gridder_cxx.h 45.6 KB
Newer Older
Martin Reinecke's avatar
Martin Reinecke committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
#ifndef GRIDDER_CXX_H
#define GRIDDER_CXX_H

/*
 *  This file is part of nifty_gridder.
 *
 *  nifty_gridder is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License as published by
 *  the Free Software Foundation; either version 2 of the License, or
 *  (at your option) any later version.
 *
 *  nifty_gridder is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with nifty_gridder; if not, write to the Free Software
 *  Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
 */

/* Copyright (C) 2019 Max-Planck-Society
   Author: Martin Reinecke */

#include <iostream>
#include <algorithm>
#include <cstdlib>
#include <cmath>
#include <vector>

#include "pocketfft_hdronly.h"
#include <array>
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
33
#define NOINLINE __attribute__((noinline))
Martin Reinecke's avatar
Martin Reinecke committed
34

Martin Reinecke's avatar
Martin Reinecke committed
35
36
namespace gridder {

37
38
namespace detail {

Martin Reinecke's avatar
Martin Reinecke committed
39
40
41
using namespace std;

template<typename T, size_t ndim> class mav
Martin Reinecke's avatar
Martin Reinecke committed
42
43
44
45
46
  {
  static_assert((ndim>0) && (ndim<3), "only supports 1D and 2D arrays");

  private:
    T *d;
Martin Reinecke's avatar
Martin Reinecke committed
47
48
    array<size_t, ndim> shp;
    array<ptrdiff_t, ndim> str;
Martin Reinecke's avatar
Martin Reinecke committed
49
50

  public:
Martin Reinecke's avatar
Martin Reinecke committed
51
52
53
54
    mav(T *d_, const array<size_t,ndim> &shp_,
        const array<ptrdiff_t,ndim> &str_)
      : d(d_), shp(shp_), str(str_) {}
    mav(T *d_, const array<size_t,ndim> &shp_)
Martin Reinecke's avatar
Martin Reinecke committed
55
56
57
58
59
60
      : d(d_), shp(shp_)
      {
      str[ndim-1]=1;
      for (size_t d=2; d<=ndim; ++d)
        str[ndim-d] = str[ndim-d+1]*shp[ndim-d+1];
      }
Martin Reinecke's avatar
Martin Reinecke committed
61
62
    operator mav<const T, ndim>() const
      { return mav<const T, ndim>(d,shp,str); }
Martin Reinecke's avatar
Martin Reinecke committed
63
64
65
66
    T &operator[](size_t i)
      { return operator()(i); }
    const T &operator[](size_t i) const
      { return operator()(i); }
67
    T &operator()(size_t i) const
Martin Reinecke's avatar
Martin Reinecke committed
68
69
      {
      static_assert(ndim==1, "ndim must be 1");
Martin Reinecke's avatar
Martin Reinecke committed
70
      return d[str[0]*i];
Martin Reinecke's avatar
Martin Reinecke committed
71
      }
72
    T &operator()(size_t i, size_t j) const
Martin Reinecke's avatar
Martin Reinecke committed
73
74
      {
      static_assert(ndim==2, "ndim must be 2");
Martin Reinecke's avatar
Martin Reinecke committed
75
      return d[str[0]*i + str[1]*j];
Martin Reinecke's avatar
Martin Reinecke committed
76
77
      }
    size_t shape(size_t i) const { return shp[i]; }
Martin Reinecke's avatar
Martin Reinecke committed
78
    const array<size_t,ndim> &shape() const { return shp; }
Martin Reinecke's avatar
Martin Reinecke committed
79
80
81
82
83
84
    size_t size() const
      {
      size_t res=1;
      for (auto v: shp) res*=v;
      return res;
      }
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
85
    ptrdiff_t stride(size_t i) const { return str[i]; }
Martin Reinecke's avatar
Martin Reinecke committed
86
    T *data()
Martin Reinecke's avatar
Martin Reinecke committed
87
      { return d; }
Martin Reinecke's avatar
Martin Reinecke committed
88
    const T *data() const
Martin Reinecke's avatar
Martin Reinecke committed
89
      { return d; }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
90
91
92
93
94
95
96
97
98
99
    bool contiguous() const
      {
      ptrdiff_t stride=1;
      for (size_t i=0; i<ndim; ++i)
        {
        if (str[ndim-1-i]!=stride) return false;
        stride *= shp[ndim-1-i];
        }
      return true;
      }
Martin Reinecke's avatar
Martin Reinecke committed
100
101
  };

Martin Reinecke's avatar
updates    
Martin Reinecke committed
102
103
template<typename T, size_t ndim> using const_mav = mav<const T, ndim>;

Martin Reinecke's avatar
Martin Reinecke committed
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
template<typename T, size_t ndim> class tmpStorage
  {
  private:
    vector<T> d;
    mav<T,ndim> mav_;

    static size_t prod(const array<size_t,ndim> &shp)
      {
      size_t res=1;
      for (auto v: shp) res*=v;
      return res;
      }

  public:
    tmpStorage(const array<size_t,ndim> &shp)
      : d(prod(shp)), mav_(d.data(), shp) {}
    mav<T,ndim> &getMav() { return mav_; };
    void fill(const T & val)
      { std::fill(d.begin(), d.end(), val); }
  };
Martin Reinecke's avatar
Martin Reinecke committed
124
125
126
127
128
129
130
//
// basic utilities
//

void myassert(bool cond, const char *msg)
  {
  if (cond) return;
Martin Reinecke's avatar
Martin Reinecke committed
131
  throw runtime_error(msg);
Martin Reinecke's avatar
Martin Reinecke committed
132
133
134
  }

template<size_t ndim> void checkShape
Martin Reinecke's avatar
Martin Reinecke committed
135
  (const array<size_t, ndim> &shp1, const array<size_t, ndim> &shp2)
Martin Reinecke's avatar
Martin Reinecke committed
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
  {
  for (size_t i=0; i<ndim; ++i)
    myassert(shp1[i]==shp2[i], "shape mismatch");
  }

/*! Returns the remainder of the division \a v1/v2.
    The result is non-negative.
    \a v1 can be positive or negative; \a v2 must be positive. */
template<typename T> inline T fmodulo (T v1, T v2)
  {
  if (v1>=0)
    return (v1<v2) ? v1 : fmod(v1,v2);
  T tmp=fmod(v1,v2)+v2;
  return (tmp==v2) ? T(0) : tmp;
  }

//
// Utilities for Gauss-Legendre quadrature
//

static inline double one_minus_x2 (double x)
  { return (fabs(x)>0.1) ? (1.+x)*(1.-x) : 1.-x*x; }

Martin Reinecke's avatar
Martin Reinecke committed
159
void legendre_prep(int n, vector<double> &x, vector<double> &w, size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
  {
  constexpr double pi = 3.141592653589793238462643383279502884197;
  constexpr double eps = 3e-14;
  int m = (n+1)>>1;
  x.resize(m);
  w.resize(m);

  double t0 = 1 - (1-1./n) / (8.*n*n);
  double t1 = 1./(4.*n+2.);

#pragma omp parallel num_threads(nthreads)
{
  int i;
#pragma omp for schedule(dynamic,100)
  for (i=1; i<=m; ++i)
    {
    double x0 = cos(pi * ((i<<2)-1) * t1) * t0;

    int dobreak=0;
    int j=0;
    double dpdx;
    while(1)
      {
      double P_1 = 1.0;
      double P0 = x0;
      double dx, x1;

      for (int k=2; k<=n; k++)
        {
        double P_2 = P_1;
        P_1 = P0;
//        P0 = ((2*k-1)*x0*P_1-(k-1)*P_2)/k;
        P0 = x0*P_1 + (k-1.)/k * (x0*P_1-P_2);
        }

      dpdx = (P_1 - x0*P0) * n / one_minus_x2(x0);

      /* Newton step */
      x1 = x0 - P0/dpdx;
      dx = x0-x1;
      x0 = x1;
      if (dobreak) break;

      if (abs(dx)<=eps) dobreak=1;
      myassert(++j<100, "convergence problem");
      }

    x[m-i] = x0;
    w[m-i] = 2. / (one_minus_x2(x0) * dpdx * dpdx);
    }
} // end of parallel region
  }

//
// Start of real gridder functionality
//

size_t get_supp(double epsilon)
  {
Martin Reinecke's avatar
Martin Reinecke committed
219
  static const vector<double> maxmaperr { 1e8, 0.32, 0.021, 6.2e-4,
Martin Reinecke's avatar
Martin Reinecke committed
220
221
222
223
224
225
226
    1.08e-5, 1.25e-7, 8.25e-10, 5.70e-12, 1.22e-13, 2.48e-15, 4.82e-17,
    6.74e-19, 5.41e-21, 4.41e-23, 7.88e-25, 3.9e-26 };

  double epssq = epsilon*epsilon;

  for (size_t i=1; i<maxmaperr.size(); ++i)
    if (epssq>maxmaperr[i]) return i;
Martin Reinecke's avatar
Martin Reinecke committed
227
  throw runtime_error("requested epsilon too small - minimum is 2e-13");
Martin Reinecke's avatar
Martin Reinecke committed
228
229
  }

Martin Reinecke's avatar
Martin Reinecke committed
230
231
232
233
234
struct RowChan
  {
  size_t row, chan;
  };

Martin Reinecke's avatar
Martin Reinecke committed
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
template<typename T> class EC_Kernel
  {
  protected:
    size_t supp;
    T beta, emb_;

  public:
    EC_Kernel(size_t supp_)
      : supp(supp_), beta(2.3*supp_), emb_(exp(-beta)) {}
    T operator()(T v) const { return exp(beta*(sqrt(T(1)-v*v)-T(1))); }
    T val_no_emb(T v) const { return exp(beta*sqrt(T(1)-v*v)); }
    T emb() const { return emb_; }
  };

template<typename T> class EC_Kernel_with_correction: public EC_Kernel<T>
  {
  protected:
    static constexpr T pi = T(3.141592653589793238462643383279502884197L);
    int p;
254
    vector<double> x, wgt, psi;
Martin Reinecke's avatar
Martin Reinecke committed
255
256
257
258
    using EC_Kernel<T>::supp;

  public:
    using EC_Kernel<T>::operator();
Martin Reinecke's avatar
Martin Reinecke committed
259
    EC_Kernel_with_correction(size_t supp_, size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
260
261
      : EC_Kernel<T>(supp_), p(int(1.5*supp_+2))
      {
Martin Reinecke's avatar
Martin Reinecke committed
262
      legendre_prep(2*p,x,wgt,nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
263
264
265
266
267
268
269
270
271
272
273
      psi=x;
      for (auto &v:psi)
        v=operator()(v);
      }
    /* Compute correction factors for the ES gridding kernel
       This implementation follows eqs. (3.8) to (3.10) of Barnett et al. 2018 */
    T corfac(T v) const
      {
      T tmp=0;
      for (int i=0; i<p; ++i)
        tmp += wgt[i]*psi[i]*cos(pi*supp*v*x[i]);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
274
      return T(1./(supp*tmp));
Martin Reinecke's avatar
Martin Reinecke committed
275
276
277
278
      }
  };
/* Compute correction factors for the ES gridding kernel
   This implementation follows eqs. (3.8) to (3.10) of Barnett et al. 2018 */
Martin Reinecke's avatar
Martin Reinecke committed
279
280
vector<double> correction_factors (size_t n, size_t nval, size_t supp,
  size_t nthreads)
Martin Reinecke's avatar
Martin Reinecke committed
281
  {
Martin Reinecke's avatar
Martin Reinecke committed
282
  EC_Kernel_with_correction<double> kernel(supp, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
283
  vector<double> res(nval);
Martin Reinecke's avatar
Martin Reinecke committed
284
285
286
287
288
289
290
291
292
293
294
295
296
297
  double xn = 1./n;
#pragma omp parallel for schedule(static) num_threads(nthreads)
  for (size_t k=0; k<nval; ++k)
    res[k] = kernel.corfac(k*xn);
  return res;
  }

template<typename T> struct UVW
  {
  T u, v, w;
  UVW () {}
  UVW (T u_, T v_, T w_) : u(u_), v(v_), w(w_) {}
  UVW operator* (T fct) const
    { return UVW(u*fct, v*fct, w*fct); }
298
299
300
301
302
303
304
  void Flip() { u=-u; v=-v; w=-w; }
  bool FixW()
    {
    bool flip = w<0;
    if (flip) Flip();
    return flip;
    }
Martin Reinecke's avatar
Martin Reinecke committed
305
306
307
308
309
  };

template<typename T> class Baselines
  {
  protected:
Martin Reinecke's avatar
Martin Reinecke committed
310
311
    vector<UVW<T>> coord;
    vector<T> f_over_c;
Martin Reinecke's avatar
Martin Reinecke committed
312
313
314
    size_t nrows, nchan;

  public:
Martin Reinecke's avatar
updates    
Martin Reinecke committed
315
    Baselines(const const_mav<T,2> &coord_, const const_mav<T,1> &freq)
Martin Reinecke's avatar
Martin Reinecke committed
316
317
318
319
320
321
322
323
      {
      constexpr double speedOfLight = 299792458.;
      myassert(coord_.shape(1)==3, "dimension mismatch");
      nrows = coord_.shape(0);
      nchan = freq.shape(0);
      myassert(nrows*nchan<(size_t(1)<<32), "too many entries in MS");
      f_over_c.resize(nchan);
      for (size_t i=0; i<nchan; ++i)
324
325
        {
        myassert(freq[i]>0, "negative channel frequency encountered");
Martin Reinecke's avatar
Martin Reinecke committed
326
        f_over_c[i] = freq(i)/speedOfLight;
327
        }
Martin Reinecke's avatar
Martin Reinecke committed
328
329
330
331
332
      coord.resize(nrows);
      for (size_t i=0; i<coord.size(); ++i)
        coord[i] = UVW<T>(coord_(i,0), coord_(i,1), coord_(i,2));
      }

Martin Reinecke's avatar
Martin Reinecke committed
333
    RowChan getRowChan(uint32_t index) const
Martin Reinecke's avatar
Martin Reinecke committed
334
335
336
      {
      size_t irow = index/nchan;
      size_t ichan = index-nchan*irow;
Martin Reinecke's avatar
Martin Reinecke committed
337
      return RowChan{irow, ichan};
Martin Reinecke's avatar
Martin Reinecke committed
338
      }
Martin Reinecke's avatar
Martin Reinecke committed
339
340
341
342
343

    UVW<T> effectiveCoord(const RowChan &rc) const
      { return coord[rc.row]*f_over_c[rc.chan]; }
    UVW<T> effectiveCoord(uint32_t index) const
      { return effectiveCoord(getRowChan(index)); }
Martin Reinecke's avatar
Martin Reinecke committed
344
345
346
    size_t Nrows() const { return nrows; }
    size_t Nchannels() const { return nchan; }

Martin Reinecke's avatar
Martin Reinecke committed
347
    void effectiveUVW(const mav<const uint32_t,1> &idx, mav<T,2> &res) const
Martin Reinecke's avatar
Martin Reinecke committed
348
349
350
351
352
353
354
355
356
357
358
359
360
      {
      size_t nvis = idx.shape(0);
      myassert(res.shape(0)==nvis, "shape mismatch");
      myassert(res.shape(1)==3, "shape mismatch");
      for (size_t i=0; i<nvis; i++)
        {
        auto uvw = effectiveCoord(idx(i));
        res(i,0) = uvw.u;
        res(i,1) = uvw.v;
        res(i,2) = uvw.w;
        }
      }

Martin Reinecke's avatar
Martin Reinecke committed
361
    template<typename T2> void ms2vis(const mav<const T2,2> &ms,
Martin Reinecke's avatar
Martin Reinecke committed
362
      const mav<const uint32_t,1> &idx, mav<T2,1> &vis, size_t nthreads) const
Martin Reinecke's avatar
Martin Reinecke committed
363
364
365
366
367
368
369
370
      {
      myassert(ms.shape(0)==nrows, "shape mismatch");
      myassert(ms.shape(1)==nchan, "shape mismatch");
      size_t nvis = idx.shape(0);
      myassert(vis.shape(0)==nvis, "shape mismatch");
#pragma omp parallel for num_threads(nthreads)
      for (size_t i=0; i<nvis; ++i)
        {
Martin Reinecke's avatar
Martin Reinecke committed
371
372
        auto rc = getRowChan(idx(i));
        vis[i] = ms(rc.row, rc.chan);
Martin Reinecke's avatar
Martin Reinecke committed
373
374
375
        }
      }

Martin Reinecke's avatar
Martin Reinecke committed
376
    template<typename T2> void vis2ms(const mav<const T2,1> &vis,
Martin Reinecke's avatar
Martin Reinecke committed
377
      const mav<const uint32_t,1> &idx, mav<T2,2> &ms, size_t nthreads) const
Martin Reinecke's avatar
Martin Reinecke committed
378
379
380
381
382
383
384
385
      {
      size_t nvis = vis.shape(0);
      myassert(idx.shape(0)==nvis, "shape mismatch");
      myassert(ms.shape(0)==nrows, "shape mismatch");
      myassert(ms.shape(1)==nchan, "shape mismatch");
#pragma omp parallel for num_threads(nthreads)
      for (size_t i=0; i<nvis; ++i)
        {
Martin Reinecke's avatar
Martin Reinecke committed
386
387
        auto rc = getRowChan(idx(i));
        ms(rc.row, rc.chan) += vis(i);
Martin Reinecke's avatar
Martin Reinecke committed
388
389
390
391
392
393
394
395
        }
      }
  };

template<typename T> class GridderConfig
  {
  protected:
    size_t nx_dirty, ny_dirty;
396
    T eps, psx, psy;
Martin Reinecke's avatar
Martin Reinecke committed
397
398
    size_t supp, nsafe, nu, nv;
    T beta;
Martin Reinecke's avatar
Martin Reinecke committed
399
    vector<T> cfu, cfv;
Martin Reinecke's avatar
Martin Reinecke committed
400
    size_t nthreads;
Martin Reinecke's avatar
Martin Reinecke committed
401

Martin Reinecke's avatar
Martin Reinecke committed
402
    complex<T> wscreen(double x, double y, double w, bool adjoint) const
Martin Reinecke's avatar
Martin Reinecke committed
403
404
      {
      constexpr double pi = 3.141592653589793238462643383279502884197;
405
406
407
      double tmp = 1-x-y;
      if (tmp<0) return 0.;
      double nm1 = (-x-y)/(sqrt(tmp)+1); // more accurate form of sqrt(1-x-y)
Martin Reinecke's avatar
Martin Reinecke committed
408
409
410
411
412
413
414
415
      double n = nm1+1., xn = 1./n;
      double phase = 2*pi*w*nm1;
      if (adjoint) phase *= -1;
      return complex<T>(cos(phase)*xn, sin(phase)*xn);
      }

  public:
    GridderConfig(size_t nxdirty, size_t nydirty, double epsilon,
Martin Reinecke's avatar
Martin Reinecke committed
416
      double pixsize_x, double pixsize_y, size_t nthreads_)
Martin Reinecke's avatar
Martin Reinecke committed
417
418
419
      : nx_dirty(nxdirty), ny_dirty(nydirty), eps(epsilon),
        psx(pixsize_x), psy(pixsize_y),
        supp(get_supp(epsilon)), nsafe((supp+1)/2),
Martin Reinecke's avatar
Martin Reinecke committed
420
        nu(max(2*nsafe,2*nx_dirty)), nv(max(2*nsafe,2*ny_dirty)),
Martin Reinecke's avatar
Martin Reinecke committed
421
        beta(2.3*supp),
Martin Reinecke's avatar
Martin Reinecke committed
422
        cfu(nx_dirty), cfv(ny_dirty), nthreads(nthreads_)
Martin Reinecke's avatar
Martin Reinecke committed
423
424
425
426
427
428
429
      {
      myassert((nx_dirty&1)==0, "nx_dirty must be even");
      myassert((ny_dirty&1)==0, "ny_dirty must be even");
      myassert(epsilon>0, "epsilon must be positive");
      myassert(pixsize_x>0, "pixsize_x must be positive");
      myassert(pixsize_y>0, "pixsize_y must be positive");

Martin Reinecke's avatar
Martin Reinecke committed
430
      auto tmp = correction_factors(nu, nx_dirty/2+1, supp, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
431
432
433
434
      cfu[nx_dirty/2]=tmp[0];
      cfu[0]=tmp[nx_dirty/2];
      for (size_t i=1; i<nx_dirty/2; ++i)
        cfu[nx_dirty/2-i] = cfu[nx_dirty/2+i] = tmp[i];
Martin Reinecke's avatar
Martin Reinecke committed
435
      tmp = correction_factors(nv, ny_dirty/2+1, supp, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
      cfv[ny_dirty/2]=tmp[0];
      cfv[0]=tmp[ny_dirty/2];
      for (size_t i=1; i<ny_dirty/2; ++i)
        cfv[ny_dirty/2-i] = cfv[ny_dirty/2+i] = tmp[i];
      }
    size_t Nxdirty() const { return nx_dirty; }
    size_t Nydirty() const { return ny_dirty; }
    double Epsilon() const { return eps; }
    double Pixsize_x() const { return psx; }
    double Pixsize_y() const { return psy; }
    size_t Nu() const { return nu; }
    size_t Nv() const { return nv; }
    size_t Supp() const { return supp; }
    size_t Nsafe() const { return nsafe; }
    T Beta() const { return beta; }
Martin Reinecke's avatar
Martin Reinecke committed
451
    size_t Nthreads() const { return nthreads; }
Martin Reinecke's avatar
Martin Reinecke committed
452

Martin Reinecke's avatar
Martin Reinecke committed
453
    void apply_taper(const mav<const T,2> &img, mav<T,2> &img2, bool divide) const
Martin Reinecke's avatar
Martin Reinecke committed
454
      {
Martin Reinecke's avatar
Martin Reinecke committed
455
456
      checkShape(img.shape(), {nx_dirty, ny_dirty});
      checkShape(img2.shape(), {nx_dirty, ny_dirty});
Martin Reinecke's avatar
Martin Reinecke committed
457
458
459
      if (divide)
        for (size_t i=0; i<nx_dirty; ++i)
          for (size_t j=0; j<ny_dirty; ++j)
Martin Reinecke's avatar
Martin Reinecke committed
460
            img2(i,j) = img(i,j)/(cfu[i]*cfv[j]);
Martin Reinecke's avatar
Martin Reinecke committed
461
462
463
      else
        for (size_t i=0; i<nx_dirty; ++i)
          for (size_t j=0; j<ny_dirty; ++j)
Martin Reinecke's avatar
Martin Reinecke committed
464
            img2(i,j) = img(i,j)*cfu[i]*cfv[j];
Martin Reinecke's avatar
Martin Reinecke committed
465
466
      }

Martin Reinecke's avatar
Martin Reinecke committed
467
    void grid2dirty_c(const mav<const complex<T>,2> &grid, mav<complex<T>,2> &dirty) const
Martin Reinecke's avatar
Martin Reinecke committed
468
      {
Martin Reinecke's avatar
Martin Reinecke committed
469
470
471
      checkShape(grid.shape(), {nu,nv});
      checkShape(dirty.shape(), {nx_dirty,ny_dirty});
      vector<complex<T>> tmpdat(nu*nv);
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
472
473
      auto tmp=mav<complex<T>,2>(tmpdat.data(),{nu,nv},{ptrdiff_t(nv),ptrdiff_t(1)});
      constexpr auto sc = ptrdiff_t(sizeof(complex<T>));
Martin Reinecke's avatar
Martin Reinecke committed
474
475
476
      pocketfft::c2c({nu,nv},{grid.stride(0)*sc,grid.stride(1)*sc},
        {tmp.stride(0)*sc, tmp.stride(1)*sc}, {0,1}, pocketfft::BACKWARD,
        grid.data(), tmp.data(), T(1), nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
477
478
479
480
481
482
483
      for (size_t i=0; i<nx_dirty; ++i)
        for (size_t j=0; j<ny_dirty; ++j)
          {
          size_t i2 = nu-nx_dirty/2+i;
          if (i2>=nu) i2-=nu;
          size_t j2 = nv-ny_dirty/2+j;
          if (j2>=nv) j2-=nv;
Martin Reinecke's avatar
Martin Reinecke committed
484
          dirty(i,j) = tmp(i2,j2)*cfu[i]*cfv[j];
Martin Reinecke's avatar
Martin Reinecke committed
485
486
          }
      }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
487
488
489

    void dirty2grid_c(const const_mav<complex<T>,2> &dirty,
      mav<complex<T>,2> &grid) const
Martin Reinecke's avatar
Martin Reinecke committed
490
      {
Martin Reinecke's avatar
updates    
Martin Reinecke committed
491
492
493
494
495
      checkShape(dirty.shape(), {nx_dirty, ny_dirty});
      checkShape(grid.shape(), {nu, nv});
      for (size_t i=0; i<nu; ++i)
        for (size_t j=0; j<nv; ++j)
          grid(i,j)=0.;
Martin Reinecke's avatar
Martin Reinecke committed
496
497
498
499
500
501
502
      for (size_t i=0; i<nx_dirty; ++i)
        for (size_t j=0; j<ny_dirty; ++j)
          {
          size_t i2 = nu-nx_dirty/2+i;
          if (i2>=nu) i2-=nu;
          size_t j2 = nv-ny_dirty/2+j;
          if (j2>=nv) j2-=nv;
Martin Reinecke's avatar
updates    
Martin Reinecke committed
503
          grid(i2,j2) = dirty(i,j)*cfu[i]*cfv[j];
Martin Reinecke's avatar
Martin Reinecke committed
504
          }
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
505
      constexpr auto sc = ptrdiff_t(sizeof(complex<T>));
Martin Reinecke's avatar
updates    
Martin Reinecke committed
506
      pocketfft::stride_t strides{grid.stride(0)*sc,grid.stride(1)*sc};
Martin Reinecke's avatar
Martin Reinecke committed
507
      pocketfft::c2c({nu,nv}, strides, strides, {0,1}, pocketfft::FORWARD,
Martin Reinecke's avatar
updates    
Martin Reinecke committed
508
        grid.data(), grid.data(), T(1), nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
509
      }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
510

Martin Reinecke's avatar
Martin Reinecke committed
511
512
513
514
515
516
517
518
519
    inline void getpix(T u_in, T v_in, T &u, T &v, int &iu0, int &iv0) const
      {
      u=fmodulo(u_in*psx, T(1))*nu,
      iu0 = int(u-supp*0.5 + 1 + nu) - nu;
      if (iu0+supp>nu+nsafe) iu0 = nu+nsafe-supp;
      v=fmodulo(v_in*psy, T(1))*nv;
      iv0 = int(v-supp*0.5 + 1 + nv) - nv;
      if (iv0+supp>nv+nsafe) iv0 = nv+nsafe-supp;
      }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
520
521
522

    void apply_wscreen(const const_mav<complex<T>,2> &dirty,
      mav<complex<T>,2> &dirty2, double w, bool adjoint) const
Martin Reinecke's avatar
Martin Reinecke committed
523
      {
Martin Reinecke's avatar
updates    
Martin Reinecke committed
524
525
      checkShape(dirty.shape(), {nx_dirty, ny_dirty});
      checkShape(dirty2.shape(), {nx_dirty, ny_dirty});
Martin Reinecke's avatar
Martin Reinecke committed
526
527
      double x0 = -0.5*nx_dirty*psx,
             y0 = -0.5*ny_dirty*psy;
Martin Reinecke's avatar
updates    
Martin Reinecke committed
528
#pragma omp parallel for num_threads(nthreads) schedule(static)
Martin Reinecke's avatar
Martin Reinecke committed
529
530
531
532
533
534
535
536
      for (size_t i=0; i<=nx_dirty/2; ++i)
        {
        double fx = x0+i*psx;
        fx *= fx;
        for (size_t j=0; j<=ny_dirty/2; ++j)
          {
          double fy = y0+j*psy;
          auto ws = wscreen(fx, fy*fy, w, adjoint);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
537
          dirty2(i,j) = dirty(i,j)*ws; // lower left
Martin Reinecke's avatar
Martin Reinecke committed
538
539
540
          size_t i2 = nx_dirty-i, j2 = ny_dirty-j;
          if ((i>0)&&(i<i2))
            {
Martin Reinecke's avatar
updates    
Martin Reinecke committed
541
            dirty2(i2,j) = dirty(i2,j)*ws; // lower right
Martin Reinecke's avatar
Martin Reinecke committed
542
            if ((j>0)&&(j<j2))
Martin Reinecke's avatar
updates    
Martin Reinecke committed
543
              dirty2(i2,j2) = dirty(i2,j2)*ws; // upper right
Martin Reinecke's avatar
Martin Reinecke committed
544
545
            }
          if ((j>0)&&(j<j2))
Martin Reinecke's avatar
updates    
Martin Reinecke committed
546
            dirty2(i,j2) = dirty(i,j2)*ws; // upper left
Martin Reinecke's avatar
Martin Reinecke committed
547
548
549
          }
        }
      }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
550

Martin Reinecke's avatar
Martin Reinecke committed
551
552
553
554
  };

constexpr int logsquare=4;

Martin Reinecke's avatar
Martin Reinecke committed
555
template<typename T, typename T2=complex<T>> class Helper
Martin Reinecke's avatar
Martin Reinecke committed
556
557
558
559
560
561
562
563
564
565
566
  {
  private:
    const GridderConfig<T> &gconf;
    int nu, nv, nsafe, supp;
    T beta;
    const T2 *grid_r;
    T2 *grid_w;
    int su, sv;
    int iu0, iv0; // start index of the current visibility
    int bu0, bv0; // start index of the current buffer

Martin Reinecke's avatar
Martin Reinecke committed
567
    vector<T2> rbuf, wbuf;
Martin Reinecke's avatar
Martin Reinecke committed
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608

    void dump() const
      {
      if (bu0<-nsafe) return; // nothing written into buffer yet

#pragma omp critical (gridder_writing_to_grid)
{
      int idxu = (bu0+nu)%nu;
      int idxv0 = (bv0+nv)%nv;
      for (int iu=0; iu<su; ++iu)
        {
        int idxv = idxv0;
        for (int iv=0; iv<sv; ++iv)
          {
          grid_w[idxu*nv + idxv] += wbuf[iu*sv + iv];
          if (++idxv>=nv) idxv=0;
          }
        if (++idxu>=nu) idxu=0;
        }
}
      }

    void load()
      {
      int idxu = (bu0+nu)%nu;
      int idxv0 = (bv0+nv)%nv;
      for (int iu=0; iu<su; ++iu)
        {
        int idxv = idxv0;
        for (int iv=0; iv<sv; ++iv)
          {
          rbuf[iu*sv + iv] = grid_r[idxu*nv + idxv];
          if (++idxv>=nv) idxv=0;
          }
        if (++idxu>=nu) idxu=0;
        }
      }

  public:
    const T2 *p0r;
    T2 *p0w;
Martin Reinecke's avatar
Martin Reinecke committed
609
    vector<T> kernel;
Martin Reinecke's avatar
Martin Reinecke committed
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652

    Helper(const GridderConfig<T> &gconf_, const T2 *grid_r_, T2 *grid_w_)
      : gconf(gconf_), nu(gconf.Nu()), nv(gconf.Nv()), nsafe(gconf.Nsafe()),
        supp(gconf.Supp()), beta(gconf.Beta()), grid_r(grid_r_),
        grid_w(grid_w_), su(2*nsafe+(1<<logsquare)), sv(2*nsafe+(1<<logsquare)),
        bu0(-1000000), bv0(-1000000),
        rbuf(su*sv*(grid_r!=nullptr),T(0)),
        wbuf(su*sv*(grid_w!=nullptr),T(0)),
        kernel(2*supp)
      {}
    ~Helper() { if (grid_w) dump(); }

    int lineJump() const { return sv; }

    void prep(T u_in, T v_in)
      {
      T u, v;
      gconf.getpix(u_in, v_in, u, v, iu0, iv0);
      T xsupp=T(2)/supp;
      auto x0 = xsupp*(iu0-u);
      auto y0 = xsupp*(iv0-v);
      for (int i=0; i<supp; ++i)
        {
        auto x = x0+i*xsupp;
        kernel[i  ] = beta*sqrt(T(1)-x*x);
        auto y = y0+i*xsupp;
        kernel[i+supp] = beta*sqrt(T(1)-y*y);
        }
      for (auto &k : kernel)
        k = exp(k);

      if ((iu0<bu0) || (iv0<bv0) || (iu0+supp>bu0+su) || (iv0+supp>bv0+sv))
        {
        if (grid_w) { dump(); fill(wbuf.begin(), wbuf.end(), T(0)); }
        bu0=((((iu0+nsafe)>>logsquare)<<logsquare))-nsafe;
        bv0=((((iv0+nsafe)>>logsquare)<<logsquare))-nsafe;
        if (grid_r) load();
        }
      p0r = grid_r ? rbuf.data() + sv*(iu0-bu0) + iv0-bv0 : nullptr;
      p0w = grid_w ? wbuf.data() + sv*(iu0-bu0) + iv0-bv0 : nullptr;
      }
  };

653
template<class T, class T2, class T3> class VisServ
Martin Reinecke's avatar
updates    
Martin Reinecke committed
654
  {
655
656
657
658
659
660
661
  private:
    const Baselines<T> &baselines;
    const const_mav<uint32_t,1> &idx;
    T2 vis;
    const T3 &wgt;
    size_t nvis;
    bool have_wgt;
Martin Reinecke's avatar
updates    
Martin Reinecke committed
662

663
664
665
666
667
668
669
670
671
672
673
  public:
    VisServ(const Baselines<T> &baselines_,
      const const_mav<uint32_t,1> &idx_, T2 vis_, const T3 &wgt_)
      : baselines(baselines_), idx(idx_), vis(vis_), wgt(wgt_)
      {
      nvis = vis.shape(0);
      checkShape(idx.shape(), {nvis});
      have_wgt = wgt.size()!=0;
      if (have_wgt) checkShape(wgt.shape(), {nvis});
      }
    size_t Nvis() const { return nvis; }
674
    const Baselines<T> &getBaselines() const { return baselines; }
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
    UVW<T> getCoord(size_t i) const
      { return baselines.effectiveCoord(idx[i]); }
    complex<T> getVis(size_t i) const
      { return have_wgt ? vis(i)*wgt(i) : vis(i); }
    uint32_t getIdx(size_t i) const { return idx[i]; }
    void setVis (size_t i, const complex<T> &v) const
      { vis(i) = have_wgt ? v*wgt(i) : v; }
    void addVis (size_t i, const complex<T> &v) const
      { vis(i) += have_wgt ? v*wgt(i) : v; }
  };
template<class T, class T2, class T3> class MsServ
  {
  private:
    const Baselines<T> &baselines;
    const const_mav<uint32_t,1> &idx;
    T2 ms;
    const T3 &wgt;
    size_t nvis;
    bool have_wgt;
Martin Reinecke's avatar
updates    
Martin Reinecke committed
694

695
696
697
698
  public:
    MsServ(const Baselines<T> &baselines_,
    const const_mav<uint32_t,1> &idx_, T2 ms_, const T3 &wgt_)
      : baselines(baselines_), idx(idx_), ms(ms_), wgt(wgt_)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
699
      {
700
701
702
703
704
705
      auto nrows = baselines.Nrows();
      auto nchan = baselines.Nchannels();
      nvis = idx.shape(0);
      checkShape(ms.shape(), {nrows, nchan});
      have_wgt = wgt.size()!=0;
      if (have_wgt) checkShape(wgt.shape(), {nrows, nchan});
Martin Reinecke's avatar
updates    
Martin Reinecke committed
706
      }
707
    size_t Nvis() const { return nvis; }
708
    const Baselines<T> &getBaselines() const { return baselines; }
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
    UVW<T> getCoord(size_t i) const
      { return baselines.effectiveCoord(idx(i)); }
    complex<T> getVis(size_t i) const
      {
      auto rc = baselines.getRowChan(idx(i));
      return have_wgt ? ms(rc.row, rc.chan)*wgt(rc.row, rc.chan)
                      : ms(rc.row, rc.chan);
      }
    uint32_t getIdx(size_t i) const { return idx[i]; }
    void setVis (size_t i, const complex<T> &v) const
      {
      auto rc = baselines.getRowChan(idx(i));
      ms(rc.row, rc.chan) = have_wgt ? v*wgt(rc.row, rc.chan) : v;
      }
    void addVis (size_t i, const complex<T> &v) const
      {
      auto rc = baselines.getRowChan(idx(i));
      ms(rc.row, rc.chan) += have_wgt ? v*wgt(rc.row, rc.chan) : v;
      }
  };
Martin Reinecke's avatar
updates    
Martin Reinecke committed
729

730
731
template<typename T, typename Serv> void x2grid_c
  (const GridderConfig<T> &gconf, const Serv &srv, mav<complex<T>,2> &grid)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
732
733
734
735
736
737
  {
  size_t nu=gconf.Nu(), nv=gconf.Nv();
  checkShape(grid.shape(), {nu, nv});
  myassert(grid.contiguous(), "grid is not contiguous");
  T beta = gconf.Beta();
  size_t supp = gconf.Supp();
Martin Reinecke's avatar
Martin Reinecke committed
738
  size_t nthreads = gconf.Nthreads();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
739
740
741
742
743
744
745
746

#pragma omp parallel num_threads(nthreads)
{
  Helper<T> hlp(gconf, nullptr, grid.data());
  T emb = exp(-2*beta);
  int jump = hlp.lineJump();
  const T * ku = hlp.kernel.data();
  const T * kv = hlp.kernel.data()+supp;
747
  size_t np = srv.Nvis();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
748
749
750

  // Loop over sampling points
#pragma omp for schedule(guided,100)
751
  for (size_t ipart=0; ipart<np; ++ipart)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
752
    {
753
    UVW<T> coord = srv.getCoord(ipart);
754
    auto flip = coord.FixW();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
755
756
    hlp.prep(coord.u, coord.v);
    auto * ptr = hlp.p0w;
757
    auto v(srv.getVis(ipart)*emb);
758
    if (flip) v=conj(v);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
759
760
761
762
763
764
765
766
767
768
769
    for (size_t cu=0; cu<supp; ++cu)
      {
      complex<T> tmp(v*ku[cu]);
      for (size_t cv=0; cv<supp; ++cv)
        ptr[cv] += tmp*kv[cv];
      ptr+=jump;
      }
    }
} // end of parallel region
  }

770
template<typename T> void vis2grid_c
Martin Reinecke's avatar
updates    
Martin Reinecke committed
771
  (const Baselines<T> &baselines, const GridderConfig<T> &gconf,
772
773
  const const_mav<uint32_t,1> &idx, const const_mav<complex<T>,1> &vis,
  mav<complex<T>,2> &grid, const const_mav<T,1> &wgt)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
774
  {
775
776
777
778
779
780
781
782
783
784
  x2grid_c(gconf, VisServ<T,const_mav<complex<T>,1>,const_mav<T,1>> (baselines, idx, vis, wgt), grid);
  }

template<typename T> void ms2grid_c
  (const Baselines<T> &baselines, const GridderConfig<T> &gconf,
  const const_mav<uint32_t,1> &idx, const const_mav<complex<T>,2> &ms,
  mav<complex<T>,2> &grid, const const_mav<T,2> &wgt)
  {
  x2grid_c(gconf, MsServ<T,const_mav<complex<T>,2>,const_mav<T,2>> (baselines, idx, ms, wgt), grid);
  }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
785

786
787
788
template<typename T, typename Serv> void grid2x_c
  (const GridderConfig<T> &gconf, const const_mav<complex<T>,2> &grid, const Serv &srv)
  {
Martin Reinecke's avatar
updates    
Martin Reinecke committed
789
790
791
792
793
  size_t nu=gconf.Nu(), nv=gconf.Nv();
  checkShape(grid.shape(), {nu, nv});
  myassert(grid.contiguous(), "grid is not contiguous");
  T beta = gconf.Beta();
  size_t supp = gconf.Supp();
Martin Reinecke's avatar
Martin Reinecke committed
794
  size_t nthreads = gconf.Nthreads();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
795
796
797
798
799
800
801
802
803

  // Loop over sampling points
#pragma omp parallel num_threads(nthreads)
{
  Helper<T> hlp(gconf, grid.data(), nullptr);
  T emb = exp(-2*beta);
  int jump = hlp.lineJump();
  const T * ku = hlp.kernel.data();
  const T * kv = hlp.kernel.data()+supp;
804
  size_t np = srv.Nvis();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
805
806

#pragma omp for schedule(guided,100)
807
  for (size_t ipart=0; ipart<np; ++ipart)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
808
    {
809
    UVW<T> coord = srv.getCoord(ipart);
810
    auto flip = coord.FixW();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
811
812
813
814
815
816
817
818
819
820
821
    hlp.prep(coord.u, coord.v);
    complex<T> r = 0;
    const auto * ptr = hlp.p0r;
    for (size_t cu=0; cu<supp; ++cu)
      {
      complex<T> tmp(0);
      for (size_t cv=0; cv<supp; ++cv)
        tmp += ptr[cv] * kv[cv];
      r += tmp*ku[cu];
      ptr += jump;
      }
822
    if (flip) r=conj(r);
823
    srv.setVis(ipart, r*emb);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
824
825
826
    }
}
  }
827
828
829
830
831
832
833
template<typename T> void grid2vis_c
  (const Baselines<T> &baselines, const GridderConfig<T> &gconf,
  const const_mav<uint32_t,1> &idx, const const_mav<complex<T>,2> &grid,
  mav<complex<T>,1> &vis, const const_mav<T,1> &wgt)
  {
  grid2x_c(gconf, grid, VisServ<T,mav<complex<T>,1>,const_mav<T,1>> (baselines, idx, vis, wgt));
  }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
834
835
836
837
838
839

template<typename T> void grid2ms_c
  (const Baselines<T> &baselines, const GridderConfig<T> &gconf,
  const const_mav<uint32_t,1> &idx, const const_mav<complex<T>,2> &grid,
  mav<complex<T>,2> &ms, const const_mav<T,2> &wgt)
  {
840
  grid2x_c(gconf, grid, MsServ<T,mav<complex<T>,2>,const_mav<T,2>> (baselines, idx, ms, wgt));
Martin Reinecke's avatar
updates    
Martin Reinecke committed
841
842
  }

Martin Reinecke's avatar
Martin Reinecke committed
843
844
845
846
847
848
849
850
template<typename T> void apply_holo
  (const Baselines<T> &baselines, const GridderConfig<T> &gconf,
  const const_mav<uint32_t,1> &idx, const const_mav<complex<T>,2> &grid,
  mav<complex<T>,2> &ogrid, const const_mav<T,1> &wgt)
  {
  size_t nu=gconf.Nu(), nv=gconf.Nv();
  checkShape(grid.shape(), {nu, nv});
  size_t nvis = idx.shape(0);
Martin Reinecke's avatar
Martin Reinecke committed
851
  size_t nthreads = gconf.Nthreads();
Martin Reinecke's avatar
Martin Reinecke committed
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
  bool have_wgt = wgt.size()!=0;
  if (have_wgt) checkShape(wgt.shape(), {nvis});
  checkShape(ogrid.shape(), {nu, nv});

  for (size_t i=0; i<nu; ++i)
    for (size_t j=0; j<nv; ++j)
      ogrid(i,j) = T(0);

  T beta = gconf.Beta();
  size_t supp = gconf.Supp();

  // Loop over sampling points
#pragma omp parallel num_threads(nthreads)
{
  Helper<T> hlp(gconf, grid.data(), ogrid.data());
  T emb = exp(-2*beta);
  int jump = hlp.lineJump();
  const T * ku = hlp.kernel.data();
  const T * kv = hlp.kernel.data()+supp;

#pragma omp for schedule(guided,100)
  for (size_t ipart=0; ipart<nvis; ++ipart)
    {
875
    UVW<T> coord = baselines.effectiveCoord(idx(ipart));
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
876
    coord.FixW();
Martin Reinecke's avatar
Martin Reinecke committed
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
    hlp.prep(coord.u, coord.v);
    complex<T> r = 0;
    const auto * ptr = hlp.p0r;
    for (size_t cu=0; cu<supp; ++cu)
      {
      complex<T> tmp(0);
      for (size_t cv=0; cv<supp; ++cv)
        tmp += ptr[cv] * kv[cv];
      r += tmp*ku[cu];
      ptr += jump;
      }
    r*=emb*emb;
    if (have_wgt)
      {
      auto twgt = wgt(ipart);
      r*=twgt*twgt;
      }
    auto * wptr = hlp.p0w;
    for (size_t cu=0; cu<supp; ++cu)
      {
      complex<T> tmp(r*ku[cu]);
      for (size_t cv=0; cv<supp; ++cv)
        wptr[cv] += tmp*kv[cv];
      wptr += jump;
      }
    }
}
  }

template<typename T> void get_correlations
  (const Baselines<T> &baselines, const GridderConfig<T> &gconf,
  const const_mav<uint32_t,1> &idx, int du, int dv,
  mav<T,2> &ogrid, const const_mav<T,1> &wgt)
  {
  size_t nu=gconf.Nu(), nv=gconf.Nv();
  size_t nvis = idx.shape(0);
  bool have_wgt = wgt.size()!=0;
  if (have_wgt) checkShape(wgt.shape(), {nvis});
  checkShape(ogrid.shape(), {nu, nv});
  size_t supp = gconf.Supp();
  myassert(size_t(abs(du))<supp, "|du| must be smaller than Supp");
  myassert(size_t(abs(dv))<supp, "|dv| must be smaller than Supp");
Martin Reinecke's avatar
Martin Reinecke committed
919
  size_t nthreads = gconf.Nthreads();
Martin Reinecke's avatar
Martin Reinecke committed
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948

  T beta = gconf.Beta();
  for (size_t i=0; i<nu; ++i)
    for (size_t j=0; j<nv; ++j)
      ogrid(i,j) = T(0);

  size_t u0, u1, v0, v1;
  if (du>=0)
    { u0=0; u1=supp-du; }
  else
    { u0=-du; u1=supp; }
  if (dv>=0)
    { v0=0; v1=supp-dv; }
  else
    { v0=-dv; v1=supp; }

  // Loop over sampling points
#pragma omp parallel num_threads(nthreads)
{
  Helper<T,T> hlp(gconf, nullptr, ogrid.data());
  T emb = exp(-2*beta);
  int jump = hlp.lineJump();
  const T * ku = hlp.kernel.data();
  const T * kv = hlp.kernel.data()+supp;

#pragma omp for schedule(guided,100)
  for (size_t ipart=0; ipart<nvis; ++ipart)
    {
    UVW<T> coord = baselines.effectiveCoord(idx(ipart));
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
949
    coord.FixW();
Martin Reinecke's avatar
Martin Reinecke committed
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
    hlp.prep(coord.u, coord.v);
    auto * wptr = hlp.p0w + u0*jump;
    auto f0 = emb*emb;
    if (have_wgt)
      {
      auto twgt = wgt(ipart);
      f0*=twgt*twgt;
      }
    for (size_t cu=u0; cu<u1; ++cu)
      {
      auto f1=ku[cu]*ku[cu+du]*f0;
      for (size_t cv=v0; cv<v1; ++cv)
        wptr[cv] += f1*kv[cv]*kv[cv+dv];
      wptr += jump;
      }
    }
}
  }


970
template<typename T, typename T2> void apply_wcorr(const GridderConfig<T> &gconf,
Martin Reinecke's avatar
Martin Reinecke committed
971
  const mav<T2,2> &dirty, const EC_Kernel_with_correction<T> &kernel, double dw)
Martin Reinecke's avatar
Martin Reinecke committed
972
973
974
  {
  auto nx_dirty=gconf.Nxdirty();
  auto ny_dirty=gconf.Nydirty();
Martin Reinecke's avatar
Martin Reinecke committed
975
  size_t nthreads = gconf.Nthreads();
Martin Reinecke's avatar
Martin Reinecke committed
976
977
978
979
980
981
982
983
984
985
986
987
988
  auto psx=gconf.Pixsize_x();
  auto psy=gconf.Pixsize_y();
  double x0 = -0.5*nx_dirty*psx,
         y0 = -0.5*ny_dirty*psy;
#pragma omp parallel for schedule(static) num_threads(nthreads)
  for (size_t i=0; i<=nx_dirty/2; ++i)
    {
    double fx = x0+i*psx;
    fx *= fx;
    for (size_t j=0; j<=ny_dirty/2; ++j)
      {
      double fy = y0+j*psy;
      fy*=fy;
989
990
991
992
993
994
995
      T fct = 0;
      double tmp = 1.-fx-fy;
      if (tmp>=0)
        {
        auto nm1 = (-fx-fy)/(sqrt(1.-fx-fy)+1.); // accurate form of sqrt(1-x-y)
        fct = (nm1<=-1) ? 0. : kernel.corfac(nm1*dw);
        }
Martin Reinecke's avatar
Martin Reinecke committed
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
      size_t i2 = nx_dirty-i, j2 = ny_dirty-j;
      dirty(i,j)*=fct;
      if ((i>0)&&(i<i2))
        {
        dirty(i2,j)*=fct;
        if ((j>0)&&(j<j2))
          dirty(i2,j2)*=fct;
        }
      if ((j>0)&&(j<j2))
        dirty(i,j2)*=fct;
      }
    }
  }

1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
template<typename T, typename Serv> void x2dirty(
  const GridderConfig<T> &gconf, const Serv &srv, const mav<T,2> &dirty, size_t verbosity=0)
  {
  auto nx_dirty=gconf.Nxdirty();
  auto ny_dirty=gconf.Nydirty();
  auto nu=gconf.Nu();
  auto nv=gconf.Nv();

  if (verbosity>0)
    cout << "Gridding without w-stacking: " << srv.Nvis()
         << " visibilities" << endl;
Martin Reinecke's avatar
Martin Reinecke committed
1021
  if (verbosity>0) cout << "Using " << gconf.Nthreads() << " threads" << endl;
1022
1023
1024
1025
1026
1027
1028

  tmpStorage<complex<T>,2> grid_({nu,nv});
  auto grid=grid_.getMav();
  grid_.fill(0.);
  x2grid_c(gconf, srv, grid);
  tmpStorage<complex<T>,2> cdirty_({nx_dirty,ny_dirty});
  auto cdirty=cdirty_.getMav();
Martin Reinecke's avatar
Martin Reinecke committed
1029
  // FIXME: this could use symmetries to accelerate the FFT
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
  gconf.grid2dirty_c(grid, cdirty);
  for (size_t i=0; i<nx_dirty; ++i)
    for (size_t j=0; j<ny_dirty; ++j)
      dirty(i,j) = cdirty(i,j).real();
  }

template<typename T, typename Serv> void dirty2x(
  const GridderConfig<T> &gconf,  const const_mav<T,2> &dirty,
  const Serv &srv, size_t verbosity=0)
  {
  auto nx_dirty=gconf.Nxdirty();
  auto ny_dirty=gconf.Nydirty();
  auto nu=gconf.Nu();
  auto nv=gconf.Nv();
  size_t nvis = srv.Nvis();

Martin Reinecke's avatar
Martin Reinecke committed
1046
1047
1048
1049
1050
  if (verbosity>0)
    cout << "Degridding without w-stacking: " << srv.Nvis()
         << " visibilities" << endl;
  if (verbosity>0) cout << "Using " << gconf.Nthreads() << " threads" << endl;

1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
  for (size_t i=0; i<nvis; ++i)
    srv.setVis(i, 0.);

  tmpStorage<complex<T>,2> cdirty_({nx_dirty,ny_dirty});
  auto cdirty=cdirty_.getMav();
  for (size_t i=0; i<nx_dirty; ++i)
    for (size_t j=0; j<ny_dirty; ++j)
      cdirty(i,j) = dirty(i,j);

  tmpStorage<complex<T>,2> grid_({nu,nv});
  auto grid=grid_.getMav();
Martin Reinecke's avatar
Martin Reinecke committed
1062
  // FIXME: this could use symmetries to accelerate the FFT
1063
1064
1065
1066
  gconf.dirty2grid_c(cdirty, grid);
  grid2x_c(gconf, const_mav<complex<T>,2>(grid.data(),{nu,nv}), srv);
  }

Martin Reinecke's avatar
Martin Reinecke committed
1067
1068
template<typename T, typename Serv> void wstack_common(
  const GridderConfig<T> &gconf, const Serv &srv, T &wmin, vector<T> &wval,
Martin Reinecke's avatar
Martin Reinecke committed
1069
  double &dw, size_t &nplanes, vector<size_t> &nvis_plane, vector<int> &minplane, size_t verbosity)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1070
1071
1072
1073
1074
  {
  auto nx_dirty=gconf.Nxdirty();
  auto ny_dirty=gconf.Nydirty();
  auto psx=gconf.Pixsize_x();
  auto psy=gconf.Pixsize_y();
1075
  size_t nvis = srv.Nvis();
Martin Reinecke's avatar
Martin Reinecke committed
1076
  size_t nthreads = gconf.Nthreads();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1077
1078

  // determine w values for every visibility, and min/max w;
Martin Reinecke's avatar
Martin Reinecke committed
1079
1080
1081
  wmin=T(1e38);
  T wmax=T(-1e38);
  wval.resize(nvis);
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1082
1083

#pragma omp parallel for num_threads(nthreads) reduction(min:wmin) reduction(max:wmax)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1084
1085
  for (size_t ipart=0; ipart<nvis; ++ipart)
    {
1086
    wval[ipart] = abs(srv.getCoord(ipart).w);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1087
1088
1089
    wmin = min(wmin,wval[ipart]);
    wmax = max(wmax,wval[ipart]);
    }
Martin Reinecke's avatar
Martin Reinecke committed
1090
  if (verbosity>0) cout << "Using " << nthreads << " threads" << endl;
Martin Reinecke's avatar
Martin Reinecke committed
1091
  if (verbosity>0) cout << "W range: " << wmin << " to " << wmax << endl;
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1092
1093
1094
  double x0 = -0.5*nx_dirty*psx,
         y0 = -0.5*ny_dirty*psy;
  double nmin = sqrt(max(1.-x0*x0-y0*y0,0.))-1.;
Martin Reinecke's avatar
Martin Reinecke committed
1095
1096
  dw = 0.25/abs(nmin);
  nplanes = max<size_t>(2,((wmax-wmin)/dw+2));
1097
  dw=(wmax-wmin)/(nplanes-1);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1098

Martin Reinecke's avatar
Martin Reinecke committed
1099
  auto w_supp = gconf.Supp();
1100
1101
1102
  wmin -= (0.5*w_supp-1)*dw;
  wmax += (0.5*w_supp-1)*dw;
  nplanes += w_supp-2;
Martin Reinecke's avatar
Martin Reinecke committed
1103
1104
  if (verbosity>0) cout << "Kernel support: " << w_supp << endl;
  if (verbosity>0) cout << "nplanes: " << nplanes << endl;
Martin Reinecke's avatar
Martin Reinecke committed
1105
1106
  nvis_plane.resize(nplanes,0);
  minplane.resize(nvis);
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1107
1108
1109
1110
#pragma omp parallel num_threads(nthreads)
{
  vector<size_t> nvploc(nplanes,0);
#pragma omp for
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1111
1112
1113
1114
1115
  for (size_t ipart=0; ipart<nvis; ++ipart)
    {
    int plane0 = int((wval[ipart]-wmin)/dw+0.5*w_supp)-int(w_supp-1);
    minplane[ipart]=plane0;
    for (int i=max<int>(0,plane0); i<min<int>(nplanes,plane0+w_supp); ++i)
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1116
      ++nvploc[i];
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1117
    }
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1118
1119
1120
1121
#pragma omp critical
  for (size_t i=0; i<nplanes; ++i)
    nvis_plane[i]+=nvploc[i];
}
Martin Reinecke's avatar
Martin Reinecke committed
1122
1123
  }

1124
template<typename T, typename Serv> void x2dirty_wstack(
1125
  const GridderConfig<T> &gconf, const Serv &srv, const mav<T,2> &dirty, size_t verbosity=0)
Martin Reinecke's avatar
Martin Reinecke committed
1126
1127
1128
1129
1130
1131
1132
  {
  auto nx_dirty=gconf.Nxdirty();
  auto ny_dirty=gconf.Nydirty();
  auto nu=gconf.Nu();
  auto nv=gconf.Nv();
  size_t nvis = srv.Nvis();
  auto w_supp = gconf.Supp();
Martin Reinecke's avatar
Martin Reinecke committed
1133
1134
  size_t nthreads = gconf.Nthreads();
  EC_Kernel_with_correction<T> kernel(w_supp, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
1135
1136
1137
1138
1139
1140
  T wmin;
  vector<T> wval;
  double dw;
  size_t nplanes;
  vector<size_t> nvis_plane;
  vector<int> minplane;
Martin Reinecke's avatar
Martin Reinecke committed
1141
  if (verbosity>0) cout << "Gridding using improved w-stacking" << endl;
Martin Reinecke's avatar
Martin Reinecke committed
1142
  wstack_common(gconf, srv, wmin, wval, dw, nplanes, nvis_plane, minplane, verbosity);
Martin Reinecke's avatar
Martin Reinecke committed
1143

Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
1144
1145
  for (size_t i=0; i<nx_dirty; ++i)
    for (size_t j=0; j<ny_dirty; ++j)
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1146
      dirty(i,j) = 0.;
Martin Reinecke's avatar
Martin Reinecke committed
1147
1148
1149
  {
  vector<complex<T>> vis_loc_;
  vector<uint32_t> idx_loc_;
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1150
  vector<uint32_t> mapper;
Martin Reinecke's avatar
Martin Reinecke committed
1151
1152
1153
1154
  tmpStorage<complex<T>,2> grid_({nu,nv});
  auto grid=grid_.getMav();
  tmpStorage<complex<T>,2> tdirty_({nx_dirty,ny_dirty});
  auto tdirty=tdirty_.getMav();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1155
1156
1157
  for (size_t iw=0; iw<nplanes; ++iw)
    {
    if (nvis_plane[iw]==0) continue;
Martin Reinecke's avatar
Martin Reinecke committed
1158
1159
1160
    if (verbosity>1)
      cout << "Working on plane " << iw << " containing " << nvis_plane[iw]
           << " visibilities" << endl;
1161
    T wcur = wmin+iw*dw;
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1162
    size_t cnt=0;
Martin Reinecke's avatar
Martin Reinecke committed
1163
    vis_loc_.resize(nvis_plane[iw]);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1164
    auto vis_loc = mav<complex<T>,1>(vis_loc_.data(), {nvis_plane[iw]});
Martin Reinecke's avatar
Martin Reinecke committed
1165
    idx_loc_.resize(nvis_plane[iw]);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1166
    auto idx_loc = mav<uint32_t,1>(idx_loc_.data(), {nvis_plane[iw]});
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1167
    mapper.resize(nvis_plane[iw]);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1168
    for (size_t ipart=0; ipart<nvis; ++ipart)
Martin Reinecke's avatar
Martin Reinecke committed
1169
      if ((int(iw)>=minplane[ipart]) && (iw<minplane[ipart]+w_supp))
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1170
1171
        mapper[cnt++] = ipart;

1172
    T tmpval=T(2)/(w_supp*dw);
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1173
1174
1175
1176
#pragma omp parallel for num_threads(nthreads)
    for (size_t i=0; i<nvis_plane[iw]; ++i)
      {
      auto ipart = mapper[i];
1177
      T x=min(T(1), tmpval*abs(wcur-wval[ipart]));
Martin Reinecke's avatar
tweak    
Martin Reinecke committed
1178
1179
1180
      vis_loc[i] = srv.getVis(ipart)*kernel(x);
      idx_loc[i] = srv.getIdx(ipart);
      }
Martin Reinecke's avatar
Martin Reinecke committed
1181
    grid_.fill(0.);
1182
    vis2grid_c(srv.getBaselines(), gconf, const_mav<uint32_t,1>(idx_loc), const_mav<complex<T>,1>(vis_loc), grid, const_mav<T,1>(nullptr,{0}));
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1183
1184
    gconf.grid2dirty_c(grid, tdirty);
    gconf.apply_wscreen(tdirty, tdirty, wcur, false);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
1185
#pragma omp parallel for num_threads(nthreads)
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
1186
1187
    for (size_t i=0; i<nx_dirty; ++i)
      for (size_t j=0; j<ny_dirty; ++j)
1188
        dirty(i,j) += tdirty(i,j).real();
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1189
    }
Martin Reinecke's avatar
Martin Reinecke committed
1190
  }
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1191
  // correct for w gridding
Martin Reinecke's avatar
Martin Reinecke committed
1192
  apply_wcorr(gconf, dirty, kernel, dw);
Martin Reinecke's avatar
updates    
Martin Reinecke committed
1193
  }
1194
template<typename T> void vis2dirty_wstack(const Baselines<T> &baselines,
1195
  const GridderConfig<T> &gconf, const const_mav<uint32_t,1> &idx,
1196
  const const_mav<complex<T>,1> &vis, mav<T,2> &dirty)
1197
  {
1198
  x2dirty_wstack(gconf, VisServ<T,const_mav<complex<T>,1>,const_mav<T,1>>(baselines, idx, vis, const_mav<T,1>(nullptr,{0})), dirty);
1199
1200
  }

1201
template<typename T, typename Serv> void dirty2x_wstack(
1202
  const GridderConfig<T> &gconf,  const const_mav<T,2> &dirty,
Martin Reinecke's avatar
Martin Reinecke committed
1203
  const Serv &srv, size_t verbosity=0)
1204
1205
1206
1207
1208
  {
  auto nx_dirty=gconf.Nxdirty();
  auto ny_dirty=gconf.Nydirty();
  auto nu=gconf.Nu();
  auto nv=gconf.Nv();
1209
  size_t nvis = srv.Nvis();
1210
  auto w_supp = gconf.Supp();
Martin Reinecke's avatar
Martin Reinecke committed
1211
1212
  size_t nthreads = gconf.Nthreads();
  EC_Kernel_with_correction<T> kernel(w_supp, nthreads);
Martin Reinecke's avatar
Martin Reinecke committed
1213
1214
1215
1216
1217
1218
  T wmin;
  vector<T> wval;
  double dw;
  size_t nplanes;
  vector<size_t> nvis_plane;
  vector<int> minplane;
Martin Reinecke's avatar
Martin Reinecke committed
1219
  if (verbosity>0) cout << "Degridding using improved w-stacking" << endl;
Martin Reinecke's avatar
Martin Reinecke committed
1220
  wstack_common(gconf, srv, wmin, wval, dw, nplanes, nvis_plane, minplane, verbosity);
1221
1222

  for (size_t i=0; i<nvis; ++i)
1223
    srv.setVis(i, 0.);
1224

1225
  tmpStorage<T,2> tdirty_({nx_dirty,ny_dirty});