用户反馈两件事:弹幕想再小些;某场合并的 1:19:25 处弹幕糊在一起(把成片抽帧 出来一看,是同一弹道上两条弹幕追尾叠字)。 重叠的根因:弹幕从上游原版的 .ass 还原成 jsonl 时,lane 是按「只有内容」的 宽度反推的(原格式的 x/y 就是这么摆的),而渲染用的是「昵称(uid):内容 + 头像」 的完整宽度 —— 后者占位时间长得多,照搬原弹道必然追尾。用用户那份 3011 条的真实 jsonl 验过:按渲染公式统计,重排前 960 条(31.9%)追尾,重排后 0 条。 - 新增 relayoutLanes:渲染前按最终字号统一重排弹道并重算宽度,公式和录制时的 placeLane 完全一致,所以录制下来的 jsonl 重排后弹道不变(已加用例覆盖)。 - 空弹道用 math.MinInt64 占位:placeLane 的条件是「本条出现时间晚于上一条的让位 时刻」,零值下 0 > 0 不成立,第一条弹幕(还原出来的时间可能正好是 0)会占不上 弹道、占用状态丢失,后面又跟着乱排。录制的弹幕时间总大于 0,所以原逻辑没暴露。 - 弹道全满(100 条都占着)时退到最早让出来的那条:录制时是直接不写这条,合并时 丢一条等于凭空少一句弹幕,压一点也比丢了强。 - 新增 danmuFontSize 设置(合并分区,0 = 跟随弹幕文件):覆盖渲染字号,头像按 同一比例缩放;改了字号会在重排时一并生效,不会因为宽度变了又追尾。 验证:layout_test.go 5 个用例;用新 ffmpeg(用户已自行升级到 n9.0.2)在隔离实例上 跑了真合并 —— 字号 95 以上的对照(60px → 42px:抽帧量到的弹幕墨迹高度 48px → 36px)、 按 1:19:00 裁出的 130 秒片段合并前后追尾 25 条 → 0 条;重跑 verify_settings.py (新增第 8 节覆盖字号字段)、verify_pick.py、verify_streamer_br.py 无回归; go test ./...、gofmt、go vet 干净。
910 lines
28 KiB
Go
910 lines
28 KiB
Go
// 弹幕合成相关:读取 jsonl,下载头像并圆形裁切,调用 ffmpeg 合成带头像的弹幕层
|
||
package main
|
||
|
||
import (
|
||
"bufio"
|
||
"bytes"
|
||
"crypto/sha1"
|
||
"encoding/json"
|
||
"errors"
|
||
"fmt"
|
||
"image"
|
||
"image/color"
|
||
"image/draw"
|
||
_ "image/gif"
|
||
_ "image/jpeg"
|
||
"image/png"
|
||
"io"
|
||
"math"
|
||
"net/http"
|
||
"os"
|
||
"os/exec"
|
||
"path/filepath"
|
||
"regexp"
|
||
"sort"
|
||
"strconv"
|
||
"strings"
|
||
"sync"
|
||
"time"
|
||
|
||
xdraw "golang.org/x/image/draw"
|
||
_ "golang.org/x/image/webp"
|
||
)
|
||
|
||
// 头像缓存目录名
|
||
const avatarCacheDir = "avatar_cache"
|
||
|
||
// 命令行 -vbitrate 指定的合成目标码率(Kbps),配合 config.MergeBitrate 使用
|
||
var burnBitrate int
|
||
|
||
// 缓存头像的边长
|
||
const avatarCacheSize = 256
|
||
|
||
// jsonl 里没有弹幕记录
|
||
var errNoDanmu = errors.New("jsonl 里没有弹幕")
|
||
|
||
// 弹幕在屏幕上的停留时长(秒),与 danmu.go 里的 danmuShowDur 一致
|
||
const danmuShowSeconds = 10.0
|
||
|
||
// 每块最多包含的弹幕条数
|
||
// 每条弹幕都要生成一段 overlay 并作为 -filter_complex 的单个参数传给 ffmpeg,
|
||
// 而 Windows 的命令行总长上限是 32767 字符(Linux 是单个参数 128KB),
|
||
// 弹幕一多整条命令就会超限且报错完全看不出原因,所以超过这个数量就分块渲染再拼接。
|
||
const maxDanmuPerChunk = 80
|
||
|
||
// 分块渲染产生的临时文件前缀
|
||
const chunkTempPrefix = "acfunlive_chunk_"
|
||
|
||
// 一个待渲染的弹幕块
|
||
type danmuChunk struct {
|
||
start float64 // 在视频时间轴上的起点(秒)
|
||
end float64 // 终点(秒),math.Inf(1) 表示一直到视频结尾
|
||
records []danmuRecord // 属于这一块的弹幕
|
||
}
|
||
|
||
// 渲染分块时共用的参数
|
||
type burnContext struct {
|
||
ffmpegFile string
|
||
videoFile string // 为空表示输出透明头像层
|
||
outFile string
|
||
offset float64
|
||
meta danmuMeta
|
||
avatars map[int64]string
|
||
bitrate int
|
||
dir string
|
||
encoder string // 已经验证可用的编码器,分块时避免每块都重新回退
|
||
}
|
||
|
||
// 头像直径,与文字的实际墨迹高度一致(字号是 em 框,墨迹约为 0.8 倍)
|
||
func avatarSize(fontSize int) int {
|
||
return fontSize * 4 / 5
|
||
}
|
||
|
||
// 头像和弹幕文字之间的间隔
|
||
func avatarGap(fontSize int) int {
|
||
return fontSize / 6
|
||
}
|
||
|
||
// 弹幕合成入口(命令行)
|
||
// 弹幕文件可以给 .jsonl,也可以给 .ass(先还原成 jsonl 再合成)
|
||
func runDanmuburn(danmuFile, videoFile, outFile string, offset float64) {
|
||
src, err := prepareDanmuSource(danmuFile, true)
|
||
if err != nil {
|
||
if err == errNoDanmu {
|
||
lPrintWarn(err.Error())
|
||
os.Exit(0)
|
||
}
|
||
lPrintErr(err)
|
||
os.Exit(1)
|
||
}
|
||
if err := burnDanmu(src.jsonlPath, videoFile, outFile, offset); err != nil {
|
||
if err == errNoDanmu {
|
||
lPrintWarn(err.Error())
|
||
os.Exit(0)
|
||
}
|
||
lPrintErr(err)
|
||
os.Exit(1)
|
||
}
|
||
}
|
||
|
||
// 合成一个 jsonl(和可选的视频)为带头像的弹幕层
|
||
// videoFile 为空时输出透明头像层(ProRes 4444),否则把弹幕层压进视频
|
||
func burnDanmu(jsonlFile, videoFile, outFile string, offset float64) error {
|
||
jsonlFile, err := filepath.Abs(jsonlFile)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
if _, err := os.Stat(jsonlFile); err != nil {
|
||
return fmt.Errorf("jsonl 文件不存在:%s", jsonlFile)
|
||
}
|
||
|
||
// 读取 jsonl,首行是元数据
|
||
f, err := os.Open(jsonlFile)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
defer f.Close()
|
||
scanner := bufio.NewScanner(f)
|
||
scanner.Buffer(make([]byte, 1024*1024), 1024*1024)
|
||
var meta danmuMeta
|
||
var records []danmuRecord
|
||
first := true
|
||
for scanner.Scan() {
|
||
line := scanner.Bytes()
|
||
if first {
|
||
if err := json.Unmarshal(line, &meta); err != nil {
|
||
return err
|
||
}
|
||
first = false
|
||
continue
|
||
}
|
||
var r danmuRecord
|
||
if json.Unmarshal(line, &r) == nil && r.Content != "" {
|
||
records = append(records, r)
|
||
}
|
||
}
|
||
if err := scanner.Err(); err != nil {
|
||
return err
|
||
}
|
||
if meta.PlayResX == 0 || meta.FontSize == 0 {
|
||
return errors.New("jsonl 文件格式不对,首行必须包含 playResX、playResY、fontSize")
|
||
}
|
||
if len(records) == 0 {
|
||
return errNoDanmu
|
||
}
|
||
|
||
// 确定输出文件和模式:有视频就把弹幕压进视频,没有视频就输出透明头像层
|
||
burnMode := videoFile != ""
|
||
if burnMode {
|
||
videoFile, err = filepath.Abs(videoFile)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
if _, err := os.Stat(videoFile); err != nil {
|
||
return fmt.Errorf("录播视频不存在:%s", videoFile)
|
||
}
|
||
if outFile == "" {
|
||
// 成片默认放到单独的文件夹里,不再和录播、弹幕混在一起
|
||
outFile = filepath.Join(mergeSaveDir(), filepath.Base(strings.TrimSuffix(jsonlFile, ".jsonl")+"_弹幕合并.mp4"))
|
||
}
|
||
} else {
|
||
if outFile == "" {
|
||
outFile = strings.TrimSuffix(jsonlFile, ".jsonl") + "_avatar.mov"
|
||
}
|
||
}
|
||
outFile, err = filepath.Abs(outFile)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
|
||
ffmpegFile := getFFmpeg()
|
||
if ffmpegFile == "" {
|
||
return errors.New("找不到 ffmpeg,请先安装并加入 PATH,或把 ffmpeg 放在本程序所在文件夹,或在设置里填写 ffmpeg 路径")
|
||
}
|
||
|
||
// 弹幕可能录得比视频久(视频中途中断过),超出视频时长的部分没法渲染,直接丢掉
|
||
videoDuration := 0.0
|
||
if burnMode {
|
||
videoDuration = probeDuration(ffmpegFile, videoFile)
|
||
if videoDuration > 0 {
|
||
kept := records[:0]
|
||
for _, r := range records {
|
||
if r.Time+offset < videoDuration {
|
||
kept = append(kept, r)
|
||
}
|
||
}
|
||
if dropped := len(records) - len(kept); dropped > 0 {
|
||
lPrintWarnf("有 %d 条弹幕超出了录播视频的时长(%.1f 秒),已跳过", dropped, videoDuration)
|
||
}
|
||
records = kept
|
||
if len(records) == 0 {
|
||
return errNoDanmu
|
||
}
|
||
}
|
||
}
|
||
|
||
// 弹幕字号可被设置覆盖(0 = 跟随弹幕文件),头像尺寸随字号等比缩放
|
||
if config.DanmuFontSize > 0 {
|
||
lPrintf("弹幕字号按设置使用 %d(弹幕文件里是 %d)", config.DanmuFontSize, meta.FontSize)
|
||
meta.FontSize = config.DanmuFontSize
|
||
}
|
||
// 弹道按最终字号统一重排,避免还原弹幕或改字号后同弹道追尾
|
||
relayoutLanes(records, meta)
|
||
|
||
// 下载并圆形裁切头像
|
||
avatarPaths := make(map[int64]string)
|
||
dir := filepath.Dir(jsonlFile)
|
||
cacheDir := filepath.Join(dir, avatarCacheDir)
|
||
// 先去重,顺便记下每个用户自带的头像链接
|
||
urlByUID := make(map[int64]string)
|
||
uids := make([]int64, 0, 64)
|
||
for _, r := range records {
|
||
if r.UID == 0 {
|
||
continue
|
||
}
|
||
if old, ok := urlByUID[r.UID]; !ok {
|
||
uids = append(uids, r.UID)
|
||
urlByUID[r.UID] = r.Avatar
|
||
} else if old == "" && r.Avatar != "" {
|
||
// 同一个用户的前几条可能没有头像链接,后面带上就补上
|
||
urlByUID[r.UID] = r.Avatar
|
||
}
|
||
}
|
||
looked := 0
|
||
for i, uid := range uids {
|
||
url := urlByUID[uid]
|
||
if url == "" {
|
||
// 从 .ass 还原出来的弹幕没有头像链接,按 uid 反查一次
|
||
url = avatarURLByUID(uid)
|
||
looked++
|
||
// 反查比较慢,定期报一下进度,避免看起来像卡住了
|
||
if looked%50 == 0 || i+1 == len(uids) {
|
||
lPrintf("正在获取头像:已处理 %d/%d 个用户", i+1, len(uids))
|
||
}
|
||
}
|
||
if url == "" {
|
||
continue
|
||
}
|
||
if path := ensureAvatar(uid, url, cacheDir); path != "" {
|
||
avatarPaths[uid] = path
|
||
}
|
||
}
|
||
if looked > 0 {
|
||
lPrintf("有 %d 个用户没有自带头像链接(多半是从 .ass 还原的弹幕),已按 uid 向 AcFun 反查", looked)
|
||
}
|
||
if len(avatarPaths) == 0 {
|
||
lPrintWarn("没有任何头像可用,弹幕将只烧录文字,请检查网络后重试合并")
|
||
}
|
||
|
||
ctx := &burnContext{
|
||
ffmpegFile: ffmpegFile,
|
||
videoFile: videoFile,
|
||
outFile: outFile,
|
||
offset: offset,
|
||
meta: meta,
|
||
avatars: avatarPaths,
|
||
bitrate: config.MergeBitrate + burnBitrate,
|
||
dir: dir,
|
||
}
|
||
|
||
chunks := splitDanmuChunks(records, offset, videoDuration)
|
||
if len(chunks) <= 1 {
|
||
lPrintf("共 %d 条弹幕,%d 个用户的头像可用,开始合成", len(records), len(avatarPaths))
|
||
return ctx.burnSingle(records)
|
||
}
|
||
|
||
lPrintf("共 %d 条弹幕,%d 个用户的头像可用;超过单次渲染上限(%d 条),分 %d 块渲染后拼接",
|
||
len(records), len(avatarPaths), maxDanmuPerChunk, len(chunks))
|
||
return ctx.burnChunked(chunks)
|
||
}
|
||
|
||
// 渲染前按最终字号重排弹道:jsonl 里的 lane/width 是录制或 .ass 还原时算的,
|
||
// 从原格式 .ass 还原的弹幕没把昵称和头像算进宽度(占位提前放行了同弹道的下一条,
|
||
// 画面上就追尾叠字),改过字号后更是全对不上。这里用和录制时相同的碰撞公式、
|
||
// 按「昵称(uid):内容 + 头像」的完整宽度重新分配一遍,保证弹道不重叠。
|
||
func relayoutLanes(records []danmuRecord, meta danmuMeta) {
|
||
sort.SliceStable(records, func(i, j int) bool { return records[i].Time < records[j].Time })
|
||
fs := meta.FontSize
|
||
show := int64(danmuShowSeconds * 1e9)
|
||
resX := int64(meta.PlayResX)
|
||
avW, avGap := avatarSize(fs), avatarGap(fs)
|
||
// 空弹道用「远古」状态占位:placeLane 的条件是本条出现时间晚于上一条的让位时刻,
|
||
// 零值下 0 > 0 不成立,第一条弹幕(还原出来的弹幕时间可能正好是 0)会占不上弹道,
|
||
// 占用状态丢掉后后面的弹幕又会跟着落回空弹道,等于白排。录制时弹幕时间总大于 0,
|
||
// 所以原逻辑没暴露这个问题。
|
||
lanes := make([]laneState, laneNum)
|
||
for i := range lanes {
|
||
lanes[i] = laneState{appear: math.MinInt64, emerge: math.MinInt64, disappear: math.MinInt64}
|
||
}
|
||
for i := range records {
|
||
r := &records[i]
|
||
r.Width = textWidth(speakerText(r.Name, r.UID, r.Content), fs)
|
||
total := int64(avW+avGap) + int64(r.Width)
|
||
rel := int64(r.Time * 1e9)
|
||
dt := laneState{
|
||
appear: rel,
|
||
emerge: rel + total*show/(resX+total),
|
||
disappear: rel + show,
|
||
}
|
||
left := rel + resX*show/(resX+total)
|
||
j := placeLane(lanes, dt, left)
|
||
if j < 0 {
|
||
// 100 条弹道都还占着:退到最早让出来的那条。录制时是直接不写这条,
|
||
// 合并时丢一条等于凭空少一句弹幕,压一点也比丢了强。
|
||
j = earliestLane(lanes)
|
||
lanes[j] = dt
|
||
}
|
||
r.Lane = j
|
||
}
|
||
}
|
||
|
||
// 最早会空出来的弹道行
|
||
func earliestLane(lanes []laneState) int {
|
||
best := 0
|
||
for i := range lanes {
|
||
if lanes[i].disappear < lanes[best].disappear {
|
||
best = i
|
||
}
|
||
}
|
||
return best
|
||
}
|
||
|
||
// 用 ffmpeg 探测媒体文件的时长(秒),探测不到返回 0
|
||
func probeDuration(ffmpegFile, path string) float64 {
|
||
cmd := exec.Command(ffmpegFile, "-hide_banner", "-i", path)
|
||
var buf bytes.Buffer
|
||
cmd.Stderr = &buf
|
||
// 没指定输出文件时 ffmpeg 会返回非 0,但时长信息已经打印出来了
|
||
_ = cmd.Run()
|
||
|
||
m := durationRe.FindStringSubmatch(buf.String())
|
||
if len(m) != 4 {
|
||
return 0
|
||
}
|
||
h, _ := strconv.Atoi(m[1])
|
||
mi, _ := strconv.Atoi(m[2])
|
||
s, _ := strconv.ParseFloat(m[3], 64)
|
||
return float64(h)*3600 + float64(mi)*60 + s
|
||
}
|
||
|
||
var durationRe = regexp.MustCompile(`Duration: (\d+):(\d+):(\d+\.\d+)`)
|
||
|
||
// 弹幕不多时直接一次渲染完,同样从 jsonl 生成字幕,不依赖录制时留下的 .nick.ass
|
||
func (c *burnContext) burnSingle(records []danmuRecord) error {
|
||
chunk := danmuChunk{start: 0, end: math.Inf(1), records: records}
|
||
// 临时字幕一律用纯 ASCII 文件名:libass 在部分环境下打不开非 ASCII 的文件名
|
||
assName := chunkTempPrefix + "single.ass"
|
||
assPath := filepath.Join(c.dir, assName)
|
||
if err := writeChunkAss(assPath, c.meta, chunk, c.offset,
|
||
avatarSize(c.meta.FontSize), avatarGap(c.meta.FontSize)); err != nil {
|
||
return fmt.Errorf("生成字幕文件失败:%v", err)
|
||
}
|
||
defer os.Remove(assPath)
|
||
|
||
return c.renderChunk(chunk, assName, c.outFile, false)
|
||
}
|
||
|
||
// 分块渲染:每块单独渲染成一个小片,最后用 concat 拼成完整成片
|
||
func (c *burnContext) burnChunked(chunks []danmuChunk) error {
|
||
ext := ".mov"
|
||
if c.videoFile != "" {
|
||
ext = ".mp4"
|
||
}
|
||
|
||
var parts []string
|
||
defer func() {
|
||
for _, p := range parts {
|
||
_ = os.Remove(p)
|
||
}
|
||
}()
|
||
|
||
ad, gap := avatarSize(c.meta.FontSize), avatarGap(c.meta.FontSize)
|
||
for i, ch := range chunks {
|
||
// 每块的字幕单独生成,只包含这一块时间范围内的弹幕
|
||
assName := fmt.Sprintf("%schunk_%04d.ass", chunkTempPrefix, i)
|
||
assPath := filepath.Join(c.dir, assName)
|
||
if err := writeChunkAss(assPath, c.meta, ch, c.offset, ad, gap); err != nil {
|
||
return fmt.Errorf("生成第 %d 块的字幕失败:%v", i+1, err)
|
||
}
|
||
defer os.Remove(assPath)
|
||
|
||
part := filepath.Join(c.dir, fmt.Sprintf("%spart_%04d%s", chunkTempPrefix, i, ext))
|
||
lPrintf("正在渲染第 %d/%d 块(%s ~ %s,%d 条弹幕)",
|
||
i+1, len(chunks), formatSeconds(ch.start), formatSeconds(ch.end), len(ch.records))
|
||
if err := c.renderChunk(ch, assName, part, true); err != nil {
|
||
return fmt.Errorf("渲染第 %d/%d 块失败:%v", i+1, len(chunks), err)
|
||
}
|
||
parts = append(parts, part)
|
||
}
|
||
|
||
lPrintf("正在把 %d 个分块拼接成完整成片", len(parts))
|
||
if err := concatChunks(c.ffmpegFile, parts, c.outFile, c.dir); err != nil {
|
||
return err
|
||
}
|
||
return nil
|
||
}
|
||
|
||
// 把弹幕按时间切成若干块,块的时间范围首尾相接,保证视频不会缺帧
|
||
// maxEnd 大于 0 时表示视频总时长,块尾不会超过它
|
||
func splitDanmuChunks(records []danmuRecord, offset, maxEnd float64) []danmuChunk {
|
||
n := len(records)
|
||
if n == 0 {
|
||
return nil
|
||
}
|
||
if n <= maxDanmuPerChunk {
|
||
return []danmuChunk{{start: 0, end: math.Inf(1), records: records}}
|
||
}
|
||
|
||
chunks := make([]danmuChunk, 0, n/maxDanmuPerChunk+1)
|
||
i := 0
|
||
start := 0.0
|
||
for i < n {
|
||
limit := i + maxDanmuPerChunk
|
||
if limit >= n {
|
||
chunks = append(chunks, danmuChunk{start: start, end: math.Inf(1), records: records[i:]})
|
||
break
|
||
}
|
||
// 块尾取第 limit 条弹幕显示结束的时刻
|
||
end := records[limit-1].Time + offset + danmuShowSeconds
|
||
if maxEnd > 0 && end > maxEnd {
|
||
end = maxEnd
|
||
}
|
||
if end <= start {
|
||
end = start + danmuShowSeconds
|
||
}
|
||
// 起点早于块尾的弹幕(显示会跨过块尾)也并入这一块
|
||
j := limit
|
||
for j < n && records[j].Time+offset < end {
|
||
j++
|
||
}
|
||
chunks = append(chunks, danmuChunk{start: start, end: end, records: records[i:j]})
|
||
i = j
|
||
start = end
|
||
}
|
||
return chunks
|
||
}
|
||
|
||
// 为分块渲染生成只包含该块弹幕的字幕文件(相对块起点计时)
|
||
func writeChunkAss(path string, meta danmuMeta, chunk danmuChunk, offset float64, avW, avGap int) error {
|
||
f, err := os.Create(path)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
defer f.Close()
|
||
|
||
// ASS 必须带 [Script Info] 开头,否则 libass 探测格式会失败
|
||
fmt.Fprintf(f, scriptInfoTmpl, "", "", filepath.Base(path), meta.PlayResX, meta.PlayResY)
|
||
fmt.Fprintf(f, stylesTmpl, meta.FontSize)
|
||
fmt.Fprint(f, eventsTmpl)
|
||
|
||
for _, r := range chunk.records {
|
||
t0 := r.Time + offset
|
||
visStart := math.Max(t0, chunk.start)
|
||
visEnd := math.Min(t0+danmuShowSeconds, chunk.end)
|
||
if visEnd <= visStart {
|
||
continue
|
||
}
|
||
text := r.Name + "(" + strconv.FormatInt(r.UID, 10) + "):" + r.Content
|
||
y := meta.FontSize * r.Lane
|
||
fmt.Fprintf(f, dialogueNickTmpl,
|
||
assTime(int64((visStart-chunk.start)*1e9)), assTime(int64((visEnd-chunk.start)*1e9)),
|
||
strings.ReplaceAll(r.Name, ",", " "), r.UID,
|
||
meta.PlayResX+avW+avGap, y, -r.Width, y, text)
|
||
}
|
||
return nil
|
||
}
|
||
|
||
// 渲染一个块:把这一块的弹幕(文字层 + 头像层)压进对应的视频片段
|
||
// cut 为 true 时表示这是分块渲染,需要按时长切出视频片段并重编码音频
|
||
func (c *burnContext) renderChunk(chunk danmuChunk, assBase, out string, cut bool) error {
|
||
burnMode := c.videoFile != ""
|
||
|
||
// 统计这一块用到的头像,按首次出现顺序分配输入序号(输入 0 是视频或背景)
|
||
uidOrder := []int64{}
|
||
uidCount := make(map[int64]int)
|
||
uidInput := make(map[int64]int)
|
||
for _, r := range chunk.records {
|
||
if _, ok := c.avatars[r.UID]; !ok {
|
||
continue
|
||
}
|
||
if _, seen := uidCount[r.UID]; !seen {
|
||
uidOrder = append(uidOrder, r.UID)
|
||
uidInput[r.UID] = len(uidOrder)
|
||
}
|
||
uidCount[r.UID]++
|
||
}
|
||
|
||
args := []string{"-y"}
|
||
if burnMode {
|
||
if cut {
|
||
// -ss 放在 -i 之前是快速定位,配合重编码能精确切到指定时间点
|
||
args = append(args, "-ss", formatSeconds(chunk.start))
|
||
}
|
||
args = append(args, "-i", c.videoFile)
|
||
} else {
|
||
args = append(args, "-f", "lavfi", "-i",
|
||
fmt.Sprintf("color=c=black@0:s=%dx%d:d=%.2f:r=30,format=rgba",
|
||
c.meta.PlayResX, c.meta.PlayResY, c.chunkDuration(chunk)+0.5))
|
||
}
|
||
for _, uid := range uidOrder {
|
||
args = append(args, "-i", c.avatars[uid])
|
||
}
|
||
|
||
fc := &strings.Builder{}
|
||
prev := "0:v"
|
||
if burnMode {
|
||
fmt.Fprintf(fc, "[%s]subtitles=filename='%s'[b0];\n", prev, assBase)
|
||
prev = "b0"
|
||
}
|
||
// 把每个头像缩放到弹幕字号对应的直径,同一个头像被多条弹幕使用时用 split 复制
|
||
ad := avatarSize(c.meta.FontSize)
|
||
gap := avatarGap(c.meta.FontSize)
|
||
for _, uid := range uidOrder {
|
||
idx := uidInput[uid]
|
||
if uidCount[uid] == 1 {
|
||
fmt.Fprintf(fc, "[%d:v]scale=%d:%d[av%d];\n", idx, ad, ad, idx)
|
||
} else {
|
||
outs := &strings.Builder{}
|
||
for j := 0; j < uidCount[uid]; j++ {
|
||
fmt.Fprintf(outs, "[av%d_%d]", idx, j)
|
||
}
|
||
fmt.Fprintf(fc, "[%d:v]scale=%d:%d,split=%d%s;\n", idx, ad, ad, uidCount[uid], outs.String())
|
||
}
|
||
}
|
||
uidCopy := make(map[int64]int) // 每个头像当前用掉的份数
|
||
for n, r := range chunk.records {
|
||
idx, ok := uidInput[r.UID]
|
||
if !ok {
|
||
continue
|
||
}
|
||
avLabel := strconv.Itoa(idx)
|
||
if uidCount[r.UID] > 1 {
|
||
avLabel = strconv.Itoa(idx) + "_" + strconv.Itoa(uidCopy[r.UID])
|
||
uidCopy[r.UID]++
|
||
}
|
||
t0 := r.Time + c.offset
|
||
visStart := math.Max(t0, chunk.start)
|
||
visEnd := math.Min(t0+danmuShowSeconds, chunk.end)
|
||
if visEnd <= visStart {
|
||
continue
|
||
}
|
||
// 头像在文字左侧并留出间隔:ASS 里的文字左边缘是这里加 ad+gap,头像跟着文字一起滚动。
|
||
// 位置公式 x = 画面宽 - (画面宽+总宽)*(t-t0)/10,这里 t 是块内时间,把 t0 也换算到块内坐标
|
||
total := ad + gap + r.Width
|
||
shift := chunk.start - t0
|
||
x := fmt.Sprintf("%d-(%d+%d)*(t%+.3f)/10", c.meta.PlayResX, c.meta.PlayResX, total, shift)
|
||
y := c.meta.FontSize*r.Lane + (c.meta.FontSize-ad)/2 // 头像与文字墨迹垂直居中
|
||
outLabel := fmt.Sprintf("s%d", n)
|
||
fmt.Fprintf(fc, "[%s][av%s]overlay=x='%s':y=%d:enable='between(t,%.3f,%.3f)'[%s];\n",
|
||
prev, avLabel, x, y, visStart-chunk.start, visEnd-chunk.start, outLabel)
|
||
prev = outLabel
|
||
}
|
||
if burnMode {
|
||
fmt.Fprintf(fc, "[%s]format=yuv420p[v]", prev)
|
||
} else {
|
||
fmt.Fprintf(fc, "[%s]format=yuva444p10le[v]", prev)
|
||
}
|
||
|
||
args = append(args, "-filter_complex", fc.String(), "-map", "[v]")
|
||
if burnMode {
|
||
args = append(args, "-map", "0:a?")
|
||
if cut {
|
||
// 分块切割点的音频不在帧边界上,必须重编码,否则拼接后会累积音画错位
|
||
args = append(args, "-c:a", "aac", "-b:a", "192k")
|
||
} else {
|
||
args = append(args, "-c:a", "copy")
|
||
}
|
||
args = append(args, "-movflags", "+faststart")
|
||
} else {
|
||
args = append(args, "-c:v", "prores_ks", "-profile:v", "4444", "-pix_fmt", "yuva444p10le")
|
||
}
|
||
if cut && !math.IsInf(chunk.end, 1) {
|
||
args = append(args, "-t", formatSeconds(chunk.end-chunk.start))
|
||
}
|
||
|
||
if !burnMode {
|
||
// 透明头像层模式固定用 ProRes 4444,不能再追加编码器覆盖
|
||
full := append(append([]string{}, args...), out)
|
||
lPrintln("开始运行 ffmpeg 合成透明头像层,输出到 " + out)
|
||
if err := runFFmpeg(c.ffmpegFile, full, c.dir); err != nil {
|
||
return err
|
||
}
|
||
return nil
|
||
}
|
||
|
||
// 烧录模式:优先 Intel QSV 硬件编码,运行失败自动回退 libx264
|
||
encoders := buildEncoders(c.bitrate, c.ffmpegFile)
|
||
// 已经在这条流水线里验证过可用的编码器直接复用,分块时不用每块都回退一次
|
||
begin := 0
|
||
if c.encoder != "" {
|
||
for i, enc := range encoders {
|
||
if enc.name == c.encoder {
|
||
begin = i
|
||
break
|
||
}
|
||
}
|
||
}
|
||
var lastErr error
|
||
for i := begin; i < len(encoders); i++ {
|
||
enc := encoders[i]
|
||
full := append(append([]string{}, args...), "-c:v", enc.name)
|
||
full = append(full, enc.opts...)
|
||
full = append(full, out)
|
||
if i == begin {
|
||
lPrintf("开始运行 ffmpeg 合成(编码器 %s),输出到 %s", enc.name, out)
|
||
}
|
||
if err := runFFmpeg(c.ffmpegFile, full, c.dir); err != nil {
|
||
lastErr = err
|
||
if i+1 < len(encoders) {
|
||
lPrintWarnf("编码器 %s 运行失败,回退到 %s:%v", enc.name, encoders[i+1].name, err)
|
||
continue
|
||
}
|
||
return fmt.Errorf("ffmpeg 运行失败:%v", lastErr)
|
||
}
|
||
c.encoder = enc.name
|
||
return nil
|
||
}
|
||
return fmt.Errorf("ffmpeg 运行失败:%v", lastErr)
|
||
}
|
||
|
||
// 这一块需要渲染的时长(用于透明头像层模式生成等长的黑底)
|
||
func (c *burnContext) chunkDuration(chunk danmuChunk) float64 {
|
||
if !math.IsInf(chunk.end, 1) {
|
||
return chunk.end - chunk.start
|
||
}
|
||
maxT := chunk.start
|
||
for _, r := range chunk.records {
|
||
if t := r.Time + c.offset + danmuShowSeconds; t > maxT {
|
||
maxT = t
|
||
}
|
||
}
|
||
return maxT - chunk.start
|
||
}
|
||
|
||
// 合成用的视频编码器配置
|
||
type videoEncoder struct {
|
||
name string
|
||
opts []string
|
||
}
|
||
|
||
// 按目标码率构造编码器候选列表,QSV 可用时优先
|
||
func buildEncoders(bitrate int, ffmpegFile string) []videoEncoder {
|
||
var qOpts, xOpts []string
|
||
if bitrate > 0 {
|
||
br := strconv.Itoa(bitrate) + "k"
|
||
qOpts = []string{"-b:v", br}
|
||
xOpts = []string{"-b:v", br, "-preset", "medium"}
|
||
} else {
|
||
qOpts = []string{"-global_quality", "22", "-preset", "medium"}
|
||
xOpts = []string{"-crf", "18", "-preset", "medium"}
|
||
}
|
||
encoders := []videoEncoder{{"libx264", xOpts}}
|
||
if qsvAvailable(ffmpegFile) {
|
||
encoders = append([]videoEncoder{{"h264_qsv", qOpts}}, encoders...)
|
||
}
|
||
return encoders
|
||
}
|
||
|
||
// 运行 ffmpeg,输出直接透传到控制台
|
||
func runFFmpeg(ffmpegFile string, args []string, dir string) error {
|
||
cmd := exec.Command(ffmpegFile, args...)
|
||
cmd.Dir = dir
|
||
cmd.Stdout = os.Stdout
|
||
cmd.Stderr = os.Stderr
|
||
return cmd.Run()
|
||
}
|
||
|
||
// 用 concat demuxer 把各分块的成片拼接起来
|
||
func concatChunks(ffmpegFile string, files []string, out, dir string) error {
|
||
listFile := filepath.Join(dir, chunkTempPrefix+"list.txt")
|
||
var b strings.Builder
|
||
for _, p := range files {
|
||
// concat demuxer 里单引号要写成 '\''
|
||
p = filepath.ToSlash(p)
|
||
b.WriteString("file '" + strings.ReplaceAll(p, "'", `'\''`) + "'\n")
|
||
}
|
||
if err := os.WriteFile(listFile, []byte(b.String()), 0644); err != nil {
|
||
return fmt.Errorf("写入拼接清单失败:%v", err)
|
||
}
|
||
defer os.Remove(listFile)
|
||
|
||
if err := runFFmpeg(ffmpegFile, []string{"-y", "-f", "concat", "-safe", "0", "-i", listFile, "-c", "copy", out}, dir); err != nil {
|
||
return fmt.Errorf("拼接各分块失败:%v", err)
|
||
}
|
||
return nil
|
||
}
|
||
|
||
// 把秒格式化成 ffmpeg 参数,保留毫秒
|
||
func formatSeconds(sec float64) string {
|
||
return strconv.FormatFloat(sec, 'f', 3, 64)
|
||
}
|
||
|
||
// 下载头像并圆形裁切,失败自动重试,返回缓存的 PNG 路径,失败返回空字符串
|
||
func ensureAvatar(uid int64, url, cacheDir string) string {
|
||
// 文件名里带上头像链接的摘要,主播换头像后不会一直用着旧缓存
|
||
sum := sha1.Sum([]byte(url))
|
||
path := filepath.Join(cacheDir, fmt.Sprintf("%d_%x.png", uid, sum[:6]))
|
||
if _, err := os.Stat(path); err == nil {
|
||
return path
|
||
}
|
||
if err := os.MkdirAll(cacheDir, 0755); err != nil {
|
||
lPrintWarnf("创建头像缓存文件夹 %s 失败,弹幕只烧录文字不带头像:%v", cacheDir, err)
|
||
return ""
|
||
}
|
||
// 清掉这个用户已经失效的旧头像
|
||
if olds, err := filepath.Glob(filepath.Join(cacheDir, strconv.FormatInt(uid, 10)+"_*.png")); err == nil {
|
||
for _, p := range olds {
|
||
if p != path {
|
||
_ = os.Remove(p)
|
||
}
|
||
}
|
||
}
|
||
|
||
err := runThrice(func() error {
|
||
return downloadAvatar(uid, url, path)
|
||
})
|
||
if err != nil {
|
||
lPrintWarnf("uid %d 的头像不可用,该用户的弹幕只烧录文字不带头像:%v", uid, err)
|
||
return ""
|
||
}
|
||
return path
|
||
}
|
||
|
||
func downloadAvatar(uid int64, url, path string) error {
|
||
client := &http.Client{Timeout: 30 * time.Second}
|
||
req, err := http.NewRequest("GET", url, nil)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
req.Header.Set("User-Agent", "Mozilla/5.0")
|
||
resp, err := client.Do(req)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
defer resp.Body.Close()
|
||
if resp.StatusCode != http.StatusOK {
|
||
return fmt.Errorf("HTTP %d", resp.StatusCode)
|
||
}
|
||
data, err := io.ReadAll(resp.Body)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
src, err := decodeAvatar(data, resp.Header.Get("Content-Type"))
|
||
if err != nil {
|
||
return err
|
||
}
|
||
|
||
out, err := os.Create(path)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
defer out.Close()
|
||
return png.Encode(out, circleAvatar(src, avatarCacheSize))
|
||
}
|
||
|
||
// 解码头像图片。AcFun 的头像里混着 gif 和 webp:
|
||
// gif 由标准库解码,多帧的取第一帧;webp 只支持静态的,动画 webp 会报错,
|
||
// 这种情况下这个用户就不烧头像,只烧文字。
|
||
func decodeAvatar(data []byte, contentType string) (image.Image, error) {
|
||
src, _, err := image.Decode(bytes.NewReader(data))
|
||
if err != nil {
|
||
return nil, fmt.Errorf("图片解析不了(Content-Type: %s):%v", contentType, err)
|
||
}
|
||
return src, nil
|
||
}
|
||
|
||
// uid → 头像链接的记忆,包含查不到的结果,避免同一个 uid 反复去问
|
||
var avatarURLMemo struct {
|
||
sync.Mutex
|
||
m map[int64]string
|
||
}
|
||
|
||
// 实际去查头像的动作,单独拎出来方便测试替换掉
|
||
var lookupAvatarURL = fetchAvatarURL
|
||
|
||
// 按 uid 反查头像链接。原版 .ass 里没有头像链接,只能这样补回来
|
||
func avatarURLByUID(uid int64) string {
|
||
if uid <= 0 {
|
||
return ""
|
||
}
|
||
avatarURLMemo.Lock()
|
||
if avatarURLMemo.m == nil {
|
||
avatarURLMemo.m = make(map[int64]string)
|
||
}
|
||
if url, ok := avatarURLMemo.m[uid]; ok {
|
||
avatarURLMemo.Unlock()
|
||
return url
|
||
}
|
||
avatarURLMemo.Unlock()
|
||
|
||
url := lookupAvatarURL(uid)
|
||
|
||
avatarURLMemo.Lock()
|
||
avatarURLMemo.m[uid] = url
|
||
avatarURLMemo.Unlock()
|
||
return url
|
||
}
|
||
|
||
// 查 AcFun 的用户信息接口取头像地址
|
||
func fetchAvatarURL(uid int64) string {
|
||
client := &http.Client{Timeout: 15 * time.Second}
|
||
addr := fmt.Sprintf("https://www.acfun.cn/rest/pc-direct/user/userInfo?userId=%d", uid)
|
||
req, err := http.NewRequest("GET", addr, nil)
|
||
if err != nil {
|
||
return ""
|
||
}
|
||
req.Header.Set("User-Agent", "Mozilla/5.0")
|
||
resp, err := client.Do(req)
|
||
if err != nil {
|
||
return ""
|
||
}
|
||
defer resp.Body.Close()
|
||
if resp.StatusCode != http.StatusOK {
|
||
return ""
|
||
}
|
||
data, err := io.ReadAll(io.LimitReader(resp.Body, 256*1024))
|
||
if err != nil {
|
||
return ""
|
||
}
|
||
var out struct {
|
||
Result int `json:"result"`
|
||
Profile struct {
|
||
HeadURL string `json:"headUrl"`
|
||
} `json:"profile"`
|
||
}
|
||
if json.Unmarshal(data, &out) != nil || out.Result != 0 {
|
||
return ""
|
||
}
|
||
return strings.TrimSpace(out.Profile.HeadURL)
|
||
}
|
||
|
||
// QSV 可用性的探测结果,按 ffmpeg 路径缓存,避免分块时每块都试一次
|
||
var qsvCache struct {
|
||
sync.Mutex
|
||
path string
|
||
ok bool
|
||
done bool
|
||
}
|
||
|
||
// 探测 Intel QSV(h264_qsv)硬件编码是否可用
|
||
func qsvAvailable(ffmpegFile string) bool {
|
||
qsvCache.Lock()
|
||
if qsvCache.done && qsvCache.path == ffmpegFile {
|
||
ok := qsvCache.ok
|
||
qsvCache.Unlock()
|
||
return ok
|
||
}
|
||
qsvCache.Unlock()
|
||
|
||
cmd := exec.Command(ffmpegFile, "-hide_banner", "-loglevel", "error",
|
||
"-f", "lavfi", "-i", "color=c=black:s=320x240:d=0.5:r=30",
|
||
"-c:v", "h264_qsv", "-frames:v", "5", "-f", "null", "-")
|
||
ok := cmd.Run() == nil
|
||
|
||
qsvCache.Lock()
|
||
qsvCache.path, qsvCache.ok, qsvCache.done = ffmpegFile, ok, true
|
||
qsvCache.Unlock()
|
||
return ok
|
||
}
|
||
|
||
// 把头像居中裁方、缩放并圆形裁切,输出带透明通道的 PNG
|
||
func circleAvatar(src image.Image, size int) image.Image {
|
||
sb := src.Bounds()
|
||
side := sb.Dx()
|
||
if sb.Dy() < side {
|
||
side = sb.Dy()
|
||
}
|
||
cx, cy := sb.Min.X+sb.Dx()/2, sb.Min.Y+sb.Dy()/2
|
||
half := side / 2
|
||
sr := image.Rect(cx-half, cy-half, cx+half, cy+half)
|
||
|
||
scaled := image.NewRGBA(image.Rect(0, 0, size, size))
|
||
xdraw.CatmullRom.Scale(scaled, scaled.Bounds(), src, sr, xdraw.Over, nil)
|
||
|
||
// 圆形遮罩,边缘 2px 抗锯齿过渡
|
||
mask := image.NewAlpha(image.Rect(0, 0, size, size))
|
||
radius := float64(size) / 2
|
||
edge := 2.0
|
||
start := 1 - edge/radius
|
||
for y := 0; y < size; y++ {
|
||
for x := 0; x < size; x++ {
|
||
dx := float64(x-size/2) + 0.5
|
||
dy := float64(y-size/2) + 0.5
|
||
d := math.Sqrt(dx*dx+dy*dy) / radius
|
||
switch {
|
||
case d >= 1:
|
||
mask.SetAlpha(x, y, color.Alpha{A: 0})
|
||
case d > start:
|
||
a := uint8(255 * (1 - (d-start)/(edge/radius)))
|
||
mask.SetAlpha(x, y, color.Alpha{A: a})
|
||
default:
|
||
mask.SetAlpha(x, y, color.Alpha{A: 255})
|
||
}
|
||
}
|
||
}
|
||
|
||
dst := image.NewRGBA(image.Rect(0, 0, size, size))
|
||
draw.DrawMask(dst, dst.Bounds(), scaled, image.Point{}, mask, image.Point{}, draw.Over)
|
||
return dst
|
||
}
|