int main(int argc, char** argv)
{
std::vector<int> heapVector;
std::vector<std::string> commandVector;
BuildVector(commandVector, std::cin);
heapVector.reserve(500);
MinHeap<int> heap(heapVector);
std::ofstream file;
file.open("output.txt");
std::vector<std::string>::iterator iter;
for (iter = commandVector.begin(); iter != commandVector.end(); ++iter)
{
// split on the space to get the operation and command
auto command = SplitString(*iter, ' ');
auto operation = (*command.begin())[0];
RunOperation(operation, command, heap, std::cout);
}
file.close();
return 0;
}
void LoadBalance3d(Mesh *mesh){
int n,p,k,v,f;
int nprocs = mesh->nprocs;
int procid = mesh->procid;
int **EToV = mesh->EToV;
double **VX = mesh->GX;
double **VY = mesh->GY;
double **VZ = mesh->GZ;
if(!procid) printf("Root: Entering LoadBalance\n");
int Nverts = mesh->Nverts;
int *Kprocs = BuildIntVector(nprocs);
/* local number of elements */
int Klocal = mesh->K;
/* find number of elements on all processors */
MPI_Allgather(&Klocal, 1, MPI_INT, Kprocs, 1, MPI_INT, MPI_COMM_WORLD);
/* element distribution -- cumulative element count on processes */
idxtype *elmdist = idxmalloc(nprocs+1, "elmdist");
elmdist[0] = 0;
for(p=0;p<nprocs;++p)
elmdist[p+1] = elmdist[p] + Kprocs[p];
/* list of element starts */
idxtype *eptr = idxmalloc(Klocal+1, "eptr");
eptr[0] = 0;
for(k=0;k<Klocal;++k)
eptr[k+1] = eptr[k] + Nverts;
/* local element to vertex */
idxtype *eind = idxmalloc(Nverts*Klocal, "eind");
for(k=0;k<Klocal;++k)
for(n=0;n<Nverts;++n)
eind[k*Nverts+n] = EToV[k][n];
/* weight per element */
idxtype *elmwgt = idxmalloc(Klocal, "elmwgt");
for(k=0;k<Klocal;++k)
elmwgt[k] = 1.;
/* weight flag */
int wgtflag = 0;
/* number flag (1=fortran, 0=c) */
int numflag = 0;
/* ncon = 1 */
int ncon = 1;
/* nodes on element face */
int ncommonnodes = 3;
/* number of partitions */
int nparts = nprocs;
/* tpwgts */
float *tpwgts = (float*) calloc(Klocal, sizeof(float));
for(k=0;k<Klocal;++k)
tpwgts[k] = 1./(float)nprocs;
float ubvec[MAXNCON];
for (n=0; n<ncon; ++n)
ubvec[n] = UNBALANCE_FRACTION;
int options[10];
options[0] = 1;
options[PMV3_OPTION_DBGLVL] = 7;
options[PMV3_OPTION_SEED] = 0;
int edgecut;
idxtype *part = idxmalloc(Klocal, "part");
MPI_Comm comm;
MPI_Comm_dup(MPI_COMM_WORLD, &comm);
ParMETIS_V3_PartMeshKway
(elmdist,
eptr,
eind,
elmwgt,
&wgtflag,
&numflag,
&ncon,
&ncommonnodes,
&nparts,
tpwgts,
ubvec,
options,
&edgecut,
part,
&comm);
int **outlist = (int**) calloc(nprocs, sizeof(int*));
double **xoutlist = (double**) calloc(nprocs, sizeof(double*));
double **youtlist = (double**) calloc(nprocs, sizeof(double*));
double **zoutlist = (double**) calloc(nprocs, sizeof(double*));
int *outK = (int*) calloc(nprocs, sizeof(int));
int *inK = (int*) calloc(nprocs, sizeof(int));
MPI_Request *inrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *outrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *xinrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *xoutrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *yinrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *youtrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *zinrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *zoutrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
for(k=0;k<Klocal;++k)
++outK[part[k]];
/* get count of incoming elements from each process */
MPI_Alltoall(outK, 1, MPI_INT,
inK, 1, MPI_INT,
MPI_COMM_WORLD);
/* count totals on each process */
int * newKprocs = BuildIntVector(nprocs);
MPI_Allreduce(outK, newKprocs, nprocs, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
int totalinK = 0;
for(p=0;p<nprocs;++p){
totalinK += inK[p];
}
int **newEToV = BuildIntMatrix(totalinK, Nverts);
double **newVX = BuildMatrix(totalinK, Nverts);
double **newVY = BuildMatrix(totalinK, Nverts);
double **newVZ = BuildMatrix(totalinK, Nverts);
int cnt = 0;
for(p=0;p<nprocs;++p){
MPI_Irecv(newEToV[cnt], Nverts*inK[p], MPI_INT, p, 666+p, MPI_COMM_WORLD,
inrequests+p);
MPI_Irecv(newVX[cnt], Nverts*inK[p], MPI_DOUBLE, p, 1666+p, MPI_COMM_WORLD,
xinrequests+p);
MPI_Irecv(newVY[cnt], Nverts*inK[p], MPI_DOUBLE, p, 2666+p, MPI_COMM_WORLD,
yinrequests+p);
MPI_Irecv(newVZ[cnt], Nverts*inK[p], MPI_DOUBLE, p, 3666+p, MPI_COMM_WORLD,
zinrequests+p);
cnt = cnt + inK[p];
}
for(p=0;p<nprocs;++p){
int cnt = 0;
outlist[p] = BuildIntVector(Nverts*outK[p]);
xoutlist[p] = BuildVector(Nverts*outK[p]);
youtlist[p] = BuildVector(Nverts*outK[p]);
zoutlist[p] = BuildVector(Nverts*outK[p]);
for(k=0;k<Klocal;++k)
if(part[k]==p){
for(v=0;v<Nverts;++v){
outlist[p][cnt] = EToV[k][v];
xoutlist[p][cnt] = VX[k][v];
youtlist[p][cnt] = VY[k][v];
zoutlist[p][cnt] = VZ[k][v];
++cnt;
}
}
MPI_Isend(outlist[p], Nverts*outK[p], MPI_INT, p, 666+procid, MPI_COMM_WORLD,
outrequests+p);
MPI_Isend(xoutlist[p], Nverts*outK[p], MPI_DOUBLE, p, 1666+procid, MPI_COMM_WORLD,
xoutrequests+p);
MPI_Isend(youtlist[p], Nverts*outK[p], MPI_DOUBLE, p, 2666+procid, MPI_COMM_WORLD,
youtrequests+p);
MPI_Isend(zoutlist[p], Nverts*outK[p], MPI_DOUBLE, p, 3666+procid, MPI_COMM_WORLD,
zoutrequests+p);
}
MPI_Status *instatus = (MPI_Status*) calloc(nprocs, sizeof(MPI_Status));
MPI_Status *outstatus = (MPI_Status*) calloc(nprocs, sizeof(MPI_Status));
MPI_Waitall(nprocs, inrequests, instatus);
MPI_Waitall(nprocs, xinrequests, instatus);
MPI_Waitall(nprocs, yinrequests, instatus);
MPI_Waitall(nprocs, zinrequests, instatus);
MPI_Waitall(nprocs, outrequests, outstatus);
MPI_Waitall(nprocs, xoutrequests, outstatus);
MPI_Waitall(nprocs, youtrequests, outstatus);
MPI_Waitall(nprocs, zoutrequests, outstatus);
if(mesh->GX!=NULL){
DestroyMatrix(mesh->GX);
DestroyMatrix(mesh->GY);
DestroyMatrix(mesh->GZ);
DestroyIntMatrix(mesh->EToV);
}
mesh->GX = newVX;
mesh->GY = newVY;
mesh->GZ = newVZ;
mesh->EToV = newEToV;
mesh->K = totalinK;
for(p=0;p<nprocs;++p){
if(outlist[p]){
free(outlist[p]);
free(xoutlist[p]);
free(youtlist[p]);
free(zoutlist[p]);
}
}
free(outK);
free(inK);
free(inrequests);
free(outrequests);
free(xinrequests);
free(xoutrequests);
free(yinrequests);
free(youtrequests);
free(zinrequests);
free(zoutrequests);
free(instatus);
free(outstatus);
}
double InitCPU3d(Mesh *mesh, int Nfields){
printf("Np = %d, BSIZE = %d\n", p_Np, BSIZE);
/* Q */
int sz = mesh->K*(p_Np)*Nfields*sizeof(float); /* TW BLOCK */
mesh->f_Q = (float*) calloc(mesh->K*p_Np*Nfields, sizeof(float));
mesh->f_rhsQ = (float*) calloc(mesh->K*p_Np*Nfields, sizeof(float));
mesh->f_resQ = (float*) calloc(mesh->K*p_Np*Nfields, sizeof(float));
/* float LIFT */
sz = p_Np*(p_Nfp)*(p_Nfaces)*sizeof(float);
mesh->f_LIFT = (float*) malloc(sz);
int sk = 0, n, m, f, k;
for(n=0;n<p_Np;++n){
for(m=0;m<p_Nfp*p_Nfaces;++m){
mesh->f_LIFT[sk++] = mesh->LIFT[n][m];
}
}
/* float Dr & Ds */
sz = p_Np*p_Np*sizeof(float);
mesh->f_Dr = (float*) malloc(sz);
mesh->f_Ds = (float*) malloc(sz);
mesh->f_Dt = (float*) malloc(sz);
sk = 0;
for(n=0;n<p_Np;++n){
for(m=0;m<p_Np;++m){
mesh->f_Dr[sk] = mesh->Dr[n][m];
mesh->f_Ds[sk] = mesh->Ds[n][m];
mesh->f_Dt[sk] = mesh->Dt[n][m];
++sk;
}
}
/* vgeo */
double drdx, dsdx, dtdx;
double drdy, dsdy, dtdy;
double drdz, dsdz, dtdz, J;
mesh->vgeo = (float*) calloc(12*mesh->K, sizeof(float));
for(k=0;k<mesh->K;++k){
GeometricFactors3d(mesh, k,
&drdx, &dsdx, &dtdx,
&drdy, &dsdy, &dtdy,
&drdz, &dsdz, &dtdz, &J);
mesh->vgeo[k*12+0] = drdx; mesh->vgeo[k*12+1] = drdy; mesh->vgeo[k*12+2] = drdz;
mesh->vgeo[k*12+4] = dsdx; mesh->vgeo[k*12+5] = dsdy; mesh->vgeo[k*12+6] = dsdz;
mesh->vgeo[k*12+8] = dtdx; mesh->vgeo[k*12+9] = dtdy; mesh->vgeo[k*12+10] = dtdz;
}
/* surfinfo (vmapM, vmapP, Fscale, Bscale, nx, ny, nz, 0) */
sz = mesh->K*p_Nfp*p_Nfaces*7*sizeof(float);
mesh->surfinfo = (float*) malloc(sz);
/* local-local info */
sk = 0;
int skP = -1;
double *nxk = BuildVector(mesh->Nfaces);
double *nyk = BuildVector(mesh->Nfaces);
double *nzk = BuildVector(mesh->Nfaces);
double *sJk = BuildVector(mesh->Nfaces);
double dt = 1e6;
sk = 0;
for(k=0;k<mesh->K;++k){
GeometricFactors3d(mesh, k,
&drdx, &dsdx, &dtdx,
&drdy, &dsdy, &dtdy,
&drdz, &dsdz, &dtdz, &J);
Normals3d(mesh, k, nxk, nyk, nzk, sJk);
for(f=0;f<mesh->Nfaces;++f){
dt = min(dt, J/sJk[f]);
for(m=0;m<p_Nfp;++m){
int id = m + f*p_Nfp + p_Nfp*p_Nfaces*k;
int idM = mesh->vmapM[id];
int idP = mesh->vmapP[id];
int nM = idM%p_Np;
int nP = idP%p_Np;
int kM = (idM-nM)/p_Np;
int kP = (idP-nP)/p_Np;
idM = Nfields*(nM + p_Np*kM);
idP = Nfields*(nP + p_Np*kP);
/* stub resolve some other way */
if(mesh->vmapP[id]<0){
idP = mesh->vmapP[id]; /* -ve numbers */
}
mesh->surfinfo[sk++] = idM;
mesh->surfinfo[sk++] = idP;
mesh->surfinfo[sk++] = sJk[f]/(2.*J);
mesh->surfinfo[sk++] = (idM==idP)?-1.:1.;
mesh->surfinfo[sk++] = nxk[f];
mesh->surfinfo[sk++] = nyk[f];
mesh->surfinfo[sk++] = nzk[f];
}
}
}
}
double InitOCCA3d(Mesh *mesh, int Nfields){
device.setup("mode = OpenCL, platformID = 0, deviceID = 2");
/* Q */
int sz = mesh->K*(BSIZE)*p_Nfields*sizeof(float);
float *f_Q = (float*) calloc(mesh->K*BSIZE*p_Nfields, sizeof(float));
c_Q = device.malloc(sz, f_Q);
c_rhsQ = device.malloc(sz, f_Q);
c_resQ = device.malloc(sz, f_Q);
printf("sz1= %d\n", sz);
sz = mesh->parNtotalout*sizeof(float);
c_tmp = device.malloc(sz+1, f_Q); // should not use f_Q
c_partQ = device.malloc(sz+1, f_Q);
printf("sz2= %d\n", sz);
/* LIFT */
sz = p_Np*(p_Nfp)*p_Nfaces*sizeof(float);
float *f_LIFT = (float*) malloc(sz);
int skL = 0;
for(int m=0;m<p_Nfp;++m){
for(int n=0;n<p_Np;++n){
for(int f=0;f<p_Nfaces;++f){
f_LIFT[skL++] = mesh->LIFT[0][p_Nfp*p_Nfaces*n+(f+p_Nfaces*m)];
}
}
}
c_LIFT = device.malloc(sz, f_LIFT);
/* DrDsDt */
sz = BSIZE*BSIZE*4*sizeof(float);
float* h_DrDsDt = (float*) calloc(BSIZE*BSIZE*4, sizeof(float));
int sk = 0;
/* note transposed arrays to avoid "bank conflicts" */
for(int n=0;n<p_Np;++n){
for(int m=0;m<p_Np;++m){
h_DrDsDt[4*(m+n*BSIZE)+0] = mesh->Dr[0][n+m*p_Np];
h_DrDsDt[4*(m+n*BSIZE)+1] = mesh->Ds[0][n+m*p_Np];
h_DrDsDt[4*(m+n*BSIZE)+2] = mesh->Dt[0][n+m*p_Np];
}
}
c_DrDsDt = device.malloc(sz, h_DrDsDt);
free(h_DrDsDt);
/* vgeo */
double drdx, dsdx, dtdx;
double drdy, dsdy, dtdy;
double drdz, dsdz, dtdz, J;
float *vgeo = (float*) calloc(12*mesh->K, sizeof(float));
for(int k=0;k<mesh->K;++k){
GeometricFactors3d(mesh, k,
&drdx, &dsdx, &dtdx,
&drdy, &dsdy, &dtdy,
&drdz, &dsdz, &dtdz, &J);
vgeo[k*12+0] = drdx; vgeo[k*12+1] = drdy; vgeo[k*12+2] = drdz;
vgeo[k*12+4] = dsdx; vgeo[k*12+5] = dsdy; vgeo[k*12+6] = dsdz;
vgeo[k*12+8] = dtdx; vgeo[k*12+9] = dtdy; vgeo[k*12+10] = dtdz;
}
sz = mesh->K*12*sizeof(float);
c_vgeo = device.malloc(sz, vgeo);
/* surfinfo (vmapM, vmapP, Fscale, Bscale, nx, ny, nz, 0) */
int sz5 = mesh->K*p_Nfp*p_Nfaces*5*sizeof(float);
float* h_surfinfo = (float*) malloc(sz5);
int sz2 = mesh->K*p_Nfp*p_Nfaces*2*sizeof(int);
int* h_mapinfo = (int*) malloc(sz2);
/* local-local info */
sk = 0;
int skP = -1;
double *nxk = BuildVector(mesh->Nfaces);
double *nyk = BuildVector(mesh->Nfaces);
double *nzk = BuildVector(mesh->Nfaces);
double *sJk = BuildVector(mesh->Nfaces);
double dt = 1e6;
for(int k=0;k<mesh->K;++k){
GeometricFactors3d(mesh, k,
&drdx, &dsdx, &dtdx,
&drdy, &dsdy, &dtdy,
&drdz, &dsdz, &dtdz, &J);
Normals3d(mesh, k, nxk, nyk, nzk, sJk);
for(int f=0;f<mesh->Nfaces;++f){
dt = min(dt, J/sJk[f]);
for(int m=0;m<p_Nfp;++m){
int n = m + f*p_Nfp + p_Nfp*p_Nfaces*k;
int idM = mesh->vmapM[n];
int idP = mesh->vmapP[n];
int nM = idM%p_Np;
int nP = idP%p_Np;
int kM = (idM-nM)/p_Np;
int kP = (idP-nP)/p_Np;
idM = nM + Nfields*BSIZE*kM;
idP = nP + Nfields*BSIZE*kP;
/* stub resolve some other way */
if(mesh->vmapP[n]<0){
idP = mesh->vmapP[n]; /* -ve numbers */
}
sk = 2*p_Nfp*p_Nfaces*k+m+f*p_Nfp;
h_mapinfo[sk + 0*p_Nfp*p_Nfaces] = idM;
h_mapinfo[sk + 1*p_Nfp*p_Nfaces] = idP;
sk = 5*p_Nfp*p_Nfaces*k+m+f*p_Nfp;
h_surfinfo[sk + 0*p_Nfp*p_Nfaces] = sJk[f]/(2.*J);
h_surfinfo[sk + 1*p_Nfp*p_Nfaces] = (idM==idP)?-1.:1.;
h_surfinfo[sk + 2*p_Nfp*p_Nfaces] = nxk[f];
h_surfinfo[sk + 3*p_Nfp*p_Nfaces] = nyk[f];
h_surfinfo[sk + 4*p_Nfp*p_Nfaces] = nzk[f];
}
}
}
c_mapinfo = device.malloc(sz2, h_mapinfo);
c_surfinfo = device.malloc(sz5, h_surfinfo);
free(h_mapinfo);
free(h_surfinfo);
printf("mesh->parNtotalout=%d\n", mesh->parNtotalout);
sz = mesh->parNtotalout*sizeof(int);
c_parmapOUT = device.malloc(sz+1, mesh->parmapOUT);
/* now build kernels */
occa::kernelInfo dgInfo;
dgInfo.addDefine("p_Np", p_Np);
dgInfo.addDefine("p_Nfp", p_Nfp);
dgInfo.addDefine("p_Nfaces", p_Nfaces);
dgInfo.addDefine("p_Nfields", p_Nfields);
dgInfo.addDefine("BSIZE", BSIZE);
dgInfo.addDefine("p_max_NfpNfaces_Np", max(p_Nfp*p_Nfaces, p_Np));
volumeKernel = device.buildKernelFromSource("src/MaxwellsVolumeKernel3D.okl",
"MaxwellsVolumeKernel3D",
dgInfo);
surfaceKernel = device.buildKernelFromSource("src/MaxwellsSurfaceKernel3D.okl",
"MaxwellsSurfaceKernel3D",
dgInfo);
rkKernel = device.buildKernelFromSource("src/MaxwellsRKKernel3D.okl",
"MaxwellsRKKernel3D",
dgInfo);
partialGetKernel = device.buildKernelFromSource("src/MaxwellsPartialGetKernel3D.okl",
"MaxwellsPartialGetKernel3D",
dgInfo);
#if 0
diagnose_array<float>("c_DrDsDt", c_DrDsDt, 4*BSIZE*BSIZE);
diagnose_array<float>("c_LIFT", c_LIFT, p_Nfaces*p_Nfp*p_Np);
diagnose_array<float>("c_vgeo", c_vgeo, mesh->K*12);
diagnose_array<float>("c_surfinfo", c_surfinfo, p_Nfaces*p_Nfp*7*mesh->K);
diagnose_array<int> ("c_parmapOUT", c_parmapOUT, mesh->parNtotalout);
#endif
return dt;
}
void BuildMaps2d(Mesh *mesh){
int nprocs = mesh->nprocs;
int procid = mesh->procid;
int K = mesh->K;
int Nfaces = mesh->Nfaces;
mesh->vmapM = BuildIntVector(p_Nfp*p_Nfaces*K);
mesh->vmapP = BuildIntVector(p_Nfp*p_Nfaces*K);
int m;
int k1,f1,p1,n1,id1, k2,f2,p2,n2,id2;
double x1, y1, x2, y2, d12;
double *nxk = BuildVector(Nfaces);
double *nyk = BuildVector(Nfaces);
double *sJk = BuildVector(Nfaces);
/* first build local */
for(k1=0;k1<K;++k1){
/* get some information about the face geometries */
Normals2d(mesh, k1, nxk, nyk, sJk);
for(f1=0;f1<Nfaces;++f1){
/* volume -> face nodes */
for(n1=0;n1<p_Nfp;++n1){
id1 = n1+f1*p_Nfp+k1*p_Nfp*p_Nfaces;
mesh->vmapM[id1] = mesh->Fmask[f1][n1] + k1*p_Np;
}
/* find neighbor */
k2 = mesh->EToE[k1][f1];
f2 = mesh->EToF[k1][f1];
p2 = mesh->EToP[k1][f1];
if(k1==k2 || procid!=p2 ){
for(n1=0;n1<p_Nfp;++n1){
id1 = n1+f1*p_Nfp+k1*p_Nfp*p_Nfaces;
mesh->vmapP[id1] = k1*p_Np + mesh->Fmask[f1][n1];
}
}else{
/* treat as boundary for the moment */
for(n1=0;n1<p_Nfp;++n1){
id1 = n1+f1*p_Nfp+k1*p_Nfp*p_Nfaces;
x1 = mesh->x[k1][mesh->Fmask[f1][n1]];
y1 = mesh->y[k1][mesh->Fmask[f1][n1]];
for(n2=0;n2<p_Nfp;++n2){
id2 = n2+f2*p_Nfp+k2*p_Nfp*p_Nfaces;
x2 = mesh->x[k2][mesh->Fmask[f2][n2]];
y2 = mesh->y[k2][mesh->Fmask[f2][n2]];
/* find normalized distance between these nodes */
/* [ use sJk as a measure of edge length (ignore factor of 2) ] */
d12 = ((x1-x2)*(x1-x2) + (y1-y2)*(y1-y2))/(sJk[f1]*sJk[f1]);
if(d12<NODETOL){
mesh->vmapP[id1] = k2*p_Np + mesh->Fmask[f2][n2];
}
}
}
}
}
}
#if 0
int n;
for(n=0;n<p_Nfp*p_Nfaces*mesh->K;++n){
x1 = mesh->x[0][mesh->vmapM[n]];
y1 = mesh->y[0][mesh->vmapM[n]];
x2 = mesh->x[0][mesh->vmapP[n]];
y2 = mesh->y[0][mesh->vmapP[n]];
d12 = ((x1-x2)*(x1-x2) + (y1-y2)*(y1-y2));
printf("n:%d %d -> %d d=%lg\n", n, mesh->vmapM[n], mesh->vmapP[n], d12);
}
#endif
/* now build parallel maps */
double **xsend = (double**) calloc(nprocs, sizeof(double*));
double **ysend = (double**) calloc(nprocs, sizeof(double*));
double **xrecv = (double**) calloc(nprocs, sizeof(double*));
double **yrecv = (double**) calloc(nprocs, sizeof(double*));
int **Esend = (int**) calloc(nprocs, sizeof(int*));
int **Fsend = (int**) calloc(nprocs, sizeof(int*));
int **Erecv = (int**) calloc(nprocs, sizeof(int*));
int **Frecv = (int**) calloc(nprocs, sizeof(int*));
for(p2=0;p2<nprocs;++p2){
if(mesh->Npar[p2]){
xsend[p2] = BuildVector(mesh->Npar[p2]*p_Nfp);
ysend[p2] = BuildVector(mesh->Npar[p2]*p_Nfp);
Esend[p2] = BuildIntVector(mesh->Npar[p2]*p_Nfp);
Fsend[p2] = BuildIntVector(mesh->Npar[p2]*p_Nfp);
xrecv[p2] = BuildVector(mesh->Npar[p2]*p_Nfp);
yrecv[p2] = BuildVector(mesh->Npar[p2]*p_Nfp);
Erecv[p2] = BuildIntVector(mesh->Npar[p2]*p_Nfp);
Frecv[p2] = BuildIntVector(mesh->Npar[p2]*p_Nfp);
}
}
int *skP = BuildIntVector(nprocs);
/* send coordinates in local order */
int cnt = 0;
for(k1=0;k1<K;++k1){
for(f1=0;f1<p_Nfaces;++f1){
p2 = mesh->EToP[k1][f1];
if(p2!=procid){
for(n1=0;n1<p_Nfp;++n1){
xsend[p2][skP[p2]] = mesh->x[k1][mesh->Fmask[f1][n1]];
ysend[p2][skP[p2]] = mesh->y[k1][mesh->Fmask[f1][n1]];
Esend[p2][skP[p2]] = mesh->EToE[k1][f1];
Fsend[p2][skP[p2]] = mesh->EToF[k1][f1];
++(skP[p2]);
}
}
}
}
MPI_Request *xsendrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *ysendrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *xrecvrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *yrecvrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *Esendrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *Fsendrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *Erecvrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Request *Frecvrequests = (MPI_Request*) calloc(nprocs, sizeof(MPI_Request));
MPI_Status *status = (MPI_Status*) calloc(nprocs, sizeof(MPI_Status));
cnt = 0;
for(p2=0;p2<nprocs;++p2){
if(p2!=procid && mesh->Npar[p2]!=0){
int Nout = mesh->Npar[p2]*p_Nfp;
MPI_Isend(xsend[p2], Nout, MPI_DOUBLE, p2, 666+p2, MPI_COMM_WORLD, xsendrequests+cnt);
MPI_Isend(ysend[p2], Nout, MPI_DOUBLE, p2, 1666+p2, MPI_COMM_WORLD, ysendrequests+cnt);
MPI_Isend(Esend[p2], Nout, MPI_INT, p2, 2666+p2, MPI_COMM_WORLD, Esendrequests+cnt);
MPI_Isend(Fsend[p2], Nout, MPI_INT, p2, 3666+p2, MPI_COMM_WORLD, Fsendrequests+cnt);
MPI_Irecv(xrecv[p2], Nout, MPI_DOUBLE, p2, 666+procid, MPI_COMM_WORLD, xrecvrequests+cnt);
MPI_Irecv(yrecv[p2], Nout, MPI_DOUBLE, p2, 1666+procid, MPI_COMM_WORLD, yrecvrequests+cnt);
MPI_Irecv(Erecv[p2], Nout, MPI_INT, p2, 2666+procid, MPI_COMM_WORLD, Erecvrequests+cnt);
MPI_Irecv(Frecv[p2], Nout, MPI_INT, p2, 3666+procid, MPI_COMM_WORLD, Frecvrequests+cnt);
++cnt;
}
}
MPI_Waitall(cnt, xsendrequests, status);
MPI_Waitall(cnt, ysendrequests, status);
MPI_Waitall(cnt, Esendrequests, status);
MPI_Waitall(cnt, Fsendrequests, status);
MPI_Waitall(cnt, xrecvrequests, status);
MPI_Waitall(cnt, yrecvrequests, status);
MPI_Waitall(cnt, Erecvrequests, status);
MPI_Waitall(cnt, Frecvrequests, status);
/* add up the total number of outgoing/ingoing nodes */
mesh->parNtotalout = 0;
for(p2=0;p2<nprocs;++p2)
mesh->parNtotalout += skP[p2]*p_Nfields;
mesh->parmapOUT = BuildIntVector(mesh->parNtotalout);
/* now match up local nodes with the requested (recv'ed nodes) */
int idout = -1;
int sk = 0;
for(p2=0;p2<nprocs;++p2){
/* for each received face */
for(m=0;m<skP[p2];++m){
k1 = Erecv[p2][m];
f1 = Frecv[p2][m];
x2 = xrecv[p2][m];
y2 = yrecv[p2][m];
Normals2d(mesh, k1, nxk, nyk, sJk);
for(n1=0;n1<p_Nfp;++n1){
x1 = mesh->x[k1][mesh->Fmask[f1][n1]];
y1 = mesh->y[k1][mesh->Fmask[f1][n1]];
d12 = ((x1-x2)*(x1-x2) + (y1-y2)*(y1-y2))/(sJk[f1]*sJk[f1]);
if(d12<NODETOL){
int fld;
for(fld=0;fld<p_Nfields;++fld){
#ifdef CUDA
mesh->parmapOUT[sk++] = k1*BSIZE*p_Nfields+mesh->Fmask[f1][n1] + BSIZE*fld;
#else
mesh->parmapOUT[sk++] = p_Nfields*(k1*p_Np+mesh->Fmask[f1][n1]) + fld;
#endif
}
}
}
}
}
/* create incoming node map */
int parcnt = -1;
for(p2=0;p2<nprocs;++p2){
for(k1=0;k1<K;++k1){
for(f1=0;f1<p_Nfaces;++f1){
if(mesh->EToP[k1][f1]==p2 && p2!=procid){
for(n1=0;n1<p_Nfp;++n1){
id1 = n1+f1*p_Nfp+k1*p_Nfp*p_Nfaces;
mesh->vmapP[id1] = parcnt;
--parcnt;
}
}
}
}
}
/* buffers for communication */
mesh->f_outQ = (float*) calloc(mesh->parNtotalout+1, sizeof(float));
mesh->f_inQ = (float*) calloc(mesh->parNtotalout+1, sizeof(float));
}
Mesh *ReadMesh3d(char *filename){
int n;
Mesh *mesh = (Mesh*) calloc(1, sizeof(Mesh));
char buf[BUFSIZ];
FILE *fp = fopen(filename, "r");
/* assume modified Gambit neutral format */
for(n=0;n<6;++n)
fgets(buf, BUFSIZ, fp);
fgets(buf, BUFSIZ, fp);
sscanf(buf, "%d %d \n", &(mesh->Nv), &(mesh->K));
mesh->Nverts = 4; /* assume tets */
mesh->Nedges = 6; /* assume tets */
mesh->Nfaces = 4; /* assume tets */
fgets(buf, BUFSIZ, fp);
fgets(buf, BUFSIZ, fp);
/* read vertex coordinates */
double *VX = BuildVector(mesh->Nv);
double *VY = BuildVector(mesh->Nv);
double *VZ = BuildVector(mesh->Nv);
for(n=0;n<mesh->Nv;++n){
fgets(buf, BUFSIZ, fp);
sscanf(buf, "%*d %lf %lf %lf", VX+n, VY+n, VZ+n);
}
/* decide on parition */
int procid, nprocs;
MPI_Comm_rank(MPI_COMM_WORLD, &procid);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
mesh->procid = procid;
mesh->nprocs = nprocs;
/* assume this proc owns a block of elements */
int Klocal, Kstart;
int *Kprocs = (int*) calloc(nprocs, sizeof(int));
int p;
int **newEToV, *newKprocs;
double **newVX, **newVY;
Klocal = (int) ( (double)(mesh->K)/(double)nprocs );
for(p=0;p<nprocs-1;++p){
Kprocs[p] = Klocal;
}
Kprocs[p] = Klocal + mesh->K - nprocs*Klocal;
Kstart= 0;
for(p=0;p<procid;++p)
Kstart += Kprocs[p];
Klocal = Kprocs[procid];
/* read element to vertex connectivity */
fgets(buf, BUFSIZ, fp);
fgets(buf, BUFSIZ, fp);
mesh->EToV = BuildIntMatrix(Klocal, mesh->Nverts);
mesh->GX = BuildMatrix(Klocal, mesh->Nverts);
mesh->GY = BuildMatrix(Klocal, mesh->Nverts);
mesh->GZ = BuildMatrix(Klocal, mesh->Nverts);
int sk = 0, v;
for(n=0;n<mesh->K;++n){
fgets(buf, BUFSIZ, fp);
if(n>=Kstart && n<Kstart+Klocal){
sscanf(buf, "%*d %*d %*d %d %d %d %d",
mesh->EToV[sk]+0, mesh->EToV[sk]+1,
mesh->EToV[sk]+2, mesh->EToV[sk]+3);
/* correct to 0-index */
--(mesh->EToV[sk][0]);
--(mesh->EToV[sk][1]);
--(mesh->EToV[sk][2]);
--(mesh->EToV[sk][3]);
for(v=0;v<mesh->Nverts;++v){
mesh->GX[sk][v] = VX[mesh->EToV[sk][v]];
mesh->GY[sk][v] = VY[mesh->EToV[sk][v]];
mesh->GZ[sk][v] = VZ[mesh->EToV[sk][v]];
}
++sk;
}
}
fgets(buf, BUFSIZ, fp);
fgets(buf, BUFSIZ, fp);
mesh->K = Klocal;
fclose(fp);
return mesh;
}
/* TargetIntrinsicLower - To handle builtins, we want to expand the
* invocation into normal LLVM code. If the target can handle the builtin, this
* function should emit the expanded code and return true.
*/
bool TreeToLLVM::TargetIntrinsicLower(tree exp,
unsigned FnCode,
Value *DestLoc,
Value *&Result,
const Type *ResultType,
std::vector<Value*> &Ops,
SmallVector<tree, 8> &Args,
BasicBlock *CurBB,
bool ResIsSigned,
bool ExpIsSigned) {
switch (FnCode) {
default: break;
case ALTIVEC_BUILTIN_VADDFP:
case ALTIVEC_BUILTIN_VADDUBM:
case ALTIVEC_BUILTIN_VADDUHM:
case ALTIVEC_BUILTIN_VADDUWM:
Result = BinaryOperator::createAdd(Ops[0], Ops[1], "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_VSUBFP:
case ALTIVEC_BUILTIN_VSUBUBM:
case ALTIVEC_BUILTIN_VSUBUHM:
case ALTIVEC_BUILTIN_VSUBUWM:
Result = BinaryOperator::createSub(Ops[0], Ops[1], "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_VAND:
Result = BinaryOperator::createAnd(Ops[0], Ops[1], "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_VANDC:
Ops[1] = BinaryOperator::createNot(Ops[1], "tmp", CurBB);
Result = BinaryOperator::createAnd(Ops[0], Ops[1], "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_VOR:
Result = BinaryOperator::createOr(Ops[0], Ops[1], "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_VNOR:
Result = BinaryOperator::createOr(Ops[0], Ops[1], "tmp", CurBB);
Result = BinaryOperator::createNot(Result, "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_VXOR:
Result = BinaryOperator::createXor(Ops[0], Ops[1], "tmp", CurBB);
return true;
case ALTIVEC_BUILTIN_LVSL: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvsl",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_LVSR: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvsr",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_LVX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_LVXL: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvxl",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_LVEBX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvebx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_LVEHX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvehx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_LVEWX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 0, "llvm.ppc.altivec.lvewx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_STVX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 1, "llvm.ppc.altivec.stvx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_STVEBX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 1, "llvm.ppc.altivec.stvebx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_STVEHX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 1, "llvm.ppc.altivec.stvehx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_STVEWX: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 1, "llvm.ppc.altivec.stvewx",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_STVXL: {
static Constant *Cache = NULL;
MergeIntPtrOperand(this, Cache, 1, "llvm.ppc.altivec.stvxl",
ResultType, Ops, CurBB, Result);
}
return true;
case ALTIVEC_BUILTIN_VSPLTISB:
if (Constant *Elt = dyn_cast<ConstantInt>(Ops[0])) {
Elt = ConstantExpr::getIntegerCast(Elt, Type::Int8Ty, true);
Result = BuildVector(Elt, Elt, Elt, Elt, Elt, Elt, Elt, Elt,
Elt, Elt, Elt, Elt, Elt, Elt, Elt, Elt, NULL);
} else {
error("%Helement must be an immediate", &EXPR_LOCATION(exp));
Result = UndefValue::get(VectorType::get(Type::Int8Ty, 16));
}
return true;
case ALTIVEC_BUILTIN_VSPLTISH:
if (Constant *Elt = dyn_cast<ConstantInt>(Ops[0])) {
Elt = ConstantExpr::getIntegerCast(Elt, Type::Int16Ty, true);
Result = BuildVector(Elt, Elt, Elt, Elt, Elt, Elt, Elt, Elt, NULL);
} else {
error("%Helement must be an immediate", &EXPR_LOCATION(exp));
Result = UndefValue::get(VectorType::get(Type::Int16Ty, 8));
}
return true;
case ALTIVEC_BUILTIN_VSPLTISW:
if (Constant *Elt = dyn_cast<ConstantInt>(Ops[0])) {
Elt = ConstantExpr::getIntegerCast(Elt, Type::Int32Ty, true);
Result = BuildVector(Elt, Elt, Elt, Elt, NULL);
} else {
error("%Hmask must be an immediate", &EXPR_LOCATION(exp));
Result = UndefValue::get(VectorType::get(Type::Int32Ty, 4));
}
return true;
case ALTIVEC_BUILTIN_VSPLTB:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[1])) {
int EV = Elt->getZExtValue();
Result = BuildVectorShuffle(Ops[0], Ops[0],
EV, EV, EV, EV, EV, EV, EV, EV,
EV, EV, EV, EV, EV, EV, EV, EV);
} else {
error("%Helement number must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VSPLTH:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[1])) {
int EV = Elt->getZExtValue();
Result = BuildVectorShuffle(Ops[0], Ops[0],
EV, EV, EV, EV, EV, EV, EV, EV);
} else {
error("%Helement number must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VSPLTW:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[1])) {
int EV = Elt->getZExtValue();
Result = BuildVectorShuffle(Ops[0], Ops[0], EV, EV, EV, EV);
} else {
error("%Helement number must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VSLDOI_16QI:
case ALTIVEC_BUILTIN_VSLDOI_8HI:
case ALTIVEC_BUILTIN_VSLDOI_4SI:
case ALTIVEC_BUILTIN_VSLDOI_4SF:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[2])) {
/* Map all of these to a shuffle. */
unsigned Amt = Elt->getZExtValue() & 15;
VectorType *v16i8 = VectorType::get(Type::Int8Ty, 16);
Value *Op0 = Ops[0];
Instruction::CastOps opc = CastInst::getCastOpcode(Op0,
IntrinsicOpIsSigned(Args,0), ResultType, false);
Ops[0] = CastToType(opc, Op0, v16i8);
Value *Op1 = Ops[1];
opc = CastInst::getCastOpcode(Op1,
IntrinsicOpIsSigned(Args,1), ResultType, false);
Ops[1] = CastToType(opc, Op1, v16i8);
Result = BuildVectorShuffle(Ops[0], Ops[1],
Amt, Amt+1, Amt+2, Amt+3,
Amt+4, Amt+5, Amt+6, Amt+7,
Amt+8, Amt+9, Amt+10, Amt+11,
Amt+12, Amt+13, Amt+14, Amt+15);
} else {
error("%Hshift amount must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VPKUHUM: {
Value *Op0 = Ops[0];
Instruction::CastOps opc = CastInst::getCastOpcode(Op0,
IntrinsicOpIsSigned(Args,0), ResultType, ExpIsSigned);
Ops[0] = CastInst::create(opc, Op0, ResultType, Op0->getName(), CurBB);
Value *Op1 = Ops[1];
opc = CastInst::getCastOpcode(Op1,
IntrinsicOpIsSigned(Args,1), ResultType, ExpIsSigned);
Ops[1] = CastInst::create(opc, Op1, ResultType, Op1->getName(), CurBB);
Result = BuildVectorShuffle(Ops[0], Ops[1], 1, 3, 5, 7, 9, 11, 13, 15,
17, 19, 21, 23, 25, 27, 29, 31);
return true;
}
case ALTIVEC_BUILTIN_VPKUWUM: {
Value *Op0 = Ops[0];
Instruction::CastOps opc = CastInst::getCastOpcode(Op0,
IntrinsicOpIsSigned(Args,0), ResultType, ExpIsSigned);
Ops[0] = CastInst::create(opc, Op0, ResultType, Op0->getName(), CurBB);
Value *Op1 = Ops[1];
opc = CastInst::getCastOpcode(Op1,
IntrinsicOpIsSigned(Args,1), ResultType, ExpIsSigned);
Ops[1] = CastInst::create(opc, Op1, ResultType, Op1->getName(), CurBB);
Result = BuildVectorShuffle(Ops[0], Ops[1], 1, 3, 5, 7, 9, 11, 13, 15);
return true;
}
case ALTIVEC_BUILTIN_VMRGHB:
Result = BuildVectorShuffle(Ops[0], Ops[1],
0, 16, 1, 17, 2, 18, 3, 19,
4, 20, 5, 21, 6, 22, 7, 23);
return true;
case ALTIVEC_BUILTIN_VMRGHH:
Result = BuildVectorShuffle(Ops[0], Ops[1], 0, 8, 1, 9, 2, 10, 3, 11);
return true;
case ALTIVEC_BUILTIN_VMRGHW:
Result = BuildVectorShuffle(Ops[0], Ops[1], 0, 4, 1, 5);
return true;
case ALTIVEC_BUILTIN_VMRGLB:
Result = BuildVectorShuffle(Ops[0], Ops[1],
8, 24, 9, 25, 10, 26, 11, 27,
12, 28, 13, 29, 14, 30, 15, 31);
return true;
case ALTIVEC_BUILTIN_VMRGLH:
Result = BuildVectorShuffle(Ops[0], Ops[1], 4, 12, 5, 13, 6, 14, 7, 15);
return true;
case ALTIVEC_BUILTIN_VMRGLW:
Result = BuildVectorShuffle(Ops[0], Ops[1], 2, 6, 3, 7);
return true;
case ALTIVEC_BUILTIN_ABS_V4SF: {
/* and out sign bits */
VectorType *v4i32 = VectorType::get(Type::Int32Ty, 4);
Ops[0] = new BitCastInst(Ops[0], v4i32, Ops[0]->getName(),CurBB);
Constant *C = ConstantInt::get(Type::Int32Ty, 0x7FFFFFFF);
C = ConstantVector::get(std::vector<Constant*>(4, C));
Result = BinaryOperator::createAnd(Ops[0], C, "tmp", CurBB);
TargetIntrinsicCastResult(Result, ResultType,
ResIsSigned, ExpIsSigned, CurBB);
return true;
}
case ALTIVEC_BUILTIN_ABS_V4SI:
case ALTIVEC_BUILTIN_ABS_V8HI:
case ALTIVEC_BUILTIN_ABS_V16QI: { /* iabs(x) -> smax(x, 0-x) */
Result = BinaryOperator::createNeg(Ops[0], "tmp", CurBB);
/* get the right smax intrinsic. */
static Constant *smax[3];
const VectorType *PTy = cast<VectorType>(ResultType);
unsigned N = GetAltivecTypeNumFromType(PTy->getElementType());
if (smax[N] == 0) {
Module *M = CurBB->getParent()->getParent();
smax[N] = M->getOrInsertFunction(std::string("llvm.ppc.altivec.vmaxs")+
GetAltivecLetterFromType(PTy->getElementType()),
ResultType, ResultType, ResultType, NULL);
}
Result = new CallInst(smax[N], Ops[0], Result, "tmp", CurBB);
return true;
}
case ALTIVEC_BUILTIN_ABSS_V4SI:
case ALTIVEC_BUILTIN_ABSS_V8HI:
case ALTIVEC_BUILTIN_ABSS_V16QI: { /* iabss(x) -> smax(x, satsub(0,x)) */
static Constant *sxs[3], *smax[3];
/* get the right satsub intrinsic. */
const VectorType *PTy = cast<VectorType>(ResultType);
unsigned N = GetAltivecTypeNumFromType(PTy->getElementType());
if (sxs[N] == 0) {
Module *M = CurBB->getParent()->getParent();
sxs[N] = M->getOrInsertFunction(std::string("llvm.ppc.altivec.vsubs")+
GetAltivecLetterFromType(PTy->getElementType())+"s",
ResultType, ResultType, ResultType, NULL);
}
Result = Constant::getNullValue(ResultType);
Result = new CallInst(sxs[N], Result, Ops[0], "tmp", CurBB);
/* get the right smax intrinsic. */
if (smax[N] == 0) {
Module *M = CurBB->getParent()->getParent();
smax[N] = M->getOrInsertFunction(std::string("llvm.ppc.altivec.vmaxs")+
GetAltivecLetterFromType(PTy->getElementType()),
ResultType, ResultType, ResultType, NULL);
}
Result = new CallInst(smax[N], Ops[0], Result, "tmp", CurBB);
return true;
}
}
return false;
}
// TargetIntrinsicLower - To handle builtins, we want to expand the
//invocation into normal LLVM code. If the target can handle the builtin, this
//function should emit the expanded code and return true.
//
bool TreeToLLVM::TargetIntrinsicLower(tree exp,
unsigned FnCode,
const MemRef *DestLoc,
Value *&Result,
const Type *ResultType,
std::vector<Value*> &Ops) {
switch (FnCode) {
default: break;
case ALTIVEC_BUILTIN_VADDFP:
case ALTIVEC_BUILTIN_VADDUBM:
case ALTIVEC_BUILTIN_VADDUHM:
case ALTIVEC_BUILTIN_VADDUWM:
Result = Builder.CreateAdd(Ops[0], Ops[1], "tmp");
return true;
case ALTIVEC_BUILTIN_VSUBFP:
case ALTIVEC_BUILTIN_VSUBUBM:
case ALTIVEC_BUILTIN_VSUBUHM:
case ALTIVEC_BUILTIN_VSUBUWM:
Result = Builder.CreateSub(Ops[0], Ops[1], "tmp");
return true;
case ALTIVEC_BUILTIN_VAND:
Result = Builder.CreateAnd(Ops[0], Ops[1], "tmp");
return true;
case ALTIVEC_BUILTIN_VANDC:
Ops[1] = Builder.CreateNot(Ops[1], "tmp");
Result = Builder.CreateAnd(Ops[0], Ops[1], "tmp");
return true;
case ALTIVEC_BUILTIN_VOR:
Result = Builder.CreateOr(Ops[0], Ops[1], "tmp");
return true;
case ALTIVEC_BUILTIN_VNOR:
Result = Builder.CreateOr(Ops[0], Ops[1], "tmp");
Result = Builder.CreateNot(Result, "tmp");
return true;
case ALTIVEC_BUILTIN_VXOR:
Result = Builder.CreateXor(Ops[0], Ops[1], "tmp");
return true;
case ALTIVEC_BUILTIN_LVSL:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvsl,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_LVSR:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvsr,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_LVX:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_LVXL:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvxl,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_LVEBX:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvebx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_LVEHX:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvehx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_LVEWX:
MergeIntPtrOperand(this, 0, Intrinsic::ppc_altivec_lvewx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_STVX:
MergeIntPtrOperand(this, 1, Intrinsic::ppc_altivec_stvx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_STVEBX:
MergeIntPtrOperand(this, 1, Intrinsic::ppc_altivec_stvebx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_STVEHX:
MergeIntPtrOperand(this, 1, Intrinsic::ppc_altivec_stvehx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_STVEWX:
MergeIntPtrOperand(this, 1, Intrinsic::ppc_altivec_stvewx,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_STVXL:
MergeIntPtrOperand(this, 1, Intrinsic::ppc_altivec_stvxl,
ResultType, Ops, Builder, Result);
return true;
case ALTIVEC_BUILTIN_VSPLTISB:
if (Constant *Elt = dyn_cast<ConstantInt>(Ops[0])) {
Elt = ConstantExpr::getIntegerCast(Elt, Type::Int8Ty, true);
Result = BuildVector(Elt, Elt, Elt, Elt, Elt, Elt, Elt, Elt,
Elt, Elt, Elt, Elt, Elt, Elt, Elt, Elt, NULL);
} else {
error("%Helement must be an immediate", &EXPR_LOCATION(exp));
Result = UndefValue::get(VectorType::get(Type::Int8Ty, 16));
}
return true;
case ALTIVEC_BUILTIN_VSPLTISH:
if (Constant *Elt = dyn_cast<ConstantInt>(Ops[0])) {
Elt = ConstantExpr::getIntegerCast(Elt, Type::Int16Ty, true);
Result = BuildVector(Elt, Elt, Elt, Elt, Elt, Elt, Elt, Elt, NULL);
} else {
error("%Helement must be an immediate", &EXPR_LOCATION(exp));
Result = UndefValue::get(VectorType::get(Type::Int16Ty, 8));
}
return true;
case ALTIVEC_BUILTIN_VSPLTISW:
if (Constant *Elt = dyn_cast<ConstantInt>(Ops[0])) {
Elt = ConstantExpr::getIntegerCast(Elt, Type::Int32Ty, true);
Result = BuildVector(Elt, Elt, Elt, Elt, NULL);
} else {
error("%Hmask must be an immediate", &EXPR_LOCATION(exp));
Result = UndefValue::get(VectorType::get(Type::Int32Ty, 4));
}
return true;
case ALTIVEC_BUILTIN_VSPLTB:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[1])) {
int EV = Elt->getZExtValue();
Result = BuildVectorShuffle(Ops[0], Ops[0],
EV, EV, EV, EV, EV, EV, EV, EV,
EV, EV, EV, EV, EV, EV, EV, EV);
} else {
error("%Helement number must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VSPLTH:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[1])) {
int EV = Elt->getZExtValue();
// gcc accepts anything up to 31, and there is code that tests for it,
// although it doesn't seem to make sense. Hardware behaves as if mod 8.
if (EV>7 && EV<=31)
EV = EV%8;
Result = BuildVectorShuffle(Ops[0], Ops[0],
EV, EV, EV, EV, EV, EV, EV, EV);
} else {
error("%Helement number must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VSPLTW:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[1])) {
int EV = Elt->getZExtValue();
// gcc accepts anything up to 31, and there is code that tests for it,
// although it doesn't seem to make sense. Hardware behaves as if mod 4.
if (EV>3 && EV<=31)
EV = EV%4;
Result = BuildVectorShuffle(Ops[0], Ops[0], EV, EV, EV, EV);
} else {
error("%Helement number must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VSLDOI_16QI:
case ALTIVEC_BUILTIN_VSLDOI_8HI:
case ALTIVEC_BUILTIN_VSLDOI_4SI:
case ALTIVEC_BUILTIN_VSLDOI_4SF:
if (ConstantInt *Elt = dyn_cast<ConstantInt>(Ops[2])) {
/* Map all of these to a shuffle. */
unsigned Amt = Elt->getZExtValue() & 15;
VectorType *v16i8 = VectorType::get(Type::Int8Ty, 16);
Ops[0] = Builder.CreateBitCast(Ops[0], v16i8, "tmp");
Ops[1] = Builder.CreateBitCast(Ops[1], v16i8, "tmp");
Result = BuildVectorShuffle(Ops[0], Ops[1],
Amt, Amt+1, Amt+2, Amt+3,
Amt+4, Amt+5, Amt+6, Amt+7,
Amt+8, Amt+9, Amt+10, Amt+11,
Amt+12, Amt+13, Amt+14, Amt+15);
} else {
error("%Hshift amount must be an immediate", &EXPR_LOCATION(exp));
Result = Ops[0];
}
return true;
case ALTIVEC_BUILTIN_VPKUHUM:
Ops[0] = Builder.CreateBitCast(Ops[0], ResultType, "tmp");
Ops[1] = Builder.CreateBitCast(Ops[1], ResultType, "tmp");
Result = BuildVectorShuffle(Ops[0], Ops[1], 1, 3, 5, 7, 9, 11, 13, 15,
17, 19, 21, 23, 25, 27, 29, 31);
return true;
case ALTIVEC_BUILTIN_VPKUWUM:
Ops[0] = Builder.CreateBitCast(Ops[0], ResultType, "tmp");
Ops[1] = Builder.CreateBitCast(Ops[1], ResultType, "tmp");
Result = BuildVectorShuffle(Ops[0], Ops[1], 1, 3, 5, 7, 9, 11, 13, 15);
return true;
case ALTIVEC_BUILTIN_VMRGHB:
Result = BuildVectorShuffle(Ops[0], Ops[1],
0, 16, 1, 17, 2, 18, 3, 19,
4, 20, 5, 21, 6, 22, 7, 23);
return true;
case ALTIVEC_BUILTIN_VMRGHH:
Result = BuildVectorShuffle(Ops[0], Ops[1], 0, 8, 1, 9, 2, 10, 3, 11);
return true;
case ALTIVEC_BUILTIN_VMRGHW:
Result = BuildVectorShuffle(Ops[0], Ops[1], 0, 4, 1, 5);
return true;
case ALTIVEC_BUILTIN_VMRGLB:
Result = BuildVectorShuffle(Ops[0], Ops[1],
8, 24, 9, 25, 10, 26, 11, 27,
12, 28, 13, 29, 14, 30, 15, 31);
return true;
case ALTIVEC_BUILTIN_VMRGLH:
Result = BuildVectorShuffle(Ops[0], Ops[1], 4, 12, 5, 13, 6, 14, 7, 15);
return true;
case ALTIVEC_BUILTIN_VMRGLW:
Result = BuildVectorShuffle(Ops[0], Ops[1], 2, 6, 3, 7);
return true;
case ALTIVEC_BUILTIN_ABS_V4SF: {
// and out sign bits
VectorType *v4i32 = VectorType::get(Type::Int32Ty, 4);
Ops[0] = Builder.CreateBitCast(Ops[0], v4i32, "tmp");
Constant *C = ConstantInt::get(Type::Int32Ty, 0x7FFFFFFF);
C = ConstantVector::get(std::vector<Constant*>(4, C));
Result = Builder.CreateAnd(Ops[0], C, "tmp");
Result = Builder.CreateBitCast(Result, ResultType, "tmp");
return true;
}
case ALTIVEC_BUILTIN_ABS_V4SI:
case ALTIVEC_BUILTIN_ABS_V8HI:
case ALTIVEC_BUILTIN_ABS_V16QI: { // iabs(x) -> smax(x, 0-x)
Result = Builder.CreateNeg(Ops[0], "tmp");
// get the right smax intrinsic.
static const Intrinsic::ID smax_iid[3] = {
Intrinsic::ppc_altivec_vmaxsw,
Intrinsic::ppc_altivec_vmaxsh,
Intrinsic::ppc_altivec_vmaxsb
};
const VectorType *PTy = cast<VectorType>(ResultType);
unsigned N = GetAltivecTypeNumFromType(PTy->getElementType());
Function *smax = Intrinsic::getDeclaration(TheModule, smax_iid[N]);
Value *ActualOps[] = { Ops[0], Result };
Result = Builder.CreateCall(smax, ActualOps, ActualOps+2, "tmp");
return true;
}
case ALTIVEC_BUILTIN_ABSS_V4SI:
case ALTIVEC_BUILTIN_ABSS_V8HI:
case ALTIVEC_BUILTIN_ABSS_V16QI: { // iabss(x) -> smax(x, satsub(0,x))
// get the right smax/subs intrinsics.
static const Intrinsic::ID smax_iid[3] = {
Intrinsic::ppc_altivec_vmaxsw,
Intrinsic::ppc_altivec_vmaxsh,
Intrinsic::ppc_altivec_vmaxsb
};
static const Intrinsic::ID subss_iid[3] = {
Intrinsic::ppc_altivec_vsubsws,
Intrinsic::ppc_altivec_vsubshs,
Intrinsic::ppc_altivec_vsubsbs
};
// get the right satsub intrinsic.
const VectorType *PTy = cast<VectorType>(ResultType);
unsigned N = GetAltivecTypeNumFromType(PTy->getElementType());
Function *smax = Intrinsic::getDeclaration(TheModule, smax_iid[N]);
Function *subss = Intrinsic::getDeclaration(TheModule, subss_iid[N]);
Value *ActualOps[] = { Constant::getNullValue(ResultType), Ops[0] };
Result = Builder.CreateCall(subss, ActualOps, ActualOps+2, "tmp");
ActualOps[0] = Ops[0];
ActualOps[1] = Result;
Result = Builder.CreateCall(smax, ActualOps, ActualOps+2, "tmp");
return true;
}
case ALTIVEC_BUILTIN_VPERM_4SI:
case ALTIVEC_BUILTIN_VPERM_4SF:
case ALTIVEC_BUILTIN_VPERM_8HI:
case ALTIVEC_BUILTIN_VPERM_16QI: {
// Operation is identical on all types; we have a single intrinsic.
const Type *VecTy = VectorType::get(Type::Int32Ty, 4);
Value *Op0 = CastToType(Instruction::BitCast, Ops[0], VecTy);
Value *Op1 = CastToType(Instruction::BitCast, Ops[1], VecTy);
Value *ActualOps[] = { Op0, Op1, Ops[2]};
Result = Builder.CreateCall(Intrinsic::getDeclaration(TheModule,
Intrinsic::ppc_altivec_vperm),
ActualOps, ActualOps+3, "tmp");
Result = CastToType(Instruction::BitCast, Result, Ops[0]->getType());
return true;
}
case ALTIVEC_BUILTIN_VSEL_4SI:
case ALTIVEC_BUILTIN_VSEL_4SF:
case ALTIVEC_BUILTIN_VSEL_8HI:
case ALTIVEC_BUILTIN_VSEL_16QI: {
// Operation is identical on all types; we have a single intrinsic.
const Type *VecTy = VectorType::get(Type::Int32Ty, 4);
Value *Op0 = CastToType(Instruction::BitCast, Ops[0], VecTy);
Value *Op1 = CastToType(Instruction::BitCast, Ops[1], VecTy);
Value *Op2 = CastToType(Instruction::BitCast, Ops[2], VecTy);
Value *ActualOps[] = { Op0, Op1, Op2 };
Result = Builder.CreateCall(Intrinsic::getDeclaration(TheModule,
Intrinsic::ppc_altivec_vsel),
ActualOps, ActualOps+3, "tmp");
Result = CastToType(Instruction::BitCast, Result, Ops[0]->getType());
return true;
}
}
return false;
}
