From: green-hand-shandong <[email protected]>
hi_mpi_venc echoes the frame index we submitted as the stream pts, in whatever value was passed to send_frame. Track the submitted AVFrame pts in a small FIFO and re-attach the original (CLI timebase) pts to the output packet, so downstream muxers see correct timestamps. Document the faster-than-realtime caveat: CLI fps-mode drops frames to synchronize unless -fps_mode passthrough is set for offline file transcodes. End-to-end validation (Atlas 300I-Duo, CANN 9.1.0, dev7): - 1080p->720p decode(hw) -> scale_ascend -> encode: 748/750 frames @ 48 fps, 4Mbps CBR output intact - 2 concurrent transcode processes on one chip: both complete - 3x loop long-run (2243 frames): stable, no DVPP memory growth - full-feature smoke: 8/8 paths rc=0 Signed-off-by: green-hand-shandong <[email protected]> --- libavcodec/ascendenc.c | 30 ++++++++++++++++++++++++------ 1 file changed, 24 insertions(+), 6 deletions(-) diff --git a/libavcodec/ascendenc.c b/libavcodec/ascendenc.c index 430bf81eb3..1fc3e780f2 100644 --- a/libavcodec/ascendenc.c +++ b/libavcodec/ascendenc.c @@ -2,6 +2,11 @@ * Ascend NPU (Atlas 300I-Duo / Ascend 310P) hardware video encoder * via the CANN DVPP media APIs (hi_mpi_venc). * + * NOTE for CLI users: the hardware encoder is much faster than real-time; + * when transcoding offline files the CLI's default fps-mode may drop + * frames to keep sync. Pass -fps_mode passthrough for file transcodes + * (the standard practice for faster-than-realtime encoders). + * * Copyright (c) 2026 green-hand-shandong * * This file is part of FFmpeg. @@ -60,6 +65,8 @@ typedef struct AscendEncContext { uint64_t frame_count; uint64_t frame_count_out; + int64_t pts_lut[64]; /* submitted frame->pts ring, consumed FIFO */ + int pts_lut_count; /* staging DVPP input buffers; VENC references them asynchronously, * so frames are staged in rotation. */ @@ -351,10 +358,19 @@ static int ascend_collect_stream(AVCodecContext *avctx, AVPacket *pkt, pkt->size += stream.pack[i].len; } } - /* VENC echoes back the pts we submitted (in the input's container - * timebase). Downstream muxers assume a monotonically increasing - * sequence in the packet timebase, so renumber by frame index. */ - pkt->pts = pkt->dts = (int64_t)ctx->frame_count_out++; + /* VENC echoes the frame index we submitted; map it back to the + * original frame pts (in the CLI-provided encoder timebase). */ + { + int64_t orig = ctx->pts_lut_count > 0 ? ctx->pts_lut[0] : + (int64_t)ctx->frame_count_out; + if (ctx->pts_lut_count > 0) { + memmove(ctx->pts_lut, ctx->pts_lut + 1, + sizeof(ctx->pts_lut[0]) * (ctx->pts_lut_count - 1)); + ctx->pts_lut_count--; + } + pkt->pts = pkt->dts = orig; + } + ctx->frame_count_out++; { const hi_venc_h264_nalu_type h264t = stream.pack[0].data_type.h264_type; const hi_venc_h265_nalu_type h265t = stream.pack[0].data_type.h265_type; @@ -446,8 +462,10 @@ static int ascend_encode_cb(AVCodecContext *avctx, AVPacket *pkt, /* time_ref must be monotonically even; VDEC/VENC pairing uses it * for reference marking. */ fi.v_frame.time_ref = (uint32_t)(ctx->frame_count * 2); - fi.v_frame.pts = frame->pts != AV_NOPTS_VALUE ? - (uint64_t)frame->pts : ctx->frame_count; + fi.v_frame.pts = ctx->frame_count; /* VENC echoes this back */ + if (ctx->pts_lut_count < 64) + ctx->pts_lut[ctx->pts_lut_count++] = + frame->pts != AV_NOPTS_VALUE ? frame->pts : (int64_t)ctx->frame_count; ret = hi_mpi_venc_send_frame(ctx->chn, &fi, ASCEND_ENC_TIMEOUT_MS); if (ret != 0) { -- 2.25.1 _______________________________________________ ffmpeg-devel mailing list -- [email protected] To unsubscribe send an email to [email protected]
