| 273 | |
| 274 | |
| 275 | template <typename ElemType> nano_time_t |
| 276 | GemmPerformanceTest<ElemType>::clblasPerfSingle(void) |
| 277 | { |
| 278 | nano_time_t time; |
| 279 | cl_event event; |
| 280 | cl_int status; |
| 281 | cl_command_queue queue = base_->commandQueues()[0]; |
| 282 | |
| 283 | status = clEnqueueWriteBuffer(queue, mobjC_, CL_TRUE, 0, |
| 284 | params_.rowsC * params_.columnsC * |
| 285 | sizeof(ElemType), backC_, 0, NULL, &event); |
| 286 | if (status != CL_SUCCESS) { |
| 287 | cerr << "Matrix C buffer object enqueuing error, status = " << |
| 288 | status << endl; |
| 289 | |
| 290 | return NANOTIME_ERR; |
| 291 | } |
| 292 | |
| 293 | status = clWaitForEvents(1, &event); |
| 294 | if (status != CL_SUCCESS) { |
| 295 | cout << "Wait on event failed, status = " << |
| 296 | status << endl; |
| 297 | |
| 298 | return NANOTIME_ERR; |
| 299 | } |
| 300 | |
| 301 | event = NULL; |
| 302 | status = (cl_int)clMath::clblas::gemm(params_.order, |
| 303 | params_.transA, params_.transB, params_.M, params_.N, params_.K, alpha_, |
| 304 | mobjA_, params_.offA, params_.lda, mobjB_, params_.offBX, params_.ldb, |
| 305 | beta_, mobjC_, params_.offCY, params_.ldc, 1, &queue, 0, NULL, &event); |
| 306 | if (status != CL_SUCCESS) { |
| 307 | cerr << "The CLBLAS GEMM function failed, status = " << |
| 308 | status << endl; |
| 309 | |
| 310 | return NANOTIME_ERR; |
| 311 | } |
| 312 | status = flushAll(1, &queue); |
| 313 | if (status != CL_SUCCESS) { |
| 314 | cerr << "clFlush() failed, status = " << status << endl; |
| 315 | return NANOTIME_ERR; |
| 316 | } |
| 317 | |
| 318 | time = getCurrentTime(); |
| 319 | status = waitForSuccessfulFinish(1, &queue, &event); |
| 320 | if (status == CL_SUCCESS) { |
| 321 | time = getCurrentTime() - time; |
| 322 | } |
| 323 | else { |
| 324 | cerr << "Waiting for completion of commands to the queue failed, " |
| 325 | "status = " << status << endl; |
| 326 | time = NANOTIME_ERR; |
| 327 | } |
| 328 | |
| 329 | return time; |
| 330 | } |
| 331 | |
| 332 | } // namespace clMath |
nothing calls this directly
no test coverage detected