avcodec/mpegpicture: Use ThreadProgress instead of ThreadFrame API

Given that MPVPictures are already directly shared between threads
in case of frame-threaded decoding, one can simply use it to
pass decoding progress information between threads. This allows
to avoid one level of indirection; it also means avoids allocations
(of the ThreadFrameProgress structure) in case of frame-threading
and indeed makes ff_thread_release_ext_buffer() decoder-only
(actually, H.264-decoder-only).

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit is contained in:
Andreas Rheinhardt 2023-10-16 01:21:55 +02:00
parent 5475000942
commit 45cf0541cf
11 changed files with 59 additions and 47 deletions

View File

@ -34,6 +34,7 @@
#include "mpegutils.h"
#include "mpegvideo.h"
#include "threadframe.h"
#include "threadprogress.h"
/**
* @param stride the number of MVs to get to the next row
@ -409,8 +410,12 @@ static void guess_mv(ERContext *s)
set_mv_strides(s, &mot_step, &mot_stride);
num_avail = 0;
if (s->last_pic.motion_val[0])
ff_thread_await_progress(s->last_pic.tf, mb_height-1, 0);
if (s->last_pic.motion_val[0]) {
if (s->last_pic.tf)
ff_thread_await_progress(s->last_pic.tf, mb_height-1, 0);
else
ff_thread_progress_await(s->last_pic.progress, mb_height - 1);
}
for (i = 0; i < mb_width * mb_height; i++) {
const int mb_xy = s->mb_index2xy[i];
int f = 0;
@ -763,7 +768,7 @@ static int is_intra_more_likely(ERContext *s)
if (s->avctx->codec_id == AV_CODEC_ID_H264) {
// FIXME
} else {
ff_thread_await_progress(s->last_pic.tf, mb_y, 0);
ff_thread_progress_await(s->last_pic.progress, mb_y);
}
is_intra_likely += s->sad(NULL, last_mb_ptr, mb_ptr,
linesize[0], 16);
@ -1198,7 +1203,7 @@ void ff_er_frame_end(ERContext *s, int *decode_error_flags)
int time_pb = s->pb_time;
av_assert0(s->avctx->codec_id != AV_CODEC_ID_H264);
ff_thread_await_progress(s->next_pic.tf, mb_y, 0);
ff_thread_progress_await(s->next_pic.progress, mb_y);
s->mv[0][0][0] = s->next_pic.motion_val[0][xy][0] * time_pb / time_pp;
s->mv[0][0][1] = s->next_pic.motion_val[0][xy][1] * time_pb / time_pp;

View File

@ -40,6 +40,7 @@
typedef struct ERPicture {
AVFrame *f;
const struct ThreadFrame *tf;
const struct ThreadProgress *progress;
// it is the caller's responsibility to allocate these buffers
int16_t (*motion_val[2])[2];

View File

@ -45,7 +45,7 @@
#include "internal.h"
#include "profiles.h"
#include "qpeldsp.h"
#include "threadframe.h"
#include "threadprogress.h"
#include "xvididct.h"
#include "unary.h"
@ -1813,7 +1813,7 @@ static int mpeg4_decode_mb(MpegEncContext *s, int16_t block[6][64])
s->last_mv[i][1][1] = 0;
}
ff_thread_await_progress(&s->next_pic.ptr->tf, s->mb_y, 0);
ff_thread_progress_await(&s->next_pic.ptr->progress, s->mb_y);
}
/* if we skipped it in the future P-frame than skip it now too */
@ -2018,10 +2018,10 @@ end:
if (s->pict_type == AV_PICTURE_TYPE_B) {
const int delta = s->mb_x + 1 == s->mb_width ? 2 : 1;
ff_thread_await_progress(&s->next_pic.ptr->tf,
ff_thread_progress_await(&s->next_pic.ptr->progress,
(s->mb_x + delta >= s->mb_width)
? FFMIN(s->mb_y + 1, s->mb_height - 1)
: s->mb_y, 0);
: s->mb_y);
if (s->next_pic.mbskip_table[xy + delta])
return SLICE_OK;
}

View File

@ -34,7 +34,7 @@ static void set_erpic(ERPicture *dst, const MPVPicture *src)
}
dst->f = src->f;
dst->tf = &src->tf;
dst->progress = &src->progress;
for (i = 0; i < 2; i++) {
dst->motion_val[i] = src->motion_val[i];

View File

@ -28,13 +28,13 @@
#include "motion_est.h"
#include "mpegpicture.h"
#include "refstruct.h"
#include "threadframe.h"
static void mpv_pic_reset(FFRefStructOpaque unused, void *obj)
{
MPVPicture *pic = obj;
ff_thread_release_ext_buffer(&pic->tf);
av_frame_unref(pic->f);
ff_thread_progress_reset(&pic->progress);
ff_refstruct_unref(&pic->hwaccel_picture_private);
@ -65,14 +65,18 @@ static void mpv_pic_reset(FFRefStructOpaque unused, void *obj)
pic->coded_picture_number = 0;
}
static int av_cold mpv_pic_init(FFRefStructOpaque unused, void *obj)
static int av_cold mpv_pic_init(FFRefStructOpaque opaque, void *obj)
{
MPVPicture *pic = obj;
int ret, init_progress = (uintptr_t)opaque.nc;
ret = ff_thread_progress_init(&pic->progress, init_progress);
if (ret < 0)
return ret;
pic->f = av_frame_alloc();
if (!pic->f)
return AVERROR(ENOMEM);
pic->tf.f = pic->f;
return 0;
}
@ -80,12 +84,15 @@ static void av_cold mpv_pic_free(FFRefStructOpaque unused, void *obj)
{
MPVPicture *pic = obj;
ff_thread_progress_destroy(&pic->progress);
av_frame_free(&pic->f);
}
av_cold FFRefStructPool *ff_mpv_alloc_pic_pool(void)
av_cold FFRefStructPool *ff_mpv_alloc_pic_pool(int init_progress)
{
return ff_refstruct_pool_alloc_ext(sizeof(MPVPicture), 0, NULL,
return ff_refstruct_pool_alloc_ext(sizeof(MPVPicture),
FF_REFSTRUCT_POOL_FLAG_FREE_ON_INIT_ERROR,
(void*)(uintptr_t)init_progress,
mpv_pic_init, mpv_pic_reset, mpv_pic_free, NULL);
}

View File

@ -26,7 +26,7 @@
#include "avcodec.h"
#include "motion_est.h"
#include "threadframe.h"
#include "threadprogress.h"
#define MPV_MAX_PLANES 3
#define EDGE_WIDTH 16
@ -55,7 +55,6 @@ typedef struct BufferPoolContext {
*/
typedef struct MPVPicture {
struct AVFrame *f;
ThreadFrame tf;
int8_t *qscale_table_base;
int8_t *qscale_table;
@ -87,6 +86,8 @@ typedef struct MPVPicture {
int display_picture_number;
int coded_picture_number;
ThreadProgress progress;
} MPVPicture;
typedef struct MPVWorkPicture {
@ -111,7 +112,7 @@ typedef struct MPVWorkPicture {
/**
* Allocate a pool of MPVPictures.
*/
struct FFRefStructPool *ff_mpv_alloc_pic_pool(void);
struct FFRefStructPool *ff_mpv_alloc_pic_pool(int init_progress);
/**
* Allocate an MPVPicture's accessories (but not the AVFrame's buffer itself)

View File

@ -40,11 +40,13 @@
#include "mpeg4videodec.h"
#include "refstruct.h"
#include "thread.h"
#include "threadframe.h"
#include "threadprogress.h"
#include "wmv2dec.h"
int ff_mpv_decode_init(MpegEncContext *s, AVCodecContext *avctx)
{
enum ThreadingStatus thread_status;
ff_mpv_common_defaults(s);
s->avctx = avctx;
@ -59,9 +61,12 @@ int ff_mpv_decode_init(MpegEncContext *s, AVCodecContext *avctx)
ff_mpv_idct_init(s);
ff_h264chroma_init(&s->h264chroma, 8); //for lowres
if (!s->picture_pool && // VC-1 can call this multiple times
ff_thread_sync_ref(avctx, offsetof(MpegEncContext, picture_pool))) {
s->picture_pool = ff_mpv_alloc_pic_pool();
if (s->picture_pool) // VC-1 can call this multiple times
return 0;
thread_status = ff_thread_sync_ref(avctx, offsetof(MpegEncContext, picture_pool));
if (thread_status != FF_THREAD_IS_COPY) {
s->picture_pool = ff_mpv_alloc_pic_pool(thread_status != FF_THREAD_NO_FRAME_THREADING);
if (!s->picture_pool)
return AVERROR(ENOMEM);
}
@ -229,7 +234,6 @@ static int alloc_picture(MpegEncContext *s, MPVWorkPicture *dst, int reference)
dst->ptr = pic;
pic->tf.f = pic->f;
pic->reference = reference;
/* WM Image / Screen codecs allocate internal buffers with different
@ -237,8 +241,8 @@ static int alloc_picture(MpegEncContext *s, MPVWorkPicture *dst, int reference)
if (avctx->codec_id != AV_CODEC_ID_WMV3IMAGE &&
avctx->codec_id != AV_CODEC_ID_VC1IMAGE &&
avctx->codec_id != AV_CODEC_ID_MSS2) {
ret = ff_thread_get_ext_buffer(avctx, &pic->tf,
reference ? AV_GET_BUFFER_FLAG_REF : 0);
ret = ff_thread_get_buffer(avctx, pic->f,
reference ? AV_GET_BUFFER_FLAG_REF : 0);
} else {
pic->f->width = avctx->width;
pic->f->height = avctx->height;
@ -281,8 +285,7 @@ static int av_cold alloc_dummy_frame(MpegEncContext *s, MPVWorkPicture *dst)
pic = dst->ptr;
pic->dummy = 1;
ff_thread_report_progress(&pic->tf, INT_MAX, 0);
ff_thread_report_progress(&pic->tf, INT_MAX, 1);
ff_thread_progress_report(&pic->progress, INT_MAX);
return 0;
}
@ -418,7 +421,7 @@ void ff_mpv_frame_end(MpegEncContext *s)
emms_c();
if (s->cur_pic.reference)
ff_thread_report_progress(&s->cur_pic.ptr->tf, INT_MAX, 0);
ff_thread_progress_report(&s->cur_pic.ptr->progress, INT_MAX);
}
void ff_print_debug_info(const MpegEncContext *s, const MPVPicture *p, AVFrame *pict)
@ -484,7 +487,7 @@ void ff_mpeg_flush(AVCodecContext *avctx)
void ff_mpv_report_decode_progress(MpegEncContext *s)
{
if (s->pict_type != AV_PICTURE_TYPE_B && !s->partitioned_frame && !s->er.error_occurred)
ff_thread_report_progress(&s->cur_pic.ptr->tf, s->mb_y, 0);
ff_thread_progress_report(&s->cur_pic.ptr->progress, s->mb_y);
}

View File

@ -823,7 +823,7 @@ av_cold int ff_mpv_encode_init(AVCodecContext *avctx)
!FF_ALLOCZ_TYPED_ARRAY(s->input_picture, MAX_B_FRAMES + 1) ||
!FF_ALLOCZ_TYPED_ARRAY(s->reordered_input_picture, MAX_B_FRAMES + 1) ||
!(s->new_pic = av_frame_alloc()) ||
!(s->picture_pool = ff_mpv_alloc_pic_pool()))
!(s->picture_pool = ff_mpv_alloc_pic_pool(0)))
return AVERROR(ENOMEM);
/* Allocate MV tables; the MV and MB tables will be copied

View File

@ -124,12 +124,12 @@ void mpv_reconstruct_mb_internal(MpegEncContext *s, int16_t block[12][64],
if (HAVE_THREADS && is_mpeg12 != DEFINITELY_MPEG12 &&
s->avctx->active_thread_type & FF_THREAD_FRAME) {
if (s->mv_dir & MV_DIR_FORWARD) {
ff_thread_await_progress(&s->last_pic.ptr->tf,
lowest_referenced_row(s, 0), 0);
ff_thread_progress_await(&s->last_pic.ptr->progress,
lowest_referenced_row(s, 0));
}
if (s->mv_dir & MV_DIR_BACKWARD) {
ff_thread_await_progress(&s->next_pic.ptr->tf,
lowest_referenced_row(s, 1), 0);
ff_thread_progress_await(&s->next_pic.ptr->progress,
lowest_referenced_row(s, 1));
}
}

View File

@ -996,11 +996,6 @@ int ff_thread_get_ext_buffer(AVCodecContext *avctx, ThreadFrame *f, int flags)
int ret;
f->owner[0] = f->owner[1] = avctx;
/* Hint: It is possible for this function to be called with codecs
* that don't support frame threading at all, namely in case
* a frame-threaded decoder shares code with codecs that are not.
* This currently affects non-MPEG-4 mpegvideo codecs.
* The following check will always be true for them. */
if (!(avctx->active_thread_type & FF_THREAD_FRAME))
return ff_get_buffer(avctx, f->f, flags);

View File

@ -43,7 +43,7 @@
#include "qpeldsp.h"
#include "rectangle.h"
#include "thread.h"
#include "threadframe.h"
#include "threadprogress.h"
#include "rv34vlc.h"
#include "rv34data.h"
@ -721,8 +721,8 @@ static inline void rv34_mc(RV34DecContext *r, const int block_type,
if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME)) {
/* wait for the referenced mb row to be finished */
int mb_row = s->mb_y + ((yoff + my + 5 + 8 * height) >> 4);
const ThreadFrame *f = dir ? &s->next_pic.ptr->tf : &s->last_pic.ptr->tf;
ff_thread_await_progress(f, mb_row, 0);
const ThreadProgress *p = dir ? &s->next_pic.ptr->progress : &s->last_pic.ptr->progress;
ff_thread_progress_await(p, mb_row);
}
dxy = ly*4 + lx;
@ -901,7 +901,7 @@ static int rv34_decode_mv(RV34DecContext *r, int block_type)
//surprisingly, it uses motion scheme from next reference frame
/* wait for the current mb row to be finished */
if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
ff_thread_await_progress(&s->next_pic.ptr->tf, FFMAX(0, s->mb_y-1), 0);
ff_thread_progress_await(&s->next_pic.ptr->progress, FFMAX(0, s->mb_y-1));
next_bt = s->next_pic.mb_type[s->mb_x + s->mb_y * s->mb_stride];
if(IS_INTRA(next_bt) || IS_SKIP(next_bt)){
@ -1485,8 +1485,8 @@ static int rv34_decode_slice(RV34DecContext *r, int end, const uint8_t* buf, int
r->loop_filter(r, s->mb_y - 2);
if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
ff_thread_report_progress(&s->cur_pic.ptr->tf,
s->mb_y - 2, 0);
ff_thread_progress_report(&s->cur_pic.ptr->progress,
s->mb_y - 2);
}
if(s->mb_x == s->resync_mb_x)
@ -1584,7 +1584,7 @@ static int finish_frame(AVCodecContext *avctx, AVFrame *pict)
s->mb_num_left = 0;
if (HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
ff_thread_report_progress(&s->cur_pic.ptr->tf, INT_MAX, 0);
ff_thread_progress_report(&s->cur_pic.ptr->progress, INT_MAX);
if (s->pict_type == AV_PICTURE_TYPE_B) {
if ((ret = av_frame_ref(pict, s->cur_pic.ptr->f)) < 0)
@ -1812,7 +1812,7 @@ int ff_rv34_decode_frame(AVCodecContext *avctx, AVFrame *pict,
ff_er_frame_end(&s->er, NULL);
ff_mpv_frame_end(s);
s->mb_num_left = 0;
ff_thread_report_progress(&s->cur_pic.ptr->tf, INT_MAX, 0);
ff_thread_progress_report(&s->cur_pic.ptr->progress, INT_MAX);
return AVERROR_INVALIDDATA;
}
}