• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2015 Jianwei Cui <thucjw@gmail.com>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_CXX11_TENSOR_TENSOR_FFT_H
11 #define EIGEN_CXX11_TENSOR_TENSOR_FFT_H
12 
13 // This code requires the ability to initialize arrays of constant
14 // values directly inside a class.
15 #if __cplusplus >= 201103L || EIGEN_COMP_MSVC >= 1900
16 
17 namespace Eigen {
18 
19 /** \class TensorFFT
20   * \ingroup CXX11_Tensor_Module
21   *
22   * \brief Tensor FFT class.
23   *
24   * TODO:
25   * Vectorize the Cooley Tukey and the Bluestein algorithm
26   * Add support for multithreaded evaluation
27   * Improve the performance on GPU
28   */
29 
30 template <bool NeedUprade> struct MakeComplex {
31   template <typename T>
32   EIGEN_DEVICE_FUNC
operatorMakeComplex33   T operator() (const T& val) const { return val; }
34 };
35 
36 template <> struct MakeComplex<true> {
37   template <typename T>
38   EIGEN_DEVICE_FUNC
39   std::complex<T> operator() (const T& val) const { return std::complex<T>(val, 0); }
40 };
41 
42 template <> struct MakeComplex<false> {
43   template <typename T>
44   EIGEN_DEVICE_FUNC
45   std::complex<T> operator() (const std::complex<T>& val) const { return val; }
46 };
47 
48 template <int ResultType> struct PartOf {
49   template <typename T> T operator() (const T& val) const { return val; }
50 };
51 
52 template <> struct PartOf<RealPart> {
53   template <typename T> T operator() (const std::complex<T>& val) const { return val.real(); }
54 };
55 
56 template <> struct PartOf<ImagPart> {
57   template <typename T> T operator() (const std::complex<T>& val) const { return val.imag(); }
58 };
59 
60 namespace internal {
61 template <typename FFT, typename XprType, int FFTResultType, int FFTDir>
62 struct traits<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir> > : public traits<XprType> {
63   typedef traits<XprType> XprTraits;
64   typedef typename NumTraits<typename XprTraits::Scalar>::Real RealScalar;
65   typedef typename std::complex<RealScalar> ComplexScalar;
66   typedef typename XprTraits::Scalar InputScalar;
67   typedef typename conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
68   typedef typename XprTraits::StorageKind StorageKind;
69   typedef typename XprTraits::Index Index;
70   typedef typename XprType::Nested Nested;
71   typedef typename remove_reference<Nested>::type _Nested;
72   static const int NumDimensions = XprTraits::NumDimensions;
73   static const int Layout = XprTraits::Layout;
74 };
75 
76 template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
77 struct eval<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>, Eigen::Dense> {
78   typedef const TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>& type;
79 };
80 
81 template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
82 struct nested<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>, 1, typename eval<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection> >::type> {
83   typedef TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection> type;
84 };
85 
86 }  // end namespace internal
87 
88 template <typename FFT, typename XprType, int FFTResultType, int FFTDir>
89 class TensorFFTOp : public TensorBase<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir>, ReadOnlyAccessors> {
90  public:
91   typedef typename Eigen::internal::traits<TensorFFTOp>::Scalar Scalar;
92   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
93   typedef typename std::complex<RealScalar> ComplexScalar;
94   typedef typename internal::conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
95   typedef OutputScalar CoeffReturnType;
96   typedef typename Eigen::internal::nested<TensorFFTOp>::type Nested;
97   typedef typename Eigen::internal::traits<TensorFFTOp>::StorageKind StorageKind;
98   typedef typename Eigen::internal::traits<TensorFFTOp>::Index Index;
99 
100   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorFFTOp(const XprType& expr, const FFT& fft)
101       : m_xpr(expr), m_fft(fft) {}
102 
103   EIGEN_DEVICE_FUNC
104   const FFT& fft() const { return m_fft; }
105 
106   EIGEN_DEVICE_FUNC
107   const typename internal::remove_all<typename XprType::Nested>::type& expression() const {
108     return m_xpr;
109   }
110 
111  protected:
112   typename XprType::Nested m_xpr;
113   const FFT m_fft;
114 };
115 
116 // Eval as rvalue
117 template <typename FFT, typename ArgType, typename Device, int FFTResultType, int FFTDir>
118 struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, Device> {
119   typedef TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir> XprType;
120   typedef typename XprType::Index Index;
121   static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
122   typedef DSizes<Index, NumDims> Dimensions;
123   typedef typename XprType::Scalar Scalar;
124   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
125   typedef typename std::complex<RealScalar> ComplexScalar;
126   typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
127   typedef internal::traits<XprType> XprTraits;
128   typedef typename XprTraits::Scalar InputScalar;
129   typedef typename internal::conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
130   typedef OutputScalar CoeffReturnType;
131   typedef typename PacketType<OutputScalar, Device>::type PacketReturnType;
132   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
133 
134   enum {
135     IsAligned = false,
136     PacketAccess = true,
137     BlockAccess = false,
138     Layout = TensorEvaluator<ArgType, Device>::Layout,
139     CoordAccess = false,
140     RawAccess = false
141   };
142 
143   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_fft(op.fft()), m_impl(op.expression(), device), m_data(NULL), m_device(device) {
144     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
145     for (int i = 0; i < NumDims; ++i) {
146       eigen_assert(input_dims[i] > 0);
147       m_dimensions[i] = input_dims[i];
148     }
149 
150     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
151       m_strides[0] = 1;
152       for (int i = 1; i < NumDims; ++i) {
153         m_strides[i] = m_strides[i - 1] * m_dimensions[i - 1];
154       }
155     } else {
156       m_strides[NumDims - 1] = 1;
157       for (int i = NumDims - 2; i >= 0; --i) {
158         m_strides[i] = m_strides[i + 1] * m_dimensions[i + 1];
159       }
160     }
161     m_size = m_dimensions.TotalSize();
162   }
163 
164   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
165     return m_dimensions;
166   }
167 
168   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(OutputScalar* data) {
169     m_impl.evalSubExprsIfNeeded(NULL);
170     if (data) {
171       evalToBuf(data);
172       return false;
173     } else {
174       m_data = (CoeffReturnType*)m_device.allocate(sizeof(CoeffReturnType) * m_size);
175       evalToBuf(m_data);
176       return true;
177     }
178   }
179 
180   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
181     if (m_data) {
182       m_device.deallocate(m_data);
183       m_data = NULL;
184     }
185     m_impl.cleanup();
186   }
187 
188   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index) const {
189     return m_data[index];
190   }
191 
192   template <int LoadMode>
193   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType
194   packet(Index index) const {
195     return internal::ploadt<PacketReturnType, LoadMode>(m_data + index);
196   }
197 
198   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
199   costPerCoeff(bool vectorized) const {
200     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
201   }
202 
203   EIGEN_DEVICE_FUNC Scalar* data() const { return m_data; }
204 
205 
206  private:
207   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalToBuf(OutputScalar* data) {
208     const bool write_to_out = internal::is_same<OutputScalar, ComplexScalar>::value;
209     ComplexScalar* buf = write_to_out ? (ComplexScalar*)data : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * m_size);
210 
211     for (Index i = 0; i < m_size; ++i) {
212       buf[i] = MakeComplex<internal::is_same<InputScalar, RealScalar>::value>()(m_impl.coeff(i));
213     }
214 
215     for (size_t i = 0; i < m_fft.size(); ++i) {
216       Index dim = m_fft[i];
217       eigen_assert(dim >= 0 && dim < NumDims);
218       Index line_len = m_dimensions[dim];
219       eigen_assert(line_len >= 1);
220       ComplexScalar* line_buf = (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * line_len);
221       const bool is_power_of_two = isPowerOfTwo(line_len);
222       const Index good_composite = is_power_of_two ? 0 : findGoodComposite(line_len);
223       const Index log_len = is_power_of_two ? getLog2(line_len) : getLog2(good_composite);
224 
225       ComplexScalar* a = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * good_composite);
226       ComplexScalar* b = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * good_composite);
227       ComplexScalar* pos_j_base_powered = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * (line_len + 1));
228       if (!is_power_of_two) {
229         // Compute twiddle factors
230         //   t_n = exp(sqrt(-1) * pi * n^2 / line_len)
231         // for n = 0, 1,..., line_len-1.
232         // For n > 2 we use the recurrence t_n = t_{n-1}^2 / t_{n-2} * t_1^2
233         pos_j_base_powered[0] = ComplexScalar(1, 0);
234         if (line_len > 1) {
235           const RealScalar pi_over_len(EIGEN_PI / line_len);
236           const ComplexScalar pos_j_base = ComplexScalar(
237 	       std::cos(pi_over_len), std::sin(pi_over_len));
238           pos_j_base_powered[1] = pos_j_base;
239           if (line_len > 2) {
240             const ComplexScalar pos_j_base_sq = pos_j_base * pos_j_base;
241             for (int j = 2; j < line_len + 1; ++j) {
242               pos_j_base_powered[j] = pos_j_base_powered[j - 1] *
243                                       pos_j_base_powered[j - 1] /
244                                       pos_j_base_powered[j - 2] * pos_j_base_sq;
245             }
246           }
247         }
248       }
249 
250       for (Index partial_index = 0; partial_index < m_size / line_len; ++partial_index) {
251         const Index base_offset = getBaseOffsetFromIndex(partial_index, dim);
252 
253         // get data into line_buf
254         const Index stride = m_strides[dim];
255         if (stride == 1) {
256           memcpy(line_buf, &buf[base_offset], line_len*sizeof(ComplexScalar));
257         } else {
258           Index offset = base_offset;
259           for (int j = 0; j < line_len; ++j, offset += stride) {
260             line_buf[j] = buf[offset];
261           }
262         }
263 
264         // processs the line
265         if (is_power_of_two) {
266           processDataLineCooleyTukey(line_buf, line_len, log_len);
267         }
268         else {
269           processDataLineBluestein(line_buf, line_len, good_composite, log_len, a, b, pos_j_base_powered);
270         }
271 
272         // write back
273         if (FFTDir == FFT_FORWARD && stride == 1) {
274           memcpy(&buf[base_offset], line_buf, line_len*sizeof(ComplexScalar));
275         } else {
276           Index offset = base_offset;
277           const ComplexScalar div_factor =  ComplexScalar(1.0 / line_len, 0);
278           for (int j = 0; j < line_len; ++j, offset += stride) {
279              buf[offset] = (FFTDir == FFT_FORWARD) ? line_buf[j] : line_buf[j] * div_factor;
280           }
281         }
282       }
283       m_device.deallocate(line_buf);
284       if (!is_power_of_two) {
285         m_device.deallocate(a);
286         m_device.deallocate(b);
287         m_device.deallocate(pos_j_base_powered);
288       }
289     }
290 
291     if(!write_to_out) {
292       for (Index i = 0; i < m_size; ++i) {
293         data[i] = PartOf<FFTResultType>()(buf[i]);
294       }
295       m_device.deallocate(buf);
296     }
297   }
298 
299   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static bool isPowerOfTwo(Index x) {
300     eigen_assert(x > 0);
301     return !(x & (x - 1));
302   }
303 
304   // The composite number for padding, used in Bluestein's FFT algorithm
305   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static Index findGoodComposite(Index n) {
306     Index i = 2;
307     while (i < 2 * n - 1) i *= 2;
308     return i;
309   }
310 
311   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static Index getLog2(Index m) {
312     Index log2m = 0;
313     while (m >>= 1) log2m++;
314     return log2m;
315   }
316 
317   // Call Cooley Tukey algorithm directly, data length must be power of 2
318   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void processDataLineCooleyTukey(ComplexScalar* line_buf, Index line_len, Index log_len) {
319     eigen_assert(isPowerOfTwo(line_len));
320     scramble_FFT(line_buf, line_len);
321     compute_1D_Butterfly<FFTDir>(line_buf, line_len, log_len);
322   }
323 
324   // Call Bluestein's FFT algorithm, m is a good composite number greater than (2 * n - 1), used as the padding length
325   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void processDataLineBluestein(ComplexScalar* line_buf, Index line_len, Index good_composite, Index log_len, ComplexScalar* a, ComplexScalar* b, const ComplexScalar* pos_j_base_powered) {
326     Index n = line_len;
327     Index m = good_composite;
328     ComplexScalar* data = line_buf;
329 
330     for (Index i = 0; i < n; ++i) {
331       if(FFTDir == FFT_FORWARD) {
332         a[i] = data[i] * numext::conj(pos_j_base_powered[i]);
333       }
334       else {
335         a[i] = data[i] * pos_j_base_powered[i];
336       }
337     }
338     for (Index i = n; i < m; ++i) {
339       a[i] = ComplexScalar(0, 0);
340     }
341 
342     for (Index i = 0; i < n; ++i) {
343       if(FFTDir == FFT_FORWARD) {
344         b[i] = pos_j_base_powered[i];
345       }
346       else {
347         b[i] = numext::conj(pos_j_base_powered[i]);
348       }
349     }
350     for (Index i = n; i < m - n; ++i) {
351       b[i] = ComplexScalar(0, 0);
352     }
353     for (Index i = m - n; i < m; ++i) {
354       if(FFTDir == FFT_FORWARD) {
355         b[i] = pos_j_base_powered[m-i];
356       }
357       else {
358         b[i] = numext::conj(pos_j_base_powered[m-i]);
359       }
360     }
361 
362     scramble_FFT(a, m);
363     compute_1D_Butterfly<FFT_FORWARD>(a, m, log_len);
364 
365     scramble_FFT(b, m);
366     compute_1D_Butterfly<FFT_FORWARD>(b, m, log_len);
367 
368     for (Index i = 0; i < m; ++i) {
369       a[i] *= b[i];
370     }
371 
372     scramble_FFT(a, m);
373     compute_1D_Butterfly<FFT_REVERSE>(a, m, log_len);
374 
375     //Do the scaling after ifft
376     for (Index i = 0; i < m; ++i) {
377       a[i] /= m;
378     }
379 
380     for (Index i = 0; i < n; ++i) {
381       if(FFTDir == FFT_FORWARD) {
382         data[i] = a[i] * numext::conj(pos_j_base_powered[i]);
383       }
384       else {
385         data[i] = a[i] * pos_j_base_powered[i];
386       }
387     }
388   }
389 
390   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static void scramble_FFT(ComplexScalar* data, Index n) {
391     eigen_assert(isPowerOfTwo(n));
392     Index j = 1;
393     for (Index i = 1; i < n; ++i){
394       if (j > i) {
395         std::swap(data[j-1], data[i-1]);
396       }
397       Index m = n >> 1;
398       while (m >= 2 && j > m) {
399         j -= m;
400         m >>= 1;
401       }
402       j += m;
403     }
404   }
405 
406   template <int Dir>
407   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_2(ComplexScalar* data) {
408     ComplexScalar tmp = data[1];
409     data[1] = data[0] - data[1];
410     data[0] += tmp;
411   }
412 
413   template <int Dir>
414   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_4(ComplexScalar* data) {
415     ComplexScalar tmp[4];
416     tmp[0] = data[0] + data[1];
417     tmp[1] = data[0] - data[1];
418     tmp[2] = data[2] + data[3];
419     if (Dir == FFT_FORWARD) {
420       tmp[3] = ComplexScalar(0.0, -1.0) * (data[2] - data[3]);
421     } else {
422       tmp[3] = ComplexScalar(0.0, 1.0) * (data[2] - data[3]);
423     }
424     data[0] = tmp[0] + tmp[2];
425     data[1] = tmp[1] + tmp[3];
426     data[2] = tmp[0] - tmp[2];
427     data[3] = tmp[1] - tmp[3];
428   }
429 
430   template <int Dir>
431   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_8(ComplexScalar* data) {
432     ComplexScalar tmp_1[8];
433     ComplexScalar tmp_2[8];
434 
435     tmp_1[0] = data[0] + data[1];
436     tmp_1[1] = data[0] - data[1];
437     tmp_1[2] = data[2] + data[3];
438     if (Dir == FFT_FORWARD) {
439       tmp_1[3] = (data[2] - data[3]) * ComplexScalar(0, -1);
440     } else {
441       tmp_1[3] = (data[2] - data[3]) * ComplexScalar(0, 1);
442     }
443     tmp_1[4] = data[4] + data[5];
444     tmp_1[5] = data[4] - data[5];
445     tmp_1[6] = data[6] + data[7];
446     if (Dir == FFT_FORWARD) {
447       tmp_1[7] = (data[6] - data[7]) * ComplexScalar(0, -1);
448     } else {
449       tmp_1[7] = (data[6] - data[7]) * ComplexScalar(0, 1);
450     }
451     tmp_2[0] = tmp_1[0] + tmp_1[2];
452     tmp_2[1] = tmp_1[1] + tmp_1[3];
453     tmp_2[2] = tmp_1[0] - tmp_1[2];
454     tmp_2[3] = tmp_1[1] - tmp_1[3];
455     tmp_2[4] = tmp_1[4] + tmp_1[6];
456 // SQRT2DIV2 = sqrt(2)/2
457 #define SQRT2DIV2 0.7071067811865476
458     if (Dir == FFT_FORWARD) {
459       tmp_2[5] = (tmp_1[5] + tmp_1[7]) * ComplexScalar(SQRT2DIV2, -SQRT2DIV2);
460       tmp_2[6] = (tmp_1[4] - tmp_1[6]) * ComplexScalar(0, -1);
461       tmp_2[7] = (tmp_1[5] - tmp_1[7]) * ComplexScalar(-SQRT2DIV2, -SQRT2DIV2);
462     } else {
463       tmp_2[5] = (tmp_1[5] + tmp_1[7]) * ComplexScalar(SQRT2DIV2, SQRT2DIV2);
464       tmp_2[6] = (tmp_1[4] - tmp_1[6]) * ComplexScalar(0, 1);
465       tmp_2[7] = (tmp_1[5] - tmp_1[7]) * ComplexScalar(-SQRT2DIV2, SQRT2DIV2);
466     }
467     data[0] = tmp_2[0] + tmp_2[4];
468     data[1] = tmp_2[1] + tmp_2[5];
469     data[2] = tmp_2[2] + tmp_2[6];
470     data[3] = tmp_2[3] + tmp_2[7];
471     data[4] = tmp_2[0] - tmp_2[4];
472     data[5] = tmp_2[1] - tmp_2[5];
473     data[6] = tmp_2[2] - tmp_2[6];
474     data[7] = tmp_2[3] - tmp_2[7];
475   }
476 
477   template <int Dir>
478   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_1D_merge(
479       ComplexScalar* data, Index n, Index n_power_of_2) {
480     // Original code:
481     // RealScalar wtemp = std::sin(M_PI/n);
482     // RealScalar wpi =  -std::sin(2 * M_PI/n);
483     const RealScalar wtemp = m_sin_PI_div_n_LUT[n_power_of_2];
484     const RealScalar wpi = (Dir == FFT_FORWARD)
485                                ? m_minus_sin_2_PI_div_n_LUT[n_power_of_2]
486                                : -m_minus_sin_2_PI_div_n_LUT[n_power_of_2];
487 
488     const ComplexScalar wp(wtemp, wpi);
489     const ComplexScalar wp_one = wp + ComplexScalar(1, 0);
490     const ComplexScalar wp_one_2 = wp_one * wp_one;
491     const ComplexScalar wp_one_3 = wp_one_2 * wp_one;
492     const ComplexScalar wp_one_4 = wp_one_3 * wp_one;
493     const Index n2 = n / 2;
494     ComplexScalar w(1.0, 0.0);
495     for (Index i = 0; i < n2; i += 4) {
496        ComplexScalar temp0(data[i + n2] * w);
497        ComplexScalar temp1(data[i + 1 + n2] * w * wp_one);
498        ComplexScalar temp2(data[i + 2 + n2] * w * wp_one_2);
499        ComplexScalar temp3(data[i + 3 + n2] * w * wp_one_3);
500        w = w * wp_one_4;
501 
502        data[i + n2] = data[i] - temp0;
503        data[i] += temp0;
504 
505        data[i + 1 + n2] = data[i + 1] - temp1;
506        data[i + 1] += temp1;
507 
508        data[i + 2 + n2] = data[i + 2] - temp2;
509        data[i + 2] += temp2;
510 
511        data[i + 3 + n2] = data[i + 3] - temp3;
512        data[i + 3] += temp3;
513     }
514   }
515 
516  template <int Dir>
517   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_1D_Butterfly(
518       ComplexScalar* data, Index n, Index n_power_of_2) {
519     eigen_assert(isPowerOfTwo(n));
520     if (n > 8) {
521       compute_1D_Butterfly<Dir>(data, n / 2, n_power_of_2 - 1);
522       compute_1D_Butterfly<Dir>(data + n / 2, n / 2, n_power_of_2 - 1);
523       butterfly_1D_merge<Dir>(data, n, n_power_of_2);
524     } else if (n == 8) {
525       butterfly_8<Dir>(data);
526     } else if (n == 4) {
527       butterfly_4<Dir>(data);
528     } else if (n == 2) {
529       butterfly_2<Dir>(data);
530     }
531   }
532 
533   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index getBaseOffsetFromIndex(Index index, Index omitted_dim) const {
534     Index result = 0;
535 
536     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
537       for (int i = NumDims - 1; i > omitted_dim; --i) {
538         const Index partial_m_stride = m_strides[i] / m_dimensions[omitted_dim];
539         const Index idx = index / partial_m_stride;
540         index -= idx * partial_m_stride;
541         result += idx * m_strides[i];
542       }
543       result += index;
544     }
545     else {
546       for (Index i = 0; i < omitted_dim; ++i) {
547         const Index partial_m_stride = m_strides[i] / m_dimensions[omitted_dim];
548         const Index idx = index / partial_m_stride;
549         index -= idx * partial_m_stride;
550         result += idx * m_strides[i];
551       }
552       result += index;
553     }
554     // Value of index_coords[omitted_dim] is not determined to this step
555     return result;
556   }
557 
558   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index getIndexFromOffset(Index base, Index omitted_dim, Index offset) const {
559     Index result = base + offset * m_strides[omitted_dim] ;
560     return result;
561   }
562 
563  protected:
564   Index m_size;
565   const FFT& m_fft;
566   Dimensions m_dimensions;
567   array<Index, NumDims> m_strides;
568   TensorEvaluator<ArgType, Device> m_impl;
569   CoeffReturnType* m_data;
570   const Device& m_device;
571 
572   // This will support a maximum FFT size of 2^32 for each dimension
573   // m_sin_PI_div_n_LUT[i] = (-2) * std::sin(M_PI / std::pow(2,i)) ^ 2;
574   const RealScalar m_sin_PI_div_n_LUT[32] = {
575     RealScalar(0.0),
576     RealScalar(-2),
577     RealScalar(-0.999999999999999),
578     RealScalar(-0.292893218813453),
579     RealScalar(-0.0761204674887130),
580     RealScalar(-0.0192147195967696),
581     RealScalar(-0.00481527332780311),
582     RealScalar(-0.00120454379482761),
583     RealScalar(-3.01181303795779e-04),
584     RealScalar(-7.52981608554592e-05),
585     RealScalar(-1.88247173988574e-05),
586     RealScalar(-4.70619042382852e-06),
587     RealScalar(-1.17654829809007e-06),
588     RealScalar(-2.94137117780840e-07),
589     RealScalar(-7.35342821488550e-08),
590     RealScalar(-1.83835707061916e-08),
591     RealScalar(-4.59589268710903e-09),
592     RealScalar(-1.14897317243732e-09),
593     RealScalar(-2.87243293150586e-10),
594     RealScalar( -7.18108232902250e-11),
595     RealScalar(-1.79527058227174e-11),
596     RealScalar(-4.48817645568941e-12),
597     RealScalar(-1.12204411392298e-12),
598     RealScalar(-2.80511028480785e-13),
599     RealScalar(-7.01277571201985e-14),
600     RealScalar(-1.75319392800498e-14),
601     RealScalar(-4.38298482001247e-15),
602     RealScalar(-1.09574620500312e-15),
603     RealScalar(-2.73936551250781e-16),
604     RealScalar(-6.84841378126949e-17),
605     RealScalar(-1.71210344531737e-17),
606     RealScalar(-4.28025861329343e-18)
607   };
608 
609   // m_minus_sin_2_PI_div_n_LUT[i] = -std::sin(2 * M_PI / std::pow(2,i));
610   const RealScalar m_minus_sin_2_PI_div_n_LUT[32] = {
611     RealScalar(0.0),
612     RealScalar(0.0),
613     RealScalar(-1.00000000000000e+00),
614     RealScalar(-7.07106781186547e-01),
615     RealScalar(-3.82683432365090e-01),
616     RealScalar(-1.95090322016128e-01),
617     RealScalar(-9.80171403295606e-02),
618     RealScalar(-4.90676743274180e-02),
619     RealScalar(-2.45412285229123e-02),
620     RealScalar(-1.22715382857199e-02),
621     RealScalar(-6.13588464915448e-03),
622     RealScalar(-3.06795676296598e-03),
623     RealScalar(-1.53398018628477e-03),
624     RealScalar(-7.66990318742704e-04),
625     RealScalar(-3.83495187571396e-04),
626     RealScalar(-1.91747597310703e-04),
627     RealScalar(-9.58737990959773e-05),
628     RealScalar(-4.79368996030669e-05),
629     RealScalar(-2.39684498084182e-05),
630     RealScalar(-1.19842249050697e-05),
631     RealScalar(-5.99211245264243e-06),
632     RealScalar(-2.99605622633466e-06),
633     RealScalar(-1.49802811316901e-06),
634     RealScalar(-7.49014056584716e-07),
635     RealScalar(-3.74507028292384e-07),
636     RealScalar(-1.87253514146195e-07),
637     RealScalar(-9.36267570730981e-08),
638     RealScalar(-4.68133785365491e-08),
639     RealScalar(-2.34066892682746e-08),
640     RealScalar(-1.17033446341373e-08),
641     RealScalar(-5.85167231706864e-09),
642     RealScalar(-2.92583615853432e-09)
643   };
644 };
645 
646 }  // end namespace Eigen
647 
648 #endif  // EIGEN_HAS_CONSTEXPR
649 
650 
651 #endif  // EIGEN_CXX11_TENSOR_TENSOR_FFT_H
652