Enable CUDA filter graph

[mthrok]

Currently, when performing GPU decoding/encoding, filter graph is disable.
We should enable filter graph so that it is more flexible for GPU processing.

implementation direction

1. Explicitly pass HW device context and/or HW frame context to AVCodecContext here. Similar to encoder.
2. Modify the FilterGraph implementation so that HW device context is attached to AVFilterContext and HW frame context is attached to AVFilterLink.
3. Test with scale_cuda filter.

Test code

import torchaudio
from torchaudio.io import StreamReader

src = "NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4_small.mp4"
r = StreamReader(src)
torchaudio.utils.ffmpeg_utils.set_log_level(56)
r.add_video_stream(-1, decoder="h264_cuvid", filter_desc="scale_cuda=480:270", hw_accel="cuda")

Currently this causes segmentation fault at

audio/torchaudio/csrc/ffmpeg/filter_graph.cpp

Lines 168 to 174 in 9bb3507

	void FilterGraph::create_filter() {
	int ret = avfilter_graph_config(pFilterGraph, nullptr);
	TORCH_CHECK(ret >= 0, "Failed to configure the graph: " + av_err2string(ret));
	// char* desc = avfilter_graph_dump(pFilterGraph, NULL);
	// std::cerr << "Filter created:\n" << desc << std::endl;
	// av_free(static_cast<void*>(desc));
	}

• log

[h264_cuvid @ 0x6814340] Calling ctx->cudl->cuCtxPushCurrent(cuda_ctx)
[h264_cuvid @ 0x6814340] Calling ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps8)
[h264_cuvid @ 0x6814340] Calling ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps10)
[h264_cuvid @ 0x6814340] Calling ctx->cvdl->cuvidGetDecoderCaps(&ctx->caps12)
[h264_cuvid @ 0x6814340] CUVID capabilities for h264_cuvid:
[h264_cuvid @ 0x6814340] 8 bit: supported: 1, min_width: 48, max_width: 4096, min_height: 16, max_height: 4096
[h264_cuvid @ 0x6814340] 10 bit: supported: 0, min_width: 0, max_width: 0, min_height: 0, max_height: 0
[h264_cuvid @ 0x6814340] 12 bit: supported: 0, min_width: 0, max_width: 0, min_height: 0, max_height: 0
[h264_cuvid @ 0x6814340] Calling ctx->cvdl->cuvidCreateVideoParser(&ctx->cuparser, &ctx->cuparseinfo)
[h264_cuvid @ 0x6814340] Calling ctx->cvdl->cuvidParseVideoData(ctx->cuparser, &seq_pkt)
[h264_cuvid @ 0x6814340] Calling ctx->cudl->cuCtxPopCurrent(&dummy)
[h264_cuvid @ 0x6814340] Invalid pkt_timebase, passing timestamps as-is.
[in @ 0x71da9c0] Setting 'video_size' to value '960x540'
[in @ 0x71da9c0] Setting 'pix_fmt' to value 'cuda'
[in @ 0x71da9c0] Setting 'time_base' to value '1/30000'
[in @ 0x71da9c0] Setting 'frame_rate' to value '30000/1001'
[in @ 0x71da9c0] Setting 'pixel_aspect' to value '1/1'
[in @ 0x71da9c0] w:960 h:540 pixfmt:cuda tb:1/30000 fr:30000/1001 sar:1/1
[Parsed_scale_cuda_0 @ 0x71db980] Setting 'w' to value '480'
[Parsed_scale_cuda_0 @ 0x71db980] Setting 'h' to value '270'
[AVFilterGraph @ 0x71da840] query_formats: 3 queried, 2 merged, 0 already done, 0 delayed
/fsx/users/moto/torchaudio/./bu.sh: line 17: 11665 Segmentation fault      (core dumped) python repro/cuda_filter.py

The segfault happens in cudascale_config_props function of vf_scale_cuda.c, it seems that hw_device_context is required.