00001 #ifndef _VIENNACL_VECTOR_SOURCE_HPP_
00002 #define _VIENNACL_VECTOR_SOURCE_HPP_
00003
00004 namespace viennacl
00005 {
00006 namespace linalg
00007 {
00008 namespace kernels
00009 {
00010 const char * const vector_align16_inplace_sub =
00011 "__kernel void inplace_sub(\n"
00012 " __global float16 * vec1,\n"
00013 " __global const float16 * vec2,\n"
00014 " unsigned int size) \n"
00015 "{ \n"
00016 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00017 " vec1[i] -= vec2[i];\n"
00018 "}\n"
00019 ;
00020
00021 const char * const vector_align16_mult =
00022 "__kernel void mult(\n"
00023 " __global const float16 * vec,\n"
00024 " __global const float * fac, \n"
00025 " __global float16 * result,\n"
00026 " unsigned int size) \n"
00027 "{ \n"
00028 " float factor = *fac;\n"
00029 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00030 " result[i] = vec[i] * factor;\n"
00031 "}\n"
00032 ;
00033
00034 const char * const vector_align16_sub =
00035 "__kernel void sub(\n"
00036 " __global const float16 * vec1,\n"
00037 " __global const float16 * vec2, \n"
00038 " __global float16 * result,\n"
00039 " unsigned int size)\n"
00040 "{ \n"
00041 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00042 " result[i] = vec1[i] - vec2[i];\n"
00043 "}\n"
00044 ;
00045
00046 const char * const vector_align16_cpu_inplace_mul =
00047 "\n"
00048 "__kernel void cpu_inplace_mult(\n"
00049 " __global float16 * vec,\n"
00050 " float factor, \n"
00051 " unsigned int size) \n"
00052 "{ \n"
00053 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00054 " vec[i] *= factor;\n"
00055 "}\n"
00056 "\n"
00057 ;
00058
00059 const char * const vector_align16_add =
00060 "__kernel void add(\n"
00061 " __global const float16 * vec1,\n"
00062 " __global const float16 * vec2, \n"
00063 " __global float16 * result,\n"
00064 " unsigned int size)\n"
00065 "{ \n"
00066 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00067 " result[i] = vec1[i] + vec2[i];\n"
00068 "}\n"
00069 ;
00070
00071 const char * const vector_align16_cpu_mult =
00072 "__kernel void cpu_mult(\n"
00073 " __global const float16 * vec,\n"
00074 " float factor, \n"
00075 " __global float16 * result,\n"
00076 " unsigned int size) \n"
00077 "{ \n"
00078 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00079 " result[i] = vec[i] * factor;\n"
00080 "}\n"
00081 ;
00082
00083 const char * const vector_align16_inplace_divide =
00084 "__kernel void inplace_divide(\n"
00085 " __global float16 * vec,\n"
00086 " __global const float * fac, //note: CPU variant is mapped to prod_scalar\n"
00087 " unsigned int size) \n"
00088 "{ \n"
00089 " float factor = *fac;\n"
00090 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00091 " vec[i] /= factor;\n"
00092 "}\n"
00093 ;
00094
00095 const char * const vector_align16_inplace_add =
00096 "__kernel void inplace_add(\n"
00097 " __global float16 * vec1,\n"
00098 " __global const float16 * vec2,\n"
00099 " unsigned int size) \n"
00100 "{ \n"
00101 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00102 " vec1[i] += vec2[i];\n"
00103 "}\n"
00104 ;
00105
00106 const char * const vector_align16_divide =
00107 "//Note: 'div' cannot be used because of complaints by the jit-compiler\n"
00108 "__kernel void divide(\n"
00109 " __global const float16 * vec,\n"
00110 " __global const float * fac, //note: CPU variant is mapped to prod_scalar\n"
00111 " __global float16 * result,\n"
00112 " unsigned int size) \n"
00113 "{ \n"
00114 " float factor = *fac;\n"
00115 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00116 " result[i] = vec[i] / factor;\n"
00117 "}\n"
00118 ;
00119
00120 const char * const vector_align16_inplace_mult =
00121 "__kernel void inplace_mult(\n"
00122 " __global float16 * vec,\n"
00123 " __global const float * fac, \n"
00124 " unsigned int size) \n"
00125 "{ \n"
00126 " float factor = *fac;\n"
00127 " for (unsigned int i = get_global_id(0); i < size/16; i += get_global_size(0))\n"
00128 " vec[i] *= factor;\n"
00129 "}\n"
00130 ;
00131
00132 const char * const vector_align4_inplace_div_add =
00133 "__kernel void inplace_div_add(\n"
00134 " __global float4 * vec1,\n"
00135 " __global const float4 * vec2,\n"
00136 " __global const float * fac, //CPU variant is mapped to mult_add\n"
00137 " unsigned int size\n"
00138 " ) \n"
00139 "{ \n"
00140 " float factor = *fac;\n"
00141 " for (unsigned int i = get_global_id(0); i < size/4; i += get_global_size(0))\n"
00142 " vec1[i] -= vec2[i] / factor;\n"
00143 "}\n"
00144 ;
00145
00146 const char * const vector_align4_cpu_mul_add =
00147 "__kernel void cpu_mul_add(\n"
00148 " __global const float4 * vec1,\n"
00149 " float factor,\n"
00150 " __global const float4 * vec2,\n"
00151 " __global float4 * result,\n"
00152 " unsigned int size\n"
00153 " ) \n"
00154 "{ \n"
00155 " for (unsigned int i = get_global_id(0); i < size/4; i += get_global_size(0))\n"
00156 " result[i] = vec1[i] * factor + vec2[i];\n"
00157 "}\n"
00158 ;
00159
00160 const char * const vector_align4_inplace_mul_sub =
00161 "__kernel void inplace_mul_sub(\n"
00162 " __global float4 * vec1,\n"
00163 " __global const float4 * vec2,\n"
00164 " __global const float * fac, //CPU variant is mapped to mult_add\n"
00165 " unsigned int size\n"
00166 " ) \n"
00167 "{ \n"
00168 " float factor = *fac;\n"
00169 " for (unsigned int i = get_global_id(0); i < size/4; i += get_global_size(0))\n"
00170 " vec1[i] -= vec2[i] * factor;\n"
00171 "}\n"
00172 ;
00173
00174 const char * const vector_align4_inplace_mul_add =
00175 "__kernel void inplace_mul_add(\n"
00176 " __global float4 * vec1,\n"
00177 " __global const float4 * vec2,\n"
00178 " __global const float * fac,\n"
00179 " unsigned int size\n"
00180 " ) \n"
00181 "{ \n"
00182 " float factor = *fac;\n"
00183 " unsigned int size_div_4 = size >> 2;\n"
00184 " for (unsigned int i = get_global_id(0); i < size_div_4; i += get_global_size(0))\n"
00185 " vec1[i] += vec2[i] * factor;\n"
00186 "}\n"
00187 ;
00188
00189 const char * const vector_align4_mul_add =
00190 "__kernel void mul_add(\n"
00191 " __global const float4 * vec1,\n"
00192 " __global const float * fac,\n"
00193 " __global const float4 * vec2,\n"
00194 " __global float4 * result,\n"
00195 " unsigned int size\n"
00196 " ) \n"
00197 "{ \n"
00198 " float factor = *fac;\n"
00199 " for (unsigned int i = get_global_id(0); i < size/4; i += get_global_size(0))\n"
00200 " result[i] = vec1[i] * factor + vec2[i];\n"
00201 "}\n"
00202 ;
00203
00204 const char * const vector_align4_cpu_inplace_mul_add =
00205 "__kernel void cpu_inplace_mul_add(\n"
00206 " __global float4 * vec1,\n"
00207 " __global const float4 * vec2,\n"
00208 " float factor,\n"
00209 " unsigned int size\n"
00210 " ) \n"
00211 "{ \n"
00212 " for (unsigned int i = get_global_id(0); i < size/4; i += get_global_size(0))\n"
00213 " vec1[i] += vec2[i] * factor;\n"
00214 "}\n"
00215 ;
00216
00217 const char * const vector_align4_inplace_div_sub =
00218 "__kernel void inplace_div_sub(\n"
00219 " __global float4 * vec1,\n"
00220 " __global const float4 * vec2,\n"
00221 " __global const float * fac, //CPU variant is mapped to mult_add\n"
00222 " unsigned int size\n"
00223 " ) \n"
00224 "{ \n"
00225 " float factor = *fac;\n"
00226 " for (unsigned int i = get_global_id(0); i < size/4; i += get_global_size(0))\n"
00227 " vec1[i] -= vec2[i] / factor;\n"
00228 "}\n"
00229 ;
00230
00231 const char * const vector_align1_inplace_sub =
00232 "__kernel void inplace_sub(\n"
00233 " __global float * vec1,\n"
00234 " __global const float * vec2,\n"
00235 " unsigned int size) \n"
00236 "{ \n"
00237 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00238 " vec1[i] -= vec2[i];\n"
00239 "}\n"
00240 ;
00241
00242 const char * const vector_align1_norm_inf =
00243 "\n"
00244 "////// norm_inf\n"
00245 "float impl_norm_inf(\n"
00246 " __global const float * vec,\n"
00247 " unsigned int start_index,\n"
00248 " unsigned int end_index,\n"
00249 " __local float * tmp_buffer)\n"
00250 "{\n"
00251 " float tmp = 0;\n"
00252 " for (unsigned int i = start_index + get_local_id(0); i < end_index; i += get_local_size(0))\n"
00253 " tmp = fmax(fabs(vec[i]), tmp);\n"
00254 " tmp_buffer[get_local_id(0)] = tmp;\n"
00255 " \n"
00256 " //step 2: parallel reduction:\n"
00257 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00258 " {\n"
00259 " barrier(CLK_LOCAL_MEM_FENCE);\n"
00260 " if (get_global_id(0) < stride)\n"
00261 " tmp_buffer[get_global_id(0)] = fmax(tmp_buffer[get_global_id(0)], tmp_buffer[get_global_id(0)+stride]);\n"
00262 " }\n"
00263 " \n"
00264 " return tmp_buffer[0];\n"
00265 "}\n"
00266 "\n"
00267 "__kernel void norm_inf(\n"
00268 " __global const float * vec,\n"
00269 " unsigned int size,\n"
00270 " __local float * tmp_buffer,\n"
00271 " global float * group_buffer)\n"
00272 "{\n"
00273 " float tmp = impl_norm_inf(vec,\n"
00274 " ( get_group_id(0) * size) / get_num_groups(0),\n"
00275 " ((get_group_id(0) + 1) * size) / get_num_groups(0),\n"
00276 " tmp_buffer);\n"
00277 " \n"
00278 " if (get_local_id(0) == 0)\n"
00279 " group_buffer[get_group_id(0)] = tmp; \n"
00280 "}\n"
00281 ;
00282
00283 const char * const vector_align1_index_norm_inf =
00284 "//index_norm_inf:\n"
00285 "unsigned int float_vector1_index_norm_inf_impl(\n"
00286 " __global const float * vec,\n"
00287 " unsigned int size,\n"
00288 " __local float * float_buffer,\n"
00289 " __local unsigned int * index_buffer)\n"
00290 "{\n"
00291 " //step 1: fill buffer:\n"
00292 " float cur_max = 0.0f;\n"
00293 " float tmp;\n"
00294 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00295 " {\n"
00296 " tmp = fabs(vec[i]);\n"
00297 " if (cur_max < tmp)\n"
00298 " {\n"
00299 " float_buffer[get_global_id(0)] = tmp;\n"
00300 " index_buffer[get_global_id(0)] = i;\n"
00301 " cur_max = tmp;\n"
00302 " }\n"
00303 " }\n"
00304 " \n"
00305 " //step 2: parallel reduction:\n"
00306 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00307 " {\n"
00308 " barrier(CLK_LOCAL_MEM_FENCE);\n"
00309 " if (get_global_id(0) < stride)\n"
00310 " {\n"
00311 " //find the first occurring index\n"
00312 " if (float_buffer[get_global_id(0)] < float_buffer[get_global_id(0)+stride])\n"
00313 " {\n"
00314 " index_buffer[get_global_id(0)] = index_buffer[get_global_id(0)+stride];\n"
00315 " float_buffer[get_global_id(0)] = float_buffer[get_global_id(0)+stride];\n"
00316 " }\n"
00317 " \n"
00318 " //index_buffer[get_global_id(0)] = float_buffer[get_global_id(0)] < float_buffer[get_global_id(0)+stride] ? index_buffer[get_global_id(0)+stride] : index_buffer[get_global_id(0)];\n"
00319 " //float_buffer[get_global_id(0)] = max(float_buffer[get_global_id(0)], float_buffer[get_global_id(0)+stride]);\n"
00320 " }\n"
00321 " }\n"
00322 " \n"
00323 " return index_buffer[0];\n"
00324 "}\n"
00325 "\n"
00326 "__kernel void index_norm_inf(\n"
00327 " __global float * vec,\n"
00328 " unsigned int size,\n"
00329 " __local float * float_buffer,\n"
00330 " __local unsigned int * index_buffer,\n"
00331 " global unsigned int * result) \n"
00332 "{ \n"
00333 " unsigned int tmp = float_vector1_index_norm_inf_impl(vec, size, float_buffer, index_buffer);\n"
00334 " if (get_global_id(0) == 0) *result = tmp;\n"
00335 "}\n"
00336 "\n"
00337 "\n"
00338 ;
00339
00340 const char * const vector_align1_mult =
00341 "__kernel void mult(\n"
00342 " __global const float * vec,\n"
00343 " __global const float * fac, \n"
00344 " __global float * result,\n"
00345 " unsigned int size) \n"
00346 "{ \n"
00347 " float factor = *fac;\n"
00348 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00349 " result[i] = vec[i] * factor;\n"
00350 "}\n"
00351 ;
00352
00353 const char * const vector_align1_swap =
00354 "////// swap:\n"
00355 "__kernel void swap(\n"
00356 " __global float * vec1,\n"
00357 " __global float * vec2,\n"
00358 " unsigned int size\n"
00359 " ) \n"
00360 "{ \n"
00361 " float tmp;\n"
00362 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00363 " {\n"
00364 " tmp = vec2[i];\n"
00365 " vec2[i] = vec1[i];\n"
00366 " vec1[i] = tmp;\n"
00367 " }\n"
00368 "}\n"
00369 " \n"
00370 ;
00371
00372 const char * const vector_align1_inplace_div_add =
00373 "///// divide add:\n"
00374 "__kernel void inplace_div_add(\n"
00375 " __global float * vec1,\n"
00376 " __global const float * vec2,\n"
00377 " __global const float * fac, //CPU variant is mapped to mult_add\n"
00378 " unsigned int size\n"
00379 " ) \n"
00380 "{ \n"
00381 " float factor = *fac;\n"
00382 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00383 " vec1[i] -= vec2[i] / factor;\n"
00384 "}\n"
00385 ;
00386
00387 const char * const vector_align1_norm_2 =
00388 "//helper:\n"
00389 "void helper_norm2_parallel_reduction( __local float * tmp_buffer )\n"
00390 "{\n"
00391 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00392 " {\n"
00393 " barrier(CLK_LOCAL_MEM_FENCE);\n"
00394 " if (get_global_id(0) < stride)\n"
00395 " tmp_buffer[get_global_id(0)] += tmp_buffer[get_global_id(0)+stride];\n"
00396 " }\n"
00397 "}\n"
00398 "\n"
00399 "////// norm_2\n"
00400 "float impl_norm_2(\n"
00401 " __global const float * vec,\n"
00402 " unsigned int start_index,\n"
00403 " unsigned int end_index,\n"
00404 " __local float * tmp_buffer)\n"
00405 "{\n"
00406 " float tmp = 0;\n"
00407 " float vec_entry = 0;\n"
00408 " for (unsigned int i = start_index + get_local_id(0); i < end_index; i += get_local_size(0))\n"
00409 " {\n"
00410 " vec_entry = vec[i];\n"
00411 " tmp += vec_entry * vec_entry;\n"
00412 " }\n"
00413 " tmp_buffer[get_local_id(0)] = tmp;\n"
00414 " \n"
00415 " helper_norm2_parallel_reduction(tmp_buffer);\n"
00416 " \n"
00417 " return tmp_buffer[0];\n"
00418 "};\n"
00419 "\n"
00420 "__kernel void norm_2(\n"
00421 " __global const float * vec,\n"
00422 " unsigned int size,\n"
00423 " __local float * tmp_buffer,\n"
00424 " global float * group_buffer)\n"
00425 "{\n"
00426 " float tmp = impl_norm_2(vec,\n"
00427 " ( get_group_id(0) * size) / get_num_groups(0),\n"
00428 " ((get_group_id(0) + 1) * size) / get_num_groups(0),\n"
00429 " tmp_buffer);\n"
00430 " \n"
00431 " if (get_local_id(0) == 0)\n"
00432 " group_buffer[get_group_id(0)] = tmp; \n"
00433 "}\n"
00434 "\n"
00435 ;
00436
00437 const char * const vector_align1_sub =
00438 "__kernel void sub(\n"
00439 " __global const float * vec1,\n"
00440 " __global const float * vec2, \n"
00441 " __global float * result,\n"
00442 " unsigned int size)\n"
00443 "{ \n"
00444 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00445 " result[i] = vec1[i] - vec2[i];\n"
00446 "}\n"
00447 ;
00448
00449 const char * const vector_align1_cpu_mul_add =
00450 "__kernel void cpu_mul_add(\n"
00451 " __global const float * vec1,\n"
00452 " float factor,\n"
00453 " __global const float * vec2,\n"
00454 " __global float * result,\n"
00455 " unsigned int size\n"
00456 " ) \n"
00457 "{ \n"
00458 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00459 " result[i] = vec1[i] * factor + vec2[i];\n"
00460 "}\n"
00461 ;
00462
00463 const char * const vector_align1_vmax =
00464 "__kernel void vmax(\n"
00465 " __global float * vec1,\n"
00466 " __global float * result,\n"
00467 " unsigned int size) \n"
00468 "{ \n"
00469 " //parallel reduction on global memory:\n"
00470 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00471 " {\n"
00472 " if (get_global_id(0) < stride)\n"
00473 " vec1[get_global_id(0)] = fmax(vec1[get_global_id(0)+stride], vec1[get_global_id(0)]);\n"
00474 " barrier(CLK_GLOBAL_MEM_FENCE);\n"
00475 " }\n"
00476 " \n"
00477 " if (get_global_id(0) == 0)\n"
00478 " *result = vec1[0]; \n"
00479 "}\n"
00480 ;
00481
00482 const char * const vector_align1_inner_prod =
00483 "//helper:\n"
00484 "void helper_inner_prod_parallel_reduction( __local float * tmp_buffer )\n"
00485 "{\n"
00486 " for (unsigned int stride = get_local_size(0)/2; stride > 0; stride /= 2)\n"
00487 " {\n"
00488 " barrier(CLK_LOCAL_MEM_FENCE);\n"
00489 " if (get_local_id(0) < stride)\n"
00490 " tmp_buffer[get_local_id(0)] += tmp_buffer[get_local_id(0)+stride];\n"
00491 " }\n"
00492 "}\n"
00493 "//////// inner products:\n"
00494 "float impl_inner_prod(\n"
00495 " __global const float * vec1,\n"
00496 " __global const float * vec2,\n"
00497 " unsigned int start_index,\n"
00498 " unsigned int end_index,\n"
00499 " __local float * tmp_buffer)\n"
00500 "{\n"
00501 " float tmp = 0;\n"
00502 " for (unsigned int i = start_index + get_local_id(0); i < end_index; i += get_local_size(0))\n"
00503 " tmp += vec1[i] * vec2[i];\n"
00504 " tmp_buffer[get_local_id(0)] = tmp;\n"
00505 " \n"
00506 " helper_inner_prod_parallel_reduction(tmp_buffer);\n"
00507 " \n"
00508 " return tmp_buffer[0];\n"
00509 "}\n"
00510 "__kernel void inner_prod(\n"
00511 " __global const float * vec1,\n"
00512 " __global const float * vec2,\n"
00513 " unsigned int size,\n"
00514 " __local float * tmp_buffer,\n"
00515 " global float * group_buffer)\n"
00516 "{\n"
00517 " float tmp = impl_inner_prod(vec1,\n"
00518 " vec2,\n"
00519 " ( get_group_id(0) * size) / get_num_groups(0),\n"
00520 " ((get_group_id(0) + 1) * size) / get_num_groups(0),\n"
00521 " tmp_buffer);\n"
00522 " \n"
00523 " if (get_local_id(0) == 0)\n"
00524 " group_buffer[get_group_id(0)] = tmp;\n"
00525 " \n"
00526 "}\n"
00527 ;
00528
00529 const char * const vector_align1_add =
00530 "__kernel void add(\n"
00531 " __global const float * vec1,\n"
00532 " __global const float * vec2, \n"
00533 " __global float * result,\n"
00534 " unsigned int size) \n"
00535 "{ \n"
00536 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00537 " result[i] = vec1[i] + vec2[i];\n"
00538 "}\n"
00539 ;
00540
00541 const char * const vector_align1_plane_rotation =
00542 "////// plane rotation: (x,y) <- (\alpha x + \beta y, -\beta x + \alpha y)\n"
00543 "__kernel void plane_rotation(\n"
00544 " __global float * vec1,\n"
00545 " __global float * vec2, \n"
00546 " float alpha,\n"
00547 " float beta,\n"
00548 " unsigned int size) \n"
00549 "{ \n"
00550 " float tmp1 = 0;\n"
00551 " float tmp2 = 0;\n"
00552 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00553 " {\n"
00554 " tmp1 = vec1[i];\n"
00555 " tmp2 = vec2[i];\n"
00556 " \n"
00557 " vec1[i] = alpha * tmp1 + beta * tmp2;\n"
00558 " vec2[i] = alpha * tmp2 - beta * tmp1;\n"
00559 " }\n"
00560 "}\n"
00561 ;
00562
00563 const char * const vector_align1_inplace_mul_sub =
00564 "__kernel void inplace_mul_sub(\n"
00565 " __global float * vec1,\n"
00566 " __global const float * vec2,\n"
00567 " __global const float * fac, //CPU variant is mapped to mult_add\n"
00568 " unsigned int size\n"
00569 " ) \n"
00570 "{ \n"
00571 " float factor = *fac;\n"
00572 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00573 " vec1[i] -= vec2[i] * factor;\n"
00574 "}\n"
00575 ;
00576
00577 const char * const vector_align1_inplace_mul_add =
00578 "__kernel void inplace_mul_add(\n"
00579 " __global float * vec1,\n"
00580 " __global const float * vec2,\n"
00581 " __global const float * fac,\n"
00582 " unsigned int size\n"
00583 " ) \n"
00584 "{ \n"
00585 " float factor = *fac;\n"
00586 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00587 " vec1[i] += vec2[i] * factor;\n"
00588 "}\n"
00589 ;
00590
00591 const char * const vector_align1_mul_add =
00592 "__kernel void mul_add(\n"
00593 " __global const float * vec1,\n"
00594 " __global const float * fac,\n"
00595 " __global const float * vec2,\n"
00596 " __global float * result,\n"
00597 " unsigned int size\n"
00598 " ) \n"
00599 "{ \n"
00600 " float factor = *fac;\n"
00601 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00602 " result[i] = vec1[i] * factor + vec2[i];\n"
00603 "}\n"
00604 ;
00605
00606 const char * const vector_align1_cpu_mult =
00607 "__kernel void cpu_mult(\n"
00608 " __global const float * vec,\n"
00609 " float factor, \n"
00610 " __global float * result,\n"
00611 " unsigned int size) \n"
00612 "{ \n"
00613 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00614 " result[i] = vec[i] * factor;\n"
00615 "}\n"
00616 ;
00617
00618 const char * const vector_align1_inplace_divide =
00619 "__kernel void inplace_divide(\n"
00620 " __global float * vec,\n"
00621 " __global const float * fac, //note: CPU variant is mapped to prod_scalar\n"
00622 " unsigned int size) \n"
00623 "{ \n"
00624 " float factor = *fac;\n"
00625 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00626 " vec[i] /= factor;\n"
00627 "}\n"
00628 ;
00629
00630 const char * const vector_align1_sqrt_sum =
00631 "__kernel void sqrt_sum(\n"
00632 " __global float * vec1,\n"
00633 " __global float * result,\n"
00634 " unsigned int size) \n"
00635 "{ \n"
00636 " //parallel reduction on global memory: \n"
00637 " \n"
00638 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00639 " {\n"
00640 " if (get_global_id(0) < stride)\n"
00641 " vec1[get_global_id(0)] += vec1[get_global_id(0)+stride];\n"
00642 " barrier(CLK_GLOBAL_MEM_FENCE);\n"
00643 " }\n"
00644 " \n"
00645 " if (get_global_id(0) == 0)\n"
00646 " *result = sqrt(vec1[0]);\n"
00647 " \n"
00648 "}\n"
00649 ;
00650
00651 const char * const vector_align1_cpu_inplace_mul_add =
00652 "__kernel void cpu_inplace_mul_add(\n"
00653 " __global float * vec1,\n"
00654 " __global const float * vec2,\n"
00655 " float factor,\n"
00656 " unsigned int size\n"
00657 " ) \n"
00658 "{ \n"
00659 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00660 " vec1[i] += vec2[i] * factor;\n"
00661 "}\n"
00662 ;
00663
00664 const char * const vector_align1_inplace_add =
00665 "__kernel void inplace_add(\n"
00666 " __global float * vec1,\n"
00667 " __global const float * vec2,\n"
00668 " unsigned int size) \n"
00669 "{ \n"
00670 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00671 " vec1[i] += vec2[i];\n"
00672 "}\n"
00673 ;
00674
00675 const char * const vector_align1_divide =
00676 "// Note: name 'div' is not allowed by the jit-compiler\n"
00677 "__kernel void divide(\n"
00678 " __global const float * vec,\n"
00679 " __global const float * fac, //note: CPU variant is mapped to prod_scalar\n"
00680 " __global float * result,\n"
00681 " unsigned int size) \n"
00682 "{ \n"
00683 " float factor = *fac;\n"
00684 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00685 " result[i] = vec[i] / factor;\n"
00686 "}\n"
00687 ;
00688
00689 const char * const vector_align1_norm_1 =
00690 "//helper:\n"
00691 "void helper_norm1_parallel_reduction( __local float * tmp_buffer )\n"
00692 "{\n"
00693 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00694 " {\n"
00695 " barrier(CLK_LOCAL_MEM_FENCE);\n"
00696 " if (get_global_id(0) < stride)\n"
00697 " tmp_buffer[get_global_id(0)] += tmp_buffer[get_global_id(0)+stride];\n"
00698 " }\n"
00699 "}\n"
00700 "\n"
00701 "////// norm_1\n"
00702 "float impl_norm_1(\n"
00703 " __global const float * vec,\n"
00704 " unsigned int start_index,\n"
00705 " unsigned int end_index,\n"
00706 " __local float * tmp_buffer)\n"
00707 "{\n"
00708 " float tmp = 0;\n"
00709 " for (unsigned int i = start_index + get_local_id(0); i < end_index; i += get_local_size(0))\n"
00710 " tmp += fabs(vec[i]);\n"
00711 " \n"
00712 " tmp_buffer[get_local_id(0)] = tmp;\n"
00713 " \n"
00714 " helper_norm1_parallel_reduction(tmp_buffer);\n"
00715 " \n"
00716 " return tmp_buffer[0];\n"
00717 "};\n"
00718 "\n"
00719 "__kernel void norm_1(\n"
00720 " __global const float * vec,\n"
00721 " unsigned int size,\n"
00722 " __local float * tmp_buffer,\n"
00723 " global float * group_buffer)\n"
00724 "{\n"
00725 " float tmp = impl_norm_1(vec,\n"
00726 " ( get_group_id(0) * size) / get_num_groups(0),\n"
00727 " ((get_group_id(0) + 1) * size) / get_num_groups(0),\n"
00728 " tmp_buffer);\n"
00729 " \n"
00730 " if (get_local_id(0) == 0)\n"
00731 " group_buffer[get_group_id(0)] = tmp; \n"
00732 "}\n"
00733 "\n"
00734 ;
00735
00736 const char * const vector_align1_clear =
00737 "__kernel void clear(\n"
00738 " __global float * vec,\n"
00739 " unsigned int size) \n"
00740 "{ \n"
00741 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00742 " vec[i] = 0;\n"
00743 "}\n"
00744 ;
00745
00746 const char * const vector_align1_cpu_inplace_mult =
00747 "__kernel void cpu_inplace_mult(\n"
00748 " __global float * vec,\n"
00749 " float factor, \n"
00750 " unsigned int size) \n"
00751 "{ \n"
00752 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00753 " vec[i] *= factor;\n"
00754 "}\n"
00755 ;
00756
00757 const char * const vector_align1_inplace_mult =
00758 "__kernel void inplace_mult(\n"
00759 " __global float * vec,\n"
00760 " __global const float * fac, \n"
00761 " unsigned int size) \n"
00762 "{ \n"
00763 " float factor = *fac;\n"
00764 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00765 " vec[i] *= factor;\n"
00766 "}\n"
00767 ;
00768
00769 const char * const vector_align1_sum =
00770 "__kernel void sum(\n"
00771 " __global float * vec1,\n"
00772 " __global float * result) \n"
00773 "{ \n"
00774 " //parallel reduction on global memory: \n"
00775 " \n"
00776 " for (unsigned int stride = get_global_size(0)/2; stride > 0; stride /= 2)\n"
00777 " {\n"
00778 " if (get_global_id(0) < stride)\n"
00779 " vec1[get_global_id(0)] += vec1[get_global_id(0)+stride];\n"
00780 " barrier(CLK_GLOBAL_MEM_FENCE);\n"
00781 " }\n"
00782 " \n"
00783 " if (get_global_id(0) == 0)\n"
00784 " *result = vec1[0]; \n"
00785 "}\n"
00786 ;
00787
00788 const char * const vector_align1_inplace_div_sub =
00789 "///// divide substract:\n"
00790 "__kernel void inplace_div_sub(\n"
00791 " __global float * vec1,\n"
00792 " __global const float * vec2,\n"
00793 " __global const float * fac, //CPU variant is mapped to mult_add\n"
00794 " unsigned int size\n"
00795 " ) \n"
00796 "{ \n"
00797 " float factor = *fac;\n"
00798 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00799 " vec1[i] -= vec2[i] / factor;\n"
00800 "}\n"
00801 ;
00802
00803 const char * const vector_align1_diag_precond =
00804 "__kernel void diag_precond(\n"
00805 " __global const float * diag_A_inv, \n"
00806 " __global float * x, \n"
00807 " unsigned int size) \n"
00808 "{ \n"
00809 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00810 " x[i] *= diag_A_inv[i];\n"
00811 "}\n"
00812 ;
00813
00814 const char * const vector_align1_mul_sub =
00815 "///// multiply subtract:\n"
00816 "__kernel void mul_sub(\n"
00817 " __global const float * vec1,\n"
00818 " __global const float * fac,\n"
00819 " __global const float * vec2,\n"
00820 " __global float * result,\n"
00821 " unsigned int size\n"
00822 " ) \n"
00823 "{ \n"
00824 " float factor = *fac;\n"
00825 " for (unsigned int i = get_global_id(0); i < size; i += get_global_size(0))\n"
00826 " result[i] = vec1[i] * factor - vec2[i];\n"
00827 "}\n"
00828 ;
00829
00830 }
00831 }
00832 }
00833 #endif