| 1418 | } |
| 1419 | |
| 1420 | void tensor_conv::get_gradient_for_data ( |
| 1421 | const bool add_to_output, |
| 1422 | const tensor& gradient_input, |
| 1423 | const tensor& filters, |
| 1424 | tensor& data_gradient |
| 1425 | ) |
| 1426 | { |
| 1427 | choose_backward_data_algorithm(gradient_input, data_gradient); |
| 1428 | |
| 1429 | const float alpha = 1; |
| 1430 | const float beta = add_to_output ? 1 : 0; |
| 1431 | |
| 1432 | // Since cudnnConvolutionBackwardData() is an asynchronous call, we need to hold a |
| 1433 | // reference to the workspace buffer so we can be sure it isn't reallocated |
| 1434 | // while the function is still executing on the device. But each time we come |
| 1435 | // here, we make sure to grab the latest workspace buffer so that, globally, we |
| 1436 | // minimize the number of such buffers. |
| 1437 | backward_data_workspace = device_global_buffer(backward_data_workspace_size_in_bytes); |
| 1438 | |
| 1439 | |
| 1440 | CHECK_CUDNN(cudnnConvolutionBackwardData(context(), |
| 1441 | &alpha, |
| 1442 | (const cudnnFilterDescriptor_t)filter_handle, |
| 1443 | filters.device(), |
| 1444 | descriptor(gradient_input), |
| 1445 | gradient_input.device(), |
| 1446 | (const cudnnConvolutionDescriptor_t)conv_handle, |
| 1447 | (cudnnConvolutionBwdDataAlgo_t)backward_data_algo, |
| 1448 | backward_data_workspace, |
| 1449 | backward_data_workspace_size_in_bytes, |
| 1450 | &beta, |
| 1451 | descriptor(data_gradient), |
| 1452 | data_gradient.device())); |
| 1453 | } |
| 1454 | |
| 1455 | void tensor_conv:: |
| 1456 | get_gradient_for_filters ( |