Skip to content

C
cuda_nla3d
Switch branch/tag
Find file
Normal viewHistoryPermalink
ocl_kernels_coo.hpp 12.2 KB
Newer Older
Arkadiy Marinenko's avatar
библиотека решателей (1 этап)
48920dc3
 
Arkadiy Marinenko committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 
// **************************************************************************
//
//    PARALUTION   www.paralution.com
//
//    Copyright (C) 2015  PARALUTION Labs UG (haftungsbeschränkt) & Co. KG
//                        Am Hasensprung 6, 76571 Gaggenau
//                        Handelsregister: Amtsgericht Mannheim, HRA 706051
//                        Vertreten durch:
//                        PARALUTION Labs Verwaltungs UG (haftungsbeschränkt)
//                        Am Hasensprung 6, 76571 Gaggenau
//                        Handelsregister: Amtsgericht Mannheim, HRB 721277
//                        Geschäftsführer: Dimitar Lukarski, Nico Trost
//
//    This program is free software: you can redistribute it and/or modify
//    it under the terms of the GNU General Public License as published by
//    the Free Software Foundation, either version 3 of the License, or
//    (at your option) any later version.
//
//    This program is distributed in the hope that it will be useful,
//    but WITHOUT ANY WARRANTY; without even the implied warranty of
//    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
//    GNU General Public License for more details.
//
//    You should have received a copy of the GNU General Public License
//    along with this program.  If not, see <http://www.gnu.org/licenses/>.
//
// **************************************************************************



// PARALUTION version 1.1.0 


#ifndef PARALUTION_OCL_KERNELS_COO_HPP_
#define PARALUTION_OCL_KERNELS_COO_HPP_

namespace paralution {

const char *ocl_kernels_coo =
	"__kernel void kernel_coo_permute(         const IndexType  nnz,\n"
	"                                 __global const IndexType *in_row,\n"
	"                                 __global const IndexType *in_col,\n"
	"                                 __global const IndexType *perm,\n"
	"                                 __global       IndexType *out_row,\n"
	"                                 __global       IndexType *out_col) {\n"
	"\n"
	"  IndexType ind = get_global_id(0);\n"
	"\n"
	"  for (IndexType i=ind; i<nnz; i+=get_local_size(0)) {\n"
	"\n"
	"    out_row[i] = perm[ in_row[i] ];\n"
	"    out_col[i] = perm[ in_col[i] ];\n"
	"\n"
	"  }\n"
	"\n"
	"}\n"
	"\n"
// ----------------------------------------------------------
// function segreduce_warp(...)
// ----------------------------------------------------------
// Modified and adapted from CUSP 0.3.1, 
// http://code.google.com/p/cusp-library/
// NVIDIA, APACHE LICENSE 2.0
// ----------------------------------------------------------
// CHANGELOG
// - adapted interface
// ----------------------------------------------------------
	"inline ValueType segreduce_warp(const IndexType thread_lane, IndexType row, ValueType val,\n"
	"                                __local IndexType *rows, __local ValueType *vals) {\n"
	"\n"
	"  IndexType tid = get_local_id(0);\n"
	"\n"
	"  rows[tid] = row;\n"
	"  vals[tid] = val;\n"
	"  \n"
	"  if( thread_lane >=  1 && row == rows[tid -  1] ) { vals[tid] = val = val + vals[tid -  1]; } \n"
	"  if( thread_lane >=  2 && row == rows[tid -  2] ) { vals[tid] = val = val + vals[tid -  2]; }\n"
	"  if( thread_lane >=  4 && row == rows[tid -  4] ) { vals[tid] = val = val + vals[tid -  4]; }\n"
	"  if( thread_lane >=  8 && row == rows[tid -  8] ) { vals[tid] = val = val + vals[tid -  8]; }\n"
	"  if( thread_lane >= 16 && row == rows[tid - 16] ) { vals[tid] = val = val + vals[tid - 16]; }\n"
	"  \n"
	"  return val;\n"
	"\n"
	"}\n"
	"\n"
// ----------------------------------------------------------
// function segreduce_block(...)
// ----------------------------------------------------------
// Modified and adapted from CUSP 0.3.1, 
// http://code.google.com/p/cusp-library/
// NVIDIA, APACHE LICENSE 2.0
// ----------------------------------------------------------
// CHANGELOG
// - adapted interface
// ----------------------------------------------------------
	"inline void segreduce_block(__local const IndexType *idx, __local ValueType *val) {\n"
	"\n"
	"  ValueType left = (ValueType) 0;\n"
	"  IndexType tid = get_local_id(0);\n"
	"\n"
	"  if( tid >=   1 && idx[tid] == idx[tid -   1] ) { left = val[tid -   1]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);  \n"
	"  if( tid >=   2 && idx[tid] == idx[tid -   2] ) { left = val[tid -   2]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"  if( tid >=   4 && idx[tid] == idx[tid -   4] ) { left = val[tid -   4]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"  if( tid >=   8 && idx[tid] == idx[tid -   8] ) { left = val[tid -   8]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"  if( tid >=  16 && idx[tid] == idx[tid -  16] ) { left = val[tid -  16]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"  if( tid >=  32 && idx[tid] == idx[tid -  32] ) { left = val[tid -  32]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);  \n"
	"  if( tid >=  64 && idx[tid] == idx[tid -  64] ) { left = val[tid -  64]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"  if( tid >= 128 && idx[tid] == idx[tid - 128] ) { left = val[tid - 128]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"  if( tid >= 256 && idx[tid] == idx[tid - 256] ) { left = val[tid - 256]; } barrier(CLK_LOCAL_MEM_FENCE); val[tid] += left; left = 0; barrier(CLK_LOCAL_MEM_FENCE);\n"
	"\n"
	"}\n"
	"\n"
// ----------------------------------------------------------
// function kernel_spmv_coo_flat(...)
// ----------------------------------------------------------
// Modified and adapted from CUSP 0.3.1, 
// http://code.google.com/p/cusp-library/
// NVIDIA, APACHE LICENSE 2.0
// ----------------------------------------------------------
// CHANGELOG
// - adapted interface
// ----------------------------------------------------------
	"__kernel void kernel_coo_spmv_flat(         const IndexType  num_nonzeros,\n"
	"                                            const IndexType  interval_size,\n"
	"                                   __global const IndexType *I, \n"
	"                                   __global const IndexType *J, \n"
	"                                   __global const ValueType *V, \n"
	"                                            const ValueType  scalar,\n"
	"                                   __global const ValueType *x, \n"
	"                                   __global       ValueType *y,\n"
	"                                   __global       IndexType *temp_rows,\n"
	"                                   __global       ValueType *temp_vals) {\n"
	"\n"
	"  __local volatile IndexType rows[48 * (BLOCK_SIZE/32)];\n"
	"  __local volatile ValueType vals[BLOCK_SIZE];\n"
	"\n"
	"        IndexType tid         = get_local_id(0);\n"
	"  const IndexType thread_id   = BLOCK_SIZE * get_group_id(0) + tid;\n"
	"  const IndexType thread_lane = tid & (WARP_SIZE-1);\n"
	"  const IndexType warp_id     = thread_id   / WARP_SIZE;\n"
	"\n"
	"  const IndexType interval_begin = warp_id * interval_size;\n"
	"  IndexType interval_end2 = interval_begin + interval_size;\n"
	"  if (interval_end2 > num_nonzeros)\n"
	"    interval_end2 = num_nonzeros;\n"
	"\n"
	"  const IndexType interval_end = interval_end2;\n"
	"\n"
	"  const IndexType idx = 16 * (tid/32 + 1) + tid;\n"
	"\n"
	"  rows[idx - 16] = -1;\n"
	"\n"
	"  if(interval_begin >= interval_end)\n"
	"    return;\n"
	"\n"
	"  if (thread_lane == 31) {\n"
	"    rows[idx] = I[interval_begin]; \n"
	"    vals[tid] = (ValueType) 0;\n"
	"  }\n"
	"\n"
	"  for(IndexType n = interval_begin + thread_lane; n < interval_end; n += WARP_SIZE) {\n"
	"    IndexType row = I[n];\n"
	"    ValueType val = scalar * V[n] * x[J[n]];\n"
	"    \n"
	"    if (thread_lane == 0) {\n"
	"      if(row == rows[idx + 31])\n"
	"        val += vals[tid + 31];\n"
	"      else\n"
	"        y[rows[idx + 31]] += vals[tid + 31];\n"
	"    }\n"
	"\n"
	"    rows[idx] = row;\n"
	"    vals[tid] = val;\n"
	"\n"
	"    if (row == rows[idx -  1]) { vals[tid] = val = val + vals[tid -  1]; } \n"
	"    if (row == rows[idx -  2]) { vals[tid] = val = val + vals[tid -  2]; }\n"
	"    if (row == rows[idx -  4]) { vals[tid] = val = val + vals[tid -  4]; }\n"
	"    if (row == rows[idx -  8]) { vals[tid] = val = val + vals[tid -  8]; }\n"
	"    if (row == rows[idx - 16]) { vals[tid] = val = val + vals[tid - 16]; }\n"
	"\n"
	"    if(thread_lane < 31 && row != rows[idx + 1])\n"
	"      y[row] += vals[tid];\n"
	"  }\n"
	"\n"
	"  if (thread_lane == 31) {\n"
	"    temp_rows[warp_id] = rows[idx];\n"
	"    temp_vals[warp_id] = vals[tid];\n"
	"  }\n"
	"\n"
	"}\n"
	"\n"
// ----------------------------------------------------------
// function kernel_spmv_coo_reduce_update(...)
// ----------------------------------------------------------
// Modified and adapted from CUSP 0.3.1, 
// http://code.google.com/p/cusp-library/
// NVIDIA, APACHE LICENSE 2.0
// ----------------------------------------------------------
// CHANGELOG
// - adapted interface
// ----------------------------------------------------------
	"__kernel void kernel_coo_spmv_reduce_update(         const IndexType  num_warps,\n"
	"                                            __global const IndexType *temp_rows,\n"
	"                                            __global const ValueType *temp_vals,\n"
	"                                            __global       ValueType *y) {\n"
	"\n"
	"  __local IndexType rows[BLOCK_SIZE + 1];\n"
	"  __local ValueType vals[BLOCK_SIZE + 1];\n"
	"\n"
	"  IndexType tid = get_local_id(0);\n"
	"\n"
	"  const IndexType end = num_warps - (num_warps & (BLOCK_SIZE - 1));\n"
	"\n"
	"  if (tid == 0) {\n"
	"    rows[BLOCK_SIZE] = (int) -1;\n"
	"    vals[BLOCK_SIZE] = (ValueType) 0;\n"
	"  }\n"
	"  \n"
	"  barrier(CLK_LOCAL_MEM_FENCE);\n"
	"\n"
	"  IndexType i = tid;\n"
	"\n"
	"  while (i < end) {\n"
	"    rows[tid] = temp_rows[i];\n"
	"    vals[tid] = temp_vals[i];\n"
	"\n"
	"    barrier(CLK_LOCAL_MEM_FENCE);\n"
	"\n"
	"    segreduce_block(rows, vals);\n"
	"\n"
	"    if (rows[tid] != rows[tid + 1])\n"
	"      y[rows[tid]] += vals[tid];\n"
	"\n"
	"    barrier(CLK_LOCAL_MEM_FENCE);\n"
	"\n"
	"    i += BLOCK_SIZE; \n"
	"  }\n"
	"\n"
	"  if (end < num_warps) {\n"
	"    if (i < num_warps) {\n"
	"      rows[tid] = temp_rows[i];\n"
	"      vals[tid] = temp_vals[i];\n"
	"    } else {\n"
	"      rows[tid] = (int) -1;\n"
	"      vals[tid] = (ValueType) 0;\n"
	"    }\n"
	"\n"
	"    barrier(CLK_LOCAL_MEM_FENCE);\n"
	" \n"
	"    segreduce_block(rows, vals);\n"
	"\n"
	"    if (i < num_warps)\n"
	"      if (rows[tid] != rows[tid + 1])\n"
	"        y[rows[tid]] += vals[tid];\n"
	"  }\n"
	"\n"
	"}\n"
	"\n"
// ----------------------------------------------------------
// function spmv_coo_serial_kernel(...)
// ----------------------------------------------------------
// Modified and adapted from CUSP 0.3.1, 
// http://code.google.com/p/cusp-library/
// NVIDIA, APACHE LICENSE 2.0
// ----------------------------------------------------------
// CHANGELOG
// - adapted interface
// ----------------------------------------------------------
	"__kernel void kernel_coo_spmv_serial(         const IndexType  num_entries,\n"
	"                                     __global const IndexType *I, \n"
	"                                     __global const IndexType *J, \n"
	"                                     __global const ValueType *V, \n"
	"                                              const ValueType  scalar,\n"
	"                                     __global const ValueType *x, \n"
	"                                     __global       ValueType *y,\n"
	"                                              const IndexType  shift) {\n"
	"\n"
	"  for(IndexType n = 0; n < num_entries-shift; n++)\n"
	"    y[I[n+shift]] += scalar * V[n+shift] * x[J[n+shift]];\n"
	"\n"
	"}\n"
	"\n"
	"__kernel void kernel_coo_csr_to_coo(const IndexType nrow, __global const IndexType *row_offset, __global IndexType *row) {\n"
	"\n"
	"  IndexType ai = get_global_id(0);\n"
	"  IndexType aj;\n"
	"\n"
	"  if (ai < nrow) {\n"
	"\n"
	"    for (aj=row_offset[ai]; aj<row_offset[ai+1]; ++aj)\n"
	"      row[aj] = ai;\n"
	"\n"
	"  }\n"
	"\n"
	"}\n"
	"\n"
;
}

#endif // PARALUTION_OCL_KERNELS_COO_HPP_