| 360 | } |
| 361 | |
| 362 | static int |
| 363 | getDefaultDecomposition( |
| 364 | PGranularity *pgran, |
| 365 | SubproblemDim *subdims, |
| 366 | unsigned int subdimsNum, |
| 367 | void *pArgs) |
| 368 | { |
| 369 | SolutionStep *step = container_of( pgran , pgran, SolutionStep); |
| 370 | size_t maxWorkGroupSize; |
| 371 | cl_device_id devID = step->device.id; |
| 372 | size_t wgX, wgY; |
| 373 | pArgs = pArgs; |
| 374 | |
| 375 | clGetDeviceInfo(devID, CL_DEVICE_MAX_WORK_GROUP_SIZE, |
| 376 | sizeof(size_t), &maxWorkGroupSize, NULL); |
| 377 | |
| 378 | if( step->args.order == clblasColumnMajor ) |
| 379 | { |
| 380 | wgY = 16; // BH preferably 16(quarter wave-front) |
| 381 | subdims[0].y = wgY; |
| 382 | wgX = maxWorkGroupSize / wgY; // BW is left upto maxWorkGroupSize of the device |
| 383 | wgX = szmin( wgX, 16 ); |
| 384 | subdims[0].x = wgX; |
| 385 | } |
| 386 | else { |
| 387 | wgX = 16; |
| 388 | subdims[0].x = wgX; |
| 389 | wgY = maxWorkGroupSize / wgX; |
| 390 | wgY = szmin( wgY, 16 ); |
| 391 | subdims[0].y = wgY; |
| 392 | } |
| 393 | |
| 394 | pgran->wgDim = 1; //1D blocking |
| 395 | pgran->wgSize[0] = (unsigned int)(wgX * wgY); |
| 396 | pgran->wgSize[1] = 1; |
| 397 | |
| 398 | if(subdimsNum > 0) |
| 399 | { |
| 400 | subdims[0].itemX = subdims[0].x; |
| 401 | subdims[0].itemY = subdims[0].y; |
| 402 | subdims[0].bwidth = 1; |
| 403 | } |
| 404 | if(subdimsNum > 1) |
| 405 | { |
| 406 | subdims[1].itemY = 1; |
| 407 | subdims[1].itemX = 1; |
| 408 | subdims[1].y = subdims[1].itemY; |
| 409 | subdims[1].x = subdims[1].itemX; |
| 410 | subdims[1].bwidth = 1; |
| 411 | } |
| 412 | |
| 413 | return 0; |
| 414 | } |