| Message ID | ZsDKRC5xDTMBBvXZ@vicerveza.homeunix.net |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel] Less CPU use in hwaccel MJPEG decoding | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/commit_msg_loongarch64 | warning | The first line of the commit message must start with a context terminated by a colon and a space, for example "lavu/opt: " or "doc: ". |
v2 attached. I had crippled the indentation. On Sat, Aug 17, 2024 at 06:05:24PM +0200, Lluís Batlle i Rossell wrote: > attached From ea2d702daa88bf3b1bde1f03c26c69bb8ab9e2b8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Llu=C3=ADs=20Batlle=20i=20Rossell?= <viric@viric.name> Date: Sat, 17 Aug 2024 18:03:37 +0200 Subject: [PATCH] Less CPU use in hwaccel MJPEG decoding It skips the unnecessary unescaping of the entropy-encoded data in case of using a hardware accelerator. This is 50% less of CPU use in my benchmark. --- libavcodec/mjpegdec.c | 11 ++++++++++- libavcodec/mjpegdec.h | 1 + libavcodec/mxpegdec.c | 2 +- 3 files changed, 12 insertions(+), 2 deletions(-) diff --git a/libavcodec/mjpegdec.c b/libavcodec/mjpegdec.c index 8676155ecf..43583c1ccf 100644 --- a/libavcodec/mjpegdec.c +++ b/libavcodec/mjpegdec.c @@ -2235,6 +2235,7 @@ found: } int ff_mjpeg_find_marker(MJpegDecodeContext *s, + const struct AVHWAccel *hwaccel, const uint8_t **buf_ptr, const uint8_t *buf_end, const uint8_t **unescaped_buf_ptr, int *unescaped_buf_size) @@ -2242,6 +2243,14 @@ int ff_mjpeg_find_marker(MJpegDecodeContext *s, int start_code; start_code = find_marker(buf_ptr, buf_end); + if (start_code == SOS && hwaccel) + { + /* hardware accelerators don't need unescaping */ + *unescaped_buf_ptr = *buf_ptr; + *unescaped_buf_size = buf_end - *buf_ptr; + return start_code; + } + av_fast_padded_malloc(&s->buffer, &s->buffer_size, buf_end - *buf_ptr); if (!s->buffer) return AVERROR(ENOMEM); @@ -2399,7 +2408,7 @@ redo_for_pal8: buf_end = buf + buf_size; while (buf_ptr < buf_end) { /* find start next marker */ - start_code = ff_mjpeg_find_marker(s, &buf_ptr, buf_end, + start_code = ff_mjpeg_find_marker(s, avctx->hwaccel, &buf_ptr, buf_end, &unescaped_buf_ptr, &unescaped_buf_size); /* EOF */ diff --git a/libavcodec/mjpegdec.h b/libavcodec/mjpegdec.h index 13c524d597..bc21eab647 100644 --- a/libavcodec/mjpegdec.h +++ b/libavcodec/mjpegdec.h @@ -184,6 +184,7 @@ int ff_mjpeg_decode_sos(MJpegDecodeContext *s, const uint8_t *mb_bitmask,int mb_bitmask_size, const AVFrame *reference); int ff_mjpeg_find_marker(MJpegDecodeContext *s, + const struct AVHWAccel *hwaccel, const uint8_t **buf_ptr, const uint8_t *buf_end, const uint8_t **unescaped_buf_ptr, int *unescaped_buf_size); diff --git a/libavcodec/mxpegdec.c b/libavcodec/mxpegdec.c index 73df2ff9ff..d18d1541ed 100644 --- a/libavcodec/mxpegdec.c +++ b/libavcodec/mxpegdec.c @@ -202,7 +202,7 @@ static int mxpeg_decode_frame(AVCodecContext *avctx, AVFrame *rframe, s->got_mxm_bitmask = 0; s->got_sof_data = !!s->got_sof_data; while (buf_ptr < buf_end) { - start_code = ff_mjpeg_find_marker(jpg, &buf_ptr, buf_end, + start_code = ff_mjpeg_find_marker(jpg, NULL, &buf_ptr, buf_end, &unescaped_buf_ptr, &unescaped_buf_size); if (start_code < 0) goto the_end;
diff --git a/libavcodec/mjpegdec.c b/libavcodec/mjpegdec.c index 8676155ecf..759a01b981 100644 --- a/libavcodec/mjpegdec.c +++ b/libavcodec/mjpegdec.c @@ -2235,12 +2235,21 @@ found: } int ff_mjpeg_find_marker(MJpegDecodeContext *s, - const uint8_t **buf_ptr, const uint8_t *buf_end, - const uint8_t **unescaped_buf_ptr, - int *unescaped_buf_size) + const struct AVHWAccel *hwaccel, + const uint8_t **buf_ptr, const uint8_t *buf_end, + const uint8_t **unescaped_buf_ptr, + int *unescaped_buf_size) { int start_code; - start_code = find_marker(buf_ptr, buf_end); + start_code = find_marker(buf_ptr, buf_end); + + if (start_code == SOS && hwaccel) + { + /* hardware accelerators don't need unescaping */ + *unescaped_buf_ptr = *buf_ptr; + *unescaped_buf_size = buf_end - *buf_ptr; + return start_code; + } av_fast_padded_malloc(&s->buffer, &s->buffer_size, buf_end - *buf_ptr); if (!s->buffer) @@ -2399,7 +2408,7 @@ redo_for_pal8: buf_end = buf + buf_size; while (buf_ptr < buf_end) { /* find start next marker */ - start_code = ff_mjpeg_find_marker(s, &buf_ptr, buf_end, + start_code = ff_mjpeg_find_marker(s, avctx->hwaccel, &buf_ptr, buf_end, &unescaped_buf_ptr, &unescaped_buf_size); /* EOF */ diff --git a/libavcodec/mjpegdec.h b/libavcodec/mjpegdec.h index 13c524d597..bc21eab647 100644 --- a/libavcodec/mjpegdec.h +++ b/libavcodec/mjpegdec.h @@ -184,6 +184,7 @@ int ff_mjpeg_decode_sos(MJpegDecodeContext *s, const uint8_t *mb_bitmask,int mb_bitmask_size, const AVFrame *reference); int ff_mjpeg_find_marker(MJpegDecodeContext *s, + const struct AVHWAccel *hwaccel, const uint8_t **buf_ptr, const uint8_t *buf_end, const uint8_t **unescaped_buf_ptr, int *unescaped_buf_size); diff --git a/libavcodec/mxpegdec.c b/libavcodec/mxpegdec.c index 73df2ff9ff..d18d1541ed 100644 --- a/libavcodec/mxpegdec.c +++ b/libavcodec/mxpegdec.c @@ -202,7 +202,7 @@ static int mxpeg_decode_frame(AVCodecContext *avctx, AVFrame *rframe, s->got_mxm_bitmask = 0; s->got_sof_data = !!s->got_sof_data; while (buf_ptr < buf_end) { - start_code = ff_mjpeg_find_marker(jpg, &buf_ptr, buf_end, + start_code = ff_mjpeg_find_marker(jpg, NULL, &buf_ptr, buf_end, &unescaped_buf_ptr, &unescaped_buf_size); if (start_code < 0) goto the_end;
attached From 9e312c20d10839cb2c731acd23bc1b00177cd7dd Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Llu=C3=ADs=20Batlle=20i=20Rossell?= <viric@viric.name> Date: Sat, 17 Aug 2024 18:03:37 +0200 Subject: [PATCH] Less CPU use in hwaccel MJPEG decoding It skips the unnecessary unescaping the entropy-encoded data in case of using a hardware accelerator. --- libavcodec/mjpegdec.c | 19 ++++++++++++++----- libavcodec/mjpegdec.h | 1 + libavcodec/mxpegdec.c | 2 +- 3 files changed, 16 insertions(+), 6 deletions(-)