Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / mdlib / nbnxn_cuda / nbnxn_cuda.cu

diff --git a/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda.cu b/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda.cu
index 09e53c24b3425da48727e3677f425b92100f16a1..27ac6045c005580d195faf174e665b795d2abdf8 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda.cu
+++ b/src/gromacs/mdlib/nbnxn_cuda/nbnxn_cuda.cu
@@ -199,6 +199,8 @@ static inline int calc_shmem_required(int kver)
         /* 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);
 #ifdef IATYPE_SHMEM
         /* i-atom types in shared memory */
         shmem += NCL_PER_SUPERCL * CL_SIZE * sizeof(int);