21 #ifndef ROCRAND_MRG31K3P_H_
22 #define ROCRAND_MRG31K3P_H_
25 #define FQUALIFIERS __forceinline__ __device__
28 #include "rocrand/rocrand_common.h"
29 #include "rocrand/rocrand_mrg31k3p_precomputed.h"
31 #define ROCRAND_MRG31K3P_M1 2147483647U
32 #define ROCRAND_MRG31K3P_M2 2147462579U
33 #define ROCRAND_MRG31K3P_MASK12 511U
34 #define ROCRAND_MRG31K3P_MASK13 16777215U
35 #define ROCRAND_MRG31K3P_MASK21 65535U
36 #define ROCRAND_MRG31K3P_NORM_DOUBLE (4.656612875245796923e-10)
37 #define ROCRAND_MRG31K3P_UINT32_NORM \
38 (2.000000001396983862)
48 #define ROCRAND_MRG31K3P_DEFAULT_SEED 12345ULL
51 namespace rocrand_device
62 #ifndef ROCRAND_DETAIL_MRG31K3P_BM_NOT_IN_STATE
68 unsigned int boxmuller_float_state;
69 unsigned int boxmuller_double_state;
70 float boxmuller_float;
71 double boxmuller_double;
88 FQUALIFIERS mrg31k3p_engine(
const unsigned long long seed,
89 const unsigned long long subsequence,
90 const unsigned long long offset)
92 this->seed(seed, subsequence, offset);
103 FQUALIFIERS void seed(
unsigned long long seed_value,
104 const unsigned long long subsequence,
105 const unsigned long long offset)
111 unsigned int x =
static_cast<unsigned int>(seed_value ^ 0x55555555U);
112 unsigned int y =
static_cast<unsigned int>((seed_value >> 32) ^ 0xAAAAAAAAU);
113 m_state.x1[0] = mod_mul_m1(x, seed_value);
114 m_state.x1[1] = mod_mul_m1(y, seed_value);
115 m_state.x1[2] = mod_mul_m1(x, seed_value);
116 m_state.x2[0] = mod_mul_m2(y, seed_value);
117 m_state.x2[1] = mod_mul_m2(x, seed_value);
118 m_state.x2[2] = mod_mul_m2(y, seed_value);
119 this->restart(subsequence, offset);
123 FQUALIFIERS void discard(
unsigned long long offset)
125 this->discard_impl(offset);
130 FQUALIFIERS void discard_subsequence(
unsigned long long subsequence)
132 this->discard_subsequence_impl(subsequence);
137 FQUALIFIERS void discard_sequence(
unsigned long long sequence)
139 this->discard_sequence_impl(sequence);
142 FQUALIFIERS void restart(
const unsigned long long subsequence,
const unsigned long long offset)
144 #ifndef ROCRAND_DETAIL_MRG31K3P_BM_NOT_IN_STATE
145 m_state.boxmuller_float_state = 0;
146 m_state.boxmuller_double_state = 0;
148 this->discard_subsequence_impl(subsequence);
149 this->discard_impl(offset);
162 = (((m_state.x1[1] & ROCRAND_MRG31K3P_MASK12) << 22) + (m_state.x1[1] >> 9))
163 + (((m_state.x1[2] & ROCRAND_MRG31K3P_MASK13) << 7) + (m_state.x1[2] >> 24));
164 tmp -= (tmp >= ROCRAND_MRG31K3P_M1) ? ROCRAND_MRG31K3P_M1 : 0;
165 tmp += m_state.x1[2];
166 tmp -= (tmp >= ROCRAND_MRG31K3P_M1) ? ROCRAND_MRG31K3P_M1 : 0;
167 m_state.x1[2] = m_state.x1[1];
168 m_state.x1[1] = m_state.x1[0];
172 tmp = (((m_state.x2[0] & ROCRAND_MRG31K3P_MASK21) << 15) + 21069 * (m_state.x2[0] >> 16));
173 tmp -= (tmp >= ROCRAND_MRG31K3P_M2) ? ROCRAND_MRG31K3P_M2 : 0;
174 tmp += ((m_state.x2[2] & ROCRAND_MRG31K3P_MASK21) << 15);
175 tmp -= (tmp >= ROCRAND_MRG31K3P_M2) ? ROCRAND_MRG31K3P_M2 : 0;
176 tmp += 21069 * (m_state.x2[2] >> 16);
177 tmp -= (tmp >= ROCRAND_MRG31K3P_M2) ? ROCRAND_MRG31K3P_M2 : 0;
178 tmp += m_state.x2[2];
179 tmp -= (tmp >= ROCRAND_MRG31K3P_M2) ? ROCRAND_MRG31K3P_M2 : 0;
180 m_state.x2[2] = m_state.x2[1];
181 m_state.x2[1] = m_state.x2[0];
185 return m_state.x1[0] - m_state.x2[0]
186 + (m_state.x1[0] <= m_state.x2[0] ? ROCRAND_MRG31K3P_M1 : 0);
191 FQUALIFIERS void discard_impl(
unsigned long long offset)
193 discard_state(offset);
197 FQUALIFIERS void discard_subsequence_impl(
unsigned long long subsequence)
201 while(subsequence > 0)
205 #if defined(__HIP_DEVICE_COMPILE__)
206 mod_mat_vec_m1(d_mrg31k3p_A1P72 + i, m_state.x1);
207 mod_mat_vec_m2(d_mrg31k3p_A2P72 + i, m_state.x2);
209 mod_mat_vec_m1(h_mrg31k3p_A1P72 + i, m_state.x1);
210 mod_mat_vec_m2(h_mrg31k3p_A2P72 + i, m_state.x2);
219 FQUALIFIERS void discard_sequence_impl(
unsigned long long sequence)
227 #if defined(__HIP_DEVICE_COMPILE__)
228 mod_mat_vec_m1(d_mrg31k3p_A1P134 + i, m_state.x1);
229 mod_mat_vec_m2(d_mrg31k3p_A2P134 + i, m_state.x2);
231 mod_mat_vec_m1(h_mrg31k3p_A1P134 + i, m_state.x1);
232 mod_mat_vec_m2(h_mrg31k3p_A2P134 + i, m_state.x2);
241 FQUALIFIERS void discard_state(
unsigned long long offset)
249 #if defined(__HIP_DEVICE_COMPILE__)
250 mod_mat_vec_m1(d_mrg31k3p_A1 + i, m_state.x1);
251 mod_mat_vec_m2(d_mrg31k3p_A2 + i, m_state.x2);
253 mod_mat_vec_m1(h_mrg31k3p_A1 + i, m_state.x1);
254 mod_mat_vec_m2(h_mrg31k3p_A2 + i, m_state.x2);
269 FQUALIFIERS static void mod_mat_vec_m1(
const unsigned int* A,
unsigned int* s)
271 unsigned long long x[3] = {s[0], s[1], s[2]};
273 s[0] = mod_m1(mod_m1(A[0] * x[0]) + mod_m1(A[1] * x[1]) + mod_m1(A[2] * x[2]));
275 s[1] = mod_m1(mod_m1(A[3] * x[0]) + mod_m1(A[4] * x[1]) + mod_m1(A[5] * x[2]));
277 s[2] = mod_m1(mod_m1(A[6] * x[0]) + mod_m1(A[7] * x[1]) + mod_m1(A[8] * x[2]));
280 FQUALIFIERS static void mod_mat_vec_m2(
const unsigned int* A,
unsigned int* s)
282 unsigned long long x[3] = {s[0], s[1], s[2]};
284 s[0] = mod_m2(mod_m2(A[0] * x[0]) + mod_m2(A[1] * x[1]) + mod_m2(A[2] * x[2]));
286 s[1] = mod_m2(mod_m2(A[3] * x[0]) + mod_m2(A[4] * x[1]) + mod_m2(A[5] * x[2]));
288 s[2] = mod_m2(mod_m2(A[6] * x[0]) + mod_m2(A[7] * x[1]) + mod_m2(A[8] * x[2]));
291 FQUALIFIERS static unsigned long long mod_mul_m1(
unsigned int i,
unsigned long long j)
293 return mod_m1(i * j);
296 FQUALIFIERS static unsigned long long mod_m1(
unsigned long long p)
298 return p % ROCRAND_MRG31K3P_M1;
301 FQUALIFIERS static unsigned long long mod_mul_m2(
unsigned int i,
unsigned long long j)
303 return mod_m2(i * j);
306 FQUALIFIERS static unsigned long long mod_m2(
unsigned long long p)
308 return p % ROCRAND_MRG31K3P_M2;
313 mrg31k3p_state m_state;
315 #ifndef ROCRAND_DETAIL_MRG31K3P_BM_NOT_IN_STATE
316 friend struct detail::engine_boxmuller_helper<mrg31k3p_engine>;
328 typedef rocrand_device::mrg31k3p_engine rocrand_state_mrg31k3p;
343 const unsigned long long subsequence,
344 const unsigned long long offset,
345 rocrand_state_mrg31k3p* state)
347 *state = rocrand_state_mrg31k3p(seed, subsequence, offset);
365 return static_cast<unsigned int>((state->next() - 1) * ROCRAND_MRG31K3P_UINT32_NORM);
378 return state->discard(offset);
391 rocrand_state_mrg31k3p* state)
393 return state->discard_subsequence(subsequence);
407 return state->discard_sequence(sequence);
FQUALIFIERS unsigned int rocrand(rocrand_state_mrg31k3p *state)
Returns uniformly distributed random unsigned int value from [0; 2^32 - 1] range.
Definition: rocrand_mrg31k3p.h:362
#define ROCRAND_MRG31K3P_DEFAULT_SEED
Default seed for MRG31K3P PRNG.
Definition: rocrand_mrg31k3p.h:48
FQUALIFIERS void rocrand_init(const unsigned long long seed, const unsigned long long subsequence, const unsigned long long offset, rocrand_state_mrg31k3p *state)
Initializes MRG31K3P state.
Definition: rocrand_mrg31k3p.h:342
FQUALIFIERS void skipahead(unsigned long long offset, rocrand_state_mrg31k3p *state)
Updates MRG31K3P state to skip ahead by offset elements.
Definition: rocrand_mrg31k3p.h:376
FQUALIFIERS void skipahead_sequence(unsigned long long sequence, rocrand_state_mrg31k3p *state)
Updates MRG31K3P state to skip ahead by sequence sequences.
Definition: rocrand_mrg31k3p.h:405
FQUALIFIERS void skipahead_subsequence(unsigned long long subsequence, rocrand_state_mrg31k3p *state)
Updates MRG31K3P state to skip ahead by subsequence subsequences.
Definition: rocrand_mrg31k3p.h:390
#define FQUALIFIERS
Shorthand for commonly used function qualifiers.
Definition: rocrand_uniform.h:31