| 68 | } |
| 69 | |
| 70 | void VideoDecoderMixed::RunImpl(dali::Workspace &ws) { |
| 71 | auto &output = ws.Output<dali::GPUBackend>(0); |
| 72 | const auto &input = ws.Input<dali::CPUBackend>(0); |
| 73 | int batch_size = input.num_samples(); |
| 74 | |
| 75 | CUcontext cuContext = nullptr; |
| 76 | CUstream cuStream = ws.stream(); |
| 77 | cuCtxGetCurrent(&cuContext); |
| 78 | |
| 79 | if (!cuContext) { |
| 80 | throw std::runtime_error("Failed to create a cuda context"); |
| 81 | } |
| 82 | |
| 83 | for (int i = 0; i < batch_size; i++) { |
| 84 | auto output_sample = output[i]; |
| 85 | uint8_t *output_data = output_sample.template mutable_data<uint8_t>(); |
| 86 | |
| 87 | auto &sample = samples_[i]; |
| 88 | sample.decoder_ = std::make_unique<NvDecoder>( |
| 89 | cuStream, cuContext, true, FFmpeg2NvCodecId(sample.demuxer_->GetVideoCodec()), false, |
| 90 | false /*_enableasyncallocations*/, false); |
| 91 | |
| 92 | |
| 93 | uint8_t *pVideo = NULL, *pFrame = nullptr; |
| 94 | int nVideoBytes = 0, nFrameReturned = 0, nFrame = 0; |
| 95 | |
| 96 | int num_frames = 0; |
| 97 | |
| 98 | do { |
| 99 | sample.demuxer_->Demux(&pVideo, &nVideoBytes); |
| 100 | nFrameReturned = sample.decoder_->Decode(pVideo, nVideoBytes); |
| 101 | |
| 102 | for (int i = 0; i < nFrameReturned; i++) { |
| 103 | pFrame = sample.decoder_->GetFrame(); |
| 104 | CUDA_CALL(cudaStreamSynchronize(cuStream)); |
| 105 | |
| 106 | uint8_t *dpFrame = output_data + num_frames * sample.demuxer_->GetHeight() * |
| 107 | sample.demuxer_->GetWidth() * 3; |
| 108 | int nWidth = sample.decoder_->GetWidth(); |
| 109 | int nPitch = sample.decoder_->GetWidth(); |
| 110 | int iMatrix = |
| 111 | sample.decoder_->GetVideoFormatInfo().video_signal_description.matrix_coefficients; |
| 112 | bool full_range = |
| 113 | sample.decoder_->GetVideoFormatInfo().video_signal_description.video_full_range_flag; |
| 114 | |
| 115 | |
| 116 | yuv_to_rgb(pFrame, nPitch, reinterpret_cast<uint8_t *>(dpFrame), |
| 117 | sample.decoder_->GetWidth() * 3, sample.decoder_->GetWidth(), |
| 118 | sample.decoder_->GetHeight(), full_range, cuStream); |
| 119 | |
| 120 | ++num_frames; |
| 121 | if (end_frame_ > 0 && num_frames >= end_frame_) { |
| 122 | return; |
| 123 | } |
| 124 | } |
| 125 | } while (nVideoBytes); |
| 126 | } |
| 127 | } |
nothing calls this directly
no test coverage detected