10 #ifndef EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H 11 #define EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H 34 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
int count_leading_zeros(
const T val)
37 return (
sizeof(T) == 8) ? __clzll(val) : __clz(val);
39 DWORD leading_zeros = 0;
41 _BitScanReverse64(&leading_zero, val);
44 _BitScanReverse(&leading_zero, val);
48 return (
sizeof(T) == 8) ?
49 __builtin_clzl(static_cast<uint64_t>(val)) :
50 __builtin_clz(static_cast<uint32_t>(val));
55 struct UnsignedTraits {
56 typedef typename conditional<sizeof(T) == 8, uint64_t, uint32_t>::type type;
60 struct DividerTraits {
61 typedef typename UnsignedTraits<T>::type type;
62 static const int N =
sizeof(T) * 8;
66 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(
const uint32_t a,
const T b) {
67 #if defined(__CUDA_ARCH__) 68 return __umulhi(a, b);
70 return (static_cast<uint64_t>(a) * b) >> 32;
75 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(
const uint64_t a,
const T b) {
76 #if defined(__CUDA_ARCH__) 77 return __umul64hi(a, b);
78 #elif defined(__SIZEOF_INT128__) 79 __uint128_t v =
static_cast<__uint128_t
>(a) * static_cast<__uint128_t>(b);
80 return static_cast<uint64_t
>(v >> 64);
82 return (TensorUInt128<static_val<0>, uint64_t>(a) * TensorUInt128<static_val<0>, uint64_t>(b)).upper();
86 template <
int N,
typename T>
87 struct DividerHelper {
88 static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier(
const int log_div,
const T divider) {
89 EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE);
90 return static_cast<uint32_t
>((
static_cast<uint64_t
>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1);
95 struct DividerHelper<64, T> {
96 static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(
const int log_div,
const T divider) {
97 #if defined(__SIZEOF_INT128__) && !defined(__CUDA_ARCH__) 98 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);
100 const uint64_t shift = 1ULL << log_div;
101 TensorUInt128<uint64_t, uint64_t> result = (TensorUInt128<uint64_t, static_val<0> >(shift, 0) / TensorUInt128<static_val<0>, uint64_t>(divider) - TensorUInt128<static_val<1>, static_val<0> >(1, 0) + TensorUInt128<static_val<0>, static_val<1> >(1));
102 return static_cast<uint64_t
>(result);
109 template <
typename T,
bool div_gt_one = false>
110 struct TensorIntDivisor {
112 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
121 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor(
const T divider) {
122 const int N = DividerTraits<T>::N;
123 eigen_assert(
static_cast<typename UnsignedTraits<T>::type
>(divider) < NumTraits<UnsignedType>::highest()/2);
124 eigen_assert(divider > 0);
127 const int leading_zeros = count_leading_zeros(static_cast<UnsignedType>(divider));
128 int log_div = N - leading_zeros;
130 if ((
static_cast<typename UnsignedTraits<T>::type
>(1) << (log_div-1)) ==
static_cast<typename UnsignedTraits<T>::type
>(divider))
133 multiplier = DividerHelper<N, T>::computeMultiplier(log_div, divider);
134 shift1 = log_div > 1 ? 1 : log_div;
135 shift2 = log_div > 1 ? log_div-1 : 0;
140 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T divide(
const T numerator)
const {
141 eigen_assert(
static_cast<typename UnsignedTraits<T>::type
>(numerator) < NumTraits<UnsignedType>::highest()/2);
144 UnsignedType t1 = muluh(multiplier, numerator);
145 UnsignedType t = (
static_cast<UnsignedType
>(numerator) - t1) >> shift1;
146 return (t1 + t) >> shift2;
150 typedef typename DividerTraits<T>::type UnsignedType;
151 UnsignedType multiplier;
161 class TensorIntDivisor<int32_t, true> {
163 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
168 EIGEN_DEVICE_FUNC TensorIntDivisor(int32_t divider) {
169 eigen_assert(divider >= 2);
173 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
int divide(
const int32_t n)
const {
175 return (__umulhi(magic, n) >> shift);
177 uint64_t v =
static_cast<uint64_t
>(magic) * static_cast<uint64_t>(n);
178 return (static_cast<uint32_t>(v >> 32) >> shift);
185 EIGEN_DEVICE_FUNC
void calcMagic(int32_t d) {
186 const unsigned two31 = 0x80000000;
188 unsigned t = two31 + (ad >> 31);
189 unsigned anc = t - 1 - t%ad;
191 unsigned q1 = two31/anc;
192 unsigned r1 = two31 - q1*anc;
193 unsigned q2 = two31/ad;
194 unsigned r2 = two31 - q2*ad;
209 }
while (q1 < delta || (q1 == delta && r1 == 0));
211 magic = (unsigned)(q2 + 1);
220 template <
typename T,
bool div_gt_one>
221 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator / (
const T& numerator,
const TensorIntDivisor<T, div_gt_one>& divisor) {
222 return divisor.divide(numerator);
229 #endif // EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H Namespace containing all symbols from the Eigen library.
Definition: CXX11Meta.h:13