* To help us fund GROMACS development, we humbly ask that you cite
* the research papers on the package. Check out http://www.gromacs.org.
*/
+#include "gmxpre.h"
+
+#include "nbnxn_cuda.h"
-#include <stdlib.h>
#include <assert.h>
+#include <stdlib.h>
+
+#include "config.h"
#if defined(_MSVC)
#include <limits>
#include <cuda.h>
-#include "types/simple.h"
-#include "types/nbnxn_pairlist.h"
-#include "types/nb_verlet.h"
-#include "types/ishift.h"
-#include "types/force_flags.h"
-#include "../nbnxn_consts.h"
-
#ifdef TMPI_ATOMICS
#include "thread_mpi/atomic.h"
#endif
-#include "nbnxn_cuda_types.h"
-#include "../../gmxlib/cuda_tools/cudautils.cuh"
-#include "nbnxn_cuda.h"
-#include "nbnxn_cuda_data_mgmt.h"
+#include "gromacs/gmxlib/cuda_tools/cudautils.cuh"
+#include "gromacs/legacyheaders/types/force_flags.h"
+#include "gromacs/legacyheaders/types/simple.h"
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_consts.h"
+#include "gromacs/mdlib/nbnxn_pairlist.h"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.h"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_types.h"
+#include "gromacs/pbcutil/ishift.h"
+#include "gromacs/utility/cstringutil.h"
#if defined TEXOBJ_SUPPORTED && __CUDA_ARCH__ >= 300
#define USE_TEXOBJ
#endif
-/*! Texture reference for nonbonded parameters; bound to cu_nbparam_t.nbfp*/
+/*! Texture reference for LJ C6/C12 parameters; bound to cu_nbparam_t.nbfp */
texture<float, 1, cudaReadModeElementType> nbfp_texref;
+/*! Texture reference for LJ-PME parameters; bound to cu_nbparam_t.nbfp_comb */
+texture<float, 1, cudaReadModeElementType> nbfp_comb_texref;
+
/*! Texture reference for Ewald coulomb force table; bound to cu_nbparam_t.coulomb_tab */
texture<float, 1, cudaReadModeElementType> coulomb_tab_texref;
#define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
/***** The kernels come here *****/
-#include "nbnxn_cuda_kernel_utils.cuh"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernel_utils.cuh"
/* Top-level kernel generation: will generate through multiple inclusion the
* following flavors for all kernels:
* - force and energy output with pair list pruning.
*/
/** Force only **/
-#include "nbnxn_cuda_kernels.cuh"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
/** Force & energy **/
#define CALC_ENERGIES
-#include "nbnxn_cuda_kernels.cuh"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
#undef CALC_ENERGIES
/*** Pair-list pruning kernels ***/
/** Force only **/
#define PRUNE_NBL
-#include "nbnxn_cuda_kernels.cuh"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
/** Force & energy **/
#define CALC_ENERGIES
-#include "nbnxn_cuda_kernels.cuh"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_kernels.cuh"
#undef CALC_ENERGIES
#undef PRUNE_NBL
/*! Force-only kernel function pointers. */
static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_noprune_ptr[eelCuNR][evdwCuNR] =
{
- { nbnxn_kernel_ElecCut_VdwLJ_F_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_cuda },
- { nbnxn_kernel_ElecRF_VdwLJ_F_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_cuda },
- { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_cuda },
- { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_cuda },
- { nbnxn_kernel_ElecEw_VdwLJ_F_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_cuda },
- { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_cuda }
+ { nbnxn_kernel_ElecCut_VdwLJ_F_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_cuda },
+ { nbnxn_kernel_ElecRF_VdwLJ_F_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_cuda },
+ { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_cuda },
+ { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_cuda },
+ { nbnxn_kernel_ElecEw_VdwLJ_F_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_cuda },
+ { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_cuda }
};
/*! Force + energy kernel function pointers. */
static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_noprune_ptr[eelCuNR][evdwCuNR] =
{
- { nbnxn_kernel_ElecCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_cuda },
- { nbnxn_kernel_ElecRF_VdwLJ_VF_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_cuda },
- { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_cuda },
- { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_cuda },
- { nbnxn_kernel_ElecEw_VdwLJ_VF_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_cuda },
- { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_cuda }
+ { nbnxn_kernel_ElecCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_cuda },
+ { nbnxn_kernel_ElecRF_VdwLJ_VF_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_cuda },
+ { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_cuda },
+ { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_cuda },
+ { nbnxn_kernel_ElecEw_VdwLJ_VF_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_cuda },
+ { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_cuda }
};
/*! Force + pruning kernel function pointers. */
static const nbnxn_cu_kfunc_ptr_t nb_kfunc_noener_prune_ptr[eelCuNR][evdwCuNR] =
{
- { nbnxn_kernel_ElecCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_cuda },
- { nbnxn_kernel_ElecRF_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_cuda },
- { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_cuda },
- { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_cuda },
- { nbnxn_kernel_ElecEw_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_cuda },
- { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_cuda }
+ { nbnxn_kernel_ElecCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_cuda },
+ { nbnxn_kernel_ElecRF_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_cuda },
+ { nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_cuda },
+ { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_cuda },
+ { nbnxn_kernel_ElecEw_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_cuda },
+ { nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_cuda }
};
/*! Force + energy + pruning kernel function pointers. */
static const nbnxn_cu_kfunc_ptr_t nb_kfunc_ener_prune_ptr[eelCuNR][evdwCuNR] =
{
- { nbnxn_kernel_ElecCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_cuda },
- { nbnxn_kernel_ElecRF_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_cuda },
- { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_cuda },
- { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_cuda },
- { nbnxn_kernel_ElecEw_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_cuda },
- { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_cuda }
+ { nbnxn_kernel_ElecCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_cuda },
+ { nbnxn_kernel_ElecRF_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_cuda },
+ { nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_cuda },
+ { nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_cuda },
+ { nbnxn_kernel_ElecEw_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_cuda },
+ { nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_cuda, nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_cuda }
};
/*! Return a pointer to the kernel version to be executed at the current step. */
return nbfp_texref;
}
+/*! Return the reference to the nbfp_comb texture. */
+const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_nbfp_comb_texref()
+{
+ return nbfp_comb_texref;
+}
+
/*! Return the reference to the coulomb_tab. */
const struct texture<float, 1, cudaReadModeElementType> &nbnxn_cuda_get_coulomb_tab_texref()
{
if (devinfo->prop.major >= 3)
{
/* Default kernel on sm 3.x 48/16 kB Shared/L1 */
- stat = cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferShared);
- stat = cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferShared);
- stat = cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferShared);
+ cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferShared);
+ cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferShared);
+ cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferShared);
stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferShared);
}
else
{
/* On Fermi prefer L1 gives 2% higher performance */
/* Default kernel on sm_2.x 16/48 kB Shared/L1 */
- stat = cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferL1);
- stat = cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferL1);
- stat = cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferL1);
+ cudaFuncSetCacheConfig(nb_kfunc_ener_prune_ptr[i][j], cudaFuncCachePreferL1);
+ cudaFuncSetCacheConfig(nb_kfunc_ener_noprune_ptr[i][j], cudaFuncCachePreferL1);
+ cudaFuncSetCacheConfig(nb_kfunc_noener_prune_ptr[i][j], cudaFuncCachePreferL1);
stat = cudaFuncSetCacheConfig(nb_kfunc_noener_noprune_ptr[i][j], cudaFuncCachePreferL1);
}
CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");