Message ID | AM7PR03MB666089E7D33FF266CFBE659E8F049@AM7PR03MB6660.eurprd03.prod.outlook.com |
---|---|
State | New |
Headers | show |
Series | [FFmpeg-devel,v2] avcodec/ffv1dec: Don't set ThreadFrame properties, fix race | expand |
Context | Check | Description |
---|---|---|
yinshiyou/make_loongarch64 | success | Make finished |
yinshiyou/make_fate_loongarch64 | success | Make fate finished |
andriy/make_aarch64_jetson | success | Make finished |
andriy/make_fate_aarch64_jetson | success | Make fate finished |
andriy/make_armv7_RPi4 | success | Make finished |
andriy/make_fate_armv7_RPi4 | success | Make fate finished |
On Thu, Mar 03, 2022 at 08:55:08AM +0100, Andreas Rheinhardt wrote: > Each FFV1 slice has its own SAR and picture structure encoded; > when a slice header was parsed, the relevant fields of a ThreadFrame's > AVFrame were directly set based upon the parsed values. This is > a data race in case slice threading is in use because of the concurrent > writes. In case of frame threading, it is also a data race because > the writes happen after ff_thread_finish_setup(), so that the reads > performed by ff_thread_ref_frame() are unsynchronized with the writes > performed when parsing the header. > > This commit fixes these issues by not writing to the ThreadFrame at all; > instead the raw data is read into the each SliceContext first; after > decoding the current frame and creating the actual output frame these > values are compared to each other. If they are valid and coincide, the > derived value is written directly to the output frame, not to the > ThreadFrame, thereby avoiding data races; in case they are not valid > or inconsistent the most commonly used valid value is used instead. > > This fixes most FFV1 FATE-tests completely when using slice threading; > the exceptions are fate-vsynth3-ffv1, vsynth3-ffv1-v3-yuv420p and > vsynth3-ffv1-v3-yuv422p10. (In these tests the partitioning into slices > does not honour chroma subsampling; as a result, chroma pixels at slice > borders get set by more than one thread without any synchronization.) > > Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com> > --- > libavcodec/ffv1.h | 4 ++ > libavcodec/ffv1dec.c | 119 +++++++++++++++++++++++++++++++++++-------- > 2 files changed, 103 insertions(+), 20 deletions(-) > > diff --git a/libavcodec/ffv1.h b/libavcodec/ffv1.h > index ac80fa85ce..f640d5a710 100644 > --- a/libavcodec/ffv1.h > +++ b/libavcodec/ffv1.h > @@ -91,6 +91,8 @@ typedef struct FFV1Context { > struct FFV1Context *fsrc; > > AVFrame *cur; > + int picture_structure; > + AVRational sample_aspect_ratio; > int plane_count; > int ac; ///< 1=range coder <-> 0=golomb rice > int ac_byte_count; ///< number of bytes used for AC coding > @@ -132,6 +134,8 @@ typedef struct FFV1Context { > int slice_coding_mode; > int slice_rct_by_coef; > int slice_rct_ry_coef; > + > + AVRational slice_sample_aspect_ratios[MAX_SLICES]; > } FFV1Context; > > int ff_ffv1_common_init(AVCodecContext *avctx); > diff --git a/libavcodec/ffv1dec.c b/libavcodec/ffv1dec.c > index 201630167d..bda8a5a2f1 100644 > --- a/libavcodec/ffv1dec.c > +++ b/libavcodec/ffv1dec.c > @@ -167,7 +167,7 @@ static int decode_slice_header(const FFV1Context *f, FFV1Context *fs) > { > RangeCoder *c = &fs->c; > uint8_t state[CONTEXT_SIZE]; > - unsigned ps, i, context_count; > + unsigned i, context_count; > memset(state, 128, sizeof(state)); > > av_assert0(f->version > 2); > @@ -205,25 +205,17 @@ static int decode_slice_header(const FFV1Context *f, FFV1Context *fs) > p->context_count = context_count; > } > > - ps = get_symbol(c, state, 0); > - if (ps == 1) { > - f->cur->interlaced_frame = 1; > - f->cur->top_field_first = 1; > - } else if (ps == 2) { > - f->cur->interlaced_frame = 1; > - f->cur->top_field_first = 0; > - } else if (ps == 3) { > - f->cur->interlaced_frame = 0; > - } > - f->cur->sample_aspect_ratio.num = get_symbol(c, state, 0); > - f->cur->sample_aspect_ratio.den = get_symbol(c, state, 0); > - > - if (av_image_check_sar(f->width, f->height, > - f->cur->sample_aspect_ratio) < 0) { > + fs->picture_structure = get_symbol(c, state, 0); > + fs->sample_aspect_ratio.num = get_symbol(c, state, 0); > + fs->sample_aspect_ratio.den = get_symbol(c, state, 0); > + /* Num or den being zero means unknown for FFV1; our unknown is 0 / 1. */ 0/0 is unknown in FFV1, 0/1 and 1/0 are treated as unknown because thats the only reasonable thing one really could do if they are encountered > + if (fs->sample_aspect_ratio.num == 0 || fs->sample_aspect_ratio.den == 0) { > + fs->sample_aspect_ratio = (AVRational) { 0, 1 }; > + } else if (av_image_check_sar(f->width, f->height, > + fs->sample_aspect_ratio) < 0) { > av_log(f->avctx, AV_LOG_WARNING, "ignoring invalid SAR: %u/%u\n", > - f->cur->sample_aspect_ratio.num, > - f->cur->sample_aspect_ratio.den); > - f->cur->sample_aspect_ratio = (AVRational){ 0, 1 }; > + fs->sample_aspect_ratio.num, fs->sample_aspect_ratio.den); > + fs->sample_aspect_ratio = (AVRational) { 0, 0 }; > } > > if (fs->version > 3) { > @@ -251,6 +243,9 @@ static int decode_slice(AVCodecContext *c, void *arg) > AVFrame * const p = f->cur; > int i, si; > > + fs->picture_structure = 0; > + fs->sample_aspect_ratio = (AVRational){ 0, 0 }; > + > for( si=0; fs != f->slice_context[si]; si ++) > ; > > @@ -831,6 +826,21 @@ static av_cold int decode_init(AVCodecContext *avctx) > return 0; > } > > +static int compare_sar(const void *a, const void *b) > +{ > + const AVRational x = *(const AVRational*)a; > + const AVRational y = *(const AVRational*)b; > + int cmp; > + > + /* 0/0 means invalid and 0/1 means unknown. > + * Order the ratios so that these values are at the end. */ > + if (x.num == 0 || y.num == 0) > + return y.num - x.num; > + cmp = av_cmp_q(x, y); > + /* For definiteness, order equivalent fractions by "lower terms last". */ > + return cmp ? cmp : y.num - x.num; > +} you can probably simplify this if you map the 32/32 fraction into a 64bit int also should be faster i also wonder a bit if it would make sense to split this "find the most common element" code out or if that makes no sense. > + > static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt) > { > uint8_t *buf = avpkt->data; > @@ -969,9 +979,78 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac > > if (f->last_picture.f) > ff_thread_release_ext_buffer(avctx, &f->last_picture); > - if ((ret = av_frame_ref(data, f->picture.f)) < 0) > + p = data; > + if ((ret = av_frame_ref(p, f->picture.f)) < 0) > return ret; > > + if (f->version > 2) { > + AVRational sar = f->slice_context[0]->sample_aspect_ratio; > + int pic_structures[4] = { 0 }, picture_structure; > + int inconsistent_sar = 0, has_sar = 0; > + > + for (int i = 0; i < f->slice_count; i++) { > + const FFV1Context *const fs = f->slice_context[i]; > + int slice_picture_structure = (unsigned)fs->picture_structure > 3 ? > + 0 : fs->picture_structure; > + > + pic_structures[slice_picture_structure]++; > + > + if (fs->sample_aspect_ratio.num != sar.num || > + fs->sample_aspect_ratio.den != sar.den) > + inconsistent_sar = 1; > + sar = fs->sample_aspect_ratio; > + if (sar.num == 0) > + inconsistent_sar = 1; > + else > + has_sar = 1; > + f->slice_sample_aspect_ratios[i] = sar; > + } > + if (has_sar) { > + if (inconsistent_sar) { > + AVRational last_sar = (AVRational){ 0, 0 }; > + qsort(f->slice_sample_aspect_ratios, f->slice_count, > + sizeof(f->slice_sample_aspect_ratios[0]), compare_sar); > + > + for (int i = 0, current_run = 0, longest_run = 0; > + i < f->slice_count; i++) { > + AVRational cur_sar = f->slice_sample_aspect_ratios[i]; > + > + if (cur_sar.num == 0) > + break; > + if (av_cmp_q(cur_sar, last_sar)) > + current_run = 1; > + else > + current_run++; > + if (current_run > longest_run) { > + longest_run = current_run; > + sar = cur_sar; > + } > + last_sar = cur_sar; > + } > + av_log(avctx, AV_LOG_WARNING, "SAR inconsistent, using %u/%u\n", > + sar.num, sar.den); > + } > + p->sample_aspect_ratio = sar; > + } if there are 10 slices and 9 say aspect unknown and one says 5000:3 i would belive the 9 more, this code looks a bit like it would always favor known over unknown [...] thx
diff --git a/libavcodec/ffv1.h b/libavcodec/ffv1.h index ac80fa85ce..f640d5a710 100644 --- a/libavcodec/ffv1.h +++ b/libavcodec/ffv1.h @@ -91,6 +91,8 @@ typedef struct FFV1Context { struct FFV1Context *fsrc; AVFrame *cur; + int picture_structure; + AVRational sample_aspect_ratio; int plane_count; int ac; ///< 1=range coder <-> 0=golomb rice int ac_byte_count; ///< number of bytes used for AC coding @@ -132,6 +134,8 @@ typedef struct FFV1Context { int slice_coding_mode; int slice_rct_by_coef; int slice_rct_ry_coef; + + AVRational slice_sample_aspect_ratios[MAX_SLICES]; } FFV1Context; int ff_ffv1_common_init(AVCodecContext *avctx); diff --git a/libavcodec/ffv1dec.c b/libavcodec/ffv1dec.c index 201630167d..bda8a5a2f1 100644 --- a/libavcodec/ffv1dec.c +++ b/libavcodec/ffv1dec.c @@ -167,7 +167,7 @@ static int decode_slice_header(const FFV1Context *f, FFV1Context *fs) { RangeCoder *c = &fs->c; uint8_t state[CONTEXT_SIZE]; - unsigned ps, i, context_count; + unsigned i, context_count; memset(state, 128, sizeof(state)); av_assert0(f->version > 2); @@ -205,25 +205,17 @@ static int decode_slice_header(const FFV1Context *f, FFV1Context *fs) p->context_count = context_count; } - ps = get_symbol(c, state, 0); - if (ps == 1) { - f->cur->interlaced_frame = 1; - f->cur->top_field_first = 1; - } else if (ps == 2) { - f->cur->interlaced_frame = 1; - f->cur->top_field_first = 0; - } else if (ps == 3) { - f->cur->interlaced_frame = 0; - } - f->cur->sample_aspect_ratio.num = get_symbol(c, state, 0); - f->cur->sample_aspect_ratio.den = get_symbol(c, state, 0); - - if (av_image_check_sar(f->width, f->height, - f->cur->sample_aspect_ratio) < 0) { + fs->picture_structure = get_symbol(c, state, 0); + fs->sample_aspect_ratio.num = get_symbol(c, state, 0); + fs->sample_aspect_ratio.den = get_symbol(c, state, 0); + /* Num or den being zero means unknown for FFV1; our unknown is 0 / 1. */ + if (fs->sample_aspect_ratio.num == 0 || fs->sample_aspect_ratio.den == 0) { + fs->sample_aspect_ratio = (AVRational) { 0, 1 }; + } else if (av_image_check_sar(f->width, f->height, + fs->sample_aspect_ratio) < 0) { av_log(f->avctx, AV_LOG_WARNING, "ignoring invalid SAR: %u/%u\n", - f->cur->sample_aspect_ratio.num, - f->cur->sample_aspect_ratio.den); - f->cur->sample_aspect_ratio = (AVRational){ 0, 1 }; + fs->sample_aspect_ratio.num, fs->sample_aspect_ratio.den); + fs->sample_aspect_ratio = (AVRational) { 0, 0 }; } if (fs->version > 3) { @@ -251,6 +243,9 @@ static int decode_slice(AVCodecContext *c, void *arg) AVFrame * const p = f->cur; int i, si; + fs->picture_structure = 0; + fs->sample_aspect_ratio = (AVRational){ 0, 0 }; + for( si=0; fs != f->slice_context[si]; si ++) ; @@ -831,6 +826,21 @@ static av_cold int decode_init(AVCodecContext *avctx) return 0; } +static int compare_sar(const void *a, const void *b) +{ + const AVRational x = *(const AVRational*)a; + const AVRational y = *(const AVRational*)b; + int cmp; + + /* 0/0 means invalid and 0/1 means unknown. + * Order the ratios so that these values are at the end. */ + if (x.num == 0 || y.num == 0) + return y.num - x.num; + cmp = av_cmp_q(x, y); + /* For definiteness, order equivalent fractions by "lower terms last". */ + return cmp ? cmp : y.num - x.num; +} + static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *avpkt) { uint8_t *buf = avpkt->data; @@ -969,9 +979,78 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac if (f->last_picture.f) ff_thread_release_ext_buffer(avctx, &f->last_picture); - if ((ret = av_frame_ref(data, f->picture.f)) < 0) + p = data; + if ((ret = av_frame_ref(p, f->picture.f)) < 0) return ret; + if (f->version > 2) { + AVRational sar = f->slice_context[0]->sample_aspect_ratio; + int pic_structures[4] = { 0 }, picture_structure; + int inconsistent_sar = 0, has_sar = 0; + + for (int i = 0; i < f->slice_count; i++) { + const FFV1Context *const fs = f->slice_context[i]; + int slice_picture_structure = (unsigned)fs->picture_structure > 3 ? + 0 : fs->picture_structure; + + pic_structures[slice_picture_structure]++; + + if (fs->sample_aspect_ratio.num != sar.num || + fs->sample_aspect_ratio.den != sar.den) + inconsistent_sar = 1; + sar = fs->sample_aspect_ratio; + if (sar.num == 0) + inconsistent_sar = 1; + else + has_sar = 1; + f->slice_sample_aspect_ratios[i] = sar; + } + if (has_sar) { + if (inconsistent_sar) { + AVRational last_sar = (AVRational){ 0, 0 }; + qsort(f->slice_sample_aspect_ratios, f->slice_count, + sizeof(f->slice_sample_aspect_ratios[0]), compare_sar); + + for (int i = 0, current_run = 0, longest_run = 0; + i < f->slice_count; i++) { + AVRational cur_sar = f->slice_sample_aspect_ratios[i]; + + if (cur_sar.num == 0) + break; + if (av_cmp_q(cur_sar, last_sar)) + current_run = 1; + else + current_run++; + if (current_run > longest_run) { + longest_run = current_run; + sar = cur_sar; + } + last_sar = cur_sar; + } + av_log(avctx, AV_LOG_WARNING, "SAR inconsistent, using %u/%u\n", + sar.num, sar.den); + } + p->sample_aspect_ratio = sar; + } + + for (int i = 0, count = 0; i < FF_ARRAY_ELEMS(pic_structures); i++) { + if (pic_structures[i] > count) { + picture_structure = i; + count = pic_structures[i]; + } + } + /* Set picture structure based upon the most commonly encountered + * picture structure. */ + if (picture_structure == 1) { + p->interlaced_frame = 1; + p->top_field_first = 1; + } else if (picture_structure == 2) { + p->interlaced_frame = 1; + p->top_field_first = 0; + } else if (picture_structure == 3) + p->interlaced_frame = 0; + } + *got_frame = 1; return buf_size;
Each FFV1 slice has its own SAR and picture structure encoded; when a slice header was parsed, the relevant fields of a ThreadFrame's AVFrame were directly set based upon the parsed values. This is a data race in case slice threading is in use because of the concurrent writes. In case of frame threading, it is also a data race because the writes happen after ff_thread_finish_setup(), so that the reads performed by ff_thread_ref_frame() are unsynchronized with the writes performed when parsing the header. This commit fixes these issues by not writing to the ThreadFrame at all; instead the raw data is read into the each SliceContext first; after decoding the current frame and creating the actual output frame these values are compared to each other. If they are valid and coincide, the derived value is written directly to the output frame, not to the ThreadFrame, thereby avoiding data races; in case they are not valid or inconsistent the most commonly used valid value is used instead. This fixes most FFV1 FATE-tests completely when using slice threading; the exceptions are fate-vsynth3-ffv1, vsynth3-ffv1-v3-yuv420p and vsynth3-ffv1-v3-yuv422p10. (In these tests the partitioning into slices does not honour chroma subsampling; as a result, chroma pixels at slice borders get set by more than one thread without any synchronization.) Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com> --- libavcodec/ffv1.h | 4 ++ libavcodec/ffv1dec.c | 119 +++++++++++++++++++++++++++++++++++-------- 2 files changed, 103 insertions(+), 20 deletions(-)