Bug: Unexpectedly high system load with H/W accel activated
Posted: Sun Jul 09, 2017 2:02 pm
Hello there,
On the following system:
CPU: Intel Core i7 6700HQ, Hyperthreading and SST enabled.
GPU: Intel HD 530 integrated graphics and NVIDIA GeForce GTX 1060 6 GB (Optimus)
Storage: Samsung 960 Pro 2 TB NVMe SSD (M.2) + Corsair MX300 2 TB SSD (2.5' SATA)
RAM: 32 GB DDR4-2133MHz SODIMMs.
I observe abnormally slow encodes and very high CPU loads when handling 1080p HEVC 10-bit encodes to standard 8-bit H.264/AVC encodes, without encoding the audio stream (direct copy) through a custom profile on ConvertXtoVideo.
Hardware optimizations for NVENC and QuickSync have been turned on, and the program runs on the default NVIDIA profile for Optimus.
From my understanding, both GPUs on this platform support HEVC 10-bit decoding through DXVA, as shown in this http://imgur.com/a/PZhYS album.
Therefore, when using full hardware acceleration, I should be able to approach a similar performance and processor utilization as that observed on Linux with FFmpeg. Yet this is not the case.
From the software license page, I can confirm that this project uses FFmpeg to some extent. Perhaps the build used here should be upgraded to take advantage of the new DXVA D3D11 H/W features, particularly on Windows 10's WDDM default behavior that can have a negative impact on GPU preemption on software that does not use these new extensions. Refer to this https://msdn.microsoft.com/en-us/ie/dn9 ... eodecoding and https://msdn.microsoft.com/en-us/ie/dn9 ... processing documentation from Microsoft for more details.
Also, related: http://forums.vso-software.fr/feature-s ... 19669.html
On the following system:
CPU: Intel Core i7 6700HQ, Hyperthreading and SST enabled.
GPU: Intel HD 530 integrated graphics and NVIDIA GeForce GTX 1060 6 GB (Optimus)
Storage: Samsung 960 Pro 2 TB NVMe SSD (M.2) + Corsair MX300 2 TB SSD (2.5' SATA)
RAM: 32 GB DDR4-2133MHz SODIMMs.
I observe abnormally slow encodes and very high CPU loads when handling 1080p HEVC 10-bit encodes to standard 8-bit H.264/AVC encodes, without encoding the audio stream (direct copy) through a custom profile on ConvertXtoVideo.
Hardware optimizations for NVENC and QuickSync have been turned on, and the program runs on the default NVIDIA profile for Optimus.
From my understanding, both GPUs on this platform support HEVC 10-bit decoding through DXVA, as shown in this http://imgur.com/a/PZhYS album.
Therefore, when using full hardware acceleration, I should be able to approach a similar performance and processor utilization as that observed on Linux with FFmpeg. Yet this is not the case.
From the software license page, I can confirm that this project uses FFmpeg to some extent. Perhaps the build used here should be upgraded to take advantage of the new DXVA D3D11 H/W features, particularly on Windows 10's WDDM default behavior that can have a negative impact on GPU preemption on software that does not use these new extensions. Refer to this https://msdn.microsoft.com/en-us/ie/dn9 ... eodecoding and https://msdn.microsoft.com/en-us/ie/dn9 ... processing documentation from Microsoft for more details.
Also, related: http://forums.vso-software.fr/feature-s ... 19669.html