Merge pull request #6461 from 21pages/fix_max_codec_thread_num

fix exceeding the maximum number of threads allowed

libs/scrap/src/common/aom.rs

@@ -106,7 +106,7 @@ mod webrtc {
         // Overwrite default config with input encoder settings & RTC-relevant values.
         c.g_w = cfg.width;
         c.g_h = cfg.height;
-        c.g_threads = codec_thread_num() as _;
+        c.g_threads = codec_thread_num(64) as _;
         c.g_timebase.num = 1;
         c.g_timebase.den = kRtpTicksPerSecond;
         c.g_input_bit_depth = kBitDepth;
@@ -452,7 +452,7 @@ impl AomDecoder {
         let i = call_aom_ptr!(aom_codec_av1_dx());
         let mut ctx = Default::default();
         let cfg = aom_codec_dec_cfg_t {
-            threads: codec_thread_num() as _,
+            threads: codec_thread_num(64) as _,
             w: 0,
             h: 0,
             allow_lowbitdepth: 1,

libs/scrap/src/common/codec.rs

@@ -566,15 +566,18 @@ pub fn base_bitrate(width: u32, height: u32) -> u32 {
     base_bitrate
 }
 
-pub fn codec_thread_num() -> usize {
+pub fn codec_thread_num(limit: usize) -> usize {
     let max: usize = num_cpus::get();
     let mut res;
     let info;
+    let mut s = System::new();
+    s.refresh_memory();
+    let memory = s.available_memory() / 1024 / 1024 / 1024;
     #[cfg(windows)]
     {
         res = 0;
         let percent = hbb_common::platform::windows::cpu_uage_one_minute();
-        info = format!("cpu usage:{:?}", percent);
+        info = format!("cpu usage: {:?}", percent);
         if let Some(pecent) = percent {
             if pecent < 100.0 {
                 res = ((100.0 - pecent) * (max as f64) / 200.0).round() as usize;
@@ -583,24 +586,41 @@ pub fn codec_thread_num() -> usize {
     }
     #[cfg(not(windows))]
     {
-        let mut s = System::new();
         s.refresh_cpu_usage();
         // https://man7.org/linux/man-pages/man3/getloadavg.3.html
         let avg = s.load_average();
-        info = format!("cpu loadavg:{}", avg.one);
+        info = format!("cpu loadavg: {}", avg.one);
         res = (((max as f64) - avg.one) * 0.5).round() as usize;
     }
     res = std::cmp::min(res, max / 2);
-    if res == 0 {
-        res = 1;
-    }
+    res = std::cmp::min(res, memory as usize / 2);
+    //  Use common thread count
+    res = match res {
+        _ if res >= 64 => 64,
+        _ if res >= 32 => 32,
+        _ if res >= 16 => 16,
+        _ if res >= 8 => 8,
+        _ if res >= 4 => 4,
+        _ if res >= 2 => 2,
+        _ => 1,
+    };
+    // https://aomedia.googlesource.com/aom/+/refs/heads/main/av1/av1_cx_iface.c#677
+    // https://aomedia.googlesource.com/aom/+/refs/heads/main/aom_util/aom_thread.h#26
+    // https://chromium.googlesource.com/webm/libvpx/+/refs/heads/main/vp8/vp8_cx_iface.c#148
+    // https://chromium.googlesource.com/webm/libvpx/+/refs/heads/main/vp9/vp9_cx_iface.c#190
+    // https://github.com/FFmpeg/FFmpeg/blob/7c16bf0829802534004326c8e65fb6cdbdb634fa/libavcodec/pthread.c#L65
+    // https://github.com/FFmpeg/FFmpeg/blob/7c16bf0829802534004326c8e65fb6cdbdb634fa/libavcodec/pthread_internal.h#L26
+    // libaom: MAX_NUM_THREADS = 64
+    // libvpx: MAX_NUM_THREADS = 64
+    // ffmpeg: MAX_AUTO_THREADS = 16
+    res = std::cmp::min(res, limit);
     // avoid frequent log
     let log = match THREAD_LOG_TIME.lock().unwrap().clone() {
         Some(instant) => instant.elapsed().as_secs() > 1,
         None => true,
     };
     if log {
-        log::info!("cpu num: {max}, {info}, codec thread: {res}");
+        log::info!("cpu num: {max}, {info}, available memory: {memory}G, codec thread: {res}");
         *THREAD_LOG_TIME.lock().unwrap() = Some(Instant::now());
     }
     res

libs/scrap/src/common/hwcodec.rs

@@ -63,7 +63,7 @@ impl EncoderApi for HwEncoder {
                     gop,
                     quality: DEFAULT_HW_QUALITY,
                     rc: DEFAULT_RC,
-                    thread_count: codec_thread_num() as _, // ffmpeg's thread_count is used for cpu
+                    thread_count: codec_thread_num(16) as _, // ffmpeg's thread_count is used for cpu
                 };
                 let format = match Encoder::format_from_name(config.name.clone()) {
                     Ok(format) => format,
@@ -235,7 +235,7 @@ impl HwDecoder {
         let ctx = DecodeContext {
             name: info.name.clone(),
             device_type: info.hwdevice.clone(),
-            thread_count: codec_thread_num() as _,
+            thread_count: codec_thread_num(16) as _,
         };
         match Decoder::new(ctx) {
             Ok(decoder) => Ok(HwDecoder { decoder, info }),

libs/scrap/src/common/vpxcodec.rs

@@ -73,7 +73,7 @@ impl EncoderApi for VpxEncoder {
                 // When the data buffer falls below this percentage of fullness, a dropped frame is indicated. Set the threshold to zero (0) to disable this feature.
                 // In dynamic scenes, low bitrate gets low fps while high bitrate gets high fps.
                 c.rc_dropframe_thresh = 25;
-                c.g_threads = codec_thread_num() as _;
+                c.g_threads = codec_thread_num(64) as _;
                 c.g_error_resilient = VPX_ERROR_RESILIENT_DEFAULT;
                 // https://developers.google.com/media/vp9/bitrate-modes/
                 // Constant Bitrate mode (CBR) is recommended for live streaming with VP9.
@@ -450,7 +450,7 @@ impl VpxDecoder {
         };
         let mut ctx = Default::default();
         let cfg = vpx_codec_dec_cfg_t {
-            threads: codec_thread_num() as _,
+            threads: codec_thread_num(64) as _,
             w: 0,
             h: 0,
         };