Add hardware acceleration support to FFmpeg

This commit is contained in:
Alexey Khit
2022-12-18 01:02:12 +03:00
parent 464ea417ef
commit db255b476a
5 changed files with 521 additions and 197 deletions

View File

@@ -1,6 +1,8 @@
package ffmpeg
import (
"bytes"
"errors"
"github.com/AlexxIT/go2rtc/cmd/app"
"github.com/AlexxIT/go2rtc/cmd/exec"
"github.com/AlexxIT/go2rtc/cmd/ffmpeg/device"
@@ -17,10 +19,29 @@ func Init() {
Mod map[string]string `yaml:"ffmpeg"`
}
// defaults
cfg.Mod = defaults // will be overriden from yaml
cfg.Mod = map[string]string{
app.LoadConfig(&cfg)
if app.GetLogger("exec").GetLevel() >= 0 {
defaults["global"] += " -v error"
}
streams.HandleFunc("ffmpeg", func(url string) (streamer.Producer, error) {
args := parseArgs(url[7:]) // remove `ffmpeg:`
if args == nil {
return nil, errors.New("can't generate ffmpeg command")
}
return exec.Handle("exec:" + args.String())
})
device.Bin = defaults["bin"]
device.Init()
}
var defaults = map[string]string{
"bin": "ffmpeg",
"global": "-hide_banner",
// inputs
"file": "-re -stream_loop -1 -i {input}",
@@ -30,16 +51,13 @@ func Init() {
// output
"output": "-user_agent ffmpeg/go2rtc -rtsp_transport tcp -f rtsp {output}",
// `-g 30` - group of picture, GOP, keyframe interval
// `-preset superfast` - we can't use ultrafast because it doesn't support `-profile main -level 4.1`
// `-tune zerolatency` - for minimal latency
// `-profile main -level 4.1` - most used streaming profile
// `-pix_fmt yuv420p` - if input pix format 4:2:2
"h264": "-c:v libx264 -g:v 30 -preset:v superfast -tune:v zerolatency -profile:v main -level:v 4.1 -pix_fmt:v yuv420p",
"h264/ultra": "-c:v libx264 -g:v 30 -preset:v ultrafast -tune:v zerolatency",
"h264/high": "-c:v libx264 -g:v 30 -preset:v superfast -tune:v zerolatency",
"h265": "-c:v libx265 -g:v 30 -preset:v superfast -tune:v zerolatency -profile:v main -level:v 5.1 -pix_fmt:v yuv420p",
// `-profile high -level 4.1` - most used streaming profile
"h264": "-c:v libx264 -g 50 -profile:v high -level:v 4.1 -preset:v superfast -tune:v zerolatency",
"h265": "-c:v libx265 -g 50 -profile:v high -level:v 5.1 -preset:v superfast -tune:v zerolatency",
"mjpeg": "-c:v mjpeg -force_duplicated_matrix:v 1 -huffman:v 0 -pix_fmt:v yuvj420p",
"opus": "-c:a libopus -ar:a 48000 -ac:a 2",
"pcmu": "-c:a pcm_mulaw -ar:a 8000 -ac:a 1",
"pcmu/16000": "-c:a pcm_mulaw -ar:a 16000 -ac:a 1",
@@ -49,90 +67,107 @@ func Init() {
"pcma/48000": "-c:a pcm_alaw -ar:a 48000 -ac:a 1",
"aac": "-c:a aac", // keep sample rate and channels
"aac/16000": "-c:a aac -ar:a 16000 -ac:a 1",
// hardware Intel and AMD on Linux
"h264/vaapi": "-c:v h264_vaapi -profile:v high -level:v 4.1 -async_depth:v 1 -sei:v 0",
"h265/vaapi": "-c:v hevc_vaapi -profile:v high -level:v 5.1 -async_depth:v 1 -sei:v 0",
"mjpeg/vaapi": "-c:v mjpeg_vaapi",
// hardware Raspberry
"h264/v4l2m2m": "-c:v h264_v4l2m2m",
"h265/v4l2m2m": "-c:v hevc_v4l2m2m",
// hardware NVidia on Linux and Windows
// preset=p2 - faster, tune=ll - low latency
"h264/cuda": "-c:v h264_nvenc -g 50 -profile:v high -level:v auto -preset:v p2 -tune:v ll",
"h265/cuda": "-c:v hevc_nvenc -g 50 -profile:v high -level:v auto",
// hardware Intel on Windows
"h264/dxva2": "-c:v h264_qsv -profile:v high -level:v 4.1 -async_depth:v 1",
"h265/dxva2": "-c:v hevc_qsv -profile:v high -level:v 5.1 -async_depth:v 1",
"mjpeg/dxva2": "-c:v mjpeg_qsv -profile:v high -level:v 5.1",
// hardware macOS
"h264/videotoolbox": "-c:v h264_videotoolbox -profile:v high -level:v 4.1",
"h265/videotoolbox": "-c:v hevc_videotoolbox -profile:v high -level:v 5.1",
}
func parseArgs(s string) *Args {
// init FFmpeg arguments
args := &Args{
bin: defaults["bin"],
global: defaults["global"],
output: defaults["output"],
}
app.LoadConfig(&cfg)
tpl := cfg.Mod
cmd := "exec:" + tpl["bin"] + " -hide_banner "
if app.GetLogger("exec").GetLevel() >= 0 {
cmd += "-v error "
}
streams.HandleFunc("ffmpeg", func(s string) (streamer.Producer, error) {
s = s[7:] // remove `ffmpeg:`
var query url.Values
var queryVideo, queryAudio bool
if i := strings.IndexByte(s, '#'); i > 0 {
query = parseQuery(s[i+1:])
queryVideo = query["video"] != nil
queryAudio = query["audio"] != nil
args.video = len(query["video"])
args.audio = len(query["audio"])
s = s[:i]
} else {
// by default query both video and audio
queryVideo = true
queryAudio = true
}
var input string
// Parse input:
// 1. Input as xxxx:// link (http or rtsp or any other)
// 2. Input as stream name
// 3. Input as FFmpeg device (local USB camera)
if i := strings.Index(s, "://"); i > 0 {
switch s[:i] {
case "http", "https", "rtmp":
input = strings.Replace(tpl["http"], "{input}", s, 1)
args.input = strings.Replace(defaults["http"], "{input}", s, 1)
case "rtsp", "rtsps":
// https://ffmpeg.org/ffmpeg-protocols.html#rtsp
// skip unnecessary input tracks
switch {
case queryVideo && queryAudio:
input = "-allowed_media_types video+audio "
case queryVideo:
input = "-allowed_media_types video "
case queryAudio:
input = "-allowed_media_types audio "
case (args.video > 0 && args.audio > 0) || (args.video == 0 && args.audio == 0):
args.input = "-allowed_media_types video+audio "
case args.video > 0:
args.input = "-allowed_media_types video "
case args.audio > 0:
args.input = "-allowed_media_types audio "
}
input += strings.Replace(tpl["rtsp"], "{input}", s, 1)
args.input += strings.Replace(defaults["rtsp"], "{input}", s, 1)
default:
input = "-i " + s
args.input = "-i " + s
}
} else if streams.Get(s) != nil {
s = "rtsp://localhost:" + rtsp.Port + "/" + s
switch {
case queryVideo && !queryAudio:
case args.video > 0 && args.audio == 0:
s += "?video"
case queryAudio && !queryVideo:
case args.audio > 0 && args.video == 0:
s += "?audio"
}
input = strings.Replace(tpl["rtsp"], "{input}", s, 1)
args.input = strings.Replace(defaults["rtsp"], "{input}", s, 1)
} else if strings.HasPrefix(s, "device?") {
var err error
input, err = device.GetInput(s)
args.input, err = device.GetInput(s)
if err != nil {
return nil, err
return nil
}
} else {
input = strings.Replace(tpl["file"], "{input}", s, 1)
args.input = strings.Replace(defaults["file"], "{input}", s, 1)
}
if _, ok := query["async"]; ok {
input = "-use_wallclock_as_timestamps 1 -async 1 " + input
if query["async"] != nil {
args.input = "-use_wallclock_as_timestamps 1 -async 1 " + args.input
}
s = cmd + input
// Parse query params:
// 1. `width`/`height` params
// 2. `rotate` param
// 3. `video` params (support multiple)
// 4. `audio` params (support multiple)
// 5. `hardware` param
if query != nil {
// 1. Process raw params for FFmpeg
for _, raw := range query["raw"] {
s += " " + raw
args.AddCodec(raw)
}
// 2. Process video filters (resize and rotation)
var filters []string
if query["width"] != nil || query["height"] != nil {
filter := "scale="
if query["width"] != nil {
@@ -146,84 +181,60 @@ func Init() {
} else {
filter += "-1"
}
filters = append(filters, filter)
args.AddFilter(filter)
}
if query["rotate"] != nil {
var filter string
switch query["rotate"][0] {
case "90":
filters = append(filters, "transpose=1") // 90 degrees clockwise
filter = "transpose=1" // 90 degrees clockwise
case "180":
filters = append(filters, "transpose=1,transpose=1")
filter = "transpose=1,transpose=1"
case "-90", "270":
filters = append(filters, "transpose=2") // 90 degrees counterclockwise
filter = "transpose=2" // 90 degrees counterclockwise
}
if filter != "" {
args.AddFilter(filter)
}
if filters != nil {
s += " -vf " + strings.Join(filters, ",")
}
// 3. Process video codecs
switch len(query["video"]) {
case 0:
s += " -vn"
case 1:
if len(query["audio"]) > 1 {
s += " -map 0:v:0?"
}
if args.video > 0 {
for _, video := range query["video"] {
if video == "copy" {
s += " -c:v copy"
if video != "copy" {
args.AddCodec(defaults[video])
} else {
s += " " + tpl[video]
args.AddCodec("-c:v copy")
}
}
default:
for i, video := range query["video"] {
if video == "copy" {
s += " -map 0:v:0? -c:v:" + strconv.Itoa(i) + " copy"
} else {
s += " -map 0:v:0? " + strings.ReplaceAll(tpl[video], ":v ", ":v:"+strconv.Itoa(i)+" ")
}
}
args.AddCodec("-vn")
}
// 4. Process audio codecs
switch len(query["audio"]) {
case 0:
s += " -an"
case 1:
if len(query["video"]) > 1 {
s += " -map 0:a:0?"
}
if args.audio > 0 {
for _, audio := range query["audio"] {
if audio == "copy" {
s += " -c:a copy"
if audio != "copy" {
args.AddCodec(defaults[audio])
} else {
s += " " + tpl[audio]
}
}
default:
for i, audio := range query["audio"] {
if audio == "copy" {
s += " -map 0:a:0? -c:a:" + strconv.Itoa(i) + " copy"
} else {
s += " -map 0:a:0? " + strings.ReplaceAll(tpl[audio], ":a ", ":a:"+strconv.Itoa(i)+" ")
}
args.AddCodec("-c:a copy")
}
}
} else {
s += " -c copy"
args.AddCodec("-an")
}
s += " " + tpl["output"]
if query["hardware"] != nil {
MakeHardware(args, query["hardware"][0])
}
}
return exec.Handle(s)
})
if args.codecs == nil {
args.AddCodec("-c copy")
}
device.Bin = cfg.Mod["bin"]
device.Init()
return args
}
func parseQuery(s string) map[string][]string {
@@ -238,3 +249,76 @@ func parseQuery(s string) map[string][]string {
}
return query
}
type Args struct {
bin string // ffmpeg
global string // -hide_banner -v error
input string // -re -stream_loop -1 -i /media/bunny.mp4
codecs []string // -c:v libx264 -g:v 30 -preset:v ultrafast -tune:v zerolatency
filters []string // scale=1920:1080
output string // -f rtsp {output}
video, audio int // count of video and audio params
}
func (a *Args) AddCodec(codec string) {
a.codecs = append(a.codecs, codec)
}
func (a *Args) AddFilter(filter string) {
a.filters = append(a.filters, filter)
}
func (a *Args) InsertFilter(filter string) {
a.filters = append([]string{filter}, a.filters...)
}
func (a *Args) String() string {
b := bytes.NewBuffer(make([]byte, 0, 512))
b.WriteString(a.bin)
if a.global != "" {
b.WriteByte(' ')
b.WriteString(a.global)
}
b.WriteByte(' ')
b.WriteString(a.input)
multimode := a.video > 1 || a.audio > 1
var iv, ia int
for _, codec := range a.codecs {
// support multiple video and/or audio codecs
if multimode && len(codec) >= 5 {
switch codec[:5] {
case "-c:v ":
codec = "-map 0:v:0? " + strings.ReplaceAll(codec, ":v ", ":v:"+strconv.Itoa(iv)+" ")
iv++
case "-c:a ":
codec = "-map 0:a:0? " + strings.ReplaceAll(codec, ":a ", ":a:"+strconv.Itoa(ia)+" ")
ia++
}
}
b.WriteByte(' ')
b.WriteString(codec)
}
if a.filters != nil {
for i, filter := range a.filters {
if i == 0 {
b.WriteString(" -vf ")
} else {
b.WriteByte(',')
}
b.WriteString(filter)
}
}
b.WriteByte(' ')
b.WriteString(a.output)
return b.String()
}

112
cmd/ffmpeg/hardware.go Normal file
View File

@@ -0,0 +1,112 @@
package ffmpeg
import (
"github.com/rs/zerolog/log"
"os/exec"
"strings"
)
const (
EngineSoftware = "software"
EngineVAAPI = "vaapi" // Intel iGPU and AMD GPU
EngineV4L2 = "v4l2" // Raspberry Pi 3 and 4
EngineCUDA = "cuda" // NVidia on Windows and Linux
EngineDXVA2 = "dxva2" // Intel on Windows
EngineVideoToolbox = "videotoolbox" // macOS
)
var cache = map[string]string{}
// MakeHardware converts software FFmpeg args to hardware args
// empty engine for autoselect
func MakeHardware(args *Args, engine string) {
for i, codec := range args.codecs {
if len(codec) < 12 {
continue // skip short line (-c:v libx264...)
}
// get current codec name
name := cut(codec, ' ', 1)
switch name {
case "libx264":
name = "h264"
case "libx265":
name = "h265"
case "mjpeg":
default:
continue // skip unsupported codec
}
// temporary disable probe for H265 and MJPEG
if engine == "" && name == "h264" {
if engine = cache[name]; engine == "" {
engine = ProbeHardware(name)
cache[name] = engine
}
}
switch engine {
case EngineVAAPI:
args.input = "-hwaccel vaapi -hwaccel_output_format vaapi " + args.input
args.codecs[i] = defaults[name+"/"+engine]
for i, filter := range args.filters {
if strings.HasPrefix(filter, "scale=") {
args.filters[i] = "scale_vaapi=" + filter[6:]
}
}
// fix if input doesn't support hwaccel, do nothing when support
args.InsertFilter("format=vaapi|nv12,hwupload")
case EngineCUDA:
args.input = "-hwaccel cuda -hwaccel_output_format cuda -extra_hw_frames 2 " + args.input
args.codecs[i] = defaults[name+"/"+engine]
for i, filter := range args.filters {
if strings.HasPrefix(filter, "scale=") {
args.filters[i] = "scale_cuda=" + filter[6:]
}
}
case EngineDXVA2:
args.input = "-hwaccel dxva2 -hwaccel_output_format dxva2_vld " + args.input
args.codecs[i] = defaults[name+"/"+engine]
for i, filter := range args.filters {
if strings.HasPrefix(filter, "scale=") {
args.filters[i] = "scale_qsv=" + filter[6:]
}
}
args.InsertFilter("hwmap=derive_device=qsv,format=qsv")
case EngineVideoToolbox:
args.input = "-hwaccel videotoolbox -hwaccel_output_format videotoolbox_vld " + args.input
args.codecs[i] = defaults[name+"/"+engine]
case EngineV4L2:
args.codecs[i] = defaults[name+"/"+engine]
}
}
}
func run(arg ...string) bool {
err := exec.Command(defaults["bin"], arg...).Run()
log.Printf("%v %v", arg, err)
return err == nil
}
func cut(s string, sep byte, pos int) string {
for n := 0; n < pos; n++ {
if i := strings.IndexByte(s, sep); i > 0 {
s = s[i+1:]
} else {
return ""
}
}
if i := strings.IndexByte(s, sep); i > 0 {
return s[:i]
}
return s
}

View File

@@ -0,0 +1,21 @@
package ffmpeg
func ProbeHardware(name string) string {
switch name {
case "h264":
if run(
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "h264_videotoolbox", "-f", "null", "-") {
return EngineVideoToolbox
}
case "h265":
if run(
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "hevc_videotoolbox", "-f", "null", "-") {
return EngineVideoToolbox
}
}
return EngineSoftware
}

View File

@@ -0,0 +1,67 @@
package ffmpeg
import (
"runtime"
)
func ProbeHardware(name string) string {
if runtime.GOARCH == "arm64" || runtime.GOARCH == "arm" {
switch name {
case "h264":
if run(
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "h264_v4l2m2m", "-f", "null", "-") {
return EngineV4L2
}
case "h265":
if run(
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "hevc_v4l2m2m", "-f", "null", "-") {
return EngineV4L2
}
}
return EngineSoftware
}
switch name {
case "h264":
if run("-init_hw_device", "cuda",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "h264_nvenc", "-f", "null", "-") {
return EngineCUDA
}
if run("-init_hw_device", "vaapi",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-vf", "format=nv12,hwupload",
"-c", "h264_vaapi", "-f", "null", "-") {
return EngineVAAPI
}
case "h265":
if run("-init_hw_device", "cuda",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "hevc_nvenc", "-f", "null", "-") {
return EngineCUDA
}
if run("-init_hw_device", "vaapi",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-vf", "format=nv12,hwupload",
"-c", "hevc_vaapi", "-f", "null", "-") {
return EngineVAAPI
}
case "mjpeg":
if run("-init_hw_device", "vaapi",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-vf", "format=nv12,hwupload",
"-c", "mjpeg_vaapi", "-f", "null", "-") {
return EngineVAAPI
}
}
return EngineSoftware
}

View File

@@ -0,0 +1,40 @@
package ffmpeg
func ProbeHardware(name string) string {
switch name {
case "h264":
if run("-init_hw_device", "cuda",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "h264_nvenc", "-f", "null", "-") {
return EngineCUDA
}
if run("-init_hw_device", "dxva2",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "h264_qsv", "-f", "null", "-") {
return EngineDXVA2
}
case "h265":
if run("-init_hw_device", "cuda",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "hevc_nvenc", "-f", "null", "-") {
return EngineCUDA
}
if run("-init_hw_device", "dxva2",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "hevc_qsv", "-f", "null", "-") {
return EngineDXVA2
}
case "mjpeg":
if run("-init_hw_device", "dxva2",
"-f", "lavfi", "-i", "testsrc2", "-t", "1",
"-c", "mjpeg_qsv", "-f", "null", "-") {
return EngineDXVA2
}
}
return EngineSoftware
}