Merge 'release-4-6' into master

[alexxy/gromacs.git] / src / gromacs / mdlib / nbnxn_internal.h

diff --git a/src/gromacs/mdlib/nbnxn_internal.h b/src/gromacs/mdlib/nbnxn_internal.h
index 289d5abb004a5096aac95dbd5c8b4b4f10c17433..bfcf1277e53e71ce4944ca7aca80e39abcc856da 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_internal.h
+++ b/src/gromacs/mdlib/nbnxn_internal.h
@@ -50,6 +50,30 @@ extern "C" {
 #endif
 
 
+/* Block size for the non-bonded thread force-buffer reduction,
+ * should be a multiple of 2 in case of AVX256.
+ */
+#define NBNXN_CELLBLOCK_SIZE       4
+#define NBNXN_CELLBLOCK_SIZE_2LOG  2
+
+/* We currently store the reduction flags as bits in an unsigned int.
+ * In most cases this limits the number of flags to 32.
+ * The reduction will automatically disable the flagging and do a full
+ * reduction when the flags won't fit, but this will lead to very slow
+ * reduction. As we anyhow don't expect reasonable performance with
+ * more than 32 threads, we put in this hard limit.
+ * You can increase this number, but the reduction will be very slow.
+ */
+#define NBNXN_CELLBLOCK_MAX_THREADS  32
+
+/* Flags for telling if threads write to force output buffers */
+typedef struct {
+    int ncb;         /* The number of cell blocks                         */
+    gmx_bool bUse;   /* Should we use these flags?                        */
+    unsigned *flag;  /* Bit i is set when thread i writes to a cell-block */
+    int flag_nalloc; /* Allocation size of cxy_flag                       */
+} nbnxn_cellblock_flags;
+
 /* A pair-search grid struct for one domain decomposition zone */
 typedef struct {
     rvec c0;             /* The lower corner of the (local) grid        */
@@ -89,6 +113,8 @@ typedef struct {
     int   *flags_simple; /* flags for simple grid converted from super  */
     int   nc_nalloc_simple; /* Allocation size for the pointers above   */
 
+    nbnxn_cellblock_flags cellblock_flags; /* Flags for F output buffers */
+
     int  nsubc_tot;      /* Total number of subcell, used for printing  */
 } nbnxn_grid_t;
 
@@ -183,6 +209,9 @@ typedef struct {
     int  *sort_work;
     int  sort_work_nalloc;
 
+    nbnxn_cellblock_flags gridi_flags; /* Flags for i-grid f buffer     */
+    nbnxn_cellblock_flags gridj_flags; /* Flags for j-grid f buffer     */
+
     int  ndistc;         /* Number of distance checks for flop counting */
 
     nbnxn_cycle_t cc[enbsCCnr];