fix case when cuda hw decode falls back to sw (#2718)

* fix case when cuda hw decode falls back to sw

* use a new filter

CHANGELOG.md

@@ -53,6 +53,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 - Automatically kill playback troubleshooting ffmpeg process if it hasn't completed after two minutes
 - Fix playback of certain BT.2020 content
 - Use playlist item count when using a playlist as filler (instead of a fixed count of 1 for each playlist item)
+- NVIDIA: fix stream failure with certain content that should decode in hardware but falls back to software
 
 ### Changed
 - No longer round framerate to nearest integer when normalizing framerate

ErsatzTV.FFmpeg/Filter/Cuda/CudaSoftwareFallbackUploadFilter.cs

@@ -0,0 +1,9 @@
+namespace ErsatzTV.FFmpeg.Filter.Cuda;
+
+public class CudaSoftwareFallbackUploadFilter : BaseFilter
+{
+    public override string Filter => "hwupload";
+
+    public override FrameState NextState(FrameState currentState) =>
+        currentState with { FrameDataLocation = FrameDataLocation.Hardware };
+}

ErsatzTV.FFmpeg/Filter/Cuda/HardwareUploadCudaFilter.cs

@@ -0,0 +1,13 @@
+namespace ErsatzTV.FFmpeg.Filter.Cuda;
+
+public class HardwareUploadCudaFilter(FrameDataLocation frameDataLocation) : BaseFilter
+{
+    public override string Filter => frameDataLocation switch
+    {
+        FrameDataLocation.Hardware => string.Empty,
+        _ => "hwupload_cuda"
+    };
+
+    public override FrameState NextState(FrameState currentState) =>
+        currentState with { FrameDataLocation = FrameDataLocation.Hardware };
+}

ErsatzTV.FFmpeg/Filter/HardwareUploadCudaFilter.cs

@@ -1,17 +0,0 @@
-namespace ErsatzTV.FFmpeg.Filter;
-
-public class HardwareUploadCudaFilter : BaseFilter
-{
-    private readonly FrameState _currentState;
-
-    public HardwareUploadCudaFilter(FrameState currentState) => _currentState = currentState;
-
-    public override string Filter => _currentState.FrameDataLocation switch
-    {
-        FrameDataLocation.Hardware => string.Empty,
-        _ => "hwupload_cuda"
-    };
-
-    public override FrameState NextState(FrameState currentState) =>
-        currentState with { FrameDataLocation = FrameDataLocation.Hardware };
-}

ErsatzTV.FFmpeg/GlobalOption/HardwareAcceleration/CudaHardwareAccelerationOption.cs

@@ -11,7 +11,7 @@ public class CudaHardwareAccelerationOption(bool isVulkanHdr) : GlobalOption
                 return ["-init_hw_device", "cuda=nv", "-init_hw_device", "vulkan=vk@nv", "-hwaccel", "vulkan"];
             }
 
-            return ["-hwaccel", "cuda"];
+            return ["-init_hw_device", "cuda", "-hwaccel", "cuda"];
         }
     }
 }

ErsatzTV.FFmpeg/Pipeline/NvidiaPipelineBuilder.cs

@@ -134,6 +134,15 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
         foreach (IDecoder decoder in maybeDecoder)
         {
             videoInputFile.AddOption(decoder);
+
+            // sometimes cuda fails to decode in hardware and falls back to software
+            // in that case, we need to upload to get the frame in hardware as expected
+            // this *should* no-op when frames are already in hardware
+            if (ffmpegState.DecoderHardwareAccelerationMode is HardwareAccelerationMode.Nvenc)
+            {
+                videoInputFile.FilterSteps.Add(new CudaSoftwareFallbackUploadFilter());
+            }
+
             return Some(decoder);
         }
 
@@ -266,7 +275,7 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
             && !context.HasSubtitleText
             && (context.HasSubtitleOverlay || context.HasWatermark || context.HasGraphicsEngine))
         {
-            var hardwareUpload = new HardwareUploadCudaFilter(currentState);
+            var hardwareUpload = new HardwareUploadCudaFilter(currentState.FrameDataLocation);
             currentState = hardwareUpload.NextState(currentState);
             videoInputFile.FilterSteps.Add(hardwareUpload);
         }
@@ -562,7 +571,7 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
             else
             {
                 watermark.FilterSteps.Add(
-                    new HardwareUploadCudaFilter(currentState with { FrameDataLocation = FrameDataLocation.Software }));
+                    new HardwareUploadCudaFilter(FrameDataLocation.Software));
 
                 var watermarkFilter = new OverlayWatermarkCudaFilter(
                     watermark.DesiredState,
@@ -612,7 +621,7 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
 
                 if (context.HasWatermark || context.HasGraphicsEngine)
                 {
-                    var subtitleHardwareUpload = new HardwareUploadCudaFilter(currentState);
+                    var subtitleHardwareUpload = new HardwareUploadCudaFilter(currentState.FrameDataLocation);
                     currentState = subtitleHardwareUpload.NextState(currentState);
                     videoInputFile.FilterSteps.Add(subtitleHardwareUpload);
                 }
@@ -626,8 +635,7 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
                 {
                     if (_ffmpegCapabilities.HasFilter(FFmpegKnownFilter.ScaleNpp))
                     {
-                        var subtitleHardwareUpload = new HardwareUploadCudaFilter(
+                        var subtitleHardwareUpload = new HardwareUploadCudaFilter(FrameDataLocation.Software);
-                            currentState with { FrameDataLocation = FrameDataLocation.Software });
                         subtitle.FilterSteps.Add(subtitleHardwareUpload);
 
                         // only scale if scaling or padding was used for main video stream
@@ -648,8 +656,7 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
                             subtitle.FilterSteps.Add(scaleFilter);
                         }
 
-                        var subtitleHardwareUpload = new HardwareUploadCudaFilter(
+                        var subtitleHardwareUpload = new HardwareUploadCudaFilter(FrameDataLocation.Software);
-                            currentState with { FrameDataLocation = FrameDataLocation.Software });
                         subtitle.FilterSteps.Add(subtitleHardwareUpload);
                     }
 
@@ -700,8 +707,7 @@ public class NvidiaPipelineBuilder : SoftwarePipelineBuilder
             {
                 graphicsEngine.FilterSteps.Add(new PixelFormatFilter(new PixelFormatYuva420P()));
 
-                graphicsEngine.FilterSteps.Add(
+                graphicsEngine.FilterSteps.Add(new HardwareUploadCudaFilter(FrameDataLocation.Software));
-                    new HardwareUploadCudaFilter(currentState with { FrameDataLocation = FrameDataLocation.Software }));
 
                 var graphicsEngineFilter = new OverlayGraphicsEngineCudaFilter();
                 graphicsEngineOverlayFilterSteps.Add(graphicsEngineFilter);