/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-rocrand/checkouts/develop/projects/rocrand/library/include/rocrand/rocrand_mtgp32.h Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-rocrand/checkouts/develop/projects/rocrand/library/include/rocrand/rocrand_mtgp32.h Source File#

API library: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-rocrand/checkouts/develop/projects/rocrand/library/include/rocrand/rocrand_mtgp32.h Source File
 // Copyright (c) 2017-2025 Advanced Micro Devices, Inc. All rights reserved.
 //
 // Permission is hereby granted, free of charge, to any person obtaining a copy
 // of this software and associated documentation files (the "Software"), to deal
 // in the Software without restriction, including without limitation the rights
 // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 // copies of the Software, and to permit persons to whom the Software is
 // furnished to do so, subject to the following conditions:
 //
 // The above copyright notice and this permission notice shall be included in
 // all copies or substantial portions of the Software.
 //
 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
 // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 // THE SOFTWARE.
  
 /*
  * Copyright (c) 2009, 2010 Mutsuo Saito, Makoto Matsumoto and Hiroshima
  * University.  All rights reserved.
  * Copyright (c) 2011 Mutsuo Saito, Makoto Matsumoto, Hiroshima
  * University and University of Tokyo.  All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are
  * met:
  *
  *     * Redistributions of source code must retain the above copyright
  *       notice, this list of conditions and the following disclaimer.
  *     * Redistributions in binary form must reproduce the above
  *       copyright notice, this list of conditions and the following
  *       disclaimer in the documentation and/or other materials provided
  *       with the distribution.
  *     * Neither the name of the Hiroshima University nor the names of
  *       its contributors may be used to endorse or promote products
  *       derived from this software without specific prior written
  *       permission.
  *
  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
  
 #ifndef ROCRAND_MTGP32_H_
 #define ROCRAND_MTGP32_H_
  
 #include "rocrand/rocrand.h"
  
 #include <hip/hip_runtime.h>
  
 #include <stdlib.h>
 #include <string.h>
  
 #define MTGP_MEXP 11213
 #define MTGP_N 351
 #define MTGP_FLOOR_2P 256
 #define MTGP_CEIL_2P 512
 #define MTGP_TN MTGP_FLOOR_2P
 #define MTGP_LS (MTGP_TN * 3)
 #define MTGP_BN_MAX 512
 #define MTGP_TS 16
 #define MTGP_STATE 1024
 #define MTGP_MASK 1023
  
 // Source: https://github.com/MersenneTwister-Lab/MTGP/blob/master/mtgp32-fast.h
 struct mtgp32_params_fast_t {
     int mexp;            
     int pos;            
     int sh1;            
     int sh2;            
     uint32_t tbl[16];        
     uint32_t tmp_tbl[16];    
     uint32_t flt_tmp_tbl[16];    
     uint32_t mask;        
     unsigned char poly_sha1[21]; 
 };
  
 namespace rocrand_device {
  
 struct mtgp32_params
 {
     unsigned int pos_tbl[MTGP_BN_MAX];
     unsigned int param_tbl[MTGP_BN_MAX][MTGP_TS];
     unsigned int temper_tbl[MTGP_BN_MAX][MTGP_TS];
     unsigned int single_temper_tbl[MTGP_BN_MAX][MTGP_TS];
     unsigned int sh1_tbl[MTGP_BN_MAX];
     unsigned int sh2_tbl[MTGP_BN_MAX];
     unsigned int mask[1];
 };
  
 typedef mtgp32_params_fast_t mtgp32_fast_params;
  
 struct mtgp32_state
 {
     int offset;
     int id;
     unsigned int status[MTGP_STATE];
 };
  
 inline
 void rocrand_mtgp32_init_state(unsigned int array[],
                                const mtgp32_fast_params *para, unsigned int seed)
 {
     int i;
     int size = para->mexp / 32 + 1;
     unsigned int hidden_seed;
     unsigned int tmp;
     hidden_seed = para->tbl[4] ^ (para->tbl[8] << 16);
     tmp = hidden_seed;
     tmp += tmp >> 16;
     tmp += tmp >> 8;
     memset(array, tmp & 0xff, sizeof(unsigned int) * size);
     array[0] = seed;
     array[1] = hidden_seed;
     for (i = 1; i < size; i++)
         array[i] ^= (1812433253) * (array[i - 1] ^ (array[i - 1] >> 30)) + i;
 }
  
 class mtgp32_engine
 {
 public:
     __forceinline__ __device__ __host__
         // Initialization is not supported for __shared__ variables
         mtgp32_engine() // cppcheck-suppress uninitMemberVar
     {
  
     }
  
     __forceinline__ __device__ __host__ mtgp32_engine(const mtgp32_state&  m_state,
                                                       const mtgp32_params* params,
                                                       int                  bid)
     {
         this->m_state = m_state;
         pos_tbl = params->pos_tbl[bid];
         sh1_tbl = params->sh1_tbl[bid];
         sh2_tbl = params->sh2_tbl[bid];
         mask = params->mask[0];
         for (int j = 0; j < MTGP_TS; j++) {
             param_tbl[j] = params->param_tbl[bid][j];
             temper_tbl[j] = params->temper_tbl[bid][j];
             single_temper_tbl[j] = params->single_temper_tbl[bid][j];
         }
     }
  
     __forceinline__ __device__ __host__ void copy(const mtgp32_engine* m_engine)
     {
 #if defined(__HIP_DEVICE_COMPILE__)
         const unsigned int thread_id = threadIdx.x;
         for(int i = thread_id; i < MTGP_STATE; i += blockDim.x)
             m_state.status[i] = m_engine->m_state.status[i];
  
         if (thread_id == 0)
         {
             m_state.offset = m_engine->m_state.offset;
             m_state.id = m_engine->m_state.id;
             pos_tbl = m_engine->pos_tbl;
             sh1_tbl = m_engine->sh1_tbl;
             sh2_tbl = m_engine->sh2_tbl;
             mask = m_engine->mask;
         }
         if (thread_id < MTGP_TS)
         {
             param_tbl[thread_id] = m_engine->param_tbl[thread_id];
             temper_tbl[thread_id] = m_engine->temper_tbl[thread_id];
             single_temper_tbl[thread_id] = m_engine->single_temper_tbl[thread_id];
         }
         __syncthreads();
 #else
         this->m_state = m_engine->m_state;
         pos_tbl = m_engine->pos_tbl;
         sh1_tbl = m_engine->sh1_tbl;
         sh2_tbl = m_engine->sh2_tbl;
         mask = m_engine->mask;
         for (int j = 0; j < MTGP_TS; j++) {
             param_tbl[j] = m_engine->param_tbl[j];
             temper_tbl[j] = m_engine->temper_tbl[j];
             single_temper_tbl[j] = m_engine->single_temper_tbl[j];
         }
 #endif
     }
  
     __forceinline__ __device__ __host__ void set_params(mtgp32_params* params)
     {
         pos_tbl = params->pos_tbl[m_state.id];
         sh1_tbl = params->sh1_tbl[m_state.id];
         sh2_tbl = params->sh2_tbl[m_state.id];
         mask = params->mask[0];
         for (int j = 0; j < MTGP_TS; j++) {
             param_tbl[j] = params->param_tbl[m_state.id][j];
             temper_tbl[j] = params->temper_tbl[m_state.id][j];
             single_temper_tbl[j] = params->single_temper_tbl[m_state.id][j];
         }
     }
  
     __forceinline__ __device__ __host__ unsigned int operator()()
     {
         return this->next();
     }
  
     __forceinline__ __device__ __host__ unsigned int next()
     {
 #ifdef __HIP_DEVICE_COMPILE__
         unsigned int o = next_thread(threadIdx.x);
         __syncthreads();
         if(threadIdx.x == 0)
         {
             m_state.offset = (m_state.offset + blockDim.x) & MTGP_MASK;
         }
         __syncthreads();
         return o;
 #else
         return 0;
 #endif
     }
  
     __forceinline__ __device__ __host__ unsigned int next_single()
     {
 #if defined(__HIP_DEVICE_COMPILE__)
         unsigned int t   = threadIdx.x;
         unsigned int d   = blockDim.x;
         int pos = pos_tbl;
         unsigned int r;
         unsigned int o;
  
         r = para_rec(m_state.status[(t + m_state.offset) & MTGP_MASK],
                      m_state.status[(t + m_state.offset + 1) & MTGP_MASK],
                      m_state.status[(t + m_state.offset + pos) & MTGP_MASK]);
         m_state.status[(t + m_state.offset + MTGP_N) & MTGP_MASK] = r;
  
         o = temper_single(r, m_state.status[(t + m_state.offset + pos - 1) & MTGP_MASK]);
         __syncthreads();
         if (t == 0)
             m_state.offset = (m_state.offset + d) & MTGP_MASK;
         __syncthreads();
         return o;
 #else
         return 0;
 #endif
     }
  
 private:
     __forceinline__ __device__ __host__ unsigned int
         para_rec(unsigned int X1, unsigned int X2, unsigned int Y) const
     {
         unsigned int X = (X1 & mask) ^ X2;
         unsigned int MAT;
  
         X ^= X << sh1_tbl;
         Y = X ^ (Y >> sh2_tbl);
         MAT = param_tbl[Y & 0x0f];
         return Y ^ MAT;
     }
  
     __forceinline__ __device__ __host__ unsigned int temper(unsigned int V, unsigned int T) const
     {
         unsigned int MAT;
  
         T ^= T >> 16;
         T ^= T >> 8;
         MAT = temper_tbl[T & 0x0f];
         return V ^ MAT;
     }
  
     __forceinline__ __device__ __host__ unsigned int temper_single(unsigned int V,
                                                                    unsigned int T) const
     {
         unsigned int MAT;
         unsigned int r;
  
         T ^= T >> 16;
         T ^= T >> 8;
         MAT = single_temper_tbl[T & 0x0f];
         r = (V >> 9) ^ MAT;
         return r;
     }
  
 protected:
     __forceinline__ __device__ __host__ unsigned int next_thread(unsigned int thread_idx)
     {
         const unsigned int r
             = para_rec(m_state.status[(thread_idx + m_state.offset) & MTGP_MASK],
                        m_state.status[(thread_idx + m_state.offset + 1) & MTGP_MASK],
                        m_state.status[(thread_idx + m_state.offset + pos_tbl) & MTGP_MASK]);
         m_state.status[(thread_idx + m_state.offset + MTGP_N) & MTGP_MASK] = r;
         return temper(r, m_state.status[(thread_idx + m_state.offset + pos_tbl - 1) & MTGP_MASK]);
     }
  
 public:
     // State
     mtgp32_state m_state;
     // Parameters
     unsigned int pos_tbl;
     unsigned int param_tbl[MTGP_TS];
     unsigned int temper_tbl[MTGP_TS];
     unsigned int sh1_tbl;
     unsigned int sh2_tbl;
     unsigned int single_temper_tbl[MTGP_TS];
     unsigned int mask;
  
 }; // mtgp32_engine class
  
 } // end namespace rocrand_device
  
 typedef rocrand_device::mtgp32_engine rocrand_state_mtgp32;
 typedef rocrand_device::mtgp32_state mtgp32_state;
 typedef rocrand_device::mtgp32_fast_params mtgp32_fast_params;
 typedef rocrand_device::mtgp32_params mtgp32_params;
  
 __host__
 inline rocrand_status rocrand_make_state_mtgp32(rocrand_state_mtgp32* state,
                                                 mtgp32_fast_params    params[],
                                                 int                   n,
                                                 unsigned long long    seed)
 {
     int i;
     rocrand_state_mtgp32 * h_state = (rocrand_state_mtgp32 *) malloc(sizeof(rocrand_state_mtgp32) * n);
     seed = seed ^ (seed >> 32);
  
     if (h_state == NULL)
         return ROCRAND_STATUS_ALLOCATION_FAILED;
  
     for (i = 0; i < n; i++) {
         rocrand_device::rocrand_mtgp32_init_state(&(h_state[i].m_state.status[0]), &params[i], (unsigned int)seed + i + 1);
         h_state[i].m_state.offset = 0;
         h_state[i].m_state.id = i;
         h_state[i].pos_tbl = params[i].pos;
         h_state[i].sh1_tbl = params[i].sh1;
         h_state[i].sh2_tbl = params[i].sh2;
         h_state[i].mask = params[0].mask;
         for (int j = 0; j < MTGP_TS; j++) {
             h_state[i].param_tbl[j] = params[i].tbl[j];
             h_state[i].temper_tbl[j] = params[i].tmp_tbl[j];
             h_state[i].single_temper_tbl[j] = params[i].flt_tmp_tbl[j];
         }
     }
  
     const hipError_t error
         = hipMemcpy(state, h_state, sizeof(rocrand_state_mtgp32) * n, hipMemcpyDefault);
     free(h_state);
  
     if(error != hipSuccess)
         return ROCRAND_STATUS_ALLOCATION_FAILED;
  
     return ROCRAND_STATUS_SUCCESS;
 }
  
 __host__
 inline rocrand_status rocrand_make_constant(const mtgp32_fast_params params[], mtgp32_params* p)
 {
     const int block_num = MTGP_BN_MAX;
     const int size1 = sizeof(uint32_t) * block_num;
     const int size2 = sizeof(uint32_t) * block_num * MTGP_TS;
     uint32_t *h_pos_tbl;
     uint32_t *h_sh1_tbl;
     uint32_t *h_sh2_tbl;
     uint32_t *h_param_tbl;
     uint32_t *h_temper_tbl;
     uint32_t *h_single_temper_tbl;
     uint32_t *h_mask;
     h_pos_tbl = (uint32_t *)malloc(size1);
     h_sh1_tbl = (uint32_t *)malloc(size1);
     h_sh2_tbl = (uint32_t *)malloc(size1);
     h_param_tbl = (uint32_t *)malloc(size2);
     h_temper_tbl = (uint32_t *)malloc(size2);
     h_single_temper_tbl = (uint32_t *)malloc(size2);
     h_mask = (uint32_t *)malloc(sizeof(uint32_t));
     rocrand_status status = ROCRAND_STATUS_SUCCESS;
  
     if (h_pos_tbl == NULL || h_sh1_tbl == NULL || h_sh2_tbl == NULL
         || h_param_tbl == NULL || h_temper_tbl == NULL || h_single_temper_tbl == NULL
         || h_mask == NULL) {
         printf("failure in allocating host memory for constant table.\n");
         status = ROCRAND_STATUS_ALLOCATION_FAILED;
     }
     else {
         h_mask[0] = params[0].mask;
         for (int i = 0; i < block_num; i++) {
             h_pos_tbl[i] = params[i].pos;
             h_sh1_tbl[i] = params[i].sh1;
             h_sh2_tbl[i] = params[i].sh2;
             for (int j = 0; j < MTGP_TS; j++) {
                 h_param_tbl[i * MTGP_TS + j] = params[i].tbl[j];
                 h_temper_tbl[i * MTGP_TS + j] = params[i].tmp_tbl[j];
                 h_single_temper_tbl[i * MTGP_TS + j] = params[i].flt_tmp_tbl[j];
             }
         }
  
         if (hipMemcpy(p->pos_tbl, h_pos_tbl, size1, hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
         if (hipMemcpy(p->sh1_tbl, h_sh1_tbl, size1, hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
         if (hipMemcpy(p->sh2_tbl, h_sh2_tbl, size1, hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
         if (hipMemcpy(p->param_tbl, h_param_tbl, size2, hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
         if (hipMemcpy(p->temper_tbl, h_temper_tbl, size2, hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
         if (hipMemcpy(p->single_temper_tbl, h_single_temper_tbl, size2, hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
         if (hipMemcpy(p->mask, h_mask, sizeof(unsigned int), hipMemcpyHostToDevice) != hipSuccess)
             status = ROCRAND_STATUS_ALLOCATION_FAILED;
     }
  
     free(h_pos_tbl);
     free(h_sh1_tbl);
     free(h_sh2_tbl);
     free(h_param_tbl);
     free(h_temper_tbl);
     free(h_single_temper_tbl);
     free(h_mask);
  
     return status;
 }
  
 __forceinline__ __device__
 unsigned int rocrand(rocrand_state_mtgp32* state)
 {
     return state->next();
 }
  
 __forceinline__ __device__
 void rocrand_mtgp32_block_copy(rocrand_state_mtgp32* src, rocrand_state_mtgp32* dest)
 {
     dest->copy(src);
 }
  
 __forceinline__ __device__
 void rocrand_mtgp32_set_params(rocrand_state_mtgp32* state, mtgp32_params* params)
 {
     state->set_params(params);
 }
  // end of group rocranddevice
  
 #endif // ROCRAND_MTGP32_H_