public/gpu-screen-recorder
1
0
Fork 0
mirror of https://repo.dec05eba.com/gpu-screen-recorder synced 2026-03-31 09:07:13 +09:00
Code Issues Actions Packages Projects Releases Wiki Activity

Enable low latency encoding mode on amd (only available on amd since mesa version mesa 24.1.4)

Browse Source
This commit is contained in:
dec05eba
2024-10-16 00:35:18 +02:00
parent 601219febb
commit 19d77b8a96
2 changed files with 19 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
TODO
View File

@@ -156,9 +156,6 @@ Restart replay/update video resolution if monitor resolution changes.
Fix pure vaapi copy on intel. Fix pure vaapi copy on intel.
ffmpeg supports vulkan encoding now (h264!). Doesn't work on amd yet because mesa is missing VK_KHR_video_maintenance1, see https://gitlab.freedesktop.org/mesa/mesa/-/issues/11857. Test on nvidia!
Test vaapi low latency mode (setenv("AMD_DEBUG", "lowlatencyenc", true);), added in mesa 24.1.4, released on july 17, 2024. Note that this forces gpu power usage to max at all times, even when recording at 2 fps.
Use nvidia low latency options for better encoding times. Use nvidia low latency options for better encoding times.
Test ideal async_depth value. Increasing async_depth also increased gpu memory usage a lot (from 100mb to 500mb when moving from async_depth 2 to 16) at 4k resolution. Setting it to 8 increases it by 200mb which might be ok. Test ideal async_depth value. Increasing async_depth also increased gpu memory usage a lot (from 100mb to 500mb when moving from async_depth 2 to 16) at 4k resolution. Setting it to 8 increases it by 200mb which might be ok.
@@ -179,4 +176,6 @@ Always use direct capture with NvFBC once the capture issue in mpv fullscreen ha
Support ROI (AV_FRAME_DATA_REGIONS_OF_INTEREST). Support ROI (AV_FRAME_DATA_REGIONS_OF_INTEREST).
Default to hevc if capture size is larger than 4096 in width or height. Default to hevc if capture size is larger than 4096 in width or height.
Set low latency mode on vulkan encoding.

20
src/main.cpp
View File

@@ -810,10 +810,9 @@ static void open_video_software(AVCodecContext *codec_context, VideoQuality vide
if(bitrate_mode == BitrateMode::QP) if(bitrate_mode == BitrateMode::QP)
video_software_set_qp(codec_context, video_quality, hdr, &options); video_software_set_qp(codec_context, video_quality, hdr, &options);
av_dict_set(&options, "preset", "medium", 0); av_dict_set(&options, "preset", "fast", 0);
dict_set_profile(codec_context, GSR_GPU_VENDOR_INTEL, color_depth, &options); dict_set_profile(codec_context, GSR_GPU_VENDOR_INTEL, color_depth, &options);
// TODO: If streaming or piping output set this to zerolatency av_dict_set(&options, "tune", "zerolatency", 0);
av_dict_set(&options, "tune", "fastdecode", 0);
if(codec_context->codec_id == AV_CODEC_ID_H264) { if(codec_context->codec_id == AV_CODEC_ID_H264) {
av_dict_set(&options, "coder", "cabac", 0); // TODO: cavlc is faster than cabac but worse compression. Which to use? av_dict_set(&options, "coder", "cabac", 0); // TODO: cavlc is faster than cabac but worse compression. Which to use?
@@ -990,6 +989,14 @@ static void open_video_hardware(AVCodecContext *codec_context, VideoQuality vide
// TODO: Enable multipass // TODO: Enable multipass
if(vendor == GSR_GPU_VENDOR_NVIDIA) { if(vendor == GSR_GPU_VENDOR_NVIDIA) {
// TODO: Test these, if they are needed, if they should be used
// av_dict_set_int(&options, "zerolatency", 1, 0);
// if(codec_context->codec_id == AV_CODEC_ID_AV1) {
// av_dict_set(&options, "tune", "ll", 0);
// } else if(codec_context->codec_id == AV_CODEC_ID_H264 || codec_context->codec_id == AV_CODEC_ID_HEVC) {
// av_dict_set(&options, "preset", "llhq", 0);
// av_dict_set(&options, "tune", "ll", 0);
// }
av_dict_set(&options, "tune", "hq", 0); av_dict_set(&options, "tune", "hq", 0);
dict_set_profile(codec_context, vendor, color_depth, &options); dict_set_profile(codec_context, vendor, color_depth, &options);
@@ -2600,6 +2607,11 @@ int main(int argc, char **argv) {
// If this is set to 1 then cuGraphicsGLRegisterImage will fail for egl context with error: invalid OpenGL or DirectX context, // If this is set to 1 then cuGraphicsGLRegisterImage will fail for egl context with error: invalid OpenGL or DirectX context,
// so we overwrite it // so we overwrite it
setenv("__GL_THREADED_OPTIMIZATIONS", "0", true); setenv("__GL_THREADED_OPTIMIZATIONS", "0", true);
// Forces low latency encoding mode. Use this environment variable until vaapi supports setting this as a parameter.
// The downside of this is that it always uses maximum power, which is not ideal for replay mode that runs on system startup.
// This option was added in mesa 24.1.4, released on july 17, 2024.
// TODO: Add an option to enable/disable this?
setenv("AMD_DEBUG", "lowlatencyenc", true);
// Some people set this to nvidia (for nvdec) or vdpau (for nvidia vdpau), which breaks gpu screen recorder since // Some people set this to nvidia (for nvdec) or vdpau (for nvidia vdpau), which breaks gpu screen recorder since
// nvidia doesn't support vaapi and nvidia-vaapi-driver doesn't support encoding yet. // nvidia doesn't support vaapi and nvidia-vaapi-driver doesn't support encoding yet.
// Let vaapi find the match vaapi driver instead of forcing a specific one. // Let vaapi find the match vaapi driver instead of forcing a specific one.
@@ -3795,7 +3807,7 @@ int main(int argc, char **argv) {
av_usleep(time_to_next_frame * 1000.0 * 1000.0); av_usleep(time_to_next_frame * 1000.0 * 1000.0);
else { else {
if(paused) if(paused)
av_usleep(20.0 * 1000.0); // 10 milliseconds av_usleep(20.0 * 1000.0); // 20 milliseconds
else if(frame_deadline_missed) else if(frame_deadline_missed)
{} {}
else if(framerate_mode == FramerateMode::CONTENT || !frame_captured) else if(framerate_mode == FramerateMode::CONTENT || !frame_captured)