• 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) 2014 Benoit Steiner <benoit.steiner.goog@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_INTDIV_H
11 #define EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H
12 
13 
14 namespace Eigen {
15 
16 /** \internal
17   *
18   * \class TensorIntDiv
19   * \ingroup CXX11_Tensor_Module
20   *
21   * \brief Fast integer division by a constant.
22   *
23   * See the paper from Granlund and Montgomery for explanation.
24   *   (at https://doi.org/10.1145/773473.178249)
25   *
26   * \sa Tensor
27   */
28 
29 namespace internal {
30 
31 namespace {
32 
33   // Note: result is undefined if val == 0
34   template <typename T>
35   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
count_leading_zeros(const T val)36   typename internal::enable_if<sizeof(T)==4,int>::type count_leading_zeros(const T val)
37   {
38 #ifdef EIGEN_GPU_COMPILE_PHASE
39     return __clz(val);
40 #elif defined(SYCL_DEVICE_ONLY)
41     return cl::sycl::clz(val);
42 #elif EIGEN_COMP_MSVC
43     unsigned long index;
44     _BitScanReverse(&index, val);
45     return 31 - index;
46 #else
47     EIGEN_STATIC_ASSERT(sizeof(unsigned long long) == 8, YOU_MADE_A_PROGRAMMING_MISTAKE);
48     return __builtin_clz(static_cast<uint32_t>(val));
49 #endif
50   }
51 
52   template <typename T>
53   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
count_leading_zeros(const T val)54   typename internal::enable_if<sizeof(T)==8,int>::type count_leading_zeros(const T val)
55   {
56 #ifdef EIGEN_GPU_COMPILE_PHASE
57     return __clzll(val);
58 #elif defined(SYCL_DEVICE_ONLY)
59     return static_cast<int>(cl::sycl::clz(val));
60 #elif EIGEN_COMP_MSVC && EIGEN_ARCH_x86_64
61     unsigned long index;
62     _BitScanReverse64(&index, val);
63     return 63 - index;
64 #elif EIGEN_COMP_MSVC
65     // MSVC's _BitScanReverse64 is not available for 32bits builds.
66     unsigned int lo = (unsigned int)(val&0xffffffff);
67     unsigned int hi = (unsigned int)((val>>32)&0xffffffff);
68     int n;
69     if(hi==0)
70       n = 32 + count_leading_zeros<unsigned int>(lo);
71     else
72       n = count_leading_zeros<unsigned int>(hi);
73     return n;
74 #else
75     EIGEN_STATIC_ASSERT(sizeof(unsigned long long) == 8, YOU_MADE_A_PROGRAMMING_MISTAKE);
76     return __builtin_clzll(static_cast<uint64_t>(val));
77 #endif
78   }
79 
80   template <typename T>
81   struct UnsignedTraits {
82     typedef typename conditional<sizeof(T) == 8, uint64_t, uint32_t>::type type;
83   };
84 
85   template <typename T>
86   struct DividerTraits {
87     typedef typename UnsignedTraits<T>::type type;
88     static const int N = sizeof(T) * 8;
89   };
90 
91   template <typename T>
muluh(const uint32_t a,const T b)92   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(const uint32_t a, const T b) {
93 #if defined(EIGEN_GPU_COMPILE_PHASE)
94     return __umulhi(a, b);
95 #elif defined(SYCL_DEVICE_ONLY)
96     return cl::sycl::mul_hi(a, static_cast<uint32_t>(b));
97 #else
98     return (static_cast<uint64_t>(a) * b) >> 32;
99 #endif
100   }
101 
102   template <typename T>
muluh(const uint64_t a,const T b)103   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
104 #if defined(EIGEN_GPU_COMPILE_PHASE)
105     return __umul64hi(a, b);
106 #elif defined(SYCL_DEVICE_ONLY)
107     return cl::sycl::mul_hi(a, static_cast<uint64_t>(b));
108 #elif EIGEN_HAS_BUILTIN_INT128
109     __uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
110     return static_cast<uint64_t>(v >> 64);
111 #else
112     return (TensorUInt128<static_val<0>, uint64_t>(a) * TensorUInt128<static_val<0>, uint64_t>(b)).upper();
113 #endif
114   }
115 
116   template <int N, typename T>
117   struct DividerHelper {
computeMultiplierDividerHelper118     static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier(const int log_div, const T divider) {
119       EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE);
120       return static_cast<uint32_t>((static_cast<uint64_t>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1);
121     }
122   };
123 
124   template <typename T>
125   struct DividerHelper<64, T> {
126     static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
127 #if EIGEN_HAS_BUILTIN_INT128 && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
128       return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
129 #else
130       const uint64_t shift = 1ULL << log_div;
131       TensorUInt128<uint64_t, uint64_t> result = TensorUInt128<uint64_t, static_val<0> >(shift, 0) / TensorUInt128<static_val<0>, uint64_t>(divider)
132                                                - TensorUInt128<static_val<1>, static_val<0> >(1, 0)
133                                                + TensorUInt128<static_val<0>, static_val<1> >(1);
134       return static_cast<uint64_t>(result);
135 #endif
136     }
137   };
138 }
139 
140 
141 template <typename T, bool div_gt_one = false>
142 struct TensorIntDivisor {
143  public:
144   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
145     multiplier = 0;
146     shift1 = 0;
147     shift2 = 0;
148   }
149 
150   // Must have 0 < divider < 2^31. This is relaxed to
151   // 0 < divider < 2^63 when using 64-bit indices on platforms that support
152   // the __uint128_t type.
153   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor(const T divider) {
154     const int N = DividerTraits<T>::N;
155     eigen_assert(static_cast<typename UnsignedTraits<T>::type>(divider) < NumTraits<UnsignedType>::highest()/2);
156     eigen_assert(divider > 0);
157 
158     // fast ln2
159     const int leading_zeros = count_leading_zeros(static_cast<UnsignedType>(divider));
160     int log_div = N - leading_zeros;
161     // if divider is a power of two then log_div is 1 more than it should be.
162     if ((static_cast<typename UnsignedTraits<T>::type>(1) << (log_div-1)) == static_cast<typename UnsignedTraits<T>::type>(divider))
163       log_div--;
164 
165     multiplier = DividerHelper<N, T>::computeMultiplier(log_div, divider);
166     shift1 = log_div > 1 ? 1 : log_div;
167     shift2 = log_div > 1 ? log_div-1 : 0;
168   }
169 
170   // Must have 0 <= numerator. On platforms that don't support the __uint128_t
171   // type numerator should also be less than 2^32-1.
172   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T divide(const T numerator) const {
173     eigen_assert(static_cast<typename UnsignedTraits<T>::type>(numerator) < NumTraits<UnsignedType>::highest()/2);
174     //eigen_assert(numerator >= 0); // this is implicitly asserted by the line above
175 
176     UnsignedType t1 = muluh(multiplier, numerator);
177     UnsignedType t = (static_cast<UnsignedType>(numerator) - t1) >> shift1;
178     return (t1 + t) >> shift2;
179   }
180 
181  private:
182   typedef typename DividerTraits<T>::type UnsignedType;
183   UnsignedType multiplier;
184   int32_t shift1;
185   int32_t shift2;
186 };
187 
188 
189 // Optimized version for signed 32 bit integers.
190 // Derived from Hacker's Delight.
191 // Only works for divisors strictly greater than one
192 template <>
193 class TensorIntDivisor<int32_t, true> {
194  public:
195   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
196     magic = 0;
197     shift = 0;
198   }
199   // Must have 2 <= divider
200   EIGEN_DEVICE_FUNC TensorIntDivisor(int32_t divider)  {
201     eigen_assert(divider >= 2);
202     calcMagic(divider);
203   }
204 
205   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE int divide(const int32_t n) const {
206 #ifdef EIGEN_GPU_COMPILE_PHASE
207     return (__umulhi(magic, n) >> shift);
208 #elif defined(SYCL_DEVICE_ONLY)
209     return (cl::sycl::mul_hi(magic, static_cast<uint32_t>(n)) >> shift);
210 #else
211     uint64_t v = static_cast<uint64_t>(magic) * static_cast<uint64_t>(n);
212     return (static_cast<uint32_t>(v >> 32) >> shift);
213 #endif
214   }
215 
216 private:
217   // Compute the magic numbers. See Hacker's Delight section 10 for an in
218   // depth explanation.
219   EIGEN_DEVICE_FUNC void calcMagic(int32_t d) {
220    const unsigned two31 = 0x80000000;     // 2**31.
221    unsigned ad = d;
222    unsigned t = two31 + (ad >> 31);
223    unsigned anc = t - 1 - t%ad;     // Absolute value of nc.
224    int p = 31;                      // Init. p.
225    unsigned q1 = two31/anc;         // Init. q1 = 2**p/|nc|.
226    unsigned r1 = two31 - q1*anc;    // Init. r1 = rem(2**p, |nc|).
227    unsigned q2 = two31/ad;          // Init. q2 = 2**p/|d|.
228    unsigned r2 = two31 - q2*ad;     // Init. r2 = rem(2**p, |d|).
229    unsigned delta = 0;
230    do {
231       p = p + 1;
232       q1 = 2*q1;           // Update q1 = 2**p/|nc|.
233       r1 = 2*r1;           // Update r1 = rem(2**p, |nc|).
234       if (r1 >= anc) {     // (Must be an unsigned
235          q1 = q1 + 1;      // comparison here).
236          r1 = r1 - anc;}
237       q2 = 2*q2;           // Update q2 = 2**p/|d|.
238       r2 = 2*r2;           // Update r2 = rem(2**p, |d|).
239       if (r2 >= ad) {      // (Must be an unsigned
240          q2 = q2 + 1;      // comparison here).
241          r2 = r2 - ad;}
242       delta = ad - r2;
243    } while (q1 < delta || (q1 == delta && r1 == 0));
244 
245    magic = (unsigned)(q2 + 1);
246    shift = p - 32;
247   }
248 
249   uint32_t magic;
250   int32_t shift;
251 };
252 
253 
254 template <typename T, bool div_gt_one>
255 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator / (const T& numerator, const TensorIntDivisor<T, div_gt_one>& divisor) {
256   return divisor.divide(numerator);
257 }
258 
259 
260 } // end namespace internal
261 } // end namespace Eigen
262 
263 #endif // EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H
264