| 250 | |
| 251 | |
| 252 | template <typename ElemType> nano_time_t |
| 253 | HbmvPerformanceTest<ElemType>::clblasPerfSingle(void) |
| 254 | { |
| 255 | nano_time_t time; |
| 256 | cl_event event; |
| 257 | cl_int status; |
| 258 | size_t lenY; |
| 259 | cl_command_queue queue = base_->commandQueues()[0]; |
| 260 | |
| 261 | lenY = ((params_.N) - 1)* params_.incy + 1 + params_.offCY; |
| 262 | |
| 263 | status = clEnqueueWriteBuffer(queue, mobjY_, CL_TRUE, 0, |
| 264 | lenY * sizeof(ElemType), backY_, 0, NULL, &event); |
| 265 | |
| 266 | if (status != CL_SUCCESS) { |
| 267 | cerr << "Vector Y buffer object enqueuing error, status = " << |
| 268 | status << endl; |
| 269 | |
| 270 | return NANOTIME_ERR; |
| 271 | } |
| 272 | |
| 273 | status = clWaitForEvents(1, &event); |
| 274 | if (status != CL_SUCCESS) { |
| 275 | cout << "Wait on event failed, status = " << |
| 276 | status << endl; |
| 277 | |
| 278 | return NANOTIME_ERR; |
| 279 | } |
| 280 | |
| 281 | event = NULL; |
| 282 | time = getCurrentTime(); |
| 283 | int iter = 20; |
| 284 | for ( int i = 1; i <= iter; i++) |
| 285 | { |
| 286 | status = clMath::clblas::hbmv(params_.order, params_.uplo, params_.N, params_.K, |
| 287 | alpha, mobjA_, params_.offA, params_.lda, mobjX_, params_.offBX, params_.incx, |
| 288 | beta, mobjY_, params_.offCY, params_.incy, 1, &queue, 0, NULL, &event); |
| 289 | |
| 290 | if (status != CL_SUCCESS) { |
| 291 | cerr << "The CLBLAS GBMV function failed, status = " << |
| 292 | status << endl; |
| 293 | return NANOTIME_ERR; |
| 294 | } |
| 295 | } |
| 296 | clFinish( queue ); |
| 297 | time = getCurrentTime() - time; |
| 298 | time /= iter; |
| 299 | |
| 300 | return time; |
| 301 | } |
| 302 | |
| 303 | } // namespace clMath |
| 304 |
nothing calls this directly
no test coverage detected