64__)
+#ifdef GMX_TARGET_X86
bX86 = true;
#else
bX86 = false;
gmx_fatal(FARGS, "Conflicting environment variables: both GMX_CUDA_STREAMSYNC and GMX_NO_CUDA_STREAMSYNC defined");
}
- if (nb->dev_info->prop.ECCEnabled == 1)
+ stat = cudaDriverGetVersion(&cuda_drv_ver);
+ CU_RET_ERR(stat, "cudaDriverGetVersion failed");
+
+ bOldDriver = (cuda_drv_ver < 5000);
+
+ if ((nb->dev_info->prop.ECCEnabled == 1) && bOldDriver)
{
+ /* Polling wait should be used instead of cudaStreamSynchronize only if:
+ * - ECC is ON & driver is old (checked above),
+ * - we're on x86/x86_64,
+ * - atomics are available, and
+ * - GPUs are not being shared.
+ */
+ bool bShouldUsePollSync = (bX86 && bTMPIAtomics &&
+ (gmx_count_gpu_dev_shared(gpu_opt) < 1));
+
if (bStreamSync)
{
nb->bUseStreamSync = true;
- sprintf(sbuf,
- "NOTE: Using a GPU with ECC enabled, but cudaStreamSynchronize-based waiting is\n"
- " forced by the GMX_CUDA_STREAMSYNC env. var. Due to a CUDA bug, this \n"
- " combination causes performance loss.");
- fprintf(stderr, "\n%s\n", sbuf);
- if (fplog)
+ /* only warn if polling should be used */
+ if (bShouldUsePollSync)
{
- fprintf(fplog, "\n%s\n", sbuf);
+ md_print_warn(fplog,
+ "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, but\n"
+ " cudaStreamSynchronize waiting is forced by the GMX_CUDA_STREAMSYNC env. var.\n");
}
}
else
{
- /* can use polling wait only on x86/x86_64 *if* atomics are available */
- nb->bUseStreamSync = ((bX86 && bTMPIAtomics) == false);
-
- if (!bX86)
- {
- sprintf(sbuf,
- "Using a GPU with ECC on; the standard cudaStreamSynchronize waiting, due to a\n"
- " CUDA bug, causes performance loss when used in combination with ECC.\n"
- " However, the polling waiting workaround can not be used as it is only\n"
- " supported on x86/x86_64, but not on the current architecture.");
- gmx_warning("%s\n", sbuf);
- if (fplog)
- {
- fprintf(fplog, "\n%s\n", sbuf);
- }
+ nb->bUseStreamSync = !bShouldUsePollSync;
- }
- else if (bTMPIAtomics)
+ if (bShouldUsePollSync)
{
- if (fplog)
- {
- fprintf(fplog,
- "NOTE: Using a GPU with ECC enabled; will use polling waiting.\n");
- }
+ md_print_warn(fplog,
+ "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0, known to\n"
+ " cause performance loss. Switching to the alternative polling GPU wait.\n"
+ " If you encounter issues, switch back to standard GPU waiting by setting\n"
+ " the GMX_CUDA_STREAMSYNC environment variable.\n");
}
else
{
+ /* Tell the user that the ECC+old driver combination can be bad */
sprintf(sbuf,
- "Using a GPU with ECC on; the standard cudaStreamSynchronize waiting, due to a\n"
- " CUDA bug, causes performance loss when used in combination with ECC.\n"
- " However, the polling waiting workaround can not be used as atomic\n"
- " operations are not supported by the current CPU+compiler combination.");
- gmx_warning("%s\n", sbuf);
- if (fplog)
- {
- fprintf(fplog, "\n%s\n", sbuf);
- }
+ "NOTE: Using a GPU with ECC enabled and CUDA driver API version <5.0.\n"
+ " A known bug in this driver version can cause performance loss.\n"
+ " However, the polling wait workaround can not be used because\n%s\n"
+ " Consider updating the driver or turning ECC off.",
+ (bX86 && bTMPIAtomics) ?
+ " GPU(s) are being oversubscribed." :
+ " atomic operations are not supported by the platform/CPU+compiler.");
+ md_print_warn(fplog, sbuf);
}
}
}
{
nb->bUseStreamSync = false;
- sprintf(sbuf,
- "NOTE: Using a GPU with no/disabled ECC, but cudaStreamSynchronize-based waiting\n"
- " is turned off and polling turned on by the GMX_NO_CUDA_STREAMSYNC env. var.");
- fprintf(stderr, "\n%s\n", sbuf);
- if (fplog)
- {
- fprintf(fplog, "\n%s\n", sbuf);
- }
+ md_print_warn(fplog,
+ "NOTE: Polling wait for GPU synchronization requested by GMX_NO_CUDA_STREAMSYNC\n");
}
else
{
}
/* set the kernel type for the current GPU */
- nb->kernel_ver = pick_nbnxn_kernel_version();
/* pick L1 cache configuration */
nbnxn_cuda_set_cacheconfig(nb->dev_info);
}
}
-void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
- const interaction_const_t *ic,
- const nonbonded_verlet_t *nbv)
+void nbnxn_cuda_init_const(nbnxn_cuda_ptr_t cu_nb,
+ const interaction_const_t *ic,
+ const nonbonded_verlet_group_t *nbv_group)
{
- init_atomdata_first(cu_nb->atdat, nbv->grp[0].nbat->ntype);
- init_nbparam(cu_nb->nbparam, ic, nbv);
+ init_atomdata_first(cu_nb->atdat, nbv_group[0].nbat->ntype);
+ init_nbparam(cu_nb->nbparam, ic, nbv_group[0].nbat, cu_nb->dev_info);
/* clear energy and shift force outputs */
nbnxn_cuda_clear_e_fshift(cu_nb);
}
-void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
+void nbnxn_cuda_init_pairlist(nbnxn_cuda_ptr_t cu_nb,
const nbnxn_pairlist_t *h_plist,
- int iloc)
+ int iloc)
{
- char sbuf[STRLEN];
- cudaError_t stat;
- bool bDoTime = cu_nb->bDoTime;
- cudaStream_t stream = cu_nb->stream[iloc];
- cu_plist_t *d_plist = cu_nb->plist[iloc];
+ char sbuf[STRLEN];
+ cudaError_t stat;
+ bool bDoTime = cu_nb->bDoTime;
+ cudaStream_t stream = cu_nb->stream[iloc];
+ cu_plist_t *d_plist = cu_nb->plist[iloc];
if (d_plist->na_c < 0)
{
}
cu_realloc_buffered((void **)&d_plist->sci, h_plist->sci, sizeof(*d_plist->sci),
- &d_plist->nsci, &d_plist->sci_nalloc,
- h_plist->nsci,
- stream, true);
+ &d_plist->nsci, &d_plist->sci_nalloc,
+ h_plist->nsci,
+ stream, true);
cu_realloc_buffered((void **)&d_plist->cj4, h_plist->cj4, sizeof(*d_plist->cj4),
- &d_plist->ncj4, &d_plist->cj4_nalloc,
- h_plist->ncj4,
- stream, true);
+ &d_plist->ncj4, &d_plist->cj4_nalloc,
+ h_plist->ncj4,
+ stream, true);
cu_realloc_buffered((void **)&d_plist->excl, h_plist->excl, sizeof(*d_plist->excl),
- &d_plist->nexcl, &d_plist->excl_nalloc,
- h_plist->nexcl,
- stream, true);
+ &d_plist->nexcl, &d_plist->excl_nalloc,
+ h_plist->nexcl,
+ stream, true);
if (bDoTime)
{
d_plist->bDoPrune = true;
}
-void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
+void nbnxn_cuda_upload_shiftvec(nbnxn_cuda_ptr_t cu_nb,
const nbnxn_atomdata_t *nbatom)
{
- cu_atomdata_t *adat = cu_nb->atdat;
- cudaStream_t ls = cu_nb->stream[eintLocal];
+ cu_atomdata_t *adat = cu_nb->atdat;
+ cudaStream_t ls = cu_nb->stream[eintLocal];
/* only if we have a dynamic box */
if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
{
- cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
+ cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec,
SHIFTS * sizeof(*adat->shift_vec), ls);
adat->bShiftVecUploaded = true;
}
/*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
static void nbnxn_cuda_clear_f(nbnxn_cuda_ptr_t cu_nb, int natoms_clear)
{
- cudaError_t stat;
- cu_atomdata_t *adat = cu_nb->atdat;
- cudaStream_t ls = cu_nb->stream[eintLocal];
+ cudaError_t stat;
+ cu_atomdata_t *adat = cu_nb->atdat;
+ cudaStream_t ls = cu_nb->stream[eintLocal];
stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
/*! Clears nonbonded shift force output array and energy outputs on the GPU. */
static void nbnxn_cuda_clear_e_fshift(nbnxn_cuda_ptr_t cu_nb)
{
- cudaError_t stat;
- cu_atomdata_t *adat = cu_nb->atdat;
- cudaStream_t ls = cu_nb->stream[eintLocal];
+ cudaError_t stat;
+ cu_atomdata_t *adat = cu_nb->atdat;
+ cudaStream_t ls = cu_nb->stream[eintLocal];
stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
void nbnxn_cuda_clear_outputs(nbnxn_cuda_ptr_t cu_nb, int flags)
{
nbnxn_cuda_clear_f(cu_nb, cu_nb->atdat->natoms);
- /* clear shift force array and energies if the outputs were
+ /* clear shift force array and energies if the outputs were
used in the current step */
if (flags & GMX_FORCE_VIRIAL)
{
}
}
-void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
+void nbnxn_cuda_init_atomdata(nbnxn_cuda_ptr_t cu_nb,
const nbnxn_atomdata_t *nbat)
{
- cudaError_t stat;
- int nalloc, natoms;
- bool realloced;
- bool bDoTime = cu_nb->bDoTime;
- cu_timers_t *timers = cu_nb->timers;
- cu_atomdata_t *d_atdat = cu_nb->atdat;
- cudaStream_t ls = cu_nb->stream[eintLocal];
-
- natoms = nbat->natoms;
+ cudaError_t stat;
+ int nalloc, natoms;
+ bool realloced;
+ bool bDoTime = cu_nb->bDoTime;
+ cu_timers_t *timers = cu_nb->timers;
+ cu_atomdata_t *d_atdat = cu_nb->atdat;
+ cudaStream_t ls = cu_nb->stream[eintLocal];
+
+ natoms = nbat->natoms;
realloced = false;
if (bDoTime)
CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
d_atdat->nalloc = nalloc;
- realloced = true;
+ realloced = true;
}
- d_atdat->natoms = natoms;
+ d_atdat->natoms = natoms;
d_atdat->natoms_local = nbat->natoms_local;
/* need to clear GPU f output if realloc happened */
}
}
-void nbnxn_cuda_free(FILE *fplog, nbnxn_cuda_ptr_t cu_nb)
+void nbnxn_cuda_free(nbnxn_cuda_ptr_t cu_nb)
{
- cudaError_t stat;
+ cudaError_t stat;
cu_atomdata_t *atdat;
cu_nbparam_t *nbparam;
cu_plist_t *plist, *plist_nl;
cu_timers_t *timers;
- if (cu_nb == NULL) return;
+ if (cu_nb == NULL)
+ {
+ return;
+ }
atdat = cu_nb->atdat;
nbparam = cu_nb->nbparam;
plist_nl = cu_nb->plist[eintNonlocal];
timers = cu_nb->timers;
- if (nbparam->eeltype == eelCuEWALD || nbparam->eeltype == eelCuEWALD_TWIN)
+ if (nbparam->eeltype == eelCuEWALD_TAB || nbparam->eeltype == eelCuEWALD_TAB_TWIN)
{
- stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
- CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab failed");
- cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
+
+#ifdef TEXOBJ_SUPPORTED
+ /* Only device CC >= 3.0 (Kepler and later) support texture objects */
+ if (cu_nb->dev_info->prop.major >= 3)
+ {
+ stat = cudaDestroyTextureObject(nbparam->coulomb_tab_texobj);
+ CU_RET_ERR(stat, "cudaDestroyTextureObject on coulomb_tab_texobj failed");
+ }
+ else
+#endif
+ {
+ stat = cudaUnbindTexture(nbnxn_cuda_get_coulomb_tab_texref());
+ CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab_texref failed");
+ }
+ cu_free_buffered(nbparam->coulomb_tab, &nbparam->coulomb_tab_size);
}
stat = cudaEventDestroy(cu_nb->nonlocal_done);
}
}
- stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
- CU_RET_ERR(stat, "cudaUnbindTexture on coulomb_tab failed");
+#ifdef TEXOBJ_SUPPORTED
+ /* Only device CC >= 3.0 (Kepler and later) support texture objects */
+ if (cu_nb->dev_info->prop.major >= 3)
+ {
+ stat = cudaDestroyTextureObject(nbparam->nbfp_texobj);
+ CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_texobj failed");
+ }
+ else
+#endif
+ {
+ stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_texref());
+ CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_texref failed");
+ }
cu_free_buffered(nbparam->nbfp);
+ if (nbparam->vdwtype == evdwCuEWALDGEOM || nbparam->vdwtype == evdwCuEWALDLB)
+ {
+#ifdef TEXOBJ_SUPPORTED
+ /* Only device CC >= 3.0 (Kepler and later) support texture objects */
+ if (cu_nb->dev_info->prop.major >= 3)
+ {
+ stat = cudaDestroyTextureObject(nbparam->nbfp_comb_texobj);
+ CU_RET_ERR(stat, "cudaDestroyTextureObject on nbfp_comb_texobj failed");
+ }
+ else
+#endif
+ {
+ stat = cudaUnbindTexture(nbnxn_cuda_get_nbfp_comb_texref());
+ CU_RET_ERR(stat, "cudaUnbindTexture on nbfp_comb_texref failed");
+ }
+ cu_free_buffered(nbparam->nbfp_comb);
+ }
+
stat = cudaFree(atdat->shift_vec);
CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
stat = cudaFree(atdat->fshift);
cu_free_buffered(plist_nl->excl, &plist_nl->nexcl, &plist->excl_nalloc);
}
+ sfree(atdat);
+ sfree(nbparam);
+ sfree(plist);
+ if (cu_nb->bUseTwoStreams)
+ {
+ sfree(plist_nl);
+ }
+ sfree(timers);
+ sfree(cu_nb->timings);
+ sfree(cu_nb);
+
if (debug)
{
fprintf(debug, "Cleaned up CUDA data structures.\n");
void cu_synchstream_atdat(nbnxn_cuda_ptr_t cu_nb, int iloc)
{
- cudaError_t stat;
+ cudaError_t stat;
cudaStream_t stream = cu_nb->stream[iloc];
stat = cudaStreamWaitEvent(stream, cu_nb->timers->stop_atdat, 0);
return (cu_nb != NULL && cu_nb->bDoTime) ? cu_nb->timings : NULL;
}
-void nbnxn_cuda_reset_timings(nbnxn_cuda_ptr_t cu_nb)
+void nbnxn_cuda_reset_timings(nonbonded_verlet_t* nbv)
{
- if (cu_nb->bDoTime)
+ if (nbv->cu_nbv && nbv->cu_nbv->bDoTime)
{
- init_timings(cu_nb->timings);
+ init_timings(nbv->cu_nbv->timings);
}
}
int nbnxn_cuda_min_ci_balanced(nbnxn_cuda_ptr_t cu_nb)
{
return cu_nb != NULL ?
- gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;
+ gpu_min_ci_balanced_factor*cu_nb->dev_info->prop.multiProcessorCount : 0;
}
+
+gmx_bool nbnxn_cuda_is_kernel_ewald_analytical(const nbnxn_cuda_ptr_t cu_nb)
+{
+ return ((cu_nb->nbparam->eeltype == eelCuEWALD_ANA) ||
+ (cu_nb->nbparam->eeltype == eelCuEWALD_ANA_TWIN));
+}