interpol_ng.h 21.5 KB
Newer Older
1
2
3
4
5
/*
 *  Copyright (C) 2020 Max-Planck-Society
 *  Author: Martin Reinecke
 */

Martin Reinecke's avatar
Martin Reinecke committed
6
7
8
#ifndef MRUTIL_INTERPOL_NG_H
#define MRUTIL_INTERPOL_NG_H

9
10
#include <vector>
#include <complex>
Martin Reinecke's avatar
Martin Reinecke committed
11
#include <cmath>
12
13
14
15
16
17
18
19
20
21
#include "mr_util/math/constants.h"
#include "mr_util/math/gl_integrator.h"
#include "mr_util/math/es_kernel.h"
#include "mr_util/infra/mav.h"
#include "mr_util/sharp/sharp.h"
#include "mr_util/sharp/sharp_almhelpers.h"
#include "mr_util/sharp/sharp_geomhelpers.h"
#include "alm.h"
#include "mr_util/math/fft.h"
#include "mr_util/bindings/pybind_utils.h"
Martin Reinecke's avatar
Martin Reinecke committed
22

Martin Reinecke's avatar
Martin Reinecke committed
23
namespace mr {
24

Martin Reinecke's avatar
Martin Reinecke committed
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
namespace detail_fft {

using std::vector;

template<typename T, typename T0> aligned_array<T> alloc_tmp_conv
  (const fmav_info &info, size_t axis, size_t len)
  {
  auto othersize = info.size()/info.shape(axis);
  constexpr auto vlen = native_simd<T0>::size();
  auto tmpsize = len*((othersize>=vlen) ? vlen : 1);
  return aligned_array<T>(tmpsize);
  }

template<typename Tplan, typename T, typename T0, typename Exec>
MRUTIL_NOINLINE void general_convolve(const fmav<T> &in, fmav<T> &out,
  const size_t axis, const vector<T0> &kernel, size_t nthreads,
Martin Reinecke's avatar
Martin Reinecke committed
41
  const Exec &exec)
Martin Reinecke's avatar
Martin Reinecke committed
42
43
44
45
46
47
48
49
50
51
52
53
54
  {
  std::shared_ptr<Tplan> plan1, plan2;

  size_t l_in=in.shape(axis), l_out=out.shape(axis);
  size_t l_min=std::min(l_in, l_out), l_max=std::max(l_in, l_out);
  MR_assert(kernel.size()==l_min/2+1, "bad kernel size");
  plan1 = get_plan<Tplan>(l_in);
  plan2 = get_plan<Tplan>(l_out);

  execParallel(
    util::thread_count(nthreads, in, axis, native_simd<T0>::size()),
    [&](Scheduler &sched) {
      constexpr auto vlen = native_simd<T0>::size();
Martin Reinecke's avatar
Martin Reinecke committed
55
      auto storage = alloc_tmp_conv<T,T0>(in, axis, l_max);
Martin Reinecke's avatar
Martin Reinecke committed
56
57
58
59
60
61
62
63
64
65
66
67
68
      multi_iter<vlen> it(in, out, axis, sched.num_threads(), sched.thread_num());
#ifndef MRUTIL_NO_SIMD
      if (vlen>1)
        while (it.remaining()>=vlen)
          {
          it.advance(vlen);
          auto tdatav = reinterpret_cast<add_vec_t<T> *>(storage.data());
          exec(it, in, out, tdatav, *plan1, *plan2, kernel);
          }
#endif
      while (it.remaining()>0)
        {
        it.advance(1);
Martin Reinecke's avatar
Martin Reinecke committed
69
        auto buf = reinterpret_cast<T *>(storage.data());
Martin Reinecke's avatar
Martin Reinecke committed
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
        exec(it, in, out, buf, *plan1, *plan2, kernel);
        }
    });  // end of parallel region
  }

struct ExecConvR1
  {
  template <typename T0, typename T, size_t vlen> void operator() (
    const multi_iter<vlen> &it, const fmav<T0> &in, fmav<T0> &out,
    T * buf, const pocketfft_r<T0> &plan1, const pocketfft_r<T0> &plan2,
    const vector<T0> &kernel) const
    {
    size_t l_in = plan1.length(),
           l_out = plan2.length(),
           l_min = std::min(l_in, l_out);
    copy_input(it, in, buf);
    plan1.exec(buf, T0(1), true);
Martin Reinecke's avatar
Martin Reinecke committed
87
    for (size_t i=0; i<l_min; ++i) buf[i]*=kernel[(i+1)/2];
Martin Reinecke's avatar
Martin Reinecke committed
88
89
90
91
92
93
94
95
96
97
98
99
    for (size_t i=l_in; i<l_out; ++i) buf[i] = T(0);
    plan2.exec(buf, T0(1), false);
    copy_output(it, buf, out);
    }
  };

template<typename T> void convolve_1d(const fmav<T> &in,
  fmav<T> &out, size_t axis, const vector<T> &kernel, size_t nthreads=1)
  {
  MR_assert(axis<in.ndim(), "bad axis number");
  MR_assert(in.ndim()==out.ndim(), "dimensionality mismatch");
  if (in.data()==out.data())
Martin Reinecke's avatar
Martin Reinecke committed
100
    MR_assert(in.stride()==out.stride(), "strides mismatch");
Martin Reinecke's avatar
Martin Reinecke committed
101
102
103
  for (size_t i=0; i<in.ndim(); ++i)
    if (i!=axis)
      MR_assert(in.shape(i)==out.shape(i), "shape mismatch");
Martin Reinecke's avatar
Martin Reinecke committed
104
105
  MR_assert(!((in.shape(axis)&1) || (out.shape(axis)&1)),
    "input and output axis lengths must be even");
Martin Reinecke's avatar
Martin Reinecke committed
106
107
108
109
110
111
  if (in.size()==0) return;
  general_convolve<pocketfft_r<T>>(in, out, axis, kernel, nthreads,
    ExecConvR1());
  }

}
Martin Reinecke's avatar
Martin Reinecke committed
112
113
114

using detail_fft::convolve_1d;

Martin Reinecke's avatar
Martin Reinecke committed
115
namespace detail_interpol_ng {
116

Martin Reinecke's avatar
Martin Reinecke committed
117
using namespace std;
118
119
120
121

template<typename T> class Interpolator
  {
  protected:
122
    bool adjoint;
Martin Reinecke's avatar
Martin Reinecke committed
123
    size_t lmax, kmax, nphi0, ntheta0, nphi, ntheta;
Martin Reinecke's avatar
Martin Reinecke committed
124
    int nthreads;
125
    T ofactor;
Martin Reinecke's avatar
fix    
Martin Reinecke committed
126
    size_t supp;
127
    ES_Kernel kernel;
128
129
    size_t ncomp;
    mav<T,4> cube; // the data cube (theta, phi, 2*mbeam+1, TGC)
130

131
    void correct(mav<T,2> &arr, int spin)
132
      {
133
      T sfct = (spin&1) ? -1 : 1;
Martin Reinecke's avatar
Martin Reinecke committed
134
135
136
137
      mav<T,2> tmp({nphi,nphi0});
      // copy and extend to second half
      for (size_t j=0; j<nphi0; ++j)
        tmp.v(0,j) = arr(0,j);
Martin Reinecke's avatar
Martin Reinecke committed
138
      for (size_t i=1, i2=nphi0-1; i+1<ntheta0; ++i,--i2)
Martin Reinecke's avatar
Martin Reinecke committed
139
        for (size_t j=0,j2=nphi0/2; j<nphi0; ++j,++j2)
140
          {
Martin Reinecke's avatar
Martin Reinecke committed
141
          if (j2>=nphi0) j2-=nphi0;
Martin Reinecke's avatar
Martin Reinecke committed
142
143
          tmp.v(i,j2) = arr(i,j2);
          tmp.v(i2,j) = sfct*tmp(i,j2);
144
          }
Martin Reinecke's avatar
Martin Reinecke committed
145
146
      for (size_t j=0; j<nphi0; ++j)
        tmp.v(ntheta0-1,j) = arr(ntheta0-1,j);
Martin Reinecke's avatar
Martin Reinecke committed
147
      auto fct = kernel.correction_factors(nphi, nphi0/2+1, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
148
149
150
151
152
      for (auto &f:fct) f/=nphi0;
      fmav<T> ftmp(tmp);
      fmav<T> ftmp0(tmp.template subarray<2>({0,0},{nphi0, nphi0}));
      convolve_1d(ftmp0, ftmp, 0, fct, nthreads);
      fmav<T> ftmp2(tmp.template subarray<2>({0,0},{ntheta, nphi0}));
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
153
      fmav<T> farr(arr);
Martin Reinecke's avatar
Martin Reinecke committed
154
      convolve_1d(ftmp2, farr, 1, fct, nthreads);
155
      }
156
157
    void decorrect(mav<T,2> &arr, int spin)
      {
158
      T sfct = (spin&1) ? -1 : 1;
Martin Reinecke's avatar
Martin Reinecke committed
159
160
161
162
163
164
      mav<T,2> tmp({nphi,nphi0});
      auto fct = kernel.correction_factors(nphi, nphi0/2+1, nthreads);
      for (auto &f:fct) f/=nphi0;
      fmav<T> farr(arr);
      fmav<T> ftmp2(tmp.template subarray<2>({0,0},{ntheta, nphi0}));
      convolve_1d(farr, ftmp2, 1, fct, nthreads);
165
      // extend to second half
Martin Reinecke's avatar
Martin Reinecke committed
166
      for (size_t i=1, i2=nphi-1; i+1<ntheta; ++i,--i2)
Martin Reinecke's avatar
Martin Reinecke committed
167
        for (size_t j=0,j2=nphi0/2; j<nphi0; ++j,++j2)
168
          {
Martin Reinecke's avatar
Martin Reinecke committed
169
          if (j2>=nphi0) j2-=nphi0;
170
171
          tmp.v(i2,j) = sfct*tmp(i,j2);
          }
Martin Reinecke's avatar
Martin Reinecke committed
172
173
174
      fmav<T> ftmp(tmp);
      fmav<T> ftmp0(tmp.template subarray<2>({0,0},{nphi0, nphi0}));
      convolve_1d(ftmp, ftmp0, 0, fct, nthreads);
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
175
      for (size_t j=0; j<nphi0; ++j)
Martin Reinecke's avatar
Martin Reinecke committed
176
177
        arr.v(0,j) = 0.5*tmp(0,j);
      for (size_t i=1; i+1<ntheta0; ++i)
178
        for (size_t j=0; j<nphi0; ++j)
Martin Reinecke's avatar
Martin Reinecke committed
179
180
181
          arr.v(i,j) = tmp(i,j);
      for (size_t j=0; j<nphi0; ++j)
        arr.v(ntheta0-1,j) = 0.5*tmp(ntheta0-1,j);
182
      }
183

Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
184
185
186
187
188
189
190
191
192
193
194
    vector<size_t> getIdx(const mav<T,2> &ptg) const
      {
      vector<size_t> idx(ptg.shape(0));
      constexpr size_t cellsize=16;
      size_t nct = ntheta/cellsize+1,
             ncp = nphi/cellsize+1;
      vector<vector<size_t>> mapper(nct*ncp);
      for (size_t i=0; i<ptg.shape(0); ++i)
        {
        size_t itheta=min(nct-1,size_t(ptg(i,0)/pi*nct)),
               iphi=min(ncp-1,size_t(ptg(i,1)/(2*pi)*ncp));
195
//        MR_assert((itheta<nct)&&(iphi<ncp), "oops");
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
196
197
198
199
200
201
202
203
204
        mapper[itheta*ncp+iphi].push_back(i);
        }
      size_t cnt=0;
      for (const auto &vec: mapper)
        for (auto i:vec)
          idx[cnt++] = i;
      return idx;
      }

205
  public:
206
207
    Interpolator(const vector<Alm<complex<T>>> &slm,
                 const vector<Alm<complex<T>>> &blm,
208
                 bool separate, T epsilon, T ofmin, int nthreads_)
209
      : adjoint(false),
210
211
        lmax(slm.at(0).Lmax()),
        kmax(blm.at(0).Mmax()),
Martin Reinecke's avatar
Martin Reinecke committed
212
213
        nphi0(2*good_size_real(lmax+1)),
        ntheta0(nphi0/2+1),
Martin Reinecke's avatar
Martin Reinecke committed
214
        nphi(std::max<size_t>(20,2*good_size_real(size_t((2*lmax+1)*ofmin/2.)))),
Martin Reinecke's avatar
Martin Reinecke committed
215
        ntheta(nphi/2+1),
Martin Reinecke's avatar
Martin Reinecke committed
216
        nthreads(nthreads_),
217
        ofactor(T(nphi)/(2*lmax+1)),
Martin Reinecke's avatar
fix    
Martin Reinecke committed
218
        supp(ES_Kernel::get_supp(epsilon, ofactor)),
Martin Reinecke's avatar
Martin Reinecke committed
219
        kernel(supp, ofactor, nthreads),
220
221
        ncomp(separate ? slm.size() : 1),
        cube({ntheta+2*supp, nphi+2*supp, 2*kmax+1, ncomp})
222
      {
223
      MR_assert((ncomp==1)||(ncomp==3), "currently only 1 or 3 components allowed");
224
225
226
227
228
229
230
231
232
      MR_assert(slm.size()==blm.size(), "inconsistent slm and blm vectors");
      for (size_t i=0; i<slm.size(); ++i)
        {
        MR_assert(slm[i].Lmax()==lmax, "inconsistent Sky lmax");
        MR_assert(slm[i].Mmax()==lmax, "Sky lmax must be equal to Sky mmax");
        MR_assert(blm[i].Lmax()==lmax, "Sky and beam lmax must be equal");
        MR_assert(blm[i].Mmax()==kmax, "Inconcistent beam mmax");
        }

233
234
      MR_assert((supp<=ntheta) && (supp<=nphi), "support too large!");
      Alm<complex<T>> a1(lmax, lmax), a2(lmax,lmax);
Martin Reinecke's avatar
Martin Reinecke committed
235
      auto ginfo = sharp_make_cc_geom_info(ntheta0,nphi0,0.,cube.stride(1),cube.stride(0));
236
237
      auto ainfo = sharp_make_triangular_alm_info(lmax,lmax,1);

238
      vector<T>lnorm(lmax+1);
239
      for (size_t i=0; i<=lmax; ++i)
Martin Reinecke's avatar
Martin Reinecke committed
240
        lnorm[i]=std::sqrt(4*pi/(2*i+1.));
241

242
      for (size_t icomp=0; icomp<ncomp; ++icomp)
243
244
245
246
        {
        for (size_t m=0; m<=lmax; ++m)
          for (size_t l=m; l<=lmax; ++l)
            {
247
248
            if (separate)
              a1(l,m) = slm[icomp](l,m)*blm[icomp](l,0).real()*T(lnorm[l]);
249
250
            else
              {
251
252
253
              a1(l,m) = 0;
              for (size_t j=0; j<slm.size(); ++j)
                a1(l,m) += slm[j](l,m)*blm[j](l,0).real()*T(lnorm[l]);
254
255
              }
            }
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
        auto m1 = cube.template subarray<2>({supp,supp,0,icomp},{ntheta,nphi,0,0});
        sharp_alm2map(a1.Alms().data(), m1.vdata(), *ginfo, *ainfo, 0, nthreads);
        correct(m1,0);

        for (size_t k=1; k<=kmax; ++k)
          {
          for (size_t m=0; m<=lmax; ++m)
            for (size_t l=m; l<=lmax; ++l)
              {
              if (l<k)
                a1(l,m)=a2(l,m)=0.;
              else
                {
                if (separate)
                  {
271
                  auto tmp = blm[icomp](l,k)*T(-2*lnorm[l]);
272
273
274
275
276
277
278
279
                  a1(l,m) = slm[icomp](l,m)*tmp.real();
                  a2(l,m) = slm[icomp](l,m)*tmp.imag();
                  }
                else
                  {
                  a1(l,m) = a2(l,m) = 0;
                  for (size_t j=0; j<slm.size(); ++j)
                    {
280
                    auto tmp = blm[j](l,k)*T(-2*lnorm[l]);
281
282
283
284
                    a1(l,m) += slm[j](l,m)*tmp.real();
                    a2(l,m) += slm[j](l,m)*tmp.imag();
                    }
                  }
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
285
                }
286
              }
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
287
288
289
290
291
292
          auto m1 = cube.template subarray<2>({supp,supp,2*k-1,icomp},{ntheta,nphi,0,0});
          auto m2 = cube.template subarray<2>({supp,supp,2*k  ,icomp},{ntheta,nphi,0,0});
          sharp_alm2map_spin(k, a1.Alms().data(), a2.Alms().data(), m1.vdata(),
            m2.vdata(), *ginfo, *ainfo, 0, nthreads);
          correct(m1,k);
          correct(m2,k);
293
          }
294
        }
295

296
297
298
299
300
      // fill border regions
      for (size_t i=0; i<supp; ++i)
        for (size_t j=0, j2=nphi/2; j<nphi; ++j,++j2)
          for (size_t k=0; k<cube.shape(2); ++k)
            {
301
            T fct = (((k+1)/2)&1) ? -1 : 1;
302
            if (j2>=nphi) j2-=nphi;
303
304
305
306
307
            for (size_t l=0; l<cube.shape(3); ++l)
              {
              cube.v(supp-1-i,j2+supp,k,l) = fct*cube(supp+1+i,j+supp,k,l);
              cube.v(supp+ntheta+i,j2+supp,k,l) = fct*cube(supp+ntheta-2-i,j+supp,k,l);
              }
308
309
310
311
            }
      for (size_t i=0; i<ntheta+2*supp; ++i)
        for (size_t j=0; j<supp; ++j)
          for (size_t k=0; k<cube.shape(2); ++k)
312
            for (size_t l=0; l<cube.shape(3); ++l)
313
            {
314
315
            cube.v(i,j,k,l) = cube(i,j+nphi,k,l);
            cube.v(i,j+nphi+supp,k,l) = cube(i,j+supp,k,l);
316
317
318
            }
      }

319
    Interpolator(size_t lmax_, size_t kmax_, size_t ncomp_, T epsilon, T ofmin, int nthreads_)
320
321
322
323
324
      : adjoint(true),
        lmax(lmax_),
        kmax(kmax_),
        nphi0(2*good_size_real(lmax+1)),
        ntheta0(nphi0/2+1),
Martin Reinecke's avatar
Martin Reinecke committed
325
        nphi(std::max<size_t>(20,2*good_size_real(size_t((2*lmax+1)*ofmin/2.)))),
326
327
        ntheta(nphi/2+1),
        nthreads(nthreads_),
328
        ofactor(T(nphi)/(2*lmax+1)),
329
330
        supp(ES_Kernel::get_supp(epsilon, ofactor)),
        kernel(supp, ofactor, nthreads),
331
332
        ncomp(ncomp_),
        cube({ntheta+2*supp, nphi+2*supp, 2*kmax+1, ncomp_})
333
      {
334
      MR_assert((ncomp==1)||(ncomp==3), "currently only 1 or 3 components allowed");
335
336
337
338
      MR_assert((supp<=ntheta) && (supp<=nphi), "support too large!");
      cube.apply([](T &v){v=0.;});
      }

339
    void interpol (const mav<T,2> &ptg, mav<T,2> &res) const
340
      {
341
      MR_assert(!adjoint, "cannot be called in adjoint mode");
342
343
      MR_assert(ptg.shape(0)==res.shape(0), "dimension mismatch");
      MR_assert(ptg.shape(1)==3, "second dimension must have length 3");
344
      MR_assert(res.shape(1)==ncomp, "# of components mismatch");
345
346
347
      T delta = T(2)/supp;
      T xdtheta = T((ntheta-1)/pi),
        xdphi = T(nphi/(2*pi));
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
348
      auto idx = getIdx(ptg);
Martin Reinecke's avatar
Martin Reinecke committed
349
      execStatic(idx.size(), nthreads, 0, [&](Scheduler &sched)
350
        {
Martin Reinecke's avatar
Martin Reinecke committed
351
352
353
        vector<T> wt(supp), wp(supp);
        vector<T> psiarr(2*kmax+1);
        while (auto rng=sched.getNext()) for(auto ind=rng.lo; ind<rng.hi; ++ind)
354
          {
Martin Reinecke's avatar
Martin Reinecke committed
355
          size_t i=idx[ind];
356
357
          T f0=T(0.5*supp+ptg(i,0)*xdtheta);
          size_t i0 = size_t(f0+T(1));
Martin Reinecke's avatar
Martin Reinecke committed
358
359
          for (size_t t=0; t<supp; ++t)
            wt[t] = kernel((t+i0-f0)*delta - 1);
360
          T f1=T(0.5)*supp+ptg(i,1)*xdphi;
Martin Reinecke's avatar
Martin Reinecke committed
361
362
363
364
          size_t i1 = size_t(f1+1.);
          for (size_t t=0; t<supp; ++t)
            wp[t] = kernel((t+i1-f1)*delta - 1);
          psiarr[0]=1.;
365
366
          double psi=ptg(i,2);
          double cpsi=cos(psi), spsi=sin(psi);
Martin Reinecke's avatar
Martin Reinecke committed
367
368
369
          double cnpsi=cpsi, snpsi=spsi;
          for (size_t l=1; l<=kmax; ++l)
            {
370
371
            psiarr[2*l-1]=T(cnpsi);
            psiarr[2*l]=T(snpsi);
Martin Reinecke's avatar
Martin Reinecke committed
372
373
374
375
            const double tmp = snpsi*cpsi + cnpsi*spsi;
            cnpsi=cnpsi*cpsi - snpsi*spsi;
            snpsi=tmp;
            }
376
377
          if (ncomp==1)
            {
378
            T vv=0;
379
380
381
            for (size_t j=0; j<supp; ++j)
              for (size_t k=0; k<supp; ++k)
                for (size_t l=0; l<2*kmax+1; ++l)
382
                  vv += cube(i0+j,i1+k,l,0)*wt[j]*wp[k]*psiarr[l];
383
384
385
386
            res.v(i,0) = vv;
            }
          else // ncomp==3
            {
387
            T v0=0., v1=0., v2=0.;
388
389
390
391
            for (size_t j=0; j<supp; ++j)
              for (size_t k=0; k<supp; ++k)
                for (size_t l=0; l<2*kmax+1; ++l)
                  {
392
                  auto tmp = wt[j]*wp[k]*psiarr[l];
393
394
395
396
397
398
399
400
                  v0 += cube(i0+j,i1+k,l,0)*tmp;
                  v1 += cube(i0+j,i1+k,l,1)*tmp;
                  v2 += cube(i0+j,i1+k,l,2)*tmp;
                  }
            res.v(i,0) = v0;
            res.v(i,1) = v1;
            res.v(i,2) = v2;
            }
401
          }
Martin Reinecke's avatar
Martin Reinecke committed
402
        });
403
      }
404

405
406
407
    size_t support() const
      { return supp; }

408
    void deinterpol (const mav<T,2> &ptg, const mav<T,2> &data)
409
410
411
412
      {
      MR_assert(adjoint, "can only be called in adjoint mode");
      MR_assert(ptg.shape(0)==data.shape(0), "dimension mismatch");
      MR_assert(ptg.shape(1)==3, "second dimension must have length 3");
413
      MR_assert(data.shape(1)==ncomp, "# of components mismatch");
414
415
416
      T delta = T(2)/supp;
      T xdtheta = T((ntheta-1)/pi),
        xdphi = T(nphi/(2*pi));
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
417
      auto idx = getIdx(ptg);
418
419
420
421
422
423
424

      constexpr size_t cellsize=16;
      size_t nct = ntheta/cellsize+5,
             ncp = nphi/cellsize+5;
      mav<std::mutex,2> locks({nct,ncp});

      execStatic(idx.size(), nthreads, 0, [&](Scheduler &sched)
425
        {
426
        size_t b_theta=99999999999999, b_phi=9999999999999999;
427
428
429
430
431
        vector<T> wt(supp), wp(supp);
        vector<T> psiarr(2*kmax+1);
        while (auto rng=sched.getNext()) for(auto ind=rng.lo; ind<rng.hi; ++ind)
          {
          size_t i=idx[ind];
432
          T f0=0.5*supp+ptg(i,0)*xdtheta;
433
434
435
          size_t i0 = size_t(f0+1.);
          for (size_t t=0; t<supp; ++t)
            wt[t] = kernel((t+i0-f0)*delta - 1);
436
          T f1=0.5*supp+ptg(i,1)*xdphi;
437
438
439
440
441
442
443
444
445
          size_t i1 = size_t(f1+1.);
          for (size_t t=0; t<supp; ++t)
            wp[t] = kernel((t+i1-f1)*delta - 1);
          psiarr[0]=1.;
          double psi=ptg(i,2);
          double cpsi=cos(psi), spsi=sin(psi);
          double cnpsi=cpsi, snpsi=spsi;
          for (size_t l=1; l<=kmax; ++l)
            {
446
447
            psiarr[2*l-1]=T(cnpsi);
            psiarr[2*l]=T(snpsi);
448
449
450
451
            const double tmp = snpsi*cpsi + cnpsi*spsi;
            cnpsi=cnpsi*cpsi - snpsi*spsi;
            snpsi=tmp;
            }
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
          size_t b_theta_new = i0/cellsize,
                 b_phi_new = i1/cellsize;
          if ((b_theta_new!=b_theta) || (b_phi_new!=b_phi))
            {
            if (b_theta<locks.shape(0))  // unlock
              {
              locks.v(b_theta,b_phi).unlock();
              locks.v(b_theta,b_phi+1).unlock();
              locks.v(b_theta+1,b_phi).unlock();
              locks.v(b_theta+1,b_phi+1).unlock();
              }
            b_theta = b_theta_new;
            b_phi = b_phi_new;
            locks.v(b_theta,b_phi).lock();
            locks.v(b_theta,b_phi+1).lock();
            locks.v(b_theta+1,b_phi).lock();
            locks.v(b_theta+1,b_phi+1).lock();
            }
470
471
          if (ncomp==1)
            {
472
            T val = data(i,0);
473
474
475
476
477
478
479
            for (size_t j=0; j<supp; ++j)
              for (size_t k=0; k<supp; ++k)
                for (size_t l=0; l<2*kmax+1; ++l)
                  cube.v(i0+j,i1+k,l,0) += val*wt[j]*wp[k]*psiarr[l];
            }
          else // ncomp==3
            {
480
            T v0=data(i,0), v1=data(i,1), v2=data(i,2);
481
482
483
            for (size_t j=0; j<supp; ++j)
              for (size_t k=0; k<supp; ++k)
                {
484
                T t0 = wt[j]*wp[k];
485
486
                for (size_t l=0; l<2*kmax+1; ++l)
                  {
487
                  T tmp = t0*psiarr[l];
488
489
490
491
492
493
                  cube.v(i0+j,i1+k,l,0) += v0*tmp;
                  cube.v(i0+j,i1+k,l,1) += v1*tmp;
                  cube.v(i0+j,i1+k,l,2) += v2*tmp;
                  }
                }
            }
494
          }
495
496
497
498
499
500
501
        if (b_theta<locks.shape(0))  // unlock
          {
          locks.v(b_theta,b_phi).unlock();
          locks.v(b_theta,b_phi+1).unlock();
          locks.v(b_theta+1,b_phi).unlock();
          locks.v(b_theta+1,b_phi+1).unlock();
          }
502
503
        });
      }
504
    void getSlm (const vector<Alm<complex<T>>> &blm, vector<Alm<complex<T>>> &slm)
505
506
      {
      MR_assert(adjoint, "can only be called in adjoint mode");
507
508
      MR_assert((blm.size()==ncomp) || (ncomp==1), "incorrect number of beam a_lm sets");
      MR_assert((slm.size()==ncomp) || (ncomp==1), "incorrect number of sky a_lm sets");
509
510
511
512
513
      Alm<complex<T>> a1(lmax, lmax), a2(lmax,lmax);
      auto ginfo = sharp_make_cc_geom_info(ntheta0,nphi0,0.,cube.stride(1),cube.stride(0));
      auto ainfo = sharp_make_triangular_alm_info(lmax,lmax,1);

      // move stuff from border regions onto the main grid
514
      for (size_t i=0; i<cube.shape(0); ++i)
515
516
        for (size_t j=0; j<supp; ++j)
          for (size_t k=0; k<cube.shape(2); ++k)
517
518
519
520
521
            for (size_t l=0; l<cube.shape(3); ++l)
              {
              cube.v(i,j+nphi,k,l) += cube(i,j,k,l);
              cube.v(i,j+supp,k,l) += cube(i,j+nphi+supp,k,l);
              }
522
523
524
525
      for (size_t i=0; i<supp; ++i)
        for (size_t j=0, j2=nphi/2; j<nphi; ++j,++j2)
          for (size_t k=0; k<cube.shape(2); ++k)
            {
526
            T fct = (((k+1)/2)&1) ? -1 : 1;
527
            if (j2>=nphi) j2-=nphi;
528
529
530
531
532
            for (size_t l=0; l<cube.shape(3); ++l)
              {
              cube.v(supp+1+i,j+supp,k,l) += fct*cube(supp-1-i,j2+supp,k,l);
              cube.v(supp+ntheta-2-i, j+supp,k,l) += fct*cube(supp+ntheta+i,j2+supp,k,l);
              }
533
            }
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
534
535
536
537

      // special treatment for poles
      for (size_t j=0,j2=nphi/2; j<nphi/2; ++j,++j2)
        for (size_t k=0; k<cube.shape(2); ++k)
538
539
          for (size_t l=0; l<cube.shape(3); ++l)
            {
540
            T fct = (((k+1)/2)&1) ? -1 : 1;
541
            if (j2>=nphi) j2-=nphi;
542
            T tval = (cube(supp,j+supp,k,l) + fct*cube(supp,j2+supp,k,l));
543
544
545
546
547
548
            cube.v(supp,j+supp,k,l) = tval;
            cube.v(supp,j2+supp,k,l) = fct*tval;
            tval = (cube(supp+ntheta-1,j+supp,k,l) + fct*cube(supp+ntheta-1,j2+supp,k,l));
            cube.v(supp+ntheta-1,j+supp,k,l) = tval;
            cube.v(supp+ntheta-1,j2+supp,k,l) = fct*tval;
            }
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
549

550
      vector<T>lnorm(lmax+1);
551
      for (size_t i=0; i<=lmax; ++i)
Martin Reinecke's avatar
Martin Reinecke committed
552
        lnorm[i]=std::sqrt(4*pi/(2*i+1.));
553

554
555
556
      for (size_t j=0; j<blm.size(); ++j)
        slm[j].SetToZero();

557
      for (size_t icomp=0; icomp<ncomp; ++icomp)
558
        {
559
        bool separate = ncomp>1;
560
        {
561
        auto m1 = cube.template subarray<2>({supp,supp,0,icomp},{ntheta,nphi,0,0});
562
563
        decorrect(m1,0);
        sharp_alm2map_adjoint(a1.Alms().vdata(), m1.data(), *ginfo, *ainfo, 0, nthreads);
564
565
        for (size_t m=0; m<=lmax; ++m)
          for (size_t l=m; l<=lmax; ++l)
566
567
568
569
570
            if (separate)
              slm[icomp](l,m) += conj(a1(l,m))*blm[icomp](l,0).real()*T(lnorm[l]);
            else
              for (size_t j=0; j<blm.size(); ++j)
                slm[j](l,m) += conj(a1(l,m))*blm[j](l,0).real()*T(lnorm[l]);
571
572
573
        }
        for (size_t k=1; k<=kmax; ++k)
          {
574
575
          auto m1 = cube.template subarray<2>({supp,supp,2*k-1,icomp},{ntheta,nphi,0,0});
          auto m2 = cube.template subarray<2>({supp,supp,2*k  ,icomp},{ntheta,nphi,0,0});
576
577
578
579
580
581
582
583
          decorrect(m1,k);
          decorrect(m2,k);

          sharp_alm2map_spin_adjoint(k, a1.Alms().vdata(), a2.Alms().vdata(), m1.data(),
            m2.data(), *ginfo, *ainfo, 0, nthreads);
          for (size_t m=0; m<=lmax; ++m)
            for (size_t l=m; l<=lmax; ++l)
              if (l>=k)
584
585
                {
                if (separate)
586
                  {
587
                  auto tmp = conj(blm[icomp](l,k))*T(-2*lnorm[l]);
588
589
                  slm[icomp](l,m) += conj(a1(l,m))*tmp.real();
                  slm[icomp](l,m) -= conj(a2(l,m))*tmp.imag();
590
                  }
591
592
593
                else
                  for (size_t j=0; j<blm.size(); ++j)
                    {
594
                    auto tmp = conj(blm[j](l,k))*T(-2*lnorm[l]);
595
596
597
598
                    slm[j](l,m) += conj(a1(l,m))*tmp.real();
                    slm[j](l,m) -= conj(a2(l,m))*tmp.imag();
                    }
                }
599
          }
600
601
        }
      }
602
603
  };

604
605
606
double epsilon_guess(size_t support, double ofactor)
  { return std::sqrt(12.)*std::exp(-(support*ofactor)); }

Martin Reinecke's avatar
Martin Reinecke committed
607
}
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
608

Martin Reinecke's avatar
Martin Reinecke committed
609
using detail_interpol_ng::Interpolator;
610
using detail_interpol_ng::epsilon_guess;
611

Martin Reinecke's avatar
Martin Reinecke committed
612
}
613

Martin Reinecke's avatar
Martin Reinecke committed
614
#endif