nifty_gridder.cc 21.4 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
/*
 *  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_fridder; if not, write to the Free Software
 *  Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
 */

Martin Reinecke's avatar
import  
Martin Reinecke committed
19
20
21
22
23
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
#include <iostream>
#include <algorithm>

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
24
25
#ifdef __GNUC__
#define RESTRICT __restrict__
Martin Reinecke's avatar
Martin Reinecke committed
26
#define NOINLINE __attribute__ ((noinline))
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
27
28
29
30
#else
#define RESTRICT
#endif

Martin Reinecke's avatar
import  
Martin Reinecke committed
31
32
33
34
35
36
using namespace std;

namespace py = pybind11;

namespace {

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
void myassert(bool cond, const char *msg)
  {
  if (cond) return;
  throw runtime_error(msg);
  }

/*! Returns the largest integer \a n that fulfills \a 2^n<=arg. */
template<typename I> inline int ilog2 (I arg)
  {
#ifdef __GNUC__
  if (arg==0) return 0;
  if (sizeof(I)==sizeof(int))
    return 8*sizeof(int)-1-__builtin_clz(arg);
  if (sizeof(I)==sizeof(long))
    return 8*sizeof(long)-1-__builtin_clzl(arg);
  if (sizeof(I)==sizeof(long long))
    return 8*sizeof(long long)-1-__builtin_clzll(arg);
#endif
  int res=0;
  while (arg > 0xFFFF) { res+=16; arg>>=16; }
  if (arg > 0x00FF) { res|=8; arg>>=8; }
  if (arg > 0x000F) { res|=4; arg>>=4; }
  if (arg > 0x0003) { res|=2; arg>>=2; }
  if (arg > 0x0001) { res|=1; }
  return res;
  }

/*! Returns the number of bits needed to represent \a arg different values.
    \a arg must be >=1. */
template<typename I> inline int bits_needed (I arg)
  {
  myassert(arg>=1, "argument must be >=1");
  if (arg==1) return 0;
  return ilog2(arg-1)+1;
  }
Martin Reinecke's avatar
Martin Reinecke committed
72
73
74
//
// Utilities for converting between Cartesian coordinates and Peano index
//
Martin Reinecke's avatar
import  
Martin Reinecke committed
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160

static const uint16_t utab[] = {
#define Z(a) 0x##a##0, 0x##a##1, 0x##a##4, 0x##a##5
#define Y(a) Z(a##0), Z(a##1), Z(a##4), Z(a##5)
#define X(a) Y(a##0), Y(a##1), Y(a##4), Y(a##5)
X(0),X(1),X(4),X(5)
#undef X
#undef Y
#undef Z
};

uint32_t coord2morton2D_32 (uint32_t x, uint32_t y)
  {
  typedef uint32_t I;
  return  (I)(utab[x&0xff])     | ((I)(utab[(x>>8)&0xff])<<16)
       | ((I)(utab[y&0xff])<<1) | ((I)(utab[(y>>8)&0xff])<<17);
  }

static const uint8_t m2p2D_1[4][4] = {
{ 4, 1, 11, 2},{0,15, 5, 6},{10,9,3,12},{14,7,13,8}};
static uint8_t m2p2D_3[4][64];
static const uint8_t p2m2D_1[4][4] = {
{ 4, 1, 3, 10},{0,6,7,13},{15,9,8,2},{11,14,12,5}};
static uint8_t p2m2D_3[4][64];
static int peano2d_done=0;

static void init_peano2d (void)
  {
  peano2d_done=1;

  for (int d=0; d<4; ++d)
    for (uint32_t p=0; p<64; ++p)
      {
      unsigned rot = d;
      uint32_t v = p<<26;
      uint32_t res = 0;
      for (int i=0; i<3; ++i)
        {
        unsigned tab=m2p2D_1[rot][v>>30];
        v<<=2;
        res = (res<<2) | (tab&0x3);
        rot = tab>>2;
        }
      m2p2D_3[d][p]=res|(rot<<6);
      }
  for (int d=0; d<4; ++d)
    for (uint32_t p=0; p<64; ++p)
      {
      unsigned rot = d;
      uint32_t v = p<<26;
      uint32_t res = 0;
      for (int i=0; i<3; ++i)
        {
        unsigned tab=p2m2D_1[rot][v>>30];
        v<<=2;
        res = (res<<2) | (tab&0x3);
        rot = tab>>2;
        }
      p2m2D_3[d][p]=res|(rot<<6);
      }
  }

uint32_t morton2peano2D_32(uint32_t v, int bits)
  {
  if (!peano2d_done) init_peano2d();
  unsigned rot = 0;
  uint32_t res = 0;
  v<<=32-(bits<<1);
  int i=0;
  for (; i<bits-2; i+=3)
    {
    unsigned tab=m2p2D_3[rot][v>>26];
    v<<=6;
    res = (res<<6) | (tab&0x3fu);
    rot = tab>>6;
    }
  for (; i<bits; ++i)
    {
    unsigned tab=m2p2D_1[rot][v>>30];
    v<<=2;
    res = (res<<2) | (tab&0x3);
    rot = tab>>2;
    }
  return res;
  }

Martin Reinecke's avatar
Martin Reinecke committed
161
162
163
164
//
// Utilities for indirect sorting (argsort)
//

Martin Reinecke's avatar
import  
Martin Reinecke committed
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
template<typename It, typename Comp> class IdxComp__
  {
  private:
    It begin;
    Comp comp;
  public:
    IdxComp__ (It begin_, Comp comp_): begin(begin_), comp(comp_) {}
    bool operator() (std::size_t a, std::size_t b) const
      { return comp(*(begin+a),*(begin+b)); }
  };
/*! Performs an indirect sort on the supplied iterator range and returns in
    \a idx a \a vector containing the indices of the smallest, second smallest,
    third smallest, etc. element, according to \a comp. */
template<typename It, typename T2, typename Comp>
  inline void buildIndex (It begin, It end, std::vector<T2> &idx, Comp comp)
  {
  using namespace std;
  T2 num=end-begin;
  idx.resize(num);
  for (T2 i=0; i<num; ++i) idx[i] = i;
185
186
//  sort (idx.begin(),idx.end(),IdxComp__<It,Comp>(begin,comp));
  stable_sort (idx.begin(),idx.end(),IdxComp__<It,Comp>(begin,comp));
Martin Reinecke's avatar
import  
Martin Reinecke committed
187
188
189
190
191
192
193
194
195
196
197
198
199
  }

/*! Performs an indirect sort on the supplied iterator range and returns in
    \a idx a \a vector containing the indices of the smallest, second smallest,
    third smallest, etc. element. */
template<typename It, typename T2> inline void buildIndex (It begin, It end,
  std::vector<T2> &idx)
  {
  using namespace std;
  typedef typename iterator_traits<It>::value_type T;
  buildIndex(begin,end,idx,less<T>());
  }

200
201
202
/*! 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. */
Martin Reinecke's avatar
Martin Reinecke committed
203
template<typename T> inline T fmodulo (T v1, T v2)
204
205
206
  {
  if (v1>=0)
    return (v1<v2) ? v1 : fmod(v1,v2);
Martin Reinecke's avatar
Martin Reinecke committed
207
208
  T tmp=fmod(v1,v2)+v2;
  return (tmp==v2) ? T(0) : tmp;
209
210
  }

Martin Reinecke's avatar
Martin Reinecke committed
211
212
213
214
//
// Utilities for Gauss-Legendre quadrature
//

Martin Reinecke's avatar
Martin Reinecke committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
static inline double one_minus_x2 (double x)
  { return (fabs(x)>0.1) ? (1.+x)*(1.-x) : 1.-x*x; }

void legendre_prep(int n, vector<double> &x, vector<double> &w)
  {
  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
{
  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;
      if (++j>=100) throw runtime_error("convergence problem");
      }

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

Martin Reinecke's avatar
Martin Reinecke committed
272
273
274
275
//
// Start of real gridder functionality
//

276
277
278
279
280
281
282
283
template<typename T>
  using pyarr = py::array_t<T>;
template<typename T>
  using pyarr_c = py::array_t<T, py::array::c_style | py::array::forcecast>;
using a_u32_c = pyarr_c<uint32_t>;
using a_d_c = pyarr_c<double>;
using a_f_c = pyarr_c<float>;
using a_cd_c = pyarr_c<complex<double>>;
Martin Reinecke's avatar
import  
Martin Reinecke committed
284

285
a_u32_c peanoindex(const a_d_c &uv_, int nu, int nv)
Martin Reinecke's avatar
import  
Martin Reinecke committed
286
287
288
289
290
291
292
293
294
295
296
297
  {
  myassert(uv_.ndim()==2, "uv array must be 2D");
  myassert(uv_.shape(1)==2, "uv.shape[1] must be 2");
  int nvis = uv_.shape(0);
  auto uv = uv_.data();

  int npmax = max(nu, nv);
  int nbits = 0;
  for (int istart = npmax-1; istart!=0; istart>>=1, ++nbits);
  vector<int> ipeano(nvis);
  for (int i=0; i<nvis; ++i)
    {
298
    auto u = fmodulo(uv[2*i], 1.)*nu;
Martin Reinecke's avatar
import  
Martin Reinecke committed
299
    auto iu = min(nu-1, int(u));
300
    auto v = fmodulo(uv[2*i+1], 1.)*nv;
Martin Reinecke's avatar
import  
Martin Reinecke committed
301
302
303
304
305
306
    auto iv = min(nv-1, int(v));
    ipeano[i] = morton2peano2D_32(coord2morton2D_32(iu,iv),nbits);
    }
  vector<int> newind;
  buildIndex(ipeano.begin(), ipeano.end(), newind);
  int odim[] = {nvis};
307
  a_u32_c res(odim);
Martin Reinecke's avatar
import  
Martin Reinecke committed
308
309
310
311
312
313
  auto iout = res.mutable_data();
  for (int i=0; i<nvis; ++i)
    iout[i] = newind[i];
  return res;
  }

Martin Reinecke's avatar
Martin Reinecke committed
314
int get_w(double epsilon)
Martin Reinecke's avatar
Martin Reinecke committed
315
316
317
318
319
320
321
322
323
324
325
  {
  static const vector<double> maxmaperr { 1e8, 0.32, 0.021, 6.2e-4,
    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;
  throw runtime_error("requested epsilon too small - minimum is 2e-13");
  }
Martin Reinecke's avatar
Martin Reinecke committed
326

Martin Reinecke's avatar
Martin Reinecke committed
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
a_d_c complex2hartley (const a_cd_c &grid_)
  {
  myassert(grid_.ndim()==2, "grid array must be 2D");
  int nu = grid_.shape(0), nv = grid_.shape(1);
  auto grid = grid_.data();

  int odim[] = {nu,nv};
  a_d_c res(odim);
  auto grid2 = res.mutable_data();
  for (int u=0; u<nu; ++u)
    {
    int xu = (u==0) ? 0 : nu-u;
    for (int v=0; v<nv; ++v)
      {
      int xv = (v==0) ? 0 : nv-v;
      int i1 = u*nv+v;
      int i2 = xu*nv+xv;
      grid2[i1] = 0.5*(grid[i1].real()+grid[i1].imag()+
                       grid[i2].real()-grid[i2].imag());
      }
    }
  return res;
  }


a_cd_c hartley2complex (const a_d_c &grid_)
  {
  myassert(grid_.ndim()==2, "grid array must be 2D");
  int nu = grid_.shape(0), nv = grid_.shape(1);
  auto grid = grid_.data();

  int odim[] = {nu,nv};
  a_cd_c res(odim);
  auto grid2 = res.mutable_data();
  for (int u=0; u<nu; ++u)
    {
    int xu = (u==0) ? 0 : nu-u;
    for (int v=0; v<nv; ++v)
      {
      int xv = (v==0) ? 0 : nv-v;
      int i1 = u*nv+v;
      int i2 = xu*nv+xv;
      double v1 = 0.5*grid[i1];
      double v2 = 0.5*grid[i2];
      grid2[i1] = complex<double>(v1+v2, v1-v2);
      }
    }
  return res;
  }


/* Compute correction factors for the ES gridding kernel
   This implementation follows eqs. (3.8) to (3.10) of Barnett et al. 2018 */
a_d_c correction_factors (size_t n, size_t nval, size_t w)
  {
  constexpr double pi = 3.141592653589793238462643383279502884197;
  auto beta = 2.3*w;
  auto p = int(1.5*w+2);
  double alpha = pi*w/n;
  vector<double> x, wgt;
  legendre_prep(2*p,x,wgt);
  auto psi = x;
  for (auto &v:psi)
    v = exp(beta*(sqrt(1-v*v)-1.));
  int odim[] = {int(nval)};
  a_d_c res(odim);
  auto val = res.mutable_data();
  for (size_t k=0; k<nval; ++k)
    {
    double tmp=0;
    for (int i=0; i<p; ++i)
      tmp += wgt[i]*psi[i]*cos(alpha*k*x[i]);
    val[k] = 1./(w*tmp);
    }
  return res;
  }

struct UV
  {
  double u, v;
  UV () {}
  UV (double u_, double v_) : u(u_), v(v_) {}
  UV operator* (double fct) const
    { return UV(u*fct, v*fct); }
  };

class Baselines
  {
  private:
    vector<UV> coord;
    vector<double> scaling;
    size_t nrows;
    size_t channelbits, channelmask;

  public:
    Baselines(const vector<UV> &coord_, const vector<double> &scaling_)
      : coord(coord_), scaling(scaling_), nrows(coord.size()/scaling.size())
      {
      myassert(nrows*scaling.size()==coord.size(), "bad array dimensions");
      channelbits = bits_needed(scaling.size());
      channelmask = (size_t(1)<<channelbits)-1;
      auto rowbits = bits_needed(nrows);
      myassert(rowbits+channelbits<=8*sizeof(uint32_t), "Ti too small");
      }
    Baselines(const pyarr_c<double> &coord_, const pyarr_c<double> &scaling_)
      {
      myassert(coord_.ndim()==2, "coord array must be 2D");
      myassert(coord_.shape(1)==2, "coord.shape[1] must be 2");
      myassert(scaling_.ndim()==1, "scaling array must be 1D");
      nrows = coord_.shape(0)/scaling_.shape(0);
      myassert(nrows*size_t(scaling_.shape(0))==size_t(coord_.shape(0)),
        "bad array dimensions");
      scaling.resize(scaling_.shape(0));
      for (size_t i=0; i<scaling.size(); ++i)
        scaling[i] = scaling_.data()[i];
      coord.resize(nrows);
      for (size_t i=0; i<coord.size(); ++i)
        coord[i] = UV(coord_.data()[2*i], coord_.data()[2*i+1]);
      channelbits = bits_needed(scaling.size());
      channelmask = (size_t(1)<<channelbits)-1;
      auto rowbits = bits_needed(nrows);
      myassert(rowbits+channelbits<=8*sizeof(uint32_t), "Ti too small");
      }

    pyarr_c<uint32_t> getIndices() const
      {
      int odim[] = {int(nrows*scaling.size())};
      pyarr_c<uint32_t> res(odim);
      auto odata = res.mutable_data();
      for (size_t i=0; i<nrows; ++i)
        for (size_t j=0; j<scaling.size(); ++j)
          odata[j+scaling.size()*i] = (i<<channelbits)+j;
      return res;
      }

    UV EffectiveCoord(uint32_t index) const
      { return coord[index>>channelbits]*scaling[index&channelmask]; }
    size_t Nrows() const { return nrows; }
    size_t Nchannels() const { return scaling.size(); }
    size_t irow(uint32_t index) const
      { return index>>channelbits; }
    size_t ichannel(uint32_t index) const
      { return index&channelmask; }
    size_t offset(uint32_t index) const
      { return (index>>channelbits)*scaling.size() + (index&channelmask); }
  };

class GridderConfig
  {
  private:
    size_t nx_dirty, ny_dirty;
    double ucorr, vcorr;
    size_t w, nsafe, nu, nv;
    size_t peano_level;

  public:
    GridderConfig(size_t nxdirty, size_t nydirty, double epsilon,
      double urange, double vrange)
      : nx_dirty(nxdirty), ny_dirty(nydirty),
        ucorr(1./urange), vcorr(1./vrange),
        w(get_w(epsilon)), nsafe((w+1)/2),
        nu(max(2*nsafe,2*nx_dirty)), nv(max(2*nsafe,2*ny_dirty)),
        peano_level(bits_needed(max(nu, nv)))
      {
      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(urange>0, "urange must be positive");
      myassert(vrange>0, "vrange must be positive");
      }
    size_t Nu() const { return nu; }
    size_t Nv() const { return nv; }
    size_t W() const { return w; }
    size_t coord2peano(const UV &coord) const
      {
      double u=fmodulo(coord.u*ucorr, 1.)*nu,
             v=fmodulo(coord.v*vcorr, 1.)*nv;
      auto iu = min(nu-1, size_t(u));
      auto iv = min(nv-1, size_t(v));
      return morton2peano2D_32(coord2morton2D_32(iu,iv),peano_level);
      }
    pyarr_c<uint32_t> reorderIndices
      (const pyarr_c<uint32_t> &idx, const Baselines &baselines) const
      {
      myassert(idx.ndim()==1, "need 1D index array");
      vector<size_t> peano(idx.shape(0));
      auto pidx = idx.data();
      for (size_t i=0; i<peano.size(); ++i)
        peano[i] = coord2peano(baselines.EffectiveCoord(pidx[i]));
      vector<size_t> newind;
      buildIndex(peano.begin(), peano.end(), newind);
      peano=vector<size_t>(); // deallocate
      int odim[] = {idx.shape(0)};
      pyarr_c<uint32_t> res(odim);
      auto iout = res.mutable_data();
      for (int i=0; i<idx.shape(0); ++i)
        iout[i] = pidx[newind[i]];
      return res;
      }
    pyarr_c<double> U_corrections() const
      { return correction_factors(nu, nx_dirty/2+1, w); }
    pyarr_c<double> V_corrections() const
      { return correction_factors(nv, ny_dirty/2+1, w); }
  };


class Helper2
Martin Reinecke's avatar
import  
Martin Reinecke committed
534
  {
Martin Reinecke's avatar
Martin Reinecke committed
535
536
  protected:
    int nu, nv;
Martin Reinecke's avatar
import  
Martin Reinecke committed
537
  public:
Martin Reinecke's avatar
Martin Reinecke committed
538
539
    int w;
    double beta;
Martin Reinecke's avatar
Martin Reinecke committed
540
  protected:
Martin Reinecke's avatar
Martin Reinecke committed
541
542
543
    int nsafe, su;
  public:
    int sv;
Martin Reinecke's avatar
import  
Martin Reinecke committed
544

Martin Reinecke's avatar
Martin Reinecke committed
545
546
547
548
    vector<double> kernel;
    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
549
    void NOINLINE update(double u_in, double v_in)
Martin Reinecke's avatar
import  
Martin Reinecke committed
550
      {
551
      auto u = fmodulo(u_in, 1.)*nu;
Martin Reinecke's avatar
Martin Reinecke committed
552
553
      iu0 = int(u-w*0.5 + 1 + nu) - nu;
      if (iu0+w>nu+nsafe) iu0 = nu+nsafe-w;
554
      auto v = fmodulo(v_in, 1.)*nv;
Martin Reinecke's avatar
Martin Reinecke committed
555
556
557
558
      iv0 = int(v-w*0.5 + 1 + nv) - nv;
      if (iv0+w>nv+nsafe) iv0 = nv+nsafe-w;
      double xw=2./w;
      for (int i=0; i<w; ++i)
Martin Reinecke's avatar
import  
Martin Reinecke committed
559
        {
Martin Reinecke's avatar
Martin Reinecke committed
560
561
562
        kernel[i  ] = xw*(iu0+i-u);
        kernel[i+w] = xw*(iv0+i-v);
        }
Martin Reinecke's avatar
Martin Reinecke committed
563
564
      for (auto &k : kernel)
        k = exp(beta*sqrt(1.-k*k));
Martin Reinecke's avatar
import  
Martin Reinecke committed
565
566
      }

Martin Reinecke's avatar
Martin Reinecke committed
567
568
    bool need_to_move() const
      { return (iu0<bu0) || (iv0<bv0) || (iu0+w>bu0+su) || (iv0+w>bv0+sv); }
Martin Reinecke's avatar
import  
Martin Reinecke committed
569

Martin Reinecke's avatar
Martin Reinecke committed
570
    void update_position()
Martin Reinecke's avatar
import  
Martin Reinecke committed
571
      {
Martin Reinecke's avatar
Martin Reinecke committed
572
573
      bu0=max(-nsafe, min(nu+nsafe-su, iu0+nsafe-su/2));
      bv0=max(-nsafe, min(nv+nsafe-sv, iv0+nsafe-sv/2));
Martin Reinecke's avatar
import  
Martin Reinecke committed
574
575
      }

Martin Reinecke's avatar
Martin Reinecke committed
576
  protected:
Martin Reinecke's avatar
Martin Reinecke committed
577
578
579
    Helper2(int nu_, int nv_, int w_)
      : nu(nu_), nv(nv_), w(w_), beta(2.3*w), nsafe((w+1)/2),
        su(min(max(2*nsafe,80), nu)), sv(min(max(2*nsafe,80), nv)),
Martin Reinecke's avatar
Martin Reinecke committed
580
581
        kernel(2*w),
        bu0(-1000000), bv0(-1000000)
Martin Reinecke's avatar
import  
Martin Reinecke committed
582
      {
Martin Reinecke's avatar
Martin Reinecke committed
583
      if (min(nu,nv)<2*nsafe) throw runtime_error("grid dimensions too small");
Martin Reinecke's avatar
import  
Martin Reinecke committed
584
585
586
      }
  };

Martin Reinecke's avatar
Martin Reinecke committed
587
class WriteHelper2: public Helper2
Martin Reinecke's avatar
import  
Martin Reinecke committed
588
589
590
591
592
593
594
  {
  private:
    vector<complex<double>> data;
    complex<double> *grid;

    void dump()
      {
Martin Reinecke's avatar
Martin Reinecke committed
595
      if (bu0<-nsafe) return; // nothing written into buffer yet
Martin Reinecke's avatar
import  
Martin Reinecke committed
596
597
#pragma omp critical
{
Martin Reinecke's avatar
Martin Reinecke committed
598
599
600
      int idxu = (bu0+nu)%nu;
      int idxv0 = (bv0+nv)%nv;
      for (int iu=0; iu<su; ++iu)
Martin Reinecke's avatar
import  
Martin Reinecke committed
601
602
        {
        int idxv = idxv0;
Martin Reinecke's avatar
Martin Reinecke committed
603
        for (int iv=0; iv<sv; ++iv)
Martin Reinecke's avatar
import  
Martin Reinecke committed
604
          {
Martin Reinecke's avatar
Martin Reinecke committed
605
          grid[idxu*nv + idxv] += data[iu*sv + iv];
Martin Reinecke's avatar
import  
Martin Reinecke committed
606
607
608
609
610
611
612
613
614
          if (++idxv>=nv) idxv=0;
          }
        if (++idxu>=nu) idxu=0;
        }
}
      }

  public:
    complex<double> *p0;
Martin Reinecke's avatar
Martin Reinecke committed
615
616
617
    WriteHelper2(int nu_, int nv_, int w, complex<double> *grid_)
      : Helper2(nu_, nv_, w), data(su*sv, 0.), grid(grid_) {}
    ~WriteHelper2() { dump(); }
Martin Reinecke's avatar
import  
Martin Reinecke committed
618

Martin Reinecke's avatar
Martin Reinecke committed
619
    void prep_write(double u_in, double v_in)
Martin Reinecke's avatar
import  
Martin Reinecke committed
620
      {
Martin Reinecke's avatar
Martin Reinecke committed
621
622
      update(u_in, v_in);
      if (need_to_move())
Martin Reinecke's avatar
import  
Martin Reinecke committed
623
624
        {
        dump();
Martin Reinecke's avatar
Martin Reinecke committed
625
        update_position();
Martin Reinecke's avatar
import  
Martin Reinecke committed
626
627
        fill(data.begin(), data.end(), 0.);
        }
Martin Reinecke's avatar
Martin Reinecke committed
628
      p0 = data.data() + sv*(iu0-bu0) + iv0-bv0;
Martin Reinecke's avatar
import  
Martin Reinecke committed
629
630
631
      }
  };

Martin Reinecke's avatar
Martin Reinecke committed
632
class ReadHelper2: public Helper2
Martin Reinecke's avatar
import  
Martin Reinecke committed
633
634
635
636
637
638
639
  {
  private:
    vector<complex<double>> data;
    const complex<double> *grid;

    void load()
      {
Martin Reinecke's avatar
Martin Reinecke committed
640
641
642
      int idxu = (bu0+nu)%nu;
      int idxv0 = (bv0+nv)%nv;
      for (int iu=0; iu<su; ++iu)
Martin Reinecke's avatar
import  
Martin Reinecke committed
643
644
        {
        int idxv = idxv0;
Martin Reinecke's avatar
Martin Reinecke committed
645
        for (int iv=0; iv<sv; ++iv)
Martin Reinecke's avatar
import  
Martin Reinecke committed
646
          {
Martin Reinecke's avatar
Martin Reinecke committed
647
          data[iu*sv + iv] = grid[idxu*nv + idxv];
Martin Reinecke's avatar
import  
Martin Reinecke committed
648
649
650
651
652
653
654
655
          if (++idxv>=nv) idxv=0;
          }
        if (++idxu>=nu) idxu=0;
        }
      }

  public:
    const complex<double> *p0;
Martin Reinecke's avatar
Martin Reinecke committed
656
657
    ReadHelper2(int nu_, int nv_, int w_, const complex<double> *grid_)
      : Helper2(nu_, nv_, w_), data(su*sv,0.), grid(grid_), p0(nullptr) {}
Martin Reinecke's avatar
import  
Martin Reinecke committed
658

Martin Reinecke's avatar
Martin Reinecke committed
659
    void prep_read(double u_in, double v_in)
Martin Reinecke's avatar
import  
Martin Reinecke committed
660
      {
Martin Reinecke's avatar
Martin Reinecke committed
661
662
      update(u_in, v_in);
      if (need_to_move())
Martin Reinecke's avatar
import  
Martin Reinecke committed
663
        {
Martin Reinecke's avatar
Martin Reinecke committed
664
        update_position();
Martin Reinecke's avatar
import  
Martin Reinecke committed
665
666
        load();
        }
Martin Reinecke's avatar
Martin Reinecke committed
667
      p0 = data.data() + sv*(iu0-bu0) + iv0-bv0;
Martin Reinecke's avatar
import  
Martin Reinecke committed
668
669
670
      }
  };

Martin Reinecke's avatar
Martin Reinecke committed
671
672
pyarr_c<double> ms2grid(const Baselines &baselines, const GridderConfig &gconf,
  const pyarr_c<uint32_t> &idx_, const pyarr_c<complex<double>> &data_)
Martin Reinecke's avatar
import  
Martin Reinecke committed
673
  {
Martin Reinecke's avatar
Martin Reinecke committed
674
675
676
677
678
679
680
681
682
683
684
  myassert(idx_.ndim()==1, "idx array must be 1D");
  myassert(data_.ndim()==2, "data must be 2D");
  auto data=data_.data();
  myassert(data_.shape(0)==baselines.Nrows(), "bad data dimension");
  myassert(data_.shape(1)==baselines.Nchannels(), "bad data dimension");
  int nvis = idx_.shape(0);
  auto idx = idx_.data();

  size_t nu=gconf.Nu(), nv=gconf.Nv();
  int odim[] = {nu, nv};
  pyarr_c<complex<double>> res(odim);
Martin Reinecke's avatar
import  
Martin Reinecke committed
685
686
687
688
689
  auto grid = res.mutable_data();
  for (int i=0; i<nu*nv; ++i) grid[i] = 0.;

#pragma omp parallel
{
Martin Reinecke's avatar
Martin Reinecke committed
690
  WriteHelper2 hlp(nu, nv, gconf.W(), grid);
Martin Reinecke's avatar
Martin Reinecke committed
691
  double emb = exp(-2*hlp.beta);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
692
693
  const double * RESTRICT ku = hlp.kernel.data();
  const double * RESTRICT kv = hlp.kernel.data()+hlp.w;
Martin Reinecke's avatar
import  
Martin Reinecke committed
694
695
696
697
698

  // Loop over sampling points
#pragma omp for schedule(dynamic,10000)
  for (int ipart=0; ipart<nvis; ++ipart)
    {
Martin Reinecke's avatar
Martin Reinecke committed
699
700
    UV coord = baselines.EffectiveCoord(idx[ipart]);
    hlp.prep_write(coord.u, coord.v);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
701
702
    auto * RESTRICT ptr = hlp.p0;
    int w = hlp.w;
Martin Reinecke's avatar
Martin Reinecke committed
703
    auto v(data[baselines.offset(idx[ipart])]*emb);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
704
    for (int cu=0; cu<w; ++cu)
Martin Reinecke's avatar
import  
Martin Reinecke committed
705
      {
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
706
      complex<double> tmp(v*ku[cu]);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
707
708
709
      for (int cv=0; cv<w; ++cv)
        ptr[cv] += tmp*kv[cv];
      ptr+=hlp.sv;
Martin Reinecke's avatar
import  
Martin Reinecke committed
710
711
712
      }
    }
} // end of parallel region
Martin Reinecke's avatar
Martin Reinecke committed
713
  return complex2hartley(res);
Martin Reinecke's avatar
import  
Martin Reinecke committed
714
  }
Martin Reinecke's avatar
Martin Reinecke committed
715
716
717
pyarr_c<complex<double>> grid2ms(const Baselines &baselines,
  const GridderConfig &gconf, const pyarr_c<uint32_t> &idx_,
  const pyarr_c<double> &grid0_)
Martin Reinecke's avatar
Martin Reinecke committed
718
  {
Martin Reinecke's avatar
Martin Reinecke committed
719
720
  size_t nu=gconf.Nu(), nv=gconf.Nv();
  myassert(idx_.ndim()==1, "idx array must be 1D");
Martin Reinecke's avatar
Martin Reinecke committed
721
  auto grid_ = hartley2complex(grid0_);
Martin Reinecke's avatar
import  
Martin Reinecke committed
722
  auto grid = grid_.data();
Martin Reinecke's avatar
Martin Reinecke committed
723
724
725
726
727
  myassert(grid_.ndim()==2, "data must be 2D");
  myassert(grid_.shape(0)==nu, "bad grid dimension");
  myassert(grid_.shape(1)==nv, "bad grid dimension");
  int nvis = idx_.shape(0);
  auto idx = idx_.data();
Martin Reinecke's avatar
import  
Martin Reinecke committed
728
729

  int odim[] = {nvis};
730
  a_cd_c res(odim);
Martin Reinecke's avatar
import  
Martin Reinecke committed
731
732
733
734
735
  auto vis = res.mutable_data();

  // Loop over sampling points
#pragma omp parallel
{
Martin Reinecke's avatar
Martin Reinecke committed
736
  ReadHelper2 hlp(nu, nv, gconf.W(), grid);
Martin Reinecke's avatar
Martin Reinecke committed
737
  double emb = exp(-2*hlp.beta);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
738
739
  const double * RESTRICT ku = hlp.kernel.data();
  const double * RESTRICT kv = hlp.kernel.data()+hlp.w;
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
740

Martin Reinecke's avatar
import  
Martin Reinecke committed
741
742
743
#pragma omp for schedule(dynamic,10000)
  for (int ipart=0; ipart<nvis; ++ipart)
    {
Martin Reinecke's avatar
Martin Reinecke committed
744
745
    UV coord = baselines.EffectiveCoord(idx[ipart]);
    hlp.prep_read(coord.u, coord.v);
Martin Reinecke's avatar
import  
Martin Reinecke committed
746
    complex<double> r = 0.;
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
747
748
749
    auto * RESTRICT ptr = hlp.p0;
    int w = hlp.w;
    for (int cu=0; cu<w; ++cu)
Martin Reinecke's avatar
import  
Martin Reinecke committed
750
      {
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
751
      complex<double> tmp(0.);
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
752
753
754
755
      for (int cv=0; cv<w; ++cv)
        tmp += ptr[cv] * kv[cv];
      r += tmp*ku[cu];
      ptr += hlp.sv;
Martin Reinecke's avatar
import  
Martin Reinecke committed
756
      }
Martin Reinecke's avatar
Martin Reinecke committed
757
    vis[baselines.offset(idx[ipart])] = r*emb;
Martin Reinecke's avatar
import  
Martin Reinecke committed
758
759
760
761
762
763
764
765
    }
}
  return res;
  }
} // unnamed namespace

PYBIND11_MODULE(nifty_gridder, m)
  {
766
767
  using namespace pybind11::literals;

Martin Reinecke's avatar
Martin Reinecke committed
768
  py::class_<Baselines> (m, "Baselines")
769
    .def(py::init<a_d_c, a_d_c>(), "coord"_a, "scaling"_a)
Martin Reinecke's avatar
Martin Reinecke committed
770
    .def("getIndices", &Baselines::getIndices);
771
772
  py::class_<GridderConfig> (m, "GridderConfig")
    .def(py::init<size_t, size_t, double, double, double>(),"nxdirty"_a,
Martin Reinecke's avatar
Martin Reinecke committed
773
774
775
776
777
778
779
780
781
      "nydirty"_a, "epsilon"_a, "urange"_a, "vrange"_a)
    .def("reorderIndices", &GridderConfig::reorderIndices,
      "idx"_a, "baselines"_a)
    .def("Nu", &GridderConfig::Nu)
    .def("Nv", &GridderConfig::Nv)
    .def("U_corrections", &GridderConfig::U_corrections)
    .def("V_corrections", &GridderConfig::V_corrections);
  m.def ("ms2grid",&ms2grid);
  m.def ("grid2ms",&grid2ms);
Martin Reinecke's avatar
import  
Martin Reinecke committed
782
  }