refactor compute reduce/chunk

This commit is contained in:
Axel Kohlmeyer
2022-10-05 14:03:35 -04:00
parent 2bc0825e4a
commit cf4e133a2c
3 changed files with 112 additions and 143 deletions

View File

@ -63,8 +63,7 @@ ComputeChunkSpreadAtom(LAMMPS *lmp, int narg, char **arg) :
val.id = argi.get_name();
val.val.c = nullptr;
if ((val.which == ArgInfo::UNKNOWN) || (val.which == ArgInfo::NONE)
|| (argi.get_dim() > 1))
if ((val.which == ArgInfo::UNKNOWN) || (val.which == ArgInfo::NONE) || (argi.get_dim() > 1))
error->all(FLERR,"Illegal compute chunk/spread/atom argument: {}", arg[iarg]);
values.push_back(val);

View File

@ -30,19 +30,17 @@
using namespace LAMMPS_NS;
enum{SUM,MINN,MAXX};
enum{ SUM, MINN, MAXX };
#define BIG 1.0e20
/* ---------------------------------------------------------------------- */
ComputeReduceChunk::ComputeReduceChunk(LAMMPS *lmp, int narg, char **arg) :
Compute(lmp, narg, arg),
which(nullptr), argindex(nullptr), value2index(nullptr), idchunk(nullptr), ids(nullptr),
vlocal(nullptr), vglobal(nullptr), alocal(nullptr), aglobal(nullptr), varatom(nullptr)
Compute(lmp, narg, arg), idchunk(nullptr), vlocal(nullptr), vglobal(nullptr),
alocal(nullptr), aglobal(nullptr), varatom(nullptr), cchunk(nullptr), ichunk(nullptr)
{
if (narg < 6) error->all(FLERR,"Illegal compute reduce/chunk command");
if (narg < 6) utils::missing_cmd_args(FLERR,"compute reduce/chunk", error);
// ID of compute chunk/atom
@ -54,7 +52,7 @@ ComputeReduceChunk::ComputeReduceChunk(LAMMPS *lmp, int narg, char **arg) :
if (strcmp(arg[4],"sum") == 0) mode = SUM;
else if (strcmp(arg[4],"min") == 0) mode = MINN;
else if (strcmp(arg[4],"max") == 0) mode = MAXX;
else error->all(FLERR,"Illegal compute reduce/chunk command");
else error->all(FLERR,"Unknown compute reduce/chunk mode: {}", arg[4]);
int iarg = 5;
@ -67,30 +65,22 @@ ComputeReduceChunk::ComputeReduceChunk(LAMMPS *lmp, int narg, char **arg) :
if (earg != &arg[iarg]) expand = 1;
arg = earg;
// parse values until one isn't recognized
which = new int[nargnew];
argindex = new int[nargnew];
ids = new char*[nargnew];
value2index = new int[nargnew];
for (int i=0; i < nargnew; ++i) {
which[i] = argindex[i] = value2index[i] = ArgInfo::UNKNOWN;
ids[i] = nullptr;
}
nvalues = 0;
// parse values until
values.clear();
for (iarg = 0; iarg < nargnew; iarg++) {
ArgInfo argi(arg[iarg]);
which[nvalues] = argi.get_type();
argindex[nvalues] = argi.get_index1();
ids[nvalues] = argi.copy_name();
value_t val;
val.which = argi.get_type();
val.argindex = argi.get_index1();
val.id = argi.get_name();
val.val.c = nullptr;
if ((which[nvalues] == ArgInfo::UNKNOWN) || (which[nvalues] == ArgInfo::NONE)
|| (argi.get_dim() > 1))
error->all(FLERR,"Illegal compute reduce/chunk command");
if ((val.which == ArgInfo::UNKNOWN) || (val.which == ArgInfo::NONE) || (argi.get_dim() > 1))
error->all(FLERR,"Illegal compute reduce/chunk argument: {}", arg[iarg]);
nvalues++;
values.push_back(val);
}
// if wildcard expansion occurred, free earg memory from expand_args()
@ -102,58 +92,55 @@ ComputeReduceChunk::ComputeReduceChunk(LAMMPS *lmp, int narg, char **arg) :
// error check
for (int i = 0; i < nvalues; i++) {
if (which[i] == ArgInfo::COMPUTE) {
int icompute = modify->find_compute(ids[i]);
if (icompute < 0)
error->all(FLERR,"Compute ID for compute reduce/chunk does not exist");
if (!modify->compute[icompute]->peratom_flag)
error->all(FLERR,"Compute reduce/chunk compute does not "
"calculate per-atom values");
if (argindex[i] == 0 &&
modify->compute[icompute]->size_peratom_cols != 0)
error->all(FLERR,"Compute reduce/chunk compute does not "
"calculate a per-atom vector");
if (argindex[i] && modify->compute[icompute]->size_peratom_cols == 0)
error->all(FLERR,"Compute reduce/chunk compute does not "
"calculate a per-atom array");
if (argindex[i] &&
argindex[i] > modify->compute[icompute]->size_peratom_cols)
error->all(FLERR,
"Compute reduce/chunk compute array is accessed out-of-range");
for (auto &val : values) {
if (val.which == ArgInfo::COMPUTE) {
val.val.c = modify->get_compute_by_id(val.id);
if (!val.val.c)
error->all(FLERR,"Compute ID {} for compute reduce/chunk does not exist", val.id);
if (!val.val.c->peratom_flag)
error->all(FLERR,"Compute reduce/chunk compute {} does not calculate per-atom values",
val.id);
if ((val.argindex == 0) && (val.val.c->size_peratom_cols != 0))
error->all(FLERR,"Compute reduce/chunk compute {} does not calculate a per-atom vector",
val.id);
if (val.argindex && (val.val.c->size_peratom_cols == 0))
error->all(FLERR,"Compute reduce/chunk compute {} does not calculate a per-atom array",
val.id);
if (val.argindex && (val.argindex > val.val.c->size_peratom_cols))
error->all(FLERR, "Compute reduce/chunk compute array {} is accessed out-of-range", val.id);
} else if (which[i] == ArgInfo::FIX) {
auto ifix = modify->get_fix_by_id(ids[i]);
if (!ifix)
error->all(FLERR,"Fix ID {} for compute reduce/chunk does not exist", ids[i]);
if (!ifix->peratom_flag)
error->all(FLERR,"Compute reduce/chunk fix {} does not calculate per-atom values", ids[i]);
if ((argindex[i] == 0) && (ifix->size_peratom_cols != 0))
error->all(FLERR,"Compute reduce/chunk fix {} does not calculate a per-atom vector", ids[i]);
if (argindex[i] && (ifix->size_peratom_cols == 0))
error->all(FLERR,"Compute reduce/chunk fix {} does not calculate a per-atom array", ids[i]);
if (argindex[i] && (argindex[i] > ifix->size_peratom_cols))
error->all(FLERR,"Compute reduce/chunk fix {} array is accessed out-of-range", ids[i]);
} else if (val.which == ArgInfo::FIX) {
val.val.f = modify->get_fix_by_id(val.id);
if (!val.val.f)
error->all(FLERR,"Fix ID {} for compute reduce/chunk does not exist", val.id);
if (!val.val.f->peratom_flag)
error->all(FLERR,"Compute reduce/chunk fix {} does not calculate per-atom values", val.id);
if ((val.argindex == 0) && (val.val.f->size_peratom_cols != 0))
error->all(FLERR,"Compute reduce/chunk fix {} does not calculate a per-atom vector", val.id);
if (val.argindex && (val.val.f->size_peratom_cols == 0))
error->all(FLERR,"Compute reduce/chunk fix {} does not calculate a per-atom array", val.id);
if (val.argindex && (val.argindex > val.val.f->size_peratom_cols))
error->all(FLERR,"Compute reduce/chunk fix {} array is accessed out-of-range", val.id);
} else if (which[i] == ArgInfo::VARIABLE) {
int ivariable = input->variable->find(ids[i]);
if (ivariable < 0)
error->all(FLERR,"Variable name for compute reduce/chunk does not exist");
if (input->variable->atomstyle(ivariable) == 0)
} else if (val.which == ArgInfo::VARIABLE) {
val.val.v = input->variable->find(val.id.c_str());
if (val.val.v < 0)
error->all(FLERR,"Variable name {} for compute reduce/chunk does not exist", val.id);
if (input->variable->atomstyle(val.val.v) == 0)
error->all(FLERR,"Compute reduce/chunk variable is not atom-style variable");
}
}
// this compute produces either a vector or array
if (nvalues == 1) {
if (values.size() == 1) {
vector_flag = 1;
size_vector_variable = 1;
extvector = 0;
} else {
array_flag = 1;
size_array_rows_variable = 1;
size_array_cols = nvalues;
size_array_cols = values.size();
extarray = 0;
}
@ -175,13 +162,7 @@ ComputeReduceChunk::ComputeReduceChunk(LAMMPS *lmp, int narg, char **arg) :
ComputeReduceChunk::~ComputeReduceChunk()
{
delete [] idchunk;
delete [] which;
delete [] argindex;
for (int m = 0; m < nvalues; m++) delete [] ids[m];
delete [] ids;
delete [] value2index;
delete[] idchunk;
memory->destroy(vlocal);
memory->destroy(vglobal);
@ -199,24 +180,21 @@ void ComputeReduceChunk::init()
// set indices of all computes,fixes,variables
for (int m = 0; m < nvalues; m++) {
if (which[m] == ArgInfo::COMPUTE) {
int icompute = modify->find_compute(ids[m]);
if (icompute < 0)
error->all(FLERR,"Compute ID for compute reduce/chunk does not exist");
value2index[m] = icompute;
for (auto &val : values) {
if (val.which == ArgInfo::COMPUTE) {
val.val.c = modify->get_compute_by_id(val.id);
if (!val.val.c)
error->all(FLERR,"Compute ID {} for compute reduce/chunk does not exist", val.id);
} else if (which[m] == ArgInfo::FIX) {
int ifix = modify->find_fix(ids[m]);
if (ifix < 0)
error->all(FLERR,"Fix ID for compute reduce/chunk does not exist");
value2index[m] = ifix;
} else if (val.which == ArgInfo::FIX) {
val.val.f = modify->get_fix_by_id(val.id);
if (!val.val.f)
error->all(FLERR,"Fix ID {} for compute reduce/chunk does not exist", val.id);
} else if (which[m] == ArgInfo::VARIABLE) {
int ivariable = input->variable->find(ids[m]);
if (ivariable < 0)
error->all(FLERR,"Variable name for compute reduce/chunk does not exist");
value2index[m] = ivariable;
} else if (val.which == ArgInfo::VARIABLE) {
val.val.v = input->variable->find(val.id.c_str());
if (val.val.v < 0)
error->all(FLERR,"Variable name {} for compute reduce/chunk does not exist", val.id);
}
}
}
@ -225,13 +203,10 @@ void ComputeReduceChunk::init()
void ComputeReduceChunk::init_chunk()
{
int icompute = modify->find_compute(idchunk);
if (icompute < 0)
error->all(FLERR,"Chunk/atom compute does not exist for "
"compute reduce/chunk");
cchunk = dynamic_cast<ComputeChunkAtom *>(modify->compute[icompute]);
if (strcmp(cchunk->style,"chunk/atom") != 0)
error->all(FLERR,"Compute reduce/chunk does not use chunk/atom compute");
cchunk = dynamic_cast<ComputeChunkAtom *>(modify->get_compute_by_id(idchunk));
if (!cchunk)
error->all(FLERR,"Compute chunk/atom {} does not exist or is incorrect style for "
"compute reduce/chunk", idchunk);
}
/* ---------------------------------------------------------------------- */
@ -295,26 +270,23 @@ void ComputeReduceChunk::compute_array()
memory->destroy(alocal);
memory->destroy(aglobal);
maxchunk = nchunk;
memory->create(alocal,maxchunk,nvalues,"reduce/chunk:alocal");
memory->create(aglobal,maxchunk,nvalues,"reduce/chunk:aglobal");
memory->create(alocal,maxchunk,values.size(),"reduce/chunk:alocal");
memory->create(aglobal,maxchunk,values.size(),"reduce/chunk:aglobal");
array = aglobal;
}
// perform local reduction of all peratom values
for (int m = 0; m < nvalues; m++) compute_one(m,&alocal[0][m],nvalues);
for (std::size_t m = 0; m < values.size(); m++) compute_one(m,&alocal[0][m],values.size());
// reduce the per-chunk values across all procs
if (mode == SUM)
MPI_Allreduce(&alocal[0][0],&aglobal[0][0],nchunk*nvalues,
MPI_DOUBLE,MPI_SUM,world);
MPI_Allreduce(&alocal[0][0],&aglobal[0][0],nchunk*values.size(),MPI_DOUBLE,MPI_SUM,world);
else if (mode == MINN)
MPI_Allreduce(&alocal[0][0],&aglobal[0][0],nchunk*nvalues,
MPI_DOUBLE,MPI_MIN,world);
MPI_Allreduce(&alocal[0][0],&aglobal[0][0],nchunk*values.size(),MPI_DOUBLE,MPI_MIN,world);
else if (mode == MAXX)
MPI_Allreduce(&alocal[0][0],&aglobal[0][0],nchunk*nvalues,
MPI_DOUBLE,MPI_MAX,world);
MPI_Allreduce(&alocal[0][0],&aglobal[0][0],nchunk*values.size(),MPI_DOUBLE,MPI_MAX,world);
}
/* ---------------------------------------------------------------------- */
@ -323,7 +295,7 @@ void ComputeReduceChunk::compute_one(int m, double *vchunk, int nstride)
{
// initialize per-chunk values in accumulation vector
for (int i = 0; i < nvalues*nchunk; i += nstride) vchunk[i] = initvalue;
for (std::size_t i = 0; i < values.size()*nchunk; i += nstride) vchunk[i] = initvalue;
// loop over my atoms
// use peratom input and chunk ID of each atom to update vector
@ -331,27 +303,22 @@ void ComputeReduceChunk::compute_one(int m, double *vchunk, int nstride)
int *mask = atom->mask;
int nlocal = atom->nlocal;
auto &val = values[m];
int index = -1;
int vidx = value2index[m];
// initialization in case it has not yet been run, e.g. when
// the compute was invoked right after it has been created
if (vidx == ArgInfo::UNKNOWN) {
init();
vidx = value2index[m];
}
if (val.val.c == nullptr) init();
if (which[m] == ArgInfo::COMPUTE) {
Compute *compute = modify->compute[vidx];
if (!(compute->invoked_flag & Compute::INVOKED_PERATOM)) {
compute->compute_peratom();
compute->invoked_flag |= Compute::INVOKED_PERATOM;
if (val.which == ArgInfo::COMPUTE) {
if (!(val.val.c->invoked_flag & Compute::INVOKED_PERATOM)) {
val.val.c->compute_peratom();
val.val.c->invoked_flag |= Compute::INVOKED_PERATOM;
}
if (argindex[m] == 0) {
double *vcompute = compute->vector_atom;
if (val.argindex == 0) {
double *vcompute = val.val.c->vector_atom;
for (int i = 0; i < nlocal; i++) {
if (!(mask[i] & groupbit)) continue;
index = ichunk[i]-1;
@ -359,8 +326,8 @@ void ComputeReduceChunk::compute_one(int m, double *vchunk, int nstride)
combine(vchunk[index*nstride],vcompute[i]);
}
} else {
double **acompute = compute->array_atom;
int argindexm1 = argindex[m] - 1;
double **acompute = val.val.c->array_atom;
int argindexm1 = val.argindex - 1;
for (int i = 0; i < nlocal; i++) {
if (!(mask[i] & groupbit)) continue;
index = ichunk[i]-1;
@ -371,13 +338,12 @@ void ComputeReduceChunk::compute_one(int m, double *vchunk, int nstride)
// access fix fields, check if fix frequency is a match
} else if (which[m] == ArgInfo::FIX) {
Fix *fix = modify->fix[vidx];
if (update->ntimestep % fix->peratom_freq)
} else if (val.which == ArgInfo::FIX) {
if (update->ntimestep % val.val.f->peratom_freq)
error->all(FLERR,"Fix used in compute reduce/chunk not computed at compatible time");
if (argindex[m] == 0) {
double *vfix = fix->vector_atom;
if (val.argindex == 0) {
double *vfix = val.val.f->vector_atom;
for (int i = 0; i < nlocal; i++) {
if (!(mask[i] & groupbit)) continue;
index = ichunk[i]-1;
@ -385,8 +351,8 @@ void ComputeReduceChunk::compute_one(int m, double *vchunk, int nstride)
combine(vchunk[index*nstride],vfix[i]);
}
} else {
double **afix = fix->array_atom;
int argindexm1 = argindex[m] - 1;
double **afix = val.val.f->array_atom;
int argindexm1 = val.argindex - 1;
for (int i = 0; i < nlocal; i++) {
if (!(mask[i] & groupbit)) continue;
index = ichunk[i]-1;
@ -397,14 +363,14 @@ void ComputeReduceChunk::compute_one(int m, double *vchunk, int nstride)
// evaluate atom-style variable
} else if (which[m] == ArgInfo::VARIABLE) {
} else if (val.which == ArgInfo::VARIABLE) {
if (atom->nmax > maxatom) {
memory->destroy(varatom);
maxatom = atom->nmax;
memory->create(varatom,maxatom,"reduce/chunk:varatom");
}
input->variable->compute_atom(vidx,igroup,varatom,1,0);
input->variable->compute_atom(val.val.v,igroup,varatom,1,0);
for (int i = 0; i < nlocal; i++) {
if (!(mask[i] & groupbit)) continue;
index = ichunk[i]-1;
@ -449,11 +415,8 @@ void ComputeReduceChunk::lock_enable()
void ComputeReduceChunk::lock_disable()
{
int icompute = modify->find_compute(idchunk);
if (icompute >= 0) {
cchunk = dynamic_cast<ComputeChunkAtom *>(modify->compute[icompute]);
cchunk->lockcount--;
}
cchunk = dynamic_cast<ComputeChunkAtom *>(modify->get_compute_by_id(idchunk));
if (cchunk) cchunk->lockcount--;
}
/* ----------------------------------------------------------------------
@ -491,7 +454,7 @@ void ComputeReduceChunk::unlock(Fix *fixptr)
double ComputeReduceChunk::memory_usage()
{
double bytes = (bigint) maxatom * sizeof(double);
if (nvalues == 1) bytes += (double) maxchunk * 2 * sizeof(double);
else bytes += (double) maxchunk * nvalues * 2 * sizeof(double);
if (values.size() == 1) bytes += (double) maxchunk * 2 * sizeof(double);
else bytes += (double) maxchunk * values.size() * 2 * sizeof(double);
return bytes;
}

View File

@ -41,12 +41,21 @@ class ComputeReduceChunk : public Compute {
double memory_usage() override;
private:
int mode, nvalues;
int *which, *argindex, *value2index;
char *idchunk;
char **ids;
struct value_t {
int which;
int argindex;
std::string id;
union {
class Compute *c;
class Fix *f;
int v;
} val;
};
std::vector<value_t> values;
int nchunk;
char *idchunk;
int mode, nchunk;
int maxchunk, maxatom;
double initvalue;
double *vlocal, *vglobal;
@ -60,8 +69,6 @@ class ComputeReduceChunk : public Compute {
void compute_one(int, double *, int);
void combine(double &, double);
};
} // namespace LAMMPS_NS
#endif
#endif