questions about the Border section about ffv1 HOT 3 CLOSED

ping. Is more than one border column/row needed?

from ffv1.

1/ "The top right sample of the slice is a so with this definition I'd expect to see"

impossible: you can not compute a from a, as a is not yet known. so no copy of a on the left of the real a
FFmpeg code is:

        sample[1][-1] = sample[0][0];
        sample[0][w]  = sample[0][w - 1];

With:

• all 0 as init
• sample[0] is the previous line (full 0 if current line is the first one)
• sample[1] is the current line
• w is the width of the plane

--> It is done only once, for pixel 0, not for pixel "-1", so no a on the last line on the extreme left.
line 1 sample "-1" is line 0 sample "0"
line 0 sample "last +1" is line 0 sample "last"
line 1 sample "-2" is always 0 (I would have put a too but this is not what is coded)

I don't know how it should be in plain text, but the reference array is correct from my point of view.

Let's try sample to the left of a first coded sample of a line is identical to the first coded sample of a line that is above (if there is one)

In practice it would be:

| X | X |   | 0 | 0 | 0 |   | X |
| X | 0 |   | 0 | 0 | 0 |   | 0 |
|   |   |   |   |   |   |   |   |
| 0 | 0 |   | a | b | c |   | c |
| 0 | a |   | d |   | e |   | e |
| 0 | d |   | f | g | h |   | h |

with X impossible cases (never used)

2/ "Is more than one border column/row needed?"

for the moment I did not face a frame with the need of column "-2" and line "-2", but the FFmpeg code is:

    const int LT = last[-1];
    const int T  = last[0];
    const int RT = last[1];
    const int L  = src[-1];

    if (p->quant_table[3][127]) {
        const int TT = last2[0];
        const int LL = src[-2];
        return p->quant_table[0][(L - LT) & 0xFF] +
               p->quant_table[1][(LT - T) & 0xFF] +
               p->quant_table[2][(T - RT) & 0xFF] +
               p->quant_table[3][(LL - L) & 0xFF] +
               p->quant_table[4][(TT - T) & 0xFF];
    } else
        return p->quant_table[0][(L - LT) & 0xFF] +
               p->quant_table[1][(LT - T) & 0xFF] +
               p->quant_table[2][(T - RT) & 0xFF];

so if the 128th number of the 4th quant table is not 0, the sample 2 columns before the current column and the sample 2 lines before the current line are used, and for the first pixel of the first column you need 2 borders/rows.

With an array:

if (p->quant_table[3][127])
|   |   |   | 0 |   |
|   | 0 |   | 0 | 0 |
|   |   |   |   |   |
| 0 | 0 |   | a
else
|   |   |   |   |   |
|   | 0 |   | 0 | 0 |
|   |   |   |   |   |
|   | 0 |   | a

with 0 values used for computing the context.

So yes, we need more than one border column/row needed (actually 2)?

I think the issue is that the context computing is not yet (well) defined, we should defined it

from ffv1.

I didn’t answer, because I can see little pros/cons in both scenari, but so far I’m unable to decide.

from ffv1.