| 616 | } |
| 617 | |
| 618 | std::string |
| 619 | Step::globalWorkSize() |
| 620 | { |
| 621 | size_t globalWorkSize[MAX_WORK_DIM] = { 0, 0, 0 }; |
| 622 | std::stringstream ss; |
| 623 | SubproblemDim dims[MAX_SUBDIMS]; |
| 624 | |
| 625 | memcpy(dims, step_.subdims, sizeof(dims)); |
| 626 | |
| 627 | if (pattern_->sops->calcThreads) { |
| 628 | pattern_->sops->calcThreads(globalWorkSize, step_.subdims, |
| 629 | &step_.pgran, &step_.args, &kextra_); |
| 630 | } |
| 631 | else { |
| 632 | SubproblemDim globDim; |
| 633 | const PGranularity *pg; |
| 634 | |
| 635 | pg = (pattern_->nrLevels == 1) ? NULL : &step_.pgran; |
| 636 | kargsToProbDims(&globDim, blasFunctionID(), &step_.args, false); |
| 637 | |
| 638 | // fixup dimensions in respect with desired work dispatch order |
| 639 | if ((pgran().wgDim == 2) && pattern_->sops->innerDecompositionAxis) { |
| 640 | if (pattern_->sops->innerDecompositionAxis(&step_.args) == |
| 641 | DECOMP_AXIS_X) { |
| 642 | |
| 643 | /* |
| 644 | * these dimensions will not be used more anywhere, so we can |
| 645 | * just swap them |
| 646 | */ |
| 647 | swapDimXY(&dims[0]); |
| 648 | swapDimXY(&dims[1]); |
| 649 | swapDimXY(&globDim); |
| 650 | } |
| 651 | } |
| 652 | |
| 653 | calcGlobalThreads(globalWorkSize, dims, pg, globDim.y, globDim.x); |
| 654 | } |
| 655 | |
| 656 | for (unsigned int i = 0; i < pgran().wgDim; i++) { |
| 657 | if (i != 0) { |
| 658 | ss << ", "; |
| 659 | } |
| 660 | ss << globalWorkSize[i]; |
| 661 | } |
| 662 | |
| 663 | return ss.str(); |
| 664 | } |
| 665 | |
| 666 | void |
| 667 | Step::setKernelName(std::string name) |