Rectangular division is done according to the picture: * * +------+-----+ * | | 2 | * | | | * | 1 +--+--+ * | |3 | 4| * | | | | * +------+--+--+ * * The longest side is divided first. */
| 866 | * The longest side is divided first. |
| 867 | */ |
| 868 | static void |
| 869 | rectDivision( |
| 870 | BlasFunctionID funcID, |
| 871 | const CLBlasKargs *args, |
| 872 | ListHead *seq, |
| 873 | cl_uint totalCUs) |
| 874 | { |
| 875 | SolutionStep *step, **sortedSteps; |
| 876 | ListNode *i, *j; |
| 877 | cl_int err; |
| 878 | cl_device_id device; |
| 879 | cl_uint nrCU, k, l; |
| 880 | SubproblemDim size, offset, stepSize; |
| 881 | unsigned int nrSteps = 0; |
| 882 | |
| 883 | /* 1. Sort steps according to the number of CU they have */ |
| 884 | /* NOTE: We expect small number of steps, so simple insertion sort |
| 885 | * would be enough. |
| 886 | */ |
| 887 | |
| 888 | sortedSteps = calloc(listLength(seq), sizeof(*sortedSteps)); |
| 889 | // assert(sortedSteps != NULL); |
| 890 | |
| 891 | k = 0; |
| 892 | for (i = listNodeFirst(seq); i != seq; i = i->next, nrSteps++) { |
| 893 | step = container_of(i, node, SolutionStep); |
| 894 | err = getQueueDevice(step->cmdQueue, &device); |
| 895 | |
| 896 | sortedSteps[k] = step; |
| 897 | nrCU = deviceComputeUnits(device, &err); |
| 898 | |
| 899 | for (j = i->next; j != seq; j = j->next) { |
| 900 | step = container_of(i, node, SolutionStep); |
| 901 | err = getQueueDevice(step->cmdQueue, &device); |
| 902 | |
| 903 | if (nrCU < deviceComputeUnits(device, &err)) { |
| 904 | sortedSteps[k] = step; |
| 905 | nrCU = deviceComputeUnits(device, &err); |
| 906 | } |
| 907 | } |
| 908 | |
| 909 | k++; |
| 910 | } |
| 911 | |
| 912 | /* 2. Calculate rectangle sizes */ |
| 913 | |
| 914 | kargsToProbDims(&offset, funcID, args, true); |
| 915 | kargsToProbDims(&size, funcID, args, false); |
| 916 | stepSize = size; |
| 917 | |
| 918 | for (l = 0; l < k; l++) { |
| 919 | step = sortedSteps[l]; |
| 920 | err = getQueueDevice(step->cmdQueue, &device); |
| 921 | nrCU = deviceComputeUnits(device, &err); |
| 922 | |
| 923 | if (totalCUs == 0) { |
| 924 | step->cmdQueue = NULL; |
| 925 | continue; |
no test coverage detected