xref: /aosp_15_r20/external/eigen/unsupported/Eigen/CXX11/src/Tensor/TensorFFT.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2015 Jianwei Cui <[email protected]>
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 namespace Eigen {
14 
15 /** \class TensorFFT
16   * \ingroup CXX11_Tensor_Module
17   *
18   * \brief Tensor FFT class.
19   *
20   * TODO:
21   * Vectorize the Cooley Tukey and the Bluestein algorithm
22   * Add support for multithreaded evaluation
23   * Improve the performance on GPU
24   */
25 
26 template <bool NeedUprade> struct MakeComplex {
27   template <typename T>
28   EIGEN_DEVICE_FUNC
operatorMakeComplex29   T operator() (const T& val) const { return val; }
30 };
31 
32 template <> struct MakeComplex<true> {
33   template <typename T>
34   EIGEN_DEVICE_FUNC
35   std::complex<T> operator() (const T& val) const { return std::complex<T>(val, 0); }
36 };
37 
38 template <> struct MakeComplex<false> {
39   template <typename T>
40   EIGEN_DEVICE_FUNC
41   std::complex<T> operator() (const std::complex<T>& val) const { return val; }
42 };
43 
44 template <int ResultType> struct PartOf {
45   template <typename T> T operator() (const T& val) const { return val; }
46 };
47 
48 template <> struct PartOf<RealPart> {
49   template <typename T> T operator() (const std::complex<T>& val) const { return val.real(); }
50 };
51 
52 template <> struct PartOf<ImagPart> {
53   template <typename T> T operator() (const std::complex<T>& val) const { return val.imag(); }
54 };
55 
56 namespace internal {
57 template <typename FFT, typename XprType, int FFTResultType, int FFTDir>
58 struct traits<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir> > : public traits<XprType> {
59   typedef traits<XprType> XprTraits;
60   typedef typename NumTraits<typename XprTraits::Scalar>::Real RealScalar;
61   typedef typename std::complex<RealScalar> ComplexScalar;
62   typedef typename XprTraits::Scalar InputScalar;
63   typedef typename conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
64   typedef typename XprTraits::StorageKind StorageKind;
65   typedef typename XprTraits::Index Index;
66   typedef typename XprType::Nested Nested;
67   typedef typename remove_reference<Nested>::type _Nested;
68   static const int NumDimensions = XprTraits::NumDimensions;
69   static const int Layout = XprTraits::Layout;
70   typedef typename traits<XprType>::PointerType PointerType;
71 };
72 
73 template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
74 struct eval<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>, Eigen::Dense> {
75   typedef const TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>& type;
76 };
77 
78 template <typename FFT, typename XprType, int FFTResultType, int FFTDirection>
79 struct nested<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection>, 1, typename eval<TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection> >::type> {
80   typedef TensorFFTOp<FFT, XprType, FFTResultType, FFTDirection> type;
81 };
82 
83 }  // end namespace internal
84 
85 template <typename FFT, typename XprType, int FFTResultType, int FFTDir>
86 class TensorFFTOp : public TensorBase<TensorFFTOp<FFT, XprType, FFTResultType, FFTDir>, ReadOnlyAccessors> {
87  public:
88   typedef typename Eigen::internal::traits<TensorFFTOp>::Scalar Scalar;
89   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
90   typedef typename std::complex<RealScalar> ComplexScalar;
91   typedef typename internal::conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
92   typedef OutputScalar CoeffReturnType;
93   typedef typename Eigen::internal::nested<TensorFFTOp>::type Nested;
94   typedef typename Eigen::internal::traits<TensorFFTOp>::StorageKind StorageKind;
95   typedef typename Eigen::internal::traits<TensorFFTOp>::Index Index;
96 
97   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorFFTOp(const XprType& expr, const FFT& fft)
98       : m_xpr(expr), m_fft(fft) {}
99 
100   EIGEN_DEVICE_FUNC
101   const FFT& fft() const { return m_fft; }
102 
103   EIGEN_DEVICE_FUNC
104   const typename internal::remove_all<typename XprType::Nested>::type& expression() const {
105     return m_xpr;
106   }
107 
108  protected:
109   typename XprType::Nested m_xpr;
110   const FFT m_fft;
111 };
112 
113 // Eval as rvalue
114 template <typename FFT, typename ArgType, typename Device, int FFTResultType, int FFTDir>
115 struct TensorEvaluator<const TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir>, Device> {
116   typedef TensorFFTOp<FFT, ArgType, FFTResultType, FFTDir> XprType;
117   typedef typename XprType::Index Index;
118   static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
119   typedef DSizes<Index, NumDims> Dimensions;
120   typedef typename XprType::Scalar Scalar;
121   typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
122   typedef typename std::complex<RealScalar> ComplexScalar;
123   typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
124   typedef internal::traits<XprType> XprTraits;
125   typedef typename XprTraits::Scalar InputScalar;
126   typedef typename internal::conditional<FFTResultType == RealPart || FFTResultType == ImagPart, RealScalar, ComplexScalar>::type OutputScalar;
127   typedef OutputScalar CoeffReturnType;
128   typedef typename PacketType<OutputScalar, Device>::type PacketReturnType;
129   static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
130     typedef StorageMemory<CoeffReturnType, Device> Storage;
131   typedef typename Storage::Type EvaluatorPointerType;
132 
133   enum {
134     IsAligned = false,
135     PacketAccess = true,
136     BlockAccess = false,
137     PreferBlockAccess = false,
138     Layout = TensorEvaluator<ArgType, Device>::Layout,
139     CoordAccess = false,
140     RawAccess = false
141   };
142 
143   //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
144   typedef internal::TensorBlockNotImplemented TensorBlock;
145   //===--------------------------------------------------------------------===//
146 
147   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) {
148     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
149     for (int i = 0; i < NumDims; ++i) {
150       eigen_assert(input_dims[i] > 0);
151       m_dimensions[i] = input_dims[i];
152     }
153 
154     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
155       m_strides[0] = 1;
156       for (int i = 1; i < NumDims; ++i) {
157         m_strides[i] = m_strides[i - 1] * m_dimensions[i - 1];
158       }
159     } else {
160       m_strides[NumDims - 1] = 1;
161       for (int i = NumDims - 2; i >= 0; --i) {
162         m_strides[i] = m_strides[i + 1] * m_dimensions[i + 1];
163       }
164     }
165     m_size = m_dimensions.TotalSize();
166   }
167 
168   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const {
169     return m_dimensions;
170   }
171 
172   EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
173     m_impl.evalSubExprsIfNeeded(NULL);
174     if (data) {
175       evalToBuf(data);
176       return false;
177     } else {
178       m_data = (EvaluatorPointerType)m_device.get((CoeffReturnType*)(m_device.allocate_temp(sizeof(CoeffReturnType) * m_size)));
179       evalToBuf(m_data);
180       return true;
181     }
182   }
183 
184   EIGEN_STRONG_INLINE void cleanup() {
185     if (m_data) {
186       m_device.deallocate(m_data);
187       m_data = NULL;
188     }
189     m_impl.cleanup();
190   }
191 
192   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index) const {
193     return m_data[index];
194   }
195 
196   template <int LoadMode>
197   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType
198   packet(Index index) const {
199     return internal::ploadt<PacketReturnType, LoadMode>(m_data + index);
200   }
201 
202   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
203   costPerCoeff(bool vectorized) const {
204     return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
205   }
206 
207   EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_data; }
208 #ifdef EIGEN_USE_SYCL
209   // binding placeholder accessors to a command group handler for SYCL
210   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
211     m_data.bind(cgh);
212   }
213 #endif
214 
215  private:
216   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalToBuf(EvaluatorPointerType data) {
217     const bool write_to_out = internal::is_same<OutputScalar, ComplexScalar>::value;
218     ComplexScalar* buf = write_to_out ? (ComplexScalar*)data : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * m_size);
219 
220     for (Index i = 0; i < m_size; ++i) {
221       buf[i] = MakeComplex<internal::is_same<InputScalar, RealScalar>::value>()(m_impl.coeff(i));
222     }
223 
224     for (size_t i = 0; i < m_fft.size(); ++i) {
225       Index dim = m_fft[i];
226       eigen_assert(dim >= 0 && dim < NumDims);
227       Index line_len = m_dimensions[dim];
228       eigen_assert(line_len >= 1);
229       ComplexScalar* line_buf = (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * line_len);
230       const bool is_power_of_two = isPowerOfTwo(line_len);
231       const Index good_composite = is_power_of_two ? 0 : findGoodComposite(line_len);
232       const Index log_len = is_power_of_two ? getLog2(line_len) : getLog2(good_composite);
233 
234       ComplexScalar* a = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * good_composite);
235       ComplexScalar* b = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * good_composite);
236       ComplexScalar* pos_j_base_powered = is_power_of_two ? NULL : (ComplexScalar*)m_device.allocate(sizeof(ComplexScalar) * (line_len + 1));
237       if (!is_power_of_two) {
238         // Compute twiddle factors
239         //   t_n = exp(sqrt(-1) * pi * n^2 / line_len)
240         // for n = 0, 1,..., line_len-1.
241         // For n > 2 we use the recurrence t_n = t_{n-1}^2 / t_{n-2} * t_1^2
242 
243         // The recurrence is correct in exact arithmetic, but causes
244         // numerical issues for large transforms, especially in
245         // single-precision floating point.
246         //
247         // pos_j_base_powered[0] = ComplexScalar(1, 0);
248         // if (line_len > 1) {
249         //   const ComplexScalar pos_j_base = ComplexScalar(
250         //       numext::cos(M_PI / line_len), numext::sin(M_PI / line_len));
251         //   pos_j_base_powered[1] = pos_j_base;
252         //   if (line_len > 2) {
253         //     const ComplexScalar pos_j_base_sq = pos_j_base * pos_j_base;
254         //     for (int i = 2; i < line_len + 1; ++i) {
255         //       pos_j_base_powered[i] = pos_j_base_powered[i - 1] *
256         //           pos_j_base_powered[i - 1] /
257         //           pos_j_base_powered[i - 2] *
258         //           pos_j_base_sq;
259         //     }
260         //   }
261         // }
262         // TODO(rmlarsen): Find a way to use Eigen's vectorized sin
263         // and cosine functions here.
264         for (int j = 0; j < line_len + 1; ++j) {
265           double arg = ((EIGEN_PI * j) * j) / line_len;
266           std::complex<double> tmp(numext::cos(arg), numext::sin(arg));
267           pos_j_base_powered[j] = static_cast<ComplexScalar>(tmp);
268         }
269       }
270 
271       for (Index partial_index = 0; partial_index < m_size / line_len; ++partial_index) {
272         const Index base_offset = getBaseOffsetFromIndex(partial_index, dim);
273 
274         // get data into line_buf
275         const Index stride = m_strides[dim];
276         if (stride == 1) {
277           m_device.memcpy(line_buf, &buf[base_offset], line_len*sizeof(ComplexScalar));
278         } else {
279           Index offset = base_offset;
280           for (int j = 0; j < line_len; ++j, offset += stride) {
281             line_buf[j] = buf[offset];
282           }
283         }
284 
285         // process the line
286         if (is_power_of_two) {
287           processDataLineCooleyTukey(line_buf, line_len, log_len);
288         }
289         else {
290           processDataLineBluestein(line_buf, line_len, good_composite, log_len, a, b, pos_j_base_powered);
291         }
292 
293         // write back
294         if (FFTDir == FFT_FORWARD && stride == 1) {
295           m_device.memcpy(&buf[base_offset], line_buf, line_len*sizeof(ComplexScalar));
296         } else {
297           Index offset = base_offset;
298           const ComplexScalar div_factor =  ComplexScalar(1.0 / line_len, 0);
299           for (int j = 0; j < line_len; ++j, offset += stride) {
300              buf[offset] = (FFTDir == FFT_FORWARD) ? line_buf[j] : line_buf[j] * div_factor;
301           }
302         }
303       }
304       m_device.deallocate(line_buf);
305       if (!is_power_of_two) {
306         m_device.deallocate(a);
307         m_device.deallocate(b);
308         m_device.deallocate(pos_j_base_powered);
309       }
310     }
311 
312     if(!write_to_out) {
313       for (Index i = 0; i < m_size; ++i) {
314         data[i] = PartOf<FFTResultType>()(buf[i]);
315       }
316       m_device.deallocate(buf);
317     }
318   }
319 
320   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static bool isPowerOfTwo(Index x) {
321     eigen_assert(x > 0);
322     return !(x & (x - 1));
323   }
324 
325   // The composite number for padding, used in Bluestein's FFT algorithm
326   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static Index findGoodComposite(Index n) {
327     Index i = 2;
328     while (i < 2 * n - 1) i *= 2;
329     return i;
330   }
331 
332   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static Index getLog2(Index m) {
333     Index log2m = 0;
334     while (m >>= 1) log2m++;
335     return log2m;
336   }
337 
338   // Call Cooley Tukey algorithm directly, data length must be power of 2
339   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void processDataLineCooleyTukey(ComplexScalar* line_buf, Index line_len, Index log_len) {
340     eigen_assert(isPowerOfTwo(line_len));
341     scramble_FFT(line_buf, line_len);
342     compute_1D_Butterfly<FFTDir>(line_buf, line_len, log_len);
343   }
344 
345   // Call Bluestein's FFT algorithm, m is a good composite number greater than (2 * n - 1), used as the padding length
346   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) {
347     Index n = line_len;
348     Index m = good_composite;
349     ComplexScalar* data = line_buf;
350 
351     for (Index i = 0; i < n; ++i) {
352       if(FFTDir == FFT_FORWARD) {
353         a[i] = data[i] * numext::conj(pos_j_base_powered[i]);
354       }
355       else {
356         a[i] = data[i] * pos_j_base_powered[i];
357       }
358     }
359     for (Index i = n; i < m; ++i) {
360       a[i] = ComplexScalar(0, 0);
361     }
362 
363     for (Index i = 0; i < n; ++i) {
364       if(FFTDir == FFT_FORWARD) {
365         b[i] = pos_j_base_powered[i];
366       }
367       else {
368         b[i] = numext::conj(pos_j_base_powered[i]);
369       }
370     }
371     for (Index i = n; i < m - n; ++i) {
372       b[i] = ComplexScalar(0, 0);
373     }
374     for (Index i = m - n; i < m; ++i) {
375       if(FFTDir == FFT_FORWARD) {
376         b[i] = pos_j_base_powered[m-i];
377       }
378       else {
379         b[i] = numext::conj(pos_j_base_powered[m-i]);
380       }
381     }
382 
383     scramble_FFT(a, m);
384     compute_1D_Butterfly<FFT_FORWARD>(a, m, log_len);
385 
386     scramble_FFT(b, m);
387     compute_1D_Butterfly<FFT_FORWARD>(b, m, log_len);
388 
389     for (Index i = 0; i < m; ++i) {
390       a[i] *= b[i];
391     }
392 
393     scramble_FFT(a, m);
394     compute_1D_Butterfly<FFT_REVERSE>(a, m, log_len);
395 
396     //Do the scaling after ifft
397     for (Index i = 0; i < m; ++i) {
398       a[i] /= m;
399     }
400 
401     for (Index i = 0; i < n; ++i) {
402       if(FFTDir == FFT_FORWARD) {
403         data[i] = a[i] * numext::conj(pos_j_base_powered[i]);
404       }
405       else {
406         data[i] = a[i] * pos_j_base_powered[i];
407       }
408     }
409   }
410 
411   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE static void scramble_FFT(ComplexScalar* data, Index n) {
412     eigen_assert(isPowerOfTwo(n));
413     Index j = 1;
414     for (Index i = 1; i < n; ++i){
415       if (j > i) {
416         std::swap(data[j-1], data[i-1]);
417       }
418       Index m = n >> 1;
419       while (m >= 2 && j > m) {
420         j -= m;
421         m >>= 1;
422       }
423       j += m;
424     }
425   }
426 
427   template <int Dir>
428   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_2(ComplexScalar* data) {
429     ComplexScalar tmp = data[1];
430     data[1] = data[0] - data[1];
431     data[0] += tmp;
432   }
433 
434   template <int Dir>
435   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_4(ComplexScalar* data) {
436     ComplexScalar tmp[4];
437     tmp[0] = data[0] + data[1];
438     tmp[1] = data[0] - data[1];
439     tmp[2] = data[2] + data[3];
440     if (Dir == FFT_FORWARD) {
441       tmp[3] = ComplexScalar(0.0, -1.0) * (data[2] - data[3]);
442     } else {
443       tmp[3] = ComplexScalar(0.0, 1.0) * (data[2] - data[3]);
444     }
445     data[0] = tmp[0] + tmp[2];
446     data[1] = tmp[1] + tmp[3];
447     data[2] = tmp[0] - tmp[2];
448     data[3] = tmp[1] - tmp[3];
449   }
450 
451   template <int Dir>
452   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_8(ComplexScalar* data) {
453     ComplexScalar tmp_1[8];
454     ComplexScalar tmp_2[8];
455 
456     tmp_1[0] = data[0] + data[1];
457     tmp_1[1] = data[0] - data[1];
458     tmp_1[2] = data[2] + data[3];
459     if (Dir == FFT_FORWARD) {
460       tmp_1[3] = (data[2] - data[3]) * ComplexScalar(0, -1);
461     } else {
462       tmp_1[3] = (data[2] - data[3]) * ComplexScalar(0, 1);
463     }
464     tmp_1[4] = data[4] + data[5];
465     tmp_1[5] = data[4] - data[5];
466     tmp_1[6] = data[6] + data[7];
467     if (Dir == FFT_FORWARD) {
468       tmp_1[7] = (data[6] - data[7]) * ComplexScalar(0, -1);
469     } else {
470       tmp_1[7] = (data[6] - data[7]) * ComplexScalar(0, 1);
471     }
472     tmp_2[0] = tmp_1[0] + tmp_1[2];
473     tmp_2[1] = tmp_1[1] + tmp_1[3];
474     tmp_2[2] = tmp_1[0] - tmp_1[2];
475     tmp_2[3] = tmp_1[1] - tmp_1[3];
476     tmp_2[4] = tmp_1[4] + tmp_1[6];
477 // SQRT2DIV2 = sqrt(2)/2
478 #define SQRT2DIV2 0.7071067811865476
479     if (Dir == FFT_FORWARD) {
480       tmp_2[5] = (tmp_1[5] + tmp_1[7]) * ComplexScalar(SQRT2DIV2, -SQRT2DIV2);
481       tmp_2[6] = (tmp_1[4] - tmp_1[6]) * ComplexScalar(0, -1);
482       tmp_2[7] = (tmp_1[5] - tmp_1[7]) * ComplexScalar(-SQRT2DIV2, -SQRT2DIV2);
483     } else {
484       tmp_2[5] = (tmp_1[5] + tmp_1[7]) * ComplexScalar(SQRT2DIV2, SQRT2DIV2);
485       tmp_2[6] = (tmp_1[4] - tmp_1[6]) * ComplexScalar(0, 1);
486       tmp_2[7] = (tmp_1[5] - tmp_1[7]) * ComplexScalar(-SQRT2DIV2, SQRT2DIV2);
487     }
488     data[0] = tmp_2[0] + tmp_2[4];
489     data[1] = tmp_2[1] + tmp_2[5];
490     data[2] = tmp_2[2] + tmp_2[6];
491     data[3] = tmp_2[3] + tmp_2[7];
492     data[4] = tmp_2[0] - tmp_2[4];
493     data[5] = tmp_2[1] - tmp_2[5];
494     data[6] = tmp_2[2] - tmp_2[6];
495     data[7] = tmp_2[3] - tmp_2[7];
496   }
497 
498   template <int Dir>
499   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void butterfly_1D_merge(
500       ComplexScalar* data, Index n, Index n_power_of_2) {
501     // Original code:
502     // RealScalar wtemp = std::sin(M_PI/n);
503     // RealScalar wpi =  -std::sin(2 * M_PI/n);
504     const RealScalar wtemp = m_sin_PI_div_n_LUT[n_power_of_2];
505     const RealScalar wpi = (Dir == FFT_FORWARD)
506                                ? m_minus_sin_2_PI_div_n_LUT[n_power_of_2]
507                                : -m_minus_sin_2_PI_div_n_LUT[n_power_of_2];
508 
509     const ComplexScalar wp(wtemp, wpi);
510     const ComplexScalar wp_one = wp + ComplexScalar(1, 0);
511     const ComplexScalar wp_one_2 = wp_one * wp_one;
512     const ComplexScalar wp_one_3 = wp_one_2 * wp_one;
513     const ComplexScalar wp_one_4 = wp_one_3 * wp_one;
514     const Index n2 = n / 2;
515     ComplexScalar w(1.0, 0.0);
516     for (Index i = 0; i < n2; i += 4) {
517        ComplexScalar temp0(data[i + n2] * w);
518        ComplexScalar temp1(data[i + 1 + n2] * w * wp_one);
519        ComplexScalar temp2(data[i + 2 + n2] * w * wp_one_2);
520        ComplexScalar temp3(data[i + 3 + n2] * w * wp_one_3);
521        w = w * wp_one_4;
522 
523        data[i + n2] = data[i] - temp0;
524        data[i] += temp0;
525 
526        data[i + 1 + n2] = data[i + 1] - temp1;
527        data[i + 1] += temp1;
528 
529        data[i + 2 + n2] = data[i + 2] - temp2;
530        data[i + 2] += temp2;
531 
532        data[i + 3 + n2] = data[i + 3] - temp3;
533        data[i + 3] += temp3;
534     }
535   }
536 
537  template <int Dir>
538   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void compute_1D_Butterfly(
539       ComplexScalar* data, Index n, Index n_power_of_2) {
540     eigen_assert(isPowerOfTwo(n));
541     if (n > 8) {
542       compute_1D_Butterfly<Dir>(data, n / 2, n_power_of_2 - 1);
543       compute_1D_Butterfly<Dir>(data + n / 2, n / 2, n_power_of_2 - 1);
544       butterfly_1D_merge<Dir>(data, n, n_power_of_2);
545     } else if (n == 8) {
546       butterfly_8<Dir>(data);
547     } else if (n == 4) {
548       butterfly_4<Dir>(data);
549     } else if (n == 2) {
550       butterfly_2<Dir>(data);
551     }
552   }
553 
554   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index getBaseOffsetFromIndex(Index index, Index omitted_dim) const {
555     Index result = 0;
556 
557     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
558       for (int i = NumDims - 1; i > omitted_dim; --i) {
559         const Index partial_m_stride = m_strides[i] / m_dimensions[omitted_dim];
560         const Index idx = index / partial_m_stride;
561         index -= idx * partial_m_stride;
562         result += idx * m_strides[i];
563       }
564       result += index;
565     }
566     else {
567       for (Index i = 0; i < omitted_dim; ++i) {
568         const Index partial_m_stride = m_strides[i] / m_dimensions[omitted_dim];
569         const Index idx = index / partial_m_stride;
570         index -= idx * partial_m_stride;
571         result += idx * m_strides[i];
572       }
573       result += index;
574     }
575     // Value of index_coords[omitted_dim] is not determined to this step
576     return result;
577   }
578 
579   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index getIndexFromOffset(Index base, Index omitted_dim, Index offset) const {
580     Index result = base + offset * m_strides[omitted_dim] ;
581     return result;
582   }
583 
584  protected:
585   Index m_size;
586   const FFT EIGEN_DEVICE_REF m_fft;
587   Dimensions m_dimensions;
588   array<Index, NumDims> m_strides;
589   TensorEvaluator<ArgType, Device> m_impl;
590   EvaluatorPointerType m_data;
591   const Device EIGEN_DEVICE_REF m_device;
592 
593   // This will support a maximum FFT size of 2^32 for each dimension
594   // m_sin_PI_div_n_LUT[i] = (-2) * std::sin(M_PI / std::pow(2,i)) ^ 2;
595   const RealScalar m_sin_PI_div_n_LUT[32] = {
596     RealScalar(0.0),
597     RealScalar(-2),
598     RealScalar(-0.999999999999999),
599     RealScalar(-0.292893218813453),
600     RealScalar(-0.0761204674887130),
601     RealScalar(-0.0192147195967696),
602     RealScalar(-0.00481527332780311),
603     RealScalar(-0.00120454379482761),
604     RealScalar(-3.01181303795779e-04),
605     RealScalar(-7.52981608554592e-05),
606     RealScalar(-1.88247173988574e-05),
607     RealScalar(-4.70619042382852e-06),
608     RealScalar(-1.17654829809007e-06),
609     RealScalar(-2.94137117780840e-07),
610     RealScalar(-7.35342821488550e-08),
611     RealScalar(-1.83835707061916e-08),
612     RealScalar(-4.59589268710903e-09),
613     RealScalar(-1.14897317243732e-09),
614     RealScalar(-2.87243293150586e-10),
615     RealScalar( -7.18108232902250e-11),
616     RealScalar(-1.79527058227174e-11),
617     RealScalar(-4.48817645568941e-12),
618     RealScalar(-1.12204411392298e-12),
619     RealScalar(-2.80511028480785e-13),
620     RealScalar(-7.01277571201985e-14),
621     RealScalar(-1.75319392800498e-14),
622     RealScalar(-4.38298482001247e-15),
623     RealScalar(-1.09574620500312e-15),
624     RealScalar(-2.73936551250781e-16),
625     RealScalar(-6.84841378126949e-17),
626     RealScalar(-1.71210344531737e-17),
627     RealScalar(-4.28025861329343e-18)
628   };
629 
630   // m_minus_sin_2_PI_div_n_LUT[i] = -std::sin(2 * M_PI / std::pow(2,i));
631   const RealScalar m_minus_sin_2_PI_div_n_LUT[32] = {
632     RealScalar(0.0),
633     RealScalar(0.0),
634     RealScalar(-1.00000000000000e+00),
635     RealScalar(-7.07106781186547e-01),
636     RealScalar(-3.82683432365090e-01),
637     RealScalar(-1.95090322016128e-01),
638     RealScalar(-9.80171403295606e-02),
639     RealScalar(-4.90676743274180e-02),
640     RealScalar(-2.45412285229123e-02),
641     RealScalar(-1.22715382857199e-02),
642     RealScalar(-6.13588464915448e-03),
643     RealScalar(-3.06795676296598e-03),
644     RealScalar(-1.53398018628477e-03),
645     RealScalar(-7.66990318742704e-04),
646     RealScalar(-3.83495187571396e-04),
647     RealScalar(-1.91747597310703e-04),
648     RealScalar(-9.58737990959773e-05),
649     RealScalar(-4.79368996030669e-05),
650     RealScalar(-2.39684498084182e-05),
651     RealScalar(-1.19842249050697e-05),
652     RealScalar(-5.99211245264243e-06),
653     RealScalar(-2.99605622633466e-06),
654     RealScalar(-1.49802811316901e-06),
655     RealScalar(-7.49014056584716e-07),
656     RealScalar(-3.74507028292384e-07),
657     RealScalar(-1.87253514146195e-07),
658     RealScalar(-9.36267570730981e-08),
659     RealScalar(-4.68133785365491e-08),
660     RealScalar(-2.34066892682746e-08),
661     RealScalar(-1.17033446341373e-08),
662     RealScalar(-5.85167231706864e-09),
663     RealScalar(-2.92583615853432e-09)
664   };
665 };
666 
667 }  // end namespace Eigen
668 
669 #endif  // EIGEN_CXX11_TENSOR_TENSOR_FFT_H
670