* 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.h"
+
+#include "config.h"
-#include <stdlib.h>
#include <assert.h>
+#include <stdlib.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 "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/pbcutil/ishift.h"
+#include "gromacs/utility/cstringutil.h"
+
#include "nbnxn_cuda_types.h"
-#include "../../gmxlib/cuda_tools/cudautils.cuh"
-#include "nbnxn_cuda.h"
-#include "nbnxn_cuda_data_mgmt.h"
#if defined TEXOBJ_SUPPORTED && __CUDA_ARCH__ >= 300
#define USE_TEXOBJ
#define NCL_PER_SUPERCL (NBNXN_GPU_NCLUSTER_PER_SUPERCLUSTER)
#define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
+/* NTHREAD_Z controls the number of j-clusters processed concurrently on NTHREAD_Z
+ * warp-pairs per block.
+ *
+ * - On CC 2.0-3.5, 5.0, and 5.2, NTHREAD_Z == 1, translating to 64 th/block with 16
+ * blocks/multiproc, is the fastest even though this setup gives low occupancy.
+ * NTHREAD_Z > 1 results in excessive register spilling unless the minimum blocks
+ * per multiprocessor is reduced proportionally to get the original number of max
+ * threads in flight (and slightly lower performance).
+ * - On CC 3.7 there are enough registers to double the number of threads; using
+ * NTHREADS_Z == 2 is fastest with 16 blocks (TODO: test with RF and other kernels
+ * with low-register use).
+ *
+ * Note that the current kernel implementation only supports NTHREAD_Z > 1 with
+ * shuffle-based reduction, hence CC >= 3.0.
+ */
+
+/* Kernel launch bounds as function of NTHREAD_Z.
+ * - CC 3.5/5.2: NTHREAD_Z=1, (64, 16) bounds
+ * - CC 3.7: NTHREAD_Z=2, (128, 16) bounds
+ */
+#if __CUDA_ARCH__ == 370
+#define NTHREAD_Z (2)
+#define MIN_BLOCKS_PER_MP (16)
+#else
+#define NTHREAD_Z (1)
+#define MIN_BLOCKS_PER_MP (16)
+#endif
+#define THREADS_PER_BLOCK (CL_SIZE*CL_SIZE*NTHREAD_Z)
+
+
/***** 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
}
/*! Calculates the amount of shared memory required by the CUDA kernel in use. */
-static inline int calc_shmem_required()
+static inline int calc_shmem_required(const int num_threads_z)
{
int shmem;
/* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
/* i-atom x+q in shared memory */
shmem = NCL_PER_SUPERCL * CL_SIZE * sizeof(float4);
- /* cj in shared memory, for both warps separately */
- shmem += 2 * NBNXN_GPU_JGROUP_SIZE * sizeof(int);
+ /* cj in shared memory, for each warp separately */
+ shmem += num_threads_z * 2 * NBNXN_GPU_JGROUP_SIZE * sizeof(int);
#ifdef IATYPE_SHMEM
/* i-atom types in shared memory */
shmem += NCL_PER_SUPERCL * CL_SIZE * sizeof(int);
bCalcEner,
plist->bDoPrune || always_prune);
- /* kernel launch config */
+ /* Kernel launch config:
+ * - The thread block dimensions match the size of i-clusters, j-clusters,
+ * and j-cluster concurrency, in x, y, and z, respectively.
+ * - The 1D block-grid contains as many blocks as super-clusters.
+ */
+ int num_threads_z = 1;
+ if (cu_nb->dev_info->prop.major == 3 && cu_nb->dev_info->prop.minor == 7)
+ {
+ num_threads_z = 2;
+ }
nblock = calc_nb_kernel_nblock(plist->nsci, cu_nb->dev_info);
- dim_block = dim3(CL_SIZE, CL_SIZE, 1);
+ dim_block = dim3(CL_SIZE, CL_SIZE, num_threads_z);
dim_grid = dim3(nblock, 1, 1);
- shmem = calc_shmem_required();
+ shmem = calc_shmem_required(num_threads_z);
if (debug)
{
fprintf(debug, "GPU launch configuration:\n\tThread block: %dx%dx%d\n\t"
- "Grid: %dx%d\n\t#Super-clusters/clusters: %d/%d (%d)\n",
+ "\tGrid: %dx%d\n\t#Super-clusters/clusters: %d/%d (%d)\n"
+ "\tShMem: %d\n",
dim_block.x, dim_block.y, dim_block.z,
dim_grid.x, dim_grid.y, plist->nsci*NCL_PER_SUPERCL,
- NCL_PER_SUPERCL, plist->na_c);
+ NCL_PER_SUPERCL, plist->na_c,
+ shmem);
}
nb_kernel<<< dim_grid, dim_block, shmem, stream>>> (*adat, *nbp, *plist, bCalcFshift);
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");