#define WARPS_PER_GROUP (THREAD_BLOCK_SIZE/TILE_SIZE)

typedef struct {
    real x, y, z;
    real q;
    real fx, fy, fz;
    ATOM_PARAMETER_DATA
#ifndef PARAMETER_SIZE_IS_EVEN
    real padding;
#endif
} AtomData;

/**
 * Compute nonbonded interactions.
 */
extern "C" __global__ void computeNonbonded(
        unsigned long long* __restrict__ forceBuffers, real* __restrict__ energyBuffer, const real4* __restrict__ posq, const tileflags* __restrict__ exclusions,
        const ushort2* __restrict__ exclusionTiles, unsigned int startTileIndex, unsigned int numTileIndices
#ifdef USE_CUTOFF
        , const ushort2* __restrict__ tiles, const unsigned int* __restrict__ interactionCount, real4 periodicBoxSize, real4 invPeriodicBoxSize, unsigned int maxTiles, const real4* __restrict__ blockCenter, const unsigned int* __restrict__ interactingAtoms
#endif
        PARAMETER_ARGUMENTS) {
    const unsigned int totalWarps = (blockDim.x*gridDim.x)/TILE_SIZE;
    const unsigned int warp = (blockIdx.x*blockDim.x+threadIdx.x)/TILE_SIZE;
    const unsigned int tgx = threadIdx.x & (TILE_SIZE-1);
    const unsigned int tbx = threadIdx.x - tgx;
    real energy = 0.0f;
    __shared__ AtomData localData[THREAD_BLOCK_SIZE];

    // First loop: process tiles that contain exclusions.
    
    const unsigned int firstExclusionTile = FIRST_EXCLUSION_TILE+warp*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/totalWarps;
    const unsigned int lastExclusionTile = FIRST_EXCLUSION_TILE+(warp+1)*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/totalWarps;
    for (int pos = firstExclusionTile; pos < lastExclusionTile; pos++) {
        const ushort2 tileIndices = exclusionTiles[pos];
        const unsigned int x = tileIndices.x;
        const unsigned int y = tileIndices.y;
        real3 force = make_real3(0);
        unsigned int atom1 = x*TILE_SIZE + tgx;
        real4 posq1 = posq[atom1];
        LOAD_ATOM1_PARAMETERS
#ifdef USE_EXCLUSIONS
        tileflags excl = exclusions[pos*TILE_SIZE+tgx];
#endif
        const bool hasExclusions = true;
        if (x == y) {
            // This tile is on the diagonal.

            const unsigned int localAtomIndex = threadIdx.x;
            localData[localAtomIndex].x = posq1.x;
            localData[localAtomIndex].y = posq1.y;
            localData[localAtomIndex].z = posq1.z;
            localData[localAtomIndex].q = posq1.w;
            LOAD_LOCAL_PARAMETERS_FROM_1
            for (unsigned int j = 0; j < TILE_SIZE; j++) {
                int atom2 = tbx+j;
                real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
#ifdef USE_PERIODIC
                delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
                delta.y -= floor(delta.y*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
                delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#endif
                real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
                real invR = RSQRT(r2);
                real r = RECIP(invR);
                LOAD_ATOM2_PARAMETERS
                atom2 = y*TILE_SIZE+j;
#ifdef USE_SYMMETRIC
                real dEdR = 0.0f;
#else
                real3 dEdR1 = make_real3(0);
                real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
                bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
#endif
                real tempEnergy = 0.0f;
                COMPUTE_INTERACTION
                energy += 0.5f*tempEnergy;
#ifdef USE_SYMMETRIC
                force.x -= delta.x*dEdR;
                force.y -= delta.y*dEdR;
                force.z -= delta.z*dEdR;
#else
                force.x -= dEdR1.x;
                force.y -= dEdR1.y;
                force.z -= dEdR1.z;
#endif
#ifdef USE_EXCLUSIONS
                excl >>= 1;
#endif
            }
        }
        else {
            // This is an off-diagonal tile.
            
            const unsigned int localAtomIndex = threadIdx.x;
            unsigned int j = y*TILE_SIZE + tgx;
            real4 tempPosq = posq[j];
            localData[localAtomIndex].x = tempPosq.x;
            localData[localAtomIndex].y = tempPosq.y;
            localData[localAtomIndex].z = tempPosq.z;
            localData[localAtomIndex].q = tempPosq.w;
            LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
            localData[localAtomIndex].fx = 0.0f;
            localData[localAtomIndex].fy = 0.0f;
            localData[localAtomIndex].fz = 0.0f;
#ifdef USE_EXCLUSIONS
            excl = (excl >> tgx) | (excl << (TILE_SIZE - tgx));
#endif
            unsigned int tj = tgx;
            for (j = 0; j < TILE_SIZE; j++) {
                int atom2 = tbx+tj;
                real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
#ifdef USE_PERIODIC
                delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
                delta.y -= floor(delta.y*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
                delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#endif
                real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
#ifdef USE_CUTOFF
                if (r2 < CUTOFF_SQUARED) {
#endif
                    real invR = RSQRT(r2);
                    real r = RECIP(invR);
                    LOAD_ATOM2_PARAMETERS
                    atom2 = y*TILE_SIZE+tj;
#ifdef USE_SYMMETRIC
                    real dEdR = 0.0f;
#else
                    real3 dEdR1 = make_real3(0);
                    real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
                    bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
#endif
                    real tempEnergy = 0.0f;
                    COMPUTE_INTERACTION
                    energy += tempEnergy;
#ifdef USE_SYMMETRIC
                    delta *= dEdR;
                    force.x -= delta.x;
                    force.y -= delta.y;
                    force.z -= delta.z;
                    localData[tbx+tj].fx += delta.x;
                    localData[tbx+tj].fy += delta.y;
                    localData[tbx+tj].fz += delta.z;
#else
                    force.x -= dEdR1.x;
                    force.y -= dEdR1.y;
                    force.z -= dEdR1.z;
                    localData[tbx+tj].fx += dEdR2.x;
                    localData[tbx+tj].fy += dEdR2.y;
                    localData[tbx+tj].fz += dEdR2.z;
#endif
#ifdef USE_CUTOFF
                }
#endif
#ifdef USE_EXCLUSIONS
                excl >>= 1;
#endif
                tj = (tj + 1) & (TILE_SIZE - 1);
            }
        }

        // Write results.

        unsigned int offset = x*TILE_SIZE + tgx;
        atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (force.x*0x100000000)));
        atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.y*0x100000000)));
        atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.z*0x100000000)));
        if (x != y) {
            offset = y*TILE_SIZE + tgx;
            atomicAdd(&forceBuffers[offset], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fx*0x100000000)));
            atomicAdd(&forceBuffers[offset+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fy*0x100000000)));
            atomicAdd(&forceBuffers[offset+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fz*0x100000000)));
        }
    }

    // Second loop: tiles without exclusions, either from the neighbor list (with cutoff) or just enumerating all
    // of them (no cutoff).

#ifdef USE_CUTOFF
    const unsigned int numTiles = interactionCount[0];
    int pos = (numTiles > maxTiles ? startTileIndex+warp*numTileIndices/totalWarps : warp*numTiles/totalWarps);
    int end = (numTiles > maxTiles ? startTileIndex+(warp+1)*numTileIndices/totalWarps : (warp+1)*numTiles/totalWarps);
#else
    const unsigned int numTiles = numTileIndices;
    int pos = startTileIndex+warp*numTiles/totalWarps;
    int end = startTileIndex+(warp+1)*numTiles/totalWarps;
#endif
    int skipBase = 0;
    int currentSkipIndex = tbx;
    __shared__ int atomIndices[THREAD_BLOCK_SIZE];
    __shared__ volatile int skipTiles[THREAD_BLOCK_SIZE];
    skipTiles[threadIdx.x] = -1;
    
    while (pos < end) {
        const bool hasExclusions = false;
        real3 force = make_real3(0);
        bool includeTile = true;

        // Extract the coordinates of this tile.
        
        unsigned int x, y;
        bool singlePeriodicCopy = false;
#ifdef USE_CUTOFF
        if (numTiles <= maxTiles) {
            ushort2 tileIndices = tiles[pos];
            x = tileIndices.x;
            singlePeriodicCopy = tileIndices.y;
        }
        else
#endif
        {
            y = (unsigned int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
            x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
            if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
                y += (x < y ? -1 : 1);
                x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
            }

            // Skip over tiles that have exclusions, since they were already processed.

            while (skipTiles[tbx+TILE_SIZE-1] < pos) {
                if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
                    ushort2 tile = exclusionTiles[skipBase+tgx];
                    skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
                }
                else
                    skipTiles[threadIdx.x] = end;
                skipBase += TILE_SIZE;            
                currentSkipIndex = tbx;
            }
            while (skipTiles[currentSkipIndex] < pos)
                currentSkipIndex++;
            includeTile = (skipTiles[currentSkipIndex] != pos);
        }
        if (includeTile) {
            unsigned int atom1 = x*TILE_SIZE + tgx;

            // Load atom data for this tile.

            real4 posq1 = posq[atom1];
            LOAD_ATOM1_PARAMETERS
            const unsigned int localAtomIndex = threadIdx.x;
#ifdef USE_CUTOFF
            unsigned int j = (numTiles <= maxTiles ? interactingAtoms[pos*TILE_SIZE+tgx] : y*TILE_SIZE + tgx);
#else
            unsigned int j = y*TILE_SIZE + tgx;
#endif
            atomIndices[threadIdx.x] = j;
            if (j < PADDED_NUM_ATOMS) {
                real4 tempPosq = posq[j];
                localData[localAtomIndex].x = tempPosq.x;
                localData[localAtomIndex].y = tempPosq.y;
                localData[localAtomIndex].z = tempPosq.z;
                localData[localAtomIndex].q = tempPosq.w;
                LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
                localData[localAtomIndex].fx = 0.0f;
                localData[localAtomIndex].fy = 0.0f;
                localData[localAtomIndex].fz = 0.0f;
            }
#ifdef USE_PERIODIC
            if (singlePeriodicCopy) {
                // 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.

                real4 blockCenterX = blockCenter[x];
                posq1.x -= floor((posq1.x-blockCenterX.x)*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
                posq1.y -= floor((posq1.y-blockCenterX.y)*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
                posq1.z -= floor((posq1.z-blockCenterX.z)*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
                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;
                unsigned int tj = tgx;
                for (j = 0; j < TILE_SIZE; j++) {
                    int atom2 = tbx+tj;
                    real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                    real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
                    real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
                    if (r2 < CUTOFF_SQUARED) {
                        real invR = RSQRT(r2);
                        real r = RECIP(invR);
                        LOAD_ATOM2_PARAMETERS
                        atom2 = atomIndices[tbx+tj];
#ifdef USE_SYMMETRIC
                        real dEdR = 0.0f;
#else
                        real3 dEdR1 = make_real3(0);
                        real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
                        bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
#endif
                        real tempEnergy = 0.0f;
                        COMPUTE_INTERACTION
                        energy += tempEnergy;
#ifdef USE_SYMMETRIC
                        delta *= dEdR;
                        force.x -= delta.x;
                        force.y -= delta.y;
                        force.z -= delta.z;
                        localData[tbx+tj].fx += delta.x;
                        localData[tbx+tj].fy += delta.y;
                        localData[tbx+tj].fz += delta.z;
#else
                        force.x -= dEdR1.x;
                        force.y -= dEdR1.y;
                        force.z -= dEdR1.z;
                        localData[tbx+tj].fx += dEdR2.x;
                        localData[tbx+tj].fy += dEdR2.y;
                        localData[tbx+tj].fz += dEdR2.z;
#endif
                    }
                    tj = (tj + 1) & (TILE_SIZE - 1);
                }
            }
            else
#endif
            {
                // We need to apply periodic boundary conditions separately for each interaction.

                unsigned int tj = tgx;
                for (j = 0; j < TILE_SIZE; j++) {
                    int atom2 = tbx+tj;
                    real4 posq2 = make_real4(localData[atom2].x, localData[atom2].y, localData[atom2].z, localData[atom2].q);
                    real3 delta = make_real3(posq2.x-posq1.x, posq2.y-posq1.y, posq2.z-posq1.z);
#ifdef USE_PERIODIC
                    delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
                    delta.y -= floor(delta.y*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
                    delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
#endif
                    real r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
#ifdef USE_CUTOFF
                    if (r2 < CUTOFF_SQUARED) {
#endif
                        real invR = RSQRT(r2);
                        real r = RECIP(invR);
                        LOAD_ATOM2_PARAMETERS
                        atom2 = atomIndices[tbx+tj];
#ifdef USE_SYMMETRIC
                        real dEdR = 0.0f;
#else
                        real3 dEdR1 = make_real3(0);
                        real3 dEdR2 = make_real3(0);
#endif
#ifdef USE_EXCLUSIONS
                        bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
#endif
                        real tempEnergy = 0.0f;
                        COMPUTE_INTERACTION
                        energy += tempEnergy;
#ifdef USE_SYMMETRIC
                        delta *= dEdR;
                        force.x -= delta.x;
                        force.y -= delta.y;
                        force.z -= delta.z;
                        localData[tbx+tj].fx += delta.x;
                        localData[tbx+tj].fy += delta.y;
                        localData[tbx+tj].fz += delta.z;
#else
                        force.x -= dEdR1.x;
                        force.y -= dEdR1.y;
                        force.z -= dEdR1.z;
                        localData[tbx+tj].fx += dEdR2.x;
                        localData[tbx+tj].fy += dEdR2.y;
                        localData[tbx+tj].fz += dEdR2.z;
#endif
#ifdef USE_CUTOFF
                    }
#endif
                    tj = (tj + 1) & (TILE_SIZE - 1);
                }
            }

            // Write results.

            atomicAdd(&forceBuffers[atom1], static_cast<unsigned long long>((long long) (force.x*0x100000000)));
            atomicAdd(&forceBuffers[atom1+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.y*0x100000000)));
            atomicAdd(&forceBuffers[atom1+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (force.z*0x100000000)));
#ifdef USE_CUTOFF
            unsigned int atom2 = atomIndices[threadIdx.x];
#else
            unsigned int atom2 = y*TILE_SIZE + tgx;
#endif
            if (atom2 < PADDED_NUM_ATOMS) {
                atomicAdd(&forceBuffers[atom2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fx*0x100000000)));
                atomicAdd(&forceBuffers[atom2+PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fy*0x100000000)));
                atomicAdd(&forceBuffers[atom2+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fz*0x100000000)));
            }
        }
        pos++;
    }
    energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;
}