Merge branch 'master' into gbn-chk

platforms/opencl/src/kernels/findInteractingBlocks.cl

@@ -5,15 +5,15 @@
 /**
  * Find a bounding box for the atoms in each block.
  */
-__kernel void findBlockBounds(int numAtoms, real4 periodicBoxSize, real4 invPeriodicBoxSize, __global const real4* restrict posq,
-        __global real4* restrict blockCenter, __global real4* restrict blockBoundingBox, __global int* restrict rebuildNeighborList,
+__kernel void findBlockBounds(int numAtoms, real4 periodicBoxSize, real4 invPeriodicBoxSize, real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ,
+        __global const real4* restrict posq, __global real4* restrict blockCenter, __global real4* restrict blockBoundingBox, __global int* restrict rebuildNeighborList,
         __global real2* restrict sortedBlocks) {
     int index = get_global_id(0);
     int base = index*TILE_SIZE;
     while (base < numAtoms) {
         real4 pos = posq[base];
 #ifdef USE_PERIODIC
-        pos.xyz -= floor(pos.xyz*invPeriodicBoxSize.xyz)*periodicBoxSize.xyz;
+        APPLY_PERIODIC_TO_POS(pos)
 #endif
         real4 minPos = pos;
         real4 maxPos = pos;
@@ -22,7 +22,7 @@ __kernel void findBlockBounds(int numAtoms, real4 periodicBoxSize, real4 invPeri
             pos = posq[i];
 #ifdef USE_PERIODIC
             real4 center = 0.5f*(maxPos+minPos);
-            pos.xyz -= floor((pos.xyz-center.xyz)*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+            APPLY_PERIODIC_TO_POS_WITH_CENTER(pos, center)
 #endif
             minPos = min(minPos, pos);
             maxPos = max(maxPos, pos);
@@ -65,9 +65,10 @@ __kernel void sortBoxData(__global const real2* restrict sortedBlock, __global c
     }
 }
 
-__kernel void findBlocksWithInteractions(real4 periodicBoxSize, real4 invPeriodicBoxSize, __global unsigned int* restrict interactionCount,
-        __global int* restrict interactingTiles, __global unsigned int* restrict interactingAtoms, __global const real4* restrict posq, unsigned int maxTiles, unsigned int startBlockIndex,
-        unsigned int numBlocks, __global real2* restrict sortedBlocks, __global const real4* restrict sortedBlockCenter, __global const real4* restrict sortedBlockBoundingBox,
+__kernel void findBlocksWithInteractions(real4 periodicBoxSize, real4 invPeriodicBoxSize, real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ,
+        __global unsigned int* restrict interactionCount, __global int* restrict interactingTiles, __global unsigned int* restrict interactingAtoms,
+        __global const real4* restrict posq, unsigned int maxTiles, unsigned int startBlockIndex, unsigned int numBlocks, __global real2* restrict sortedBlocks,
+        __global const real4* restrict sortedBlockCenter, __global const real4* restrict sortedBlockBoundingBox,
         __global const unsigned int* restrict exclusionIndices, __global const unsigned int* restrict exclusionRowIndices, __global real4* restrict oldPositions,
         __global const int* restrict rebuildNeighborList) {
 
@@ -108,7 +109,7 @@ __kernel void findBlocksWithInteractions(real4 periodicBoxSize, real4 invPeriodi
             // The box is small enough that we can just translate all the atoms into a single periodic
             // box, then skip having to apply periodic boundary conditions later.
             
-            pos1.xyz -= floor((pos1.xyz-blockCenterX.xyz)*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+            APPLY_PERIODIC_TO_POS_WITH_CENTER(pos1, blockCenterX)
         }
 #endif
         posBuffer[get_local_id(0)] = pos1;
@@ -136,7 +137,7 @@ __kernel void findBlocksWithInteractions(real4 periodicBoxSize, real4 invPeriodi
                 real4 blockSizeY = (block2 < NUM_BLOCKS ? sortedBlockBoundingBox[block2] : (real4) (0));
                 real4 blockDelta = blockCenterX-blockCenterY;
 #ifdef USE_PERIODIC
-                blockDelta.xyz -= floor(blockDelta.xyz*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+                APPLY_PERIODIC_TO_DELTA(blockDelta)
 #endif
                 blockDelta.x = max((real) 0, fabs(blockDelta.x)-blockSizeX.x-blockSizeY.x);
                 blockDelta.y = max((real) 0, fabs(blockDelta.y)-blockSizeX.y-blockSizeY.y);
@@ -166,7 +167,7 @@ __kernel void findBlocksWithInteractions(real4 periodicBoxSize, real4 invPeriodi
                     real3 pos2 = posq[atom2].xyz;
 #ifdef USE_PERIODIC
                     if (singlePeriodicCopy)
-                        pos2.xyz -= floor((pos2.xyz-blockCenterX.xyz)*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+                        APPLY_PERIODIC_TO_POS_WITH_CENTER(pos2, blockCenterX)
 #endif
                     bool interacts = false;
                     if (atom2 < NUM_ATOMS) {
@@ -174,7 +175,7 @@ __kernel void findBlocksWithInteractions(real4 periodicBoxSize, real4 invPeriodi
                         if (!singlePeriodicCopy) {
                             for (int j = 0; j < TILE_SIZE; j++) {
                                 real3 delta = pos2-posBuffer[warpStart+j];
-                                delta.xyz -= floor(delta.xyz*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+                                APPLY_PERIODIC_TO_DELTA(delta)
                                 interacts |= (delta.x*delta.x+delta.y*delta.y+delta.z*delta.z < PADDED_CUTOFF_SQUARED);
                             }
                         }

platforms/opencl/src/kernels/monteCarloBarostat.cl

@@ -2,7 +2,8 @@
  * Scale the particle positions with each axis independent.
  */
 
-__kernel void scalePositions(float scaleX, float scaleY, float scaleZ, int numMolecules, real4 periodicBoxSize, real4 invPeriodicBoxSize, __global real4* restrict posq,
+__kernel void scalePositions(float scaleX, float scaleY, float scaleZ, int numMolecules, real4 periodicBoxSize, real4 invPeriodicBoxSize,
+        real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ, __global real4* restrict posq,
         __global const int* restrict moleculeAtoms, __global const int* restrict moleculeStartIndex) {
     for (int index = get_global_id(0); index < numMolecules; index += get_global_size(0)) {
         int first = moleculeStartIndex[index];
@@ -11,19 +12,17 @@ __kernel void scalePositions(float scaleX, float scaleY, float scaleZ, int numMo
 
         // Find the center of each molecule.
 
-        real4 center = (real4) 0;
+        real3 center = (real3) 0;
         for (int atom = first; atom < last; atom++)
-            center += posq[moleculeAtoms[atom]];
+            center += posq[moleculeAtoms[atom]].xyz;
         center /= (real) numAtoms;
 
         // Move it into the first periodic box.
 
-        int xcell = (int) floor(center.x*invPeriodicBoxSize.x);
-        int ycell = (int) floor(center.y*invPeriodicBoxSize.y);
-        int zcell = (int) floor(center.z*invPeriodicBoxSize.z);
-        real4 delta = (real4) (xcell*periodicBoxSize.x, ycell*periodicBoxSize.y, zcell*periodicBoxSize.z, 0);
-        real4 scaleXYZ = (real4) (scaleX, scaleY, scaleZ, 1);
-        center -= delta;
+        real3 oldCenter = center;
+        APPLY_PERIODIC_TO_POS(center)
+        real3 delta = oldCenter-center;;
+        real3 scaleXYZ = (real3) (scaleX, scaleY, scaleZ);
 
         // Now scale the position of the molecule center.
 

platforms/opencl/src/kernels/nonbonded.cl

@@ -26,7 +26,8 @@ __kernel void computeNonbonded(
         __global const ushort2* restrict exclusionTiles, unsigned int startTileIndex, unsigned int numTileIndices
 #ifdef USE_CUTOFF
         , __global const int* restrict tiles, __global const unsigned int* restrict interactionCount, real4 periodicBoxSize, real4 invPeriodicBoxSize, 
-        unsigned int maxTiles, __global const real4* restrict blockCenter, __global const real4* restrict blockSize, __global const int* restrict interactingAtoms
+        real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ, unsigned int maxTiles, __global const real4* restrict blockCenter,
+        __global const real4* restrict blockSize, __global const int* restrict interactingAtoms
 #endif
         PARAMETER_ARGUMENTS) {
     const unsigned int totalWarps = get_global_size(0)/TILE_SIZE;
@@ -67,7 +68,7 @@ __kernel void computeNonbonded(
                 real4 posq2 = (real4) (localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                 real4 delta = (real4) (posq2.xyz - posq1.xyz, 0);
 #ifdef USE_PERIODIC
-                delta.xyz -= floor(delta.xyz*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+                APPLY_PERIODIC_TO_DELTA(delta)
 #endif
                 real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
                 real invR = RSQRT(r2);
@@ -121,7 +122,7 @@ __kernel void computeNonbonded(
                 real4 posq2 = (real4) (localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                 real4 delta = (real4) (posq2.xyz - posq1.xyz, 0);
 #ifdef USE_PERIODIC
-                delta.xyz -= floor(delta.xyz*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+                APPLY_PERIODIC_TO_DELTA(delta)
 #endif
                 real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
 #ifdef PRUNE_BY_CUTOFF
@@ -289,10 +290,8 @@ __kernel void computeNonbonded(
                 // box, then skip having to apply periodic boundary conditions later.
 
                 real4 blockCenterX = blockCenter[x];
-                posq1.xyz -= floor((posq1.xyz-blockCenterX.xyz)*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
-                localData[localAtomIndex].x -= floor((localData[localAtomIndex].x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
-                localData[localAtomIndex].y -= floor((localData[localAtomIndex].y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
-                localData[localAtomIndex].z -= floor((localData[localAtomIndex].z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
+                APPLY_PERIODIC_TO_POS_WITH_CENTER(posq1, blockCenterX)
+                APPLY_PERIODIC_TO_POS_WITH_CENTER(localData[localAtomIndex], blockCenterX)
                 SYNC_WARPS;
                 unsigned int tj = tgx;
                 for (j = 0; j < TILE_SIZE; j++) {
@@ -349,7 +348,7 @@ __kernel void computeNonbonded(
                     real4 posq2 = (real4) (localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                     real4 delta = (real4) (posq2.xyz - posq1.xyz, 0);
 #ifdef USE_PERIODIC
-                    delta.xyz -= floor(delta.xyz*invPeriodicBoxSize.xyz+0.5f)*periodicBoxSize.xyz;
+                    APPLY_PERIODIC_TO_DELTA(delta)
 #endif
                     real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
 #ifdef PRUNE_BY_CUTOFF