lammps/lib/cuda/neighbor_kernel.cu

661 lines
19 KiB
Plaintext

/* ----------------------------------------------------------------------
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
Original Version:
http://lammps.sandia.gov, Sandia National Laboratories
Steve Plimpton, sjplimp@sandia.gov
See the README file in the top-level LAMMPS directory.
-----------------------------------------------------------------------
USER-CUDA Package and associated modifications:
https://sourceforge.net/projects/lammpscuda/
Christian Trott, christian.trott@tu-ilmenau.de
Lars Winterfeld, lars.winterfeld@tu-ilmenau.de
Theoretical Physics II, University of Technology Ilmenau, Germany
See the README file in the USER-CUDA directory.
This software is distributed under the GNU General Public License.
------------------------------------------------------------------------- */
#define SBBITS 30
__global__ void Binning_Kernel(int* binned_id, int bin_nmax, int bin_dim_x, int bin_dim_y, int bin_dim_z,
CUDA_FLOAT rez_bin_size_x, CUDA_FLOAT rez_bin_size_y, CUDA_FLOAT rez_bin_size_z)
{
int i = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
/*int* bin_count=(int*) _buffer;
bin_count=bin_count+20;
CUDA_FLOAT* binned_x=(CUDA_FLOAT*)(bin_count+bin_dim_x*bin_dim_y*bin_dim_z);*/
CUDA_FLOAT* binned_x = (CUDA_FLOAT*) _buffer;
binned_x = &binned_x[2];
int* bin_count = (int*) &binned_x[3 * bin_dim_x * bin_dim_y * bin_dim_z * bin_nmax];
if(i < _nall) {
// copy atom position from global device memory to local register
// in this 3 steps to get as much coalesced access as possible
X_FLOAT* my_x = _x + i;
CUDA_FLOAT x_i = *my_x;
my_x += _nmax;
CUDA_FLOAT y_i = *my_x;
my_x += _nmax;
CUDA_FLOAT z_i = *my_x;
// calculate flat bin index
int bx = __float2int_rd(rez_bin_size_x * (x_i - _sublo[0])) + 2;
int by = __float2int_rd(rez_bin_size_y * (y_i - _sublo[1])) + 2;
int bz = __float2int_rd(rez_bin_size_z * (z_i - _sublo[2])) + 2;
bx -= bx * negativCUDA(1.0f * bx);
bx -= (bx - bin_dim_x + 1) * negativCUDA(1.0f * bin_dim_x - 1.0f - 1.0f * bx);
by -= by * negativCUDA(1.0f * by);
by -= (by - bin_dim_y + 1) * negativCUDA(1.0f * bin_dim_y - 1.0f - 1.0f * by);
bz -= bz * negativCUDA(1.0f * bz);
bz -= (bz - bin_dim_z + 1) * negativCUDA(1.0f * bin_dim_z - 1.0f - 1.0f * bz);
const unsigned j = bin_dim_z * (bin_dim_y * bx + by) + bz;
// add new atom to bin, get bin-array position
const unsigned k = atomicAdd(& bin_count[j], 1);
if(k < bin_nmax) {
binned_id [bin_nmax * j + k] = i;
binned_x [3 * bin_nmax * j + k] = x_i;
binned_x [3 * bin_nmax * j + k + bin_nmax] = y_i;
binned_x [3 * bin_nmax * j + k + 2 * bin_nmax] = z_i;
} else {
// normally, this should not happen:
int errorn = atomicAdd((int*) _buffer, 1);
MYEMUDBG(printf("# CUDA: Binning_Kernel: WARNING: atom %i ignored, no place left in bin %u\n", i, j);)
}
}
}
__device__ inline int exclusion(int &i, int &j, int &itype, int &jtype)
{
int m;
if(_nex_type)
if(_ex_type[itype * _cuda_ntypes + jtype]) return 1;
if(_nex_group) {
for(m = 0; m < _nex_group; m++) {
if(_mask[i] & _ex1_bit[m] && _mask[j] & _ex2_bit[m]) return 1;
if(_mask[i] & _ex2_bit[m] && _mask[j] & _ex1_bit[m]) return 1;
}
}
if(_nex_mol) {
if(_molecule[i] == _molecule[j])
for(m = 0; m < _nex_mol; m++)
if(_mask[i] & _ex_mol_bit[m] && _mask[j] & _ex_mol_bit[m]) return 1;
}
return 0;
}
extern __shared__ CUDA_FLOAT shared[];
__device__ inline int find_special(int3 &n, int* list, int &tag, int3 flag)
{
int k = n.z;
for(int l = 0; l < n.z; l++) k = ((list[l] == tag) ? l : k);
return k < n.x ? flag.x : (k < n.y ? flag.y : (k < n.z ? flag.z : 0));
}
template <const unsigned int exclude>
__global__ void NeighborBuildFullBin_Kernel(int* binned_id, int bin_nmax, int bin_dim_x, int bin_dim_y, CUDA_FLOAT globcutoff, int block_style, bool neighall)
{
int natoms = neighall ? _nall : _nlocal;
//const bool domol=false;
int bin_dim_z = gridDim.y;
CUDA_FLOAT* binned_x = (CUDA_FLOAT*) _buffer;
binned_x = &binned_x[2];
int* bin_count = (int*) &binned_x[3 * bin_dim_x * bin_dim_y * bin_dim_z * bin_nmax];
int bin = __mul24(gridDim.y, blockIdx.x) + blockIdx.y;
int bin_x = blockIdx.x / bin_dim_y;
int bin_y = blockIdx.x - bin_x * bin_dim_y;
int bin_z = blockIdx.y;
int bin_c = bin_count[bin];
CUDA_FLOAT cut;
if(globcutoff > 0)
cut = globcutoff;
int i = _nall;
CUDA_FLOAT* my_x;
CUDA_FLOAT x_i, y_i, z_i;
for(int actOffset = 0; actOffset < bin_c; actOffset += blockDim.x) {
int actIdx = threadIdx.x + actOffset;
CUDA_FLOAT* other_x = shared;
int* other_id = (int*) &other_x[3 * blockDim.x];
if(actIdx < bin_c) {
i = binned_id[__mul24(bin, bin_nmax) + actIdx];
my_x = binned_x + __mul24(__mul24(bin, 3), bin_nmax) + actIdx;
x_i = *my_x;
my_x += bin_nmax;
y_i = *my_x;
my_x += bin_nmax;
z_i = *my_x;
} else
i = 2 * _nall;
__syncthreads();
int jnum = 0;
int itype;
if(i < natoms) {
jnum = 0;
_ilist[i] = i;
itype = _type[i];
}
//__syncthreads();
for(int otherActOffset = 0; otherActOffset < bin_c; otherActOffset += blockDim.x) {
int otherActIdx = threadIdx.x + otherActOffset;
if(otherActIdx < bin_c) {
if(otherActOffset == actOffset) {
other_id[threadIdx.x] = i;
other_x[threadIdx.x] = x_i;
other_x[threadIdx.x + blockDim.x] = y_i;
other_x[threadIdx.x + 2 * blockDim.x] = z_i;
} else {
other_id[threadIdx.x] = binned_id[__mul24(bin, bin_nmax) + otherActIdx];
my_x = binned_x + __mul24(__mul24(bin, 3), bin_nmax) + otherActIdx;
other_x[threadIdx.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + blockDim.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + __mul24(2, blockDim.x)] = *my_x;
}
}
__syncthreads();
int kk = threadIdx.x;
for(int k = 0; k < MIN(bin_c - otherActOffset, blockDim.x); ++k) {
if(i < natoms) {
kk++;
kk = kk < MIN(bin_c - otherActOffset, blockDim.x) ? kk : 0;
int j = other_id[kk];
if(exclude && exclusion(i, j, itype, _type[j])) continue;
if(globcutoff < 0) {
int jtype = _type[j];
cut = _cutneighsq[itype * _cuda_ntypes + jtype];
}
CUDA_FLOAT delx = x_i - other_x[kk];
CUDA_FLOAT dely = y_i - other_x[kk + blockDim.x];
CUDA_FLOAT delz = z_i - other_x[kk + 2 * blockDim.x];
CUDA_FLOAT rsq = delx * delx + dely * dely + delz * delz;
if(rsq <= cut && i != j) {
if(jnum < _maxneighbors) {
if(block_style)
_neighbors[i * _maxneighbors + jnum] = j;
else
_neighbors[i + jnum * natoms] = j;
}
++jnum;
}
}
}
__syncthreads();
}
for(int obin_x = bin_x - 1; obin_x < bin_x + 2; obin_x++)
for(int obin_y = bin_y - 1; obin_y < bin_y + 2; obin_y++)
for(int obin_z = bin_z - 1; obin_z < bin_z + 2; obin_z++) {
if(obin_x < 0 || obin_y < 0 || obin_z < 0) continue;
if(obin_x >= bin_dim_x || obin_y >= bin_dim_y || obin_z >= bin_dim_z) continue;
int other_bin = bin_dim_z * (bin_dim_y * obin_x + obin_y) + obin_z;
if(other_bin == bin) continue;
int obin_c = bin_count[other_bin];
for(int otherActOffset = 0; otherActOffset < obin_c; otherActOffset += blockDim.x) {
int otherActIdx = otherActOffset + threadIdx.x;
if(threadIdx.x < MIN(blockDim.x, obin_c - otherActOffset)) {
other_id[threadIdx.x] = binned_id[__mul24(other_bin, bin_nmax) + otherActIdx];
my_x = binned_x + __mul24(__mul24(other_bin, 3), bin_nmax) + otherActIdx;
other_x[threadIdx.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + blockDim.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + 2 * blockDim.x] = *my_x;
}
__syncthreads();
for(int k = 0; k < MIN(blockDim.x, obin_c - otherActOffset); ++k) {
if(i < natoms) {
int j = other_id[k];
if(exclude && exclusion(i, j, itype, _type[j])) continue;
if(globcutoff < 0) {
int jtype = _type[j];
cut = _cutneighsq[itype * _cuda_ntypes + jtype];
}
CUDA_FLOAT delx = x_i - other_x[k];
CUDA_FLOAT dely = y_i - other_x[k + blockDim.x];
CUDA_FLOAT delz = z_i - other_x[k + 2 * blockDim.x];
CUDA_FLOAT rsq = delx * delx + dely * dely + delz * delz;
if(rsq <= cut && i != j) {
if(jnum < _maxneighbors) {
if(block_style)
_neighbors[i * _maxneighbors + jnum] = j;
else
_neighbors[i + jnum * natoms] = j;
}
++jnum;
}
}
}
__syncthreads();
}
}
if(jnum > _maxneighbors)((int*)_buffer)[0] = -jnum;
if(i < natoms)
_numneigh[i] = jnum;
}
}
__global__ void FindSpecial(int block_style)
{
int ii = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
int which;
int tag_mask = 0;
int3 spec_flag;
int3 mynspecial = {0, 0, 1};
if(ii >= _nlocal) return;
int special_id[CUDA_MAX_NSPECIAL];
int i = _ilist[ii];
if(i >= _nlocal) return;
int jnum = _numneigh[i];
if(_special_flag[1] == 0) spec_flag.x = -1;
else if(_special_flag[1] == 1) spec_flag.x = 0;
else spec_flag.x = 1;
if(_special_flag[2] == 0) spec_flag.y = -1;
else if(_special_flag[2] == 1) spec_flag.y = 0;
else spec_flag.y = 2;
if(_special_flag[3] == 0) spec_flag.z = -1;
else if(_special_flag[3] == 1) spec_flag.z = 0;
else spec_flag.z = 3;
mynspecial.x = _nspecial[i];
mynspecial.y = _nspecial[i + _nmax];
mynspecial.z = _nspecial[i + 2 * _nmax];
if(i < _nlocal) {
int* list = &_special[i];
for(int k = 0; k < mynspecial.z; k++) {
special_id[k] = list[k * _nmax];
tag_mask = tag_mask | special_id[k];
}
}
for(int k = 0; k < MIN(jnum, _maxneighbors); k++) {
int j;
if(block_style)
j = _neighbors[i * _maxneighbors + k];
else
j = _neighbors[i + k * _nlocal];
int tag_j = _tag[j];
which = 0;
if((tag_mask & tag_j) == tag_j) {
which = find_special(mynspecial, special_id, tag_j, spec_flag);
if(which > 0) {
if(block_style)
_neighbors[i * _maxneighbors + k] = j ^ (which << SBBITS);
else
_neighbors[i + k * _nlocal] = j ^ (which << SBBITS);
} else if(which < 0) {
if(block_style)
_neighbors[i * _maxneighbors + k] = _neighbors[i * _maxneighbors + jnum - 1];
else
_neighbors[i + k * _nlocal] = _neighbors[i + (jnum - 1) * _nlocal];
jnum--;
k--;
}
}
}
_numneigh[i] = jnum;
}
__global__ void NeighborBuildFullBin_OverlapComm_Kernel(int* binned_id, int bin_nmax, int bin_dim_x, int bin_dim_y, CUDA_FLOAT globcutoff, int block_style)
{
int bin_dim_z = gridDim.y;
CUDA_FLOAT* binned_x = (CUDA_FLOAT*) _buffer;
binned_x = &binned_x[2];
int* bin_count = (int*) &binned_x[3 * bin_dim_x * bin_dim_y * bin_dim_z * bin_nmax];
int bin = __mul24(gridDim.y, blockIdx.x) + blockIdx.y;
int bin_x = blockIdx.x / bin_dim_y;
int bin_y = blockIdx.x - bin_x * bin_dim_y;
int bin_z = blockIdx.y;
int bin_c = bin_count[bin];
CUDA_FLOAT cut;
if(globcutoff > 0)
cut = globcutoff;
int i = _nall;
CUDA_FLOAT* my_x;
CUDA_FLOAT x_i, y_i, z_i;
for(int actOffset = 0; actOffset < bin_c; actOffset += blockDim.x) {
int actIdx = threadIdx.x + actOffset;
CUDA_FLOAT* other_x = shared;
int* other_id = (int*) &other_x[3 * blockDim.x];
if(actIdx < bin_c) {
i = binned_id[__mul24(bin, bin_nmax) + actIdx];
my_x = binned_x + __mul24(__mul24(bin, 3), bin_nmax) + actIdx;
x_i = *my_x;
my_x += bin_nmax;
y_i = *my_x;
my_x += bin_nmax;
z_i = *my_x;
} else
i = 2 * _nall;
__syncthreads();
int jnum = 0;
int jnum_border = 0;
int jnum_inner = 0;
int i_border = -1;
int itype;
if(i < _nlocal) {
jnum = 0;
_ilist[i] = i;
itype = _type[i];
}
__syncthreads();
for(int otherActOffset = 0; otherActOffset < bin_c; otherActOffset += blockDim.x) {
int otherActIdx = threadIdx.x + otherActOffset;
if(otherActIdx < bin_c) {
if(otherActOffset == actOffset) {
other_id[threadIdx.x] = i;
other_x[threadIdx.x] = x_i;
other_x[threadIdx.x + blockDim.x] = y_i;
other_x[threadIdx.x + 2 * blockDim.x] = z_i;
} else {
other_id[threadIdx.x] = binned_id[__mul24(bin, bin_nmax) + otherActIdx];
my_x = binned_x + __mul24(__mul24(bin, 3), bin_nmax) + otherActIdx;
other_x[threadIdx.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + blockDim.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + __mul24(2, blockDim.x)] = *my_x;
}
}
__syncthreads();
int kk = threadIdx.x;
for(int k = 0; k < MIN(bin_c - otherActOffset, blockDim.x); ++k) {
if(i < _nlocal) {
kk++;
kk = kk < MIN(bin_c - otherActOffset, blockDim.x) ? kk : 0;
int j = other_id[kk];
if(globcutoff < 0) {
int jtype = _type[j];
cut = _cutneighsq[itype * _cuda_ntypes + jtype];
}
CUDA_FLOAT delx = x_i - other_x[kk];
CUDA_FLOAT dely = y_i - other_x[kk + blockDim.x];
CUDA_FLOAT delz = z_i - other_x[kk + 2 * blockDim.x];
CUDA_FLOAT rsq = delx * delx + dely * dely + delz * delz;
if(rsq <= cut && i != j) {
if((j >= _nlocal) && (i_border < 0))
i_border = atomicAdd(_inum_border, 1);
if(jnum < _maxneighbors) {
if(block_style) {
_neighbors[i * _maxneighbors + jnum] = j;
if(j >= _nlocal) {
_neighbors_border[i_border * _maxneighbors + jnum_border] = j;
} else {
_neighbors_inner[i * _maxneighbors + jnum_inner] = j;
}
} else {
_neighbors[i + jnum * _nlocal] = j;
if(j >= _nlocal) {
_neighbors_border[i_border + jnum_border * _nlocal] = j;
} else {
_neighbors_inner[i + jnum_inner * _nlocal] = j;
}
}
}
++jnum;
if(j >= _nlocal)
jnum_border++;
else
jnum_inner++;
}
}
}
__syncthreads();
}
for(int obin_x = bin_x - 1; obin_x < bin_x + 2; obin_x++)
for(int obin_y = bin_y - 1; obin_y < bin_y + 2; obin_y++)
for(int obin_z = bin_z - 1; obin_z < bin_z + 2; obin_z++) {
if(obin_x < 0 || obin_y < 0 || obin_z < 0) continue;
if(obin_x >= bin_dim_x || obin_y >= bin_dim_y || obin_z >= bin_dim_z) continue;
int other_bin = bin_dim_z * (bin_dim_y * obin_x + obin_y) + obin_z;
if(other_bin == bin) continue;
int obin_c = bin_count[other_bin];
for(int otherActOffset = 0; otherActOffset < obin_c; otherActOffset += blockDim.x) {
int otherActIdx = otherActOffset + threadIdx.x;
if(threadIdx.x < MIN(blockDim.x, obin_c - otherActOffset)) {
other_id[threadIdx.x] = binned_id[__mul24(other_bin, bin_nmax) + otherActIdx];
my_x = binned_x + __mul24(__mul24(other_bin, 3), bin_nmax) + otherActIdx;
other_x[threadIdx.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + blockDim.x] = *my_x;
my_x += bin_nmax;
other_x[threadIdx.x + 2 * blockDim.x] = *my_x;
}
__syncthreads();
for(int k = 0; k < MIN(blockDim.x, obin_c - otherActOffset); ++k) {
if(i < _nlocal) {
int j = other_id[k];
if(globcutoff < 0) {
int jtype = _type[j];
cut = _cutneighsq[itype * _cuda_ntypes + jtype];
}
CUDA_FLOAT delx = x_i - other_x[k];
CUDA_FLOAT dely = y_i - other_x[k + blockDim.x];
CUDA_FLOAT delz = z_i - other_x[k + 2 * blockDim.x];
CUDA_FLOAT rsq = delx * delx + dely * dely + delz * delz;
if(rsq <= cut && i != j) {
if((j >= _nlocal) && (i_border < 0))
i_border = atomicAdd(_inum_border, 1);
if(jnum < _maxneighbors) {
if(block_style) {
_neighbors[i * _maxneighbors + jnum] = j;
if(j >= _nlocal) {
_neighbors_border[i_border * _maxneighbors + jnum_border] = j;
} else {
_neighbors_inner[i * _maxneighbors + jnum_inner] = j;
}
} else {
_neighbors[i + jnum * _nlocal] = j;
if(j >= _nlocal) {
_neighbors_border[i_border + jnum_border * _nlocal] = j;
} else {
_neighbors_inner[i + jnum_inner * _nlocal] = j;
}
}
}
++jnum;
if(j >= _nlocal)
jnum_border++;
else
jnum_inner++;
}
}
}
__syncthreads();
}
}
if(jnum > _maxneighbors)((int*)_buffer)[0] = -jnum;
if(i < _nlocal) {
_numneigh[i] = jnum;
_numneigh_inner[i] = jnum_inner;
if(i_border >= 0) _numneigh_border[i_border] = jnum_border;
if(i_border >= 0) _ilist_border[i_border] = i;
}
}
}
__global__ void NeighborBuildFullNsq_Kernel()
{
int i = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
int* buffer = (int*) _buffer;
if(i < _nlocal) {
X_FLOAT* my_x = _x + i;
CUDA_FLOAT x_i = *my_x;
my_x += _nmax;
CUDA_FLOAT y_i = *my_x;
my_x += _nmax;
CUDA_FLOAT z_i = *my_x;
int jnum = 0;
int* jlist = _firstneigh[i];
_ilist[i] = i;
int itype = _type[i];
__syncthreads();
for(int j = 0; j < _nall; ++j) {
my_x = _x + j;
CUDA_FLOAT x_j = *my_x;
my_x += _nmax;
CUDA_FLOAT y_j = *my_x;
my_x += _nmax;
CUDA_FLOAT z_j = *my_x;
CUDA_FLOAT delx = x_i - x_j;
CUDA_FLOAT dely = y_i - y_j;
CUDA_FLOAT delz = z_i - z_j;
CUDA_FLOAT rsq = delx * delx + dely * dely + delz * delz;
int jtype = _type[j];
if(rsq <= _cutneighsq[itype * _cuda_ntypes + jtype] && i != j) {
if(jnum < _maxneighbors)
jlist[jnum] = j;
if(i == 151)((int*)_buffer)[jnum + 2] = j;
++jnum;
}
__syncthreads();
}
if(jnum > _maxneighbors) buffer[0] = 0;
_numneigh[i] = jnum;
if(i == 151)((int*)_buffer)[1] = jnum;
}
}