diff mbox series

[FFmpeg-devel,3/3,RFC] avcodec/ffv1: Better rounding for slice positions

Message ID 20231007001444.31861-3-michael@niedermayer.cc
State New
Headers show
Series [FFmpeg-devel,1/3] avcodec/ffv1enc: Slice combination is unsupported | expand

Commit Message

Michael Niedermayer Oct. 7, 2023, 12:14 a.m. UTC 
This fixes green lines in some odd dimensions with some slice configurations
like Ticket 5548

This also changes the encoder and whats encoded, and would require an
update to the specification. This change attempts to limit the change
to configurations that have missing lines currently.

Testing is welcome and needed

Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
---
 libavcodec/ffv1.c    | 38 ++++++++++++++++++++++++++++++++++----
 libavcodec/ffv1.h    |  2 ++
 libavcodec/ffv1dec.c |  8 ++++----
 libavcodec/ffv1enc.c |  7 ++++---
 4 files changed, 44 insertions(+), 11 deletions(-)

Comments

Jerome Martinez Oct. 7, 2023, 4:52 a.m. UTC | #1 
On 07/10/2023 02:14, Michael Niedermayer wrote:
> This fixes green lines in some odd dimensions with some slice configurations
> like Ticket 5548
>
> This also changes the encoder and whats encoded, and would require an
> update to the specification. This change attempts to limit the change
> to configurations that have missing lines currently.

It changes a lot the count of pixels per slice, and , e.g. with 4:2:2 and 4 slices per direction (16 slices in total), 13 pixel width:
before: 3/3/3/4 for luma, 2/2/2/2 for chroma (so 1 chroma too much)
after: 4/4/2/3 for luma, 2/2/1/2 for chroma

Wouldn't it easier for spec and maths to keep the previous behavior for luma and consider extra chroma as to be not encoded?
Something like 3/3/3/4 for luma, 2/2/2/1 for chroma
Actually maybe not really a change for spec in that case, more making a part more explicit while considering the patch as a bug fix rather than a spec change.

Or did I miss another issue? I'll check more a bit later.
Michael Niedermayer Oct. 7, 2023, 5:11 p.m. UTC | #2 
On Sat, Oct 07, 2023 at 06:52:07AM +0200, Jerome Martinez wrote:
> On 07/10/2023 02:14, Michael Niedermayer wrote:
> > This fixes green lines in some odd dimensions with some slice configurations
> > like Ticket 5548
> > 
> > This also changes the encoder and whats encoded, and would require an
> > update to the specification. This change attempts to limit the change
> > to configurations that have missing lines currently.
> 
> It changes a lot the count of pixels per slice, and , e.g. with 4:2:2 and 4 slices per direction (16 slices in total), 13 pixel width:
> before: 3/3/3/4 for luma, 2/2/2/2 for chroma (so 1 chroma too much)
> after: 4/4/2/3 for luma, 2/2/1/2 for chroma
> 
> Wouldn't it easier for spec and maths to keep the previous behavior for luma and consider extra chroma as to be not encoded?
> Something like 3/3/3/4 for luma, 2/2/2/1 for chroma
> Actually maybe not really a change for spec in that case, more making a part more explicit while considering the patch as a bug fix rather than a spec change.
> 

> Or did I miss another issue? I'll check more a bit later.

The existing thing has a few odd behaviors in corner cases
1. sometimes its not encoding the rightmost & bottommost chroma (leaving a green line)
2. sometimes its encoding chroma columns/rows in more than one slice (which is a waste of bits)

iam not sure if the chrome that gets encoded in each slice is entirely
expected, i guess it depends on what one expects

The new code tries to keep slices size a multiple of the chroma subsampling
factor. That way the left top coordinate of every 4:2:0 slice would always be
even for example and only the last ones could have right / bottom odd when  the
whole image has odd width/height

But this is a RFC for exactly the reasons that you mention, it changes things
and would become the default in future versions

Any future compression between luma and chroma planes may benefit from slices
being less odd in their chroma vs luma coordinates. But noone is working on
this AFAIK. Its just hypothetical

thx

[...]
diff mbox series

Patch

diff --git a/libavcodec/ffv1.c b/libavcodec/ffv1.c
index f44fead350b..e780c910aee 100644
--- a/libavcodec/ffv1.c
+++ b/libavcodec/ffv1.c
@@ -103,6 +103,36 @@  av_cold int ff_ffv1_init_slices_state(FFV1Context *f)
     return 0;
 }
 
+int ff_need_new_slices(int width, int num_h_slices, int chroma_shift) {
+    int mpw = 1<<chroma_shift;
+    int i = width * (int64_t)(num_h_slices - 1) / num_h_slices;
+
+    return width % mpw && (width - i) % mpw == 0;
+}
+
+int ff_slice_coord(const FFV1Context *f, int width, int sx, int num_h_slices, int chroma_shift) {
+    int mpw = 1<<chroma_shift;
+    int awidth = FFALIGN(width, mpw);
+    int old = 1;
+
+    if (f->version > 4) {
+        old = 0;
+    } else if (f->version == 3 && f->micro_version > 4 ||
+               f->version == 4 && f->micro_version > 2
+    ) {
+        // If the old system will not encode all pixels we will choose the new
+        old = !ff_need_new_slices(width, num_h_slices, chroma_shift);
+    }
+
+    if (old)
+        return width * sx / num_h_slices;
+
+    sx = (awidth * (int64_t)sx + (num_h_slices * mpw / 2)) / (num_h_slices * mpw) * mpw;
+    if (sx == awidth)
+        sx = width;
+    return sx;
+}
+
 av_cold int ff_ffv1_init_slice_contexts(FFV1Context *f)
 {
     int i, max_slice_count = f->num_h_slices * f->num_v_slices;
@@ -112,10 +142,10 @@  av_cold int ff_ffv1_init_slice_contexts(FFV1Context *f)
     for (i = 0; i < max_slice_count;) {
         int sx          = i % f->num_h_slices;
         int sy          = i / f->num_h_slices;
-        int sxs         = f->avctx->width  *  sx      / f->num_h_slices;
-        int sxe         = f->avctx->width  * (sx + 1) / f->num_h_slices;
-        int sys         = f->avctx->height *  sy      / f->num_v_slices;
-        int sye         = f->avctx->height * (sy + 1) / f->num_v_slices;
+        int sxs         = ff_slice_coord(f, f->avctx->width , sx    , f->num_h_slices, f->chroma_h_shift);
+        int sxe         = ff_slice_coord(f, f->avctx->width , sx + 1, f->num_h_slices, f->chroma_h_shift);
+        int sys         = ff_slice_coord(f, f->avctx->height, sy   ,  f->num_v_slices, f->chroma_v_shift);
+        int sye         = ff_slice_coord(f, f->avctx->height, sy + 1, f->num_v_slices, f->chroma_v_shift);
         FFV1Context *fs = av_mallocz(sizeof(*fs));
 
         if (!fs)
diff --git a/libavcodec/ffv1.h b/libavcodec/ffv1.h
index 14e40936065..3f658a26522 100644
--- a/libavcodec/ffv1.h
+++ b/libavcodec/ffv1.h
@@ -143,6 +143,8 @@  int ff_ffv1_init_slice_contexts(FFV1Context *f);
 int ff_ffv1_allocate_initial_states(FFV1Context *f);
 void ff_ffv1_clear_slice_state(const FFV1Context *f, FFV1Context *fs);
 int ff_ffv1_close(AVCodecContext *avctx);
+int ff_need_new_slices(int width, int num_h_slices, int chroma_shift);
+int ff_slice_coord(const FFV1Context *f, int width, int sx, int num_h_slices, int chroma_shift);
 
 static av_always_inline int fold(int diff, int bits)
 {
diff --git a/libavcodec/ffv1dec.c b/libavcodec/ffv1dec.c
index a376a4c8170..0486938eb2b 100644
--- a/libavcodec/ffv1dec.c
+++ b/libavcodec/ffv1dec.c
@@ -184,10 +184,10 @@  static int decode_slice_header(const FFV1Context *f, FFV1Context *fs)
     if (sx > f->num_h_slices - sw || sy > f->num_v_slices - sh)
         return AVERROR_INVALIDDATA;
 
-    fs->slice_x      =  sx       * (int64_t)f->width  / f->num_h_slices;
-    fs->slice_y      =  sy       * (int64_t)f->height / f->num_v_slices;
-    fs->slice_width  = (sx + sw) * (int64_t)f->width  / f->num_h_slices - fs->slice_x;
-    fs->slice_height = (sy + sh) * (int64_t)f->height / f->num_v_slices - fs->slice_y;
+    fs->slice_x      =  ff_slice_coord(f, f->width , sx     , f->num_h_slices, f->chroma_h_shift);
+    fs->slice_y      =  ff_slice_coord(f, f->height, sy     , f->num_v_slices, f->chroma_v_shift);
+    fs->slice_width  =  ff_slice_coord(f, f->width , sx + sw, f->num_h_slices, f->chroma_h_shift) - fs->slice_x;
+    fs->slice_height =  ff_slice_coord(f, f->height, sy + sh, f->num_v_slices, f->chroma_v_shift) - fs->slice_y;
 
     av_assert0((unsigned)fs->slice_width  <= f->width &&
                 (unsigned)fs->slice_height <= f->height);
diff --git a/libavcodec/ffv1enc.c b/libavcodec/ffv1enc.c
index 11be7fe32ea..e04695efa93 100644
--- a/libavcodec/ffv1enc.c
+++ b/libavcodec/ffv1enc.c
@@ -397,7 +397,8 @@  static int write_extradata(FFV1Context *f)
     int i, j, k;
     uint8_t state2[32][CONTEXT_SIZE];
     unsigned v;
-
+    int need_new_slices = ff_need_new_slices(f->avctx->width , f->num_h_slices, f->chroma_h_shift)
+                        ||ff_need_new_slices(f->avctx->height, f->num_v_slices, f->chroma_v_shift);
     memset(state2, 128, sizeof(state2));
     memset(state, 128, sizeof(state));
 
@@ -412,9 +413,9 @@  static int write_extradata(FFV1Context *f)
     put_symbol(c, state, f->version, 0);
     if (f->version > 2) {
         if (f->version == 3) {
-            f->micro_version = 4;
+            f->micro_version = 4 + need_new_slices;
         } else if (f->version == 4)
-            f->micro_version = 2;
+            f->micro_version = 2 + need_new_slices;
         put_symbol(c, state, f->micro_version, 0);
     }