| 178 | } |
| 179 | |
| 180 | static void |
| 181 | calcNrThreads( |
| 182 | size_t threads[2], |
| 183 | const SubproblemDim *subdims, |
| 184 | const PGranularity *pgran, |
| 185 | const void *args, |
| 186 | const void *_extra) |
| 187 | { |
| 188 | DUMMY_ARG_USAGE(subdims); |
| 189 | const CLBLASKernExtra *extra = ( CLBLASKernExtra *)_extra; |
| 190 | CLBlasKargs *kargs = (CLBlasKargs *)args; |
| 191 | SolutionStep *step = container_of(kargs, args, SolutionStep); |
| 192 | TargetDevice *kDevice = &(step->device); |
| 193 | |
| 194 | cl_int err; |
| 195 | unsigned int numComputeUnits = deviceComputeUnits( (kDevice->id), &err ); |
| 196 | if(err != CL_SUCCESS) { |
| 197 | numComputeUnits = 1; |
| 198 | } |
| 199 | |
| 200 | unsigned int vecLen = extra->vecLenA; |
| 201 | unsigned int blockSize = pgran->wgSize[0] * pgran->wgSize[1]; |
| 202 | |
| 203 | unsigned int wgToSpawn = ((kargs->N + (blockSize*vecLen) - 1)/ (blockSize*vecLen)); |
| 204 | wgToSpawn = min( wgToSpawn, (numComputeUnits * WORKGROUPS_PER_CU) ); |
| 205 | |
| 206 | threads[0] = wgToSpawn * blockSize; |
| 207 | threads[1] = 1; |
| 208 | } |
| 209 | |
| 210 | // |
| 211 | // FIXME: Report correct return value - Needs change in KPRINTF |
nothing calls this directly
no test coverage detected