void gradient_standard(float *x, float *fcost, float *grad)
/*< standard velocity gradient >*/
{
int ix, iz, is, ir, it, wit, iturn;
int sx, rx;
float temp, dmax;
float **p0, **p1, **p2, **term, **tmparray, *rr, ***wave, **pp;
float *sendbuf, *recvbuf;
/* residual file */
sf_file Fres;
Fres=sf_output("Fres");
sf_putint(Fres,"n1",nt);
sf_putint(Fres,"n2",nr);
/* initialize fcost */
*fcost=0.;
/* update velocity */
pad2d(x, vv, nz, nx, nb);
/* initialize gradient */
memset(grad, 0., nzx*sizeof(float));
/* initialize data misfit */
swap=0.;
/* memory allocation */
p0=sf_floatalloc2(padnz, padnx);
p1=sf_floatalloc2(padnz, padnx);
p2=sf_floatalloc2(padnz, padnx);
term=sf_floatalloc2(padnz, padnx);
rr=sf_floatalloc(padnzx);
wave=sf_floatalloc3(nz, nx, wnt);
pp=sf_floatalloc2(nt, nr);
iturn=0;
for(is=cpuid; is<ns; is+=numprocs){
if(cpuid==0) sf_warning("###### is=%d ######", is+1);
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(pp[0], 0., nr*nt*sizeof(float));
sx=s0_v+is*ds_v;
source_map(sx, sz, frectx, frectz, padnx, padnz, padnzx, rr);
wit=0;
/* forward propagation */
for(it=0; it<nt; it++){
if(verb) sf_warning("Forward propagation is=%d; it=%d;", is+1, it);
/* output predicted data */
for(ir=0; ir<nr2[is]; ir++){
rx=r0_v[is]+ir*dr_v;
pp[r02[is]+ir][it]=p1[rx][rz];
}
/* save wavefield */
if(it%interval==0){
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(wave,p1,wit,nb,nx,nz)
#endif
for(ix=0; ix<nx; ix++)
for(iz=0; iz<nz; iz++)
wave[wit][ix][iz]=p1[ix+nb][iz+nb];
wit++;
}
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load source */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(term,rr,padnx,padnz,ww,it)
#endif
for(ix=0; ix<padnx; ix++){
for(iz=0; iz<padnz; iz++){
term[ix][iz] += rr[ix*padnz+iz]*ww[it];
}
}
/* update */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p0,p1,p2,vv,term,padnx,padnz,dt2)
#endif
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, bc, padnx, padnz, nb);
apply_sponge(p1, bc, padnx, padnz, nb);
} // end of time loop
/* check */
if(wit != wnt) sf_error("Incorrect number of wavefield snapshots");
wit--;
/* calculate data residual and data misfit */
for(ir=0; ir<nr; ir++){
for(it=0; it<nt; it++){
pp[ir][it]=dd[iturn][ir][it]-pp[ir][it];
pp[ir][it] *= weight[ir][it];
swap += 0.5*pp[ir][it]*pp[ir][it];
}
}
smooth_misfit(pp, fcost, nr, nt, drectx, drectz, nrepeat, ider);
iturn++;
/* check the data residual */
if(is==ns/2) sf_floatwrite(pp[0], nr*nt, Fres);
sf_fileclose(Fres);
/* initialization */
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(term[0], 0., padnzx*sizeof(float));
/* backward propagation */
for(it=nt-1; it>=0; it--){
if(verb) sf_warning("Backward propagation is=%d; it=%d;", is+1, it);
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load data residual */
for(ir=0; ir<nr2[is]; ir++){
rx=r0_v[is]+ir*dr_v;
term[rx][rz] += pp[r02[is]+ir][it];
}
/* update */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p0,p1,p2,vv,term,padnx,padnz,dt2)
#endif
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* calculate gradient */
if(it%interval==0){
if(wit != wnt-1 && wit != 0){ // avoid the first and last time step
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz,temp) \
shared(nx,nz,vv,wave,p1,wit,wdt2,grad)
#endif
for(ix=0; ix<nx; ix++){
for(iz=0; iz<nz; iz++){
temp=vv[ix+nb][iz+nb];
temp=temp*temp*temp;
temp=-2./temp;
grad[ix*nz+iz] += (wave[wit+1][ix][iz]-2.*wave[wit][ix][iz]+wave[wit-1][ix][iz])/wdt2*p1[ix+nb][iz+nb]*temp;
}
}
}
wit--;
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, bc, padnx, padnz, nb);
apply_sponge(p1, bc, padnx, padnz, nb);
} // end of time loop
}// end of shot loop
MPI_Barrier(comm);
/* misfit reduction */
//MPI_ALLreduce(sendbuf, recvbuf, 1, MPI_FLOAT, MPI_SUM, comm);
if(cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=fcost;
}else{
sendbuf=fcost;
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, 1, MPI_FLOAT, MPI_SUM, 0, comm);
MPI_Bcast(fcost, 1, MPI_FLOAT, 0, comm);
if(cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=&swap;
}else{
sendbuf=&swap;
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, 1, MPI_FLOAT, MPI_SUM, 0, comm);
MPI_Bcast(&swap, 1, MPI_FLOAT, 0, comm);
/* gradient reduction */
//MPI_ALLreduce(sendbuf, recvbuf, nzx, MPI_FLOAT, MPI_SUM, comm);
if(cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=grad;
}else{
sendbuf=grad;
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, nzx, MPI_FLOAT, MPI_SUM, 0, comm);
MPI_Bcast(grad, nzx, MPI_FLOAT, 0, comm);
/* scaling gradient */
if(first){
dmax=0.;
for(ix=0; ix<nzx; ix++)
if(fabsf(grad[ix])>dmax)
dmax=fabsf(grad[ix]);
scaling=0.1/dmax;
first=false;
}
/* smooth gradient */
gradient_smooth2(grectx, grectz, nx, nz, waterz, scaling, grad);
/* free allocated memory */
free(*p0); free(p0); free(*p1); free(p1);
free(*p2); free(p2); free(*pp); free(pp);
free(**wave); free(*wave); free(wave);
free(rr); free(*term); free(term);
}
void gradient_init(sf_file Fdat, sf_mpi *mpipar, sf_sou soupar, sf_acqui acpar, sf_vec_s array, sf_fwi_s fwipar, bool verb1)
/*< initialize >*/
{
int iturn, is;
verb=verb1;
first=true; // only at the first iteration (for calculating the gradient scaling parameter)
cpuid=mpipar->cpuid;
numprocs=mpipar->numprocs;
nz=acpar->nz;
nx=acpar->nx;
nzx=nz*nx;
padnz=acpar->padnz;
padnx=acpar->padnx;
padnzx=padnz*padnx;
nb=acpar->nb;
nt=acpar->nt;
ns=acpar->ns;
ds_v=acpar->ds_v;
s0_v=acpar->s0_v;
sz=acpar->sz;
nr=acpar->nr;
dr_v=acpar->dr_v;
rz=acpar->rz;
nr2=acpar->nr2;
r02=acpar->r02;
r0_v=acpar->r0_v;
frectx=soupar->frectx;
frectz=soupar->frectz;
interval=acpar->interval;
wnt=(nt-1)/interval+1;
dt=acpar->dt;
dt2=dt*dt;
wdt=dt*interval;
wdt2=wdt*wdt;
dx2=acpar->dx*acpar->dx;
dz2=acpar->dz*acpar->dz;
/* smoothing kernel parameters */
drectx=fwipar->drectx;
drectz=fwipar->drectz;
nrepeat=fwipar->nrepeat;
ider=fwipar->ider;
/* gradient preconditioning */
waterz=fwipar->waterz;
grectx=fwipar->grectx;
grectz=fwipar->grectz;
/* data residual weighting */
gain=fwipar->gain;
wt1=fwipar->wt1;
wt2=fwipar->wt2;
woff1=fwipar->woff1;
woff2=fwipar->woff2;
wtn1=(wt1-acpar->t0)/dt+0.5;
wtn2=(wt2-acpar->t0)/dt+0.5;
woffn1=(woff1-acpar->r0)/acpar->dr+0.5;
woffn2=(woff2-acpar->r0)/acpar->dr+0.5;
weight=sf_floatalloc2(nt, nr);
residual_weighting(weight, nt, nr, wtn1, wtn2, woffn1, woffn2, gain);
ww=array->ww;
bc=acpar->bc;
/* read data */
if(ns%numprocs==0) nturn=ns/numprocs;
else nturn=ns/numprocs+1;
dd=sf_floatalloc3(nt, nr, nturn);
memset(dd[0][0], 0., nt*nr*nturn*sizeof(float));
for(iturn=0; iturn<nturn; iturn++){
is=iturn*numprocs+cpuid;
if(is<ns){
sf_seek(Fdat, is*nr*nt*sizeof(float), SEEK_SET);
sf_floatread(dd[iturn][0], nr*nt, Fdat);
}
}
/* padding and convert vector to 2-d array */
vv = sf_floatalloc2(padnz, padnx);
pad2d(array->vv, vv, nz, nx, nb);
/* hard thresholding */
threshold[0]=fwipar->v1;
threshold[1]=fwipar->v2;
return;
}
void forward_modeling_a(sf_file Fdat, sf_mpi *mpipar, sf_sou soupar, sf_acqui acpar, sf_vec array, bool verb)
/*< acoustic forward modeling >*/
{
int ix, iz, is, ir, it;
int sx, rx, sz, rz, rectx, rectz;
int nz, nx, padnz, padnx, padnzx, nt, nr, nb;
float dx2, dz2, dt2, dt;
float **vv, **dd;
float **p0, **p1, **p2, **term, **tmparray, *rr;
FILE *swap;
MPI_Comm comm=MPI_COMM_WORLD;
swap=fopen("temswap.bin", "wb+");
padnz=acpar->padnz;
padnx=acpar->padnx;
padnzx=padnz*padnx;
nz=acpar->nz;
nx=acpar->nx;
nt=acpar->nt;
nr=acpar->nr;
nb=acpar->nb;
sz=acpar->sz;
rz=acpar->rz;
rectx=soupar->rectx;
rectz=soupar->rectz;
dx2=acpar->dx*acpar->dx;
dz2=acpar->dz*acpar->dz;
dt2=acpar->dt*acpar->dt;
dt=acpar->dt;
vv = sf_floatalloc2(padnz, padnx);
dd=sf_floatalloc2(nt, nr);
p0=sf_floatalloc2(padnz, padnx);
p1=sf_floatalloc2(padnz, padnx);
p2=sf_floatalloc2(padnz, padnx);
term=sf_floatalloc2(padnz, padnx);
rr=sf_floatalloc(padnzx);
/* padding and convert vector to 2-d array */
pad2d(array->vv, vv, nz, nx, nb);
for(is=mpipar->cpuid; is<acpar->ns; is+=mpipar->numprocs){
sf_warning("###### is=%d ######", is+1);
memset(dd[0], 0., nr*nt*sizeof(float));
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
sx=acpar->s0_v+is*acpar->ds_v;
source_map(sx, sz, rectx, rectz, padnx, padnz, padnzx, rr);
for(it=0; it<nt; it++){
if(verb) sf_warning("Modeling is=%d; it=%d;", is+1, it);
/* output data */
for(ir=0; ir<acpar->nr2[is]; ir++){
rx=acpar->r0_v[is]+ir*acpar->dr_v;
dd[acpar->r02[is]+ir][it]=p1[rx][rz];
}
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load source */
for(ix=0; ix<padnx; ix++){
for(iz=0; iz<padnz; iz++){
term[ix][iz] += rr[ix*padnz+iz]*array->ww[it];
}
}
/* update */
for(ix=0; ix<padnx; ix++){
for(iz=0; iz<padnz; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, acpar->bc, padnx, padnz, nb);
apply_sponge(p1, acpar->bc, padnx, padnz, nb);
} // end of time loop
fseeko(swap, is*nr*nt*sizeof(float), SEEK_SET);
fwrite(dd[0], sizeof(float), nr*nt, swap);
}// end of shot loop
fclose(swap);
MPI_Barrier(comm);
/* transfer data to Fdat */
if(mpipar->cpuid==0){
swap=fopen("temswap.bin", "rb");
for(is=0; is<acpar->ns; is++){
fseeko(swap, is*nr*nt*sizeof(float), SEEK_SET);
fread(dd[0], sizeof(float), nr*nt, swap);
sf_floatwrite(dd[0], nr*nt, Fdat);
}
fclose(swap);
remove("temswap.bin");
}
MPI_Barrier(comm);
/* release allocated memory */
free(*p0); free(p0); free(*p1); free(p1);
free(*p2); free(p2); free(*vv); free(vv);
free(*dd); free(dd);
free(rr); free(*term); free(term);
}
void rtm(sf_file Fdat, sf_file Fimg, sf_mpi *mpipar, sf_sou soupar, sf_acqui acpar, sf_vec_s array, bool verb)
/*< acoustic rtm >*/
{
int ix, iz, is, ir, it, wit;
int sx, rx, sz, rz, frectx, frectz;
int nz, nx, nzx, padnz, padnx, padnzx, nt, nr, nb, wnt;
float dx2, dz2, dt2, dt;
float **vv, **dd, **mm;
float **p0, **p1, **p2, **term, **tmparray, *rr, ***wave;
float *sendbuf, *recvbuf;
//sf_file Fwfl1, Fwfl2;
//Fwfl1=sf_output("Fwfl1");
//Fwfl2=sf_output("Fwfl2");
MPI_Comm comm=MPI_COMM_WORLD;
nz=acpar->nz;
nx=acpar->nx;
nzx=nz*nx;
padnz=acpar->padnz;
padnx=acpar->padnx;
padnzx=padnz*padnx;
nr=acpar->nr;
nb=acpar->nb;
sz=acpar->sz;
rz=acpar->rz;
frectx=soupar->frectx;
frectz=soupar->frectz;
nt=acpar->nt;
wnt=(nt-1)/acpar->interval+1;
dx2=acpar->dx*acpar->dx;
dz2=acpar->dz*acpar->dz;
dt2=acpar->dt*acpar->dt;
dt=acpar->dt;
//sf_putint(Fwfl1, "n1", padnz);
//sf_putint(Fwfl1, "n2",padnx);
//sf_putint(Fwfl1, "n3",(nt-1)/50+1);
//sf_putint(Fwfl2, "n1", padnz);
//sf_putint(Fwfl2, "n2",padnx);
//sf_putint(Fwfl2, "n3",(nt-1)/50+1);
/* memory allocation */
vv = sf_floatalloc2(padnz, padnx);
dd=sf_floatalloc2(nt, nr);
mm=sf_floatalloc2(nz, nx);
p0=sf_floatalloc2(padnz, padnx);
p1=sf_floatalloc2(padnz, padnx);
p2=sf_floatalloc2(padnz, padnx);
term=sf_floatalloc2(padnz, padnx);
rr=sf_floatalloc(padnzx);
wave=sf_floatalloc3(nz, nx, wnt);
/* padding and convert vector to 2-d array */
pad2d(array->vv, vv, nz, nx, nb);
memset(mm[0], 0., nzx*sizeof(float));
for(is=mpipar->cpuid; is<acpar->ns; is+=mpipar->numprocs){
sf_warning("###### is=%d ######", is+1);
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
sx=acpar->s0_v+is*acpar->ds_v;
source_map(sx, sz, frectx, frectz, padnx, padnz, padnzx, rr);
wit=0;
/* forward propagation */
for(it=0; it<nt; it++){
if(verb) sf_warning("Forward propagation is=%d; it=%d;", is+1, it);
/* save wavefield */
if(it%acpar->interval==0){
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(wave,p1,wit,nb,nx,nz)
#endif
for(ix=0; ix<nx; ix++)
for(iz=0; iz<nz; iz++)
wave[wit][ix][iz]=p1[ix+nb][iz+nb];
wit++;
}
//if(is==acpar->ns/2 && it%50==0) sf_floatwrite(p1[0],padnzx, Fwfl1);
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load source */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(term,rr,padnx,padnz,it)
#endif
for(ix=0; ix<padnx; ix++){
for(iz=0; iz<padnz; iz++){
term[ix][iz] += rr[ix*padnz+iz]*array->ww[it];
}
}
/* update */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p0,p1,p2,vv,term,padnx,padnz,dt2)
#endif
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, acpar->bc, padnx, padnz, nb);
apply_sponge(p1, acpar->bc, padnx, padnz, nb);
} // end of time loop
/* check */
if(wit != wnt) sf_error("Incorrect number of wavefield snapshots");
wit--;
/* read data */
sf_seek(Fdat, is*nr*nt*sizeof(float), SEEK_SET);
sf_floatread(dd[0], nr*nt, Fdat);
/* initialization */
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(term[0], 0., padnzx*sizeof(float));
/* backward propagation */
for(it=nt-1; it>=0; it--){
if(verb) sf_warning("Backward propagation is=%d; it=%d;", is+1, it);
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load data */
for(ir=0; ir<acpar->nr2[is]; ir++){
rx=acpar->r0_v[is]+ir*acpar->dr_v;
term[rx][rz] += dd[acpar->r02[is]+ir][it];
}
/* update */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p0,p1,p2,vv,term,padnx,padnz,dt2)
#endif
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
//if(is==acpar->ns/2 && it%50==0) sf_floatwrite(p1[0],padnzx, Fwfl2);
/* calculate image */
if(it%acpar->interval==0){
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p1,wave,nx,nz,wit,mm,nb)
#endif
for(ix=0; ix<nx; ix++)
for(iz=0; iz<nz; iz++)
mm[ix][iz] += wave[wit][ix][iz]*p1[ix+nb][iz+nb];
wit--;
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, acpar->bc, padnx, padnz, nb);
apply_sponge(p1, acpar->bc, padnx, padnz, nb);
} // end of time loop
}// end of shot loop
MPI_Barrier(comm);
//sf_fileclose(Fwfl1);
//sf_fileclose(Fwfl2);
if(mpipar->cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=mm[0];
}else{
sendbuf=mm[0];
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, nzx, MPI_FLOAT, MPI_SUM, 0, comm);
if(mpipar->cpuid==0) sf_floatwrite(mm[0], nzx, Fimg);
MPI_Barrier(comm);
/* release allocated memory */
free(*p0); free(p0); free(*p1); free(p1);
free(*p2); free(p2); free(*vv); free(vv);
free(*dd); free(dd); free(*mm); free(mm);
free(rr); free(*term); free(term);
free(**wave); free(*wave); free(wave);
}
void rtm_q(sf_file Fdat, sf_file Fimg, sf_mpi *mpipar, sf_sou soupar, sf_acqui acpar, sf_vec_q array, bool verb)
/*< visco-acoustic rtm >*/
{
int ix, iz, is, ir, it, wit;
int sx, rx, sz, rz, frectx, frectz;
int nz, nx, nzx, padnz, padnx, padnzx, nt, nr, nb, wnt;
float dx2, dz2, dt2, dt;
float **vv, **tau, **taus, **dd, **mm;
float **p0, **p1, **p2, **r1, **r2, **term, **tmparray, *rr, ***wave;
float *sendbuf, *recvbuf;
MPI_Comm comm=MPI_COMM_WORLD;
nz=acpar->nz;
nx=acpar->nx;
nzx=nz*nx;
padnz=acpar->padnz;
padnx=acpar->padnx;
padnzx=padnz*padnx;
nr=acpar->nr;
nb=acpar->nb;
sz=acpar->sz;
rz=acpar->rz;
frectx=soupar->frectx;
frectz=soupar->frectz;
nt=acpar->nt;
wnt=(nt-1)/acpar->interval+1;
dx2=acpar->dx*acpar->dx;
dz2=acpar->dz*acpar->dz;
dt2=acpar->dt*acpar->dt;
dt=acpar->dt;
/* memory allocation */
vv = sf_floatalloc2(padnz, padnx);
tau= sf_floatalloc2(padnz, padnx);
taus=sf_floatalloc2(padnz, padnx);
dd=sf_floatalloc2(nt, nr);
mm=sf_floatalloc2(nz, nx);
p0=sf_floatalloc2(padnz, padnx);
p1=sf_floatalloc2(padnz, padnx);
p2=sf_floatalloc2(padnz, padnx);
r1=sf_floatalloc2(padnz, padnx);
r2=sf_floatalloc2(padnz, padnx);
term=sf_floatalloc2(padnz, padnx);
rr=sf_floatalloc(padnzx);
wave=sf_floatalloc3(nz, nx, wnt);
/* padding and convert vector to 2-d array */
pad2d(array->vv, vv, nz, nx, nb);
pad2d(array->tau, tau, nz, nx, nb);
pad2d(array->taus, taus, nz, nx, nb);
memset(mm[0], 0., nzx*sizeof(float));
for(is=mpipar->cpuid; is<acpar->ns; is+=mpipar->numprocs){
sf_warning("###### is=%d ######", is+1);
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(r1[0], 0., padnzx*sizeof(float));
memset(r2[0], 0., padnzx*sizeof(float));
memset(term[0], 0., padnzx*sizeof(float));
sx=acpar->s0_v+is*acpar->ds_v;
source_map(sx, sz, frectx, frectz, padnx, padnz, padnzx, rr);
wit=0;
/* forward propagation */
for(it=0; it<nt; it++){
if(verb) sf_warning("Forward propagation is=%d; it=%d;", is+1, it);
/* save wavefield */
if(it%acpar->interval==0){
for(ix=0; ix<nx; ix++)
for(iz=0; iz<nz; iz++)
wave[wit][ix][iz]=p1[ix+nb][iz+nb];
wit++;
}
/* load source */
for(ix=0; ix<padnx; ix++){
for(iz=0; iz<padnz; iz++){
p1[ix][iz] += rr[ix*padnz+iz]*array->ww[it];
}
}
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* update */
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
r2[ix][iz]=
(-tau[ix][iz]/taus[ix][iz]*term[ix][iz]
+ (1./dt-0.5/taus[ix][iz])*r1[ix][iz])
/(1./dt+0.5/taus[ix][iz]);
term[ix][iz]=term[ix][iz]*(1.+tau[ix][iz])+(r2[ix][iz]+r1[ix][iz])*0.5;
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
tmparray=r1; r1=r2; r2=tmparray;
/* boundary condition */
apply_sponge(p0, acpar->bc, padnx, padnz, nb);
apply_sponge(p1, acpar->bc, padnx, padnz, nb);
apply_sponge(r1, acpar->bc, padnx, padnz, nb);
} // end of time loop
/* check */
if(wit != wnt) sf_error("Incorrect number of wavefield snapshots");
/* read data */
sf_seek(Fdat, is*nr*nt*sizeof(float), SEEK_SET);
sf_floatread(dd[0], nr*nt, Fdat);
/* initialization */
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(r1[0], 0., padnzx*sizeof(float));
memset(r2[0], 0., padnzx*sizeof(float));
memset(term[0], 0., padnzx*sizeof(float));
/* backward propagation */
for(it=nt-1; it>=0; it--){
if(verb) sf_warning("Backward propagation is=%d; it=%d;", is+1, it);
/* load data */
for(ir=0; ir<acpar->nr2[is]; ir++){
rx=acpar->r0_v[is]+ir*acpar->dr_v;
p1[rx][rz]=dd[acpar->r02[is]+ir][it];
}
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* update */
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
r2[ix][iz]=
(-tau[ix][iz]/taus[ix][iz]*term[ix][iz]
+ (1./dt-0.5/taus[ix][iz])*r1[ix][iz])
/(1./dt+0.5/taus[ix][iz]);
term[ix][iz]=term[ix][iz]*(1.+tau[ix][iz])+(r2[ix][iz]+r1[ix][iz])*0.5;
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* calculate image */
if(it%acpar->interval==0){
for(ix=0; ix<nx; ix++)
for(iz=0; iz<nz; iz++)
mm[ix][iz] += wave[wit-1][ix][iz]*p2[ix+nb][iz+nb];
wit--;
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
tmparray=r1; r1=r2; r2=tmparray;
/* boundary condition */
apply_sponge(p0, acpar->bc, padnx, padnz, nb);
apply_sponge(p1, acpar->bc, padnx, padnz, nb);
apply_sponge(r1, acpar->bc, padnx, padnz, nb);
} // end of time loop
}// end of shot loop
MPI_Barrier(comm);
if(mpipar->cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=mm[0];
}else{
sendbuf=mm[0];
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, nzx, MPI_FLOAT, MPI_SUM, 0, comm);
if(mpipar->cpuid==0) sf_floatwrite(mm[0], nzx, Fimg);
MPI_Barrier(comm);
}
//JS
//static int counter=0;
void gradient_pas_av(float *x, float *fcost, float *grad)
/*< acoustic velocity gradient >*/
{
int ix, iz, is, ir, it, wit, iturn;
int rx;
float temp, dmax;
float **p0, **p1, **p2, **term, **tmparray, ***wave, **pp;
float *sendbuf, *recvbuf;
/*
//JS
sf_file Fwfl1, Fwfl2, Fres;
counter++;
if (cpuid==0 && counter==3) {
Fwfl1=sf_output("Fwfl1");
Fwfl2=sf_output("Fwfl2");
Fres=sf_output("Fres");
sf_putint(Fwfl1,"n1",padnz);
sf_putint(Fwfl1,"n2",padnx);
sf_putint(Fwfl1,"n3",(nt-1)/50+1);
sf_putint(Fwfl2,"n1",padnz);
sf_putint(Fwfl2,"n2",padnx);
sf_putint(Fwfl2,"n3",(nt-1)/50+1);
sf_putint(Fres,"n1",nt);
sf_putint(Fres,"n2",nr);
}
*/
/* initialize fcost */
*fcost=0.;
/* update velocity */
pad2d(x, vv, nz, nx, nb);
/* initialize gradient */
memset(grad, 0., nzx*sizeof(float));
/* memory allocation */
p0=sf_floatalloc2(padnz, padnx);
p1=sf_floatalloc2(padnz, padnx);
p2=sf_floatalloc2(padnz, padnx);
term=sf_floatalloc2(padnz, padnx);
wave=sf_floatalloc3(nz, nx, wnt);
pp=sf_floatalloc2(nt, nr);
iturn=0;
for(is=cpuid; is<ns; is+=numprocs){
if(cpuid==0) sf_warning("###### is=%d ######", is+1);
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(pp[0], 0., nr*nt*sizeof(float));
wit=0;
/* forward propagation */
for(it=0; it<nt; it++){
if(verb) sf_warning("Forward propagation it=%d;", it);
/* output predicted data */
for(ir=0; ir<nr; ir++){
rx=r0_v[0]+ir*dr_v;
pp[ir][it]=p1[rx][rz];
}
/* save wavefield */
if(it%interval==0){
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(wave,p1,wit,nb,nx,nz)
#endif
for(ix=0; ix<nx; ix++)
for(iz=0; iz<nz; iz++)
wave[wit][ix][iz]=p1[ix+nb][iz+nb];
wit++;
}
/*
//JS
if(is==0 && counter==3 && it%50==0) sf_floatwrite(p1[0],padnzx,Fwfl1);
*/
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load source */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(term,nb,ww3,it)
#endif
for(ix=0; ix<nx; ix++){
for(iz=0; iz<nz; iz++){
term[ix+nb][iz+nb] += ww3[iturn][it][ix][iz];
}
}
/* update */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p0,p1,p2,vv,term,padnx,padnz,dt2)
#endif
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, bc, padnx, padnz, nb);
apply_sponge(p1, bc, padnx, padnz, nb);
} // end of time loop
/* check */
if(wit != wnt) sf_error("Incorrect number of wavefield snapshots");
wit--;
/* calculate data residual and data misfit */
for(ir=0; ir<nr; ir++){
for(it=0; it<nt; it++){
pp[ir][it]=dd[iturn][ir][it]-pp[ir][it];
*fcost += 0.5*pp[ir][it]*pp[ir][it];
}
}
/* window the data residual */
for(ir=0; ir<nr; ir++){
/* multiscale */
if(NULL != blo){
sf_butter_apply(blo, nt, pp[ir]);
sf_reverse(nt, pp[ir]);
sf_butter_apply(blo, nt, pp[ir]);
sf_reverse(nt, pp[ir]);
}
if(NULL != bhi){
sf_butter_apply(bhi, nt, pp[ir]);
sf_reverse(nt, pp[ir]);
sf_butter_apply(bhi, nt, pp[ir]);
sf_reverse(nt, pp[ir]);
}
for(it=0; it<nt; it++){
pp[ir][it] *= weight[ir][it];
}
}
/*
// JS
if(is==0 && counter==3) sf_floatwrite(pp[0], nr*nt, Fres);
*/
/* initialization */
memset(p0[0], 0., padnzx*sizeof(float));
memset(p1[0], 0., padnzx*sizeof(float));
memset(p2[0], 0., padnzx*sizeof(float));
memset(term[0], 0., padnzx*sizeof(float));
/* backward propagation */
for(it=nt-1; it>=0; it--){
if(verb) sf_warning("Backward propagation it=%d;", it);
/* laplacian operator */
laplace(p1, term, padnx, padnz, dx2, dz2);
/* load data residual*/
for(ir=0; ir<nr; ir++){
rx=r0_v[0]+ir*dr_v;
term[rx][rz] += pp[ir][it];
}
/* update */
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz) \
shared(p0,p1,p2,vv,term,padnx,padnz,dt2)
#endif
for(ix=4; ix<padnx-4; ix++){
for(iz=4; iz<padnz-4; iz++){
p2[ix][iz]=2*p1[ix][iz]-p0[ix][iz]+vv[ix][iz]*vv[ix][iz]*dt2*term[ix][iz];
}
}
/*
// JS
if(is==0 && counter==3 && it%50==0) sf_floatwrite(p1[0],padnzx,Fwfl2);
*/
/* calculate gradient */
if(it%interval==0){
if(wit != wnt-1 && wit != 0){ // avoid the first and last time step
#ifdef _OPENMP
#pragma omp parallel for \
private(ix,iz,temp) \
shared(nx,nz,vv,wave,p1,wit,wdt2,grad)
#endif
for(ix=0; ix<nx; ix++){
for(iz=0; iz<nz; iz++){
temp=vv[ix+nb][iz+nb];
temp=temp*temp*temp;
temp=-2./temp;
grad[ix*nz+iz] += gwt[iturn][it][ix][iz]*(wave[wit+1][ix][iz]-2.*wave[wit][ix][iz]+wave[wit-1][ix][iz])/wdt2*p1[ix+nb][iz+nb]*temp;
}
}
}
wit--;
}
/* swap wavefield pointer of different time steps */
tmparray=p0; p0=p1; p1=p2; p2=tmparray;
/* boundary condition */
apply_sponge(p0, bc, padnx, padnz, nb);
apply_sponge(p1, bc, padnx, padnz, nb);
} // end of time loop
iturn++;
/*
// JS
if(is==0 && counter==3){
sf_fileclose(Fwfl1);
sf_fileclose(Fwfl2);
sf_fileclose(Fres);
}
sf_warning("---counter=%d fcost=%3.3e---",counter, *fcost);
*/
} // end of shot loop
MPI_Barrier(comm);
/* misfit reduction */
if(cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=fcost;
}else{
sendbuf=fcost;
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, 1, MPI_FLOAT, MPI_SUM, 0, comm);
MPI_Bcast(fcost, 1, MPI_FLOAT, 0, comm);
/* gradient reduction */
if(cpuid==0){
sendbuf=MPI_IN_PLACE;
recvbuf=grad;
}else{
sendbuf=grad;
recvbuf=NULL;
}
MPI_Reduce(sendbuf, recvbuf, nzx, MPI_FLOAT, MPI_SUM, 0, comm);
MPI_Bcast(grad, nzx, MPI_FLOAT, 0, comm);
/* scaling gradient */
if(first){
dmax=0.;
for(ix=0; ix<nzx; ix++)
if(fabsf(grad[ix])>dmax)
dmax=fabsf(grad[ix]);
scaling=0.1/dmax;
first=false;
}
/* smooth gradient */
gradient_smooth2b(grectx, grectz, nx, nz, waterz, waterzb, scaling, grad);
/* free allocated memory */
free(*p0); free(p0); free(*p1); free(p1);
free(*p2); free(p2); free(*pp); free(pp);
free(**wave); free(*wave); free(wave);
free(*term); free(term);
}
/* for passive source and fwi */
void gradient_pas_init(sf_file Fdat, sf_file Fsrc, sf_file Fmwt, sf_mpi *mpipar, sf_sou soupar, sf_acqui acpar, sf_vec array, sf_fwi fwipar, sf_pas paspar, bool verb1)
/*< initialize >*/
{
float **dwt=NULL,***mwt,***wwt;
int it,ix,iz,iturn,is,rdn;
char filename[20]="tempbin",srdn[10];
FILE *temp;
verb=verb1;
first=true; // only at the first iteration, need to calculate the gradient scaling parameter
cpuid=mpipar->cpuid;
numprocs=mpipar->numprocs;
nz=acpar->nz;
nx=acpar->nx;
nzx=nz*nx;
padnz=acpar->padnz;
padnx=acpar->padnx;
padnzx=padnz*padnx;
nb=acpar->nb;
nt=acpar->nt;
ns=acpar->ns;
nr=acpar->nr;
dr_v=acpar->dr_v;
r0_v=acpar->r0_v;
rz=acpar->rz;
grectx=fwipar->rectx;
grectz=fwipar->rectz;
interval=acpar->interval;
wnt=(nt-1)/interval+1;
dt=acpar->dt;
idt=1./dt;
dt2=dt*dt;
wdt=dt*interval;
wdt2=wdt*wdt;
dx2=acpar->dx*acpar->dx;
dz2=acpar->dz*acpar->dz;
wt1=fwipar->wt1;
wt2=fwipar->wt2;
woff1=fwipar->woff1;
woff2=fwipar->woff2;
waterz=fwipar->waterz;
waterzb=fwipar->waterzb;
bc=acpar->bc;
if (cpuid==0) {
srand(time(NULL));
rdn = rand()%1000000000;
sprintf(srdn,"%d",rdn);
strcat(filename,srdn);
}
MPI_Bcast(filename, 20, MPI_CHAR, 0, comm);
if(verb && cpuid==0) sf_warning("filename=%s",filename);
temp=fopen(filename, "wb+");
if(ns%numprocs==0) nturn=ns/numprocs;
else nturn=ns/numprocs+1;
/* allocate data/source/weight */
dd = sf_floatalloc3(nt, nx, nturn);
ww3 = sf_floatalloc4(nz, nx, nt, nturn);
gwt = sf_floatalloc4(nz, nx, nt, nturn);
wwt = sf_floatalloc3(nz, nx, nt); /* temporary output var */
if (!paspar->onlyvel) {
mwt = sf_floatalloc3(nz, nx, nt); /* src model weight */
/*
dwt = sf_floatalloc2(acpar->nt, acpar->nx);
wtn1=(fwipar->wt1-acpar->t0)/acpar->dt+0.5;
wtn2=(fwipar->wt2-acpar->t0)/acpar->dt+0.5;
woffn1=(fwipar->woff1-acpar->r0)/acpar->dr+0.5;
woffn2=(fwipar->woff2-acpar->r0)/acpar->dr+0.5;
residual_weighting(dwt, acpar->nt, acpar->nx, wtn1, wtn2, woffn1, woffn2, !fwipar->oreo);
*/
} else {
mwt=NULL;
dwt=NULL;
}
/* read data/source */
for(iturn=0; iturn<nturn; iturn++){
is=iturn*numprocs+cpuid;
if(is<ns){
/* read data */
sf_seek(Fdat, is*nt*nx*sizeof(float), SEEK_SET);
sf_floatread(dd[iturn][0], nt*nx, Fdat);
if (paspar->onlyvel) {
/* read source */
sf_seek(Fsrc, is*nz*nx*nt*sizeof(float), SEEK_SET);
sf_floatread(ww3[iturn][0][0], nz*nx*nt, Fsrc);
} else {
/* linear inversion of source */
lstri_op(dd[iturn], dwt, ww3[iturn], mwt, acpar, array, paspar, verb);
/* write source */
fseeko(temp, is*nz*nx*nt*sizeof(float), SEEK_SET);
fwrite(ww3[iturn][0][0], sizeof(float), nz*nx*nt, temp);
if (NULL!=Fmwt && is==0) sf_floatwrite(mwt[0][0], nz*nx*nt, Fmwt);
}
/* calculate gradient mask */
if (!paspar->onlyvel && paspar->prec && paspar->hidesrc) {
#ifdef _OPENMP
#pragma omp parallel for default(shared) private(it,ix,iz)
#endif
for (it=0; it<nt; it++)
for (ix=0; ix<nx; ix++)
for (iz=0; iz<nz; iz++)
gwt[iturn][it][ix][iz] = mwt[it][ix][iz];
threshold(true, nz*nx*nt, paspar->hard, gwt[iturn][0][0]);
#ifdef _OPENMP
#pragma omp parallel for default(shared) private(it,ix,iz)
#endif
for (it=0; it<nt; it++)
for (ix=0; ix<nx; ix++)
for (iz=0; iz<nz; iz++)
gwt[iturn][it][ix][iz] = 1.-gwt[iturn][it][ix][iz];
} else {
#ifdef _OPENMP
#pragma omp parallel for default(shared) private(it,ix,iz)
#endif
for (it=0; it<nt; it++)
for (ix=0; ix<nx; ix++)
for (iz=0; iz<nz; iz++)
gwt[iturn][it][ix][iz] = 1.;
}
} /* if is<ns */
}
fclose(temp);
MPI_Barrier(comm);
if(!paspar->onlyvel && cpuid==0) {
temp=fopen(filename, "rb");
for(is=0; is<ns; is++){
fseeko(temp, is*nz*nx*nt*sizeof(float), SEEK_SET);
fread(wwt[0][0], sizeof(float), nz*nx*nt, temp);
sf_floatwrite(wwt[0][0], nz*nx*nt, Fsrc);
}
fclose(temp);
remove(filename);
}
MPI_Barrier(comm);
/* data residual weights */
wtn1=(wt1-acpar->t0)/dt+0.5;
wtn2=(wt2-acpar->t0)/dt+0.5;
woffn1=(woff1-acpar->r0)/acpar->dr+0.5;
woffn2=(woff2-acpar->r0)/acpar->dr+0.5;
weight=sf_floatalloc2(nt, nr);
residual_weighting(weight, nt, nr, wtn1, wtn2, woffn1, woffn2, fwipar->oreo);
/* padding and convert vector to 2-d array */
vv = sf_floatalloc2(padnz, padnx);
tau= sf_floatalloc2(padnz, padnx);
taus=sf_floatalloc2(padnz, padnx);
pad2d(array->vv, vv, nz, nx, nb);
pad2d(array->tau, tau, nz, nx, nb);
pad2d(array->taus, taus, nz, nx, nb);
/* multiscale gradient */
if(soupar->flo > 0.0001) blo=sf_butter_init(false, soupar->flo, 3);
if(soupar->fhi < 0.5-0.0001) bhi=sf_butter_init(true, soupar->fhi, 3);
free(**wwt); free(*wwt); free(wwt);
if (NULL!=mwt) { free(**mwt); free(*mwt); free(mwt); }
return;
}
