[FFmpeg-devel] [PATCH v5 3/3] avcodec/nvenc: write out user data unregistered SEI
Marton Balint
cus at passwd.hu
Mon May 17 18:52:13 EEST 2021
On Mon, 17 May 2021, Timo Rothenpieler wrote:
> On 17.05.2021 10:19, Brad Hards wrote:
>> Signed-off-by: Brad Hards <bradh at frogmouth.net>
>> ---
>> libavcodec/nvenc.c | 64 ++++++++++++++++++++++++++++++++++++++--------
>> 1 file changed, 53 insertions(+), 11 deletions(-)
>>
>> diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
>> index 0dcd93a99c..e22fdfb5a8 100644
>> --- a/libavcodec/nvenc.c
>> +++ b/libavcodec/nvenc.c
>> @@ -2170,9 +2170,10 @@ static int nvenc_send_frame(AVCodecContext *avctx,
>> const AVFrame *frame)
>> NVENCSTATUS nv_status;
>> NvencSurface *tmp_out_surf, *in_surf;
>> int res, res2;
>> - NV_ENC_SEI_PAYLOAD sei_data[8];
>> + NV_ENC_SEI_PAYLOAD *sei_data = 0;
>> int sei_count = 0;
>> int i;
>> + int total_unregistered_sei = 0;
>> NvencContext *ctx = avctx->priv_data;
>> NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
>> @@ -2185,6 +2186,8 @@ static int nvenc_send_frame(AVCodecContext *avctx,
>> const AVFrame *frame)
>> return AVERROR(EINVAL);
>> if (frame && frame->buf[0]) {
>> + void *a53_data = NULL;
>> + void *tc_data = NULL;
>> in_surf = get_free_frame(ctx);
>> if (!in_surf)
>> return AVERROR(EAGAIN);
>> @@ -2230,7 +2233,6 @@ static int nvenc_send_frame(AVCodecContext *avctx,
>> const AVFrame *frame)
>> pic_params.inputTimeStamp = frame->pts;
>> if (ctx->a53_cc && av_frame_get_side_data(frame,
>> AV_FRAME_DATA_A53_CC)) {
>> - void *a53_data = NULL;
>> size_t a53_size = 0;
>> if (ff_alloc_a53_sei(frame, 0, (void**)&a53_data,
>> &a53_size) < 0) {
>> @@ -2238,15 +2240,21 @@ static int nvenc_send_frame(AVCodecContext *avctx,
>> const AVFrame *frame)
>> }
>> if (a53_data) {
>> - sei_data[sei_count].payloadSize = (uint32_t)a53_size;
>> - sei_data[sei_count].payloadType = 4;
>> - sei_data[sei_count].payload = (uint8_t*)a53_data;
>> - sei_count ++;
>> + sei_data = av_realloc_array(sei_data, sei_count + 1,
>> sizeof(NV_ENC_SEI_PAYLOAD));
>
> Doing this realloc dance every frame is extremely inefficient.
> The sei_data array and its current size could instead be stored in the nvenc
> context, with a best-guess initial size, and then only grow when the already
> allocated size is too small.
Don't we have av_fast_realloc for this?
Thanks,
Marton
More information about the ffmpeg-devel
mailing list