177 |
twopass_stat_t * stats; |
twopass_stat_t * stats; |
178 |
|
|
179 |
/*---------------------------------- |
/*---------------------------------- |
180 |
* Histerysis helpers |
* Hysteresis helpers |
181 |
*--------------------------------*/ |
*--------------------------------*/ |
182 |
|
|
183 |
/* This field holds the int2float conversion errors of each quant per |
/* This field holds the int2float conversion errors of each quant per |
270 |
static void first_pass_stats_prepare_data(rc_2pass2_t * rc); |
static void first_pass_stats_prepare_data(rc_2pass2_t * rc); |
271 |
static void first_pass_scale_curve_internal(rc_2pass2_t *rc); |
static void first_pass_scale_curve_internal(rc_2pass2_t *rc); |
272 |
static void scaled_curve_apply_advanced_parameters(rc_2pass2_t * rc); |
static void scaled_curve_apply_advanced_parameters(rc_2pass2_t * rc); |
273 |
|
static int check_curve_for_vbv_compliancy(rc_2pass2_t * rc, const float fps); |
274 |
|
static int scale_curve_for_vbv_compliancy(rc_2pass2_t * rc, const float fps); |
275 |
#if 0 |
#if 0 |
276 |
static void stats_print(rc_2pass2_t * rc); |
static void stats_print(rc_2pass2_t * rc); |
277 |
#endif |
#endif |
428 |
* shape the curve in the BEFORE/AFTER pair of functions */ |
* shape the curve in the BEFORE/AFTER pair of functions */ |
429 |
scaled_curve_apply_advanced_parameters(rc); |
scaled_curve_apply_advanced_parameters(rc); |
430 |
|
|
431 |
|
/* Check curve for VBV compliancy and rescale if necessary */ |
432 |
|
|
433 |
|
#ifdef VBV_FORCE |
434 |
|
if (rc->param.vbv_size==0) |
435 |
|
{ |
436 |
|
rc->param.vbv_size = 3145728; |
437 |
|
rc->param.vbv_initial = 2359296; |
438 |
|
rc->param.vbv_maxrate = 4000000; |
439 |
|
rc->param.vbv_peakrate = 10000000; |
440 |
|
} |
441 |
|
#endif |
442 |
|
|
443 |
|
if (rc->param.vbv_size>0) /* vbv_size==0 switches VBV check off */ |
444 |
|
{ |
445 |
|
const double fps = (double)create->fbase/(double)create->fincr; |
446 |
|
int status = check_curve_for_vbv_compliancy(rc, fps); |
447 |
|
#ifdef VBV_DEBUG |
448 |
|
if (status) |
449 |
|
fprintf(stderr,"underflow detected\n Scaling Curve for compliancy... "); |
450 |
|
#endif |
451 |
|
|
452 |
|
status = scale_curve_for_vbv_compliancy(rc, fps); |
453 |
|
|
454 |
|
#ifdef VBV_DEBUG |
455 |
|
if (status==0) |
456 |
|
fprintf(stderr,"done.\n"); |
457 |
|
else |
458 |
|
fprintf(stderr,"impossible.\n"); |
459 |
|
#endif |
460 |
|
} |
461 |
*handle = rc; |
*handle = rc; |
462 |
return(0); |
return(0); |
463 |
} |
} |
1413 |
} |
} |
1414 |
|
|
1415 |
/***************************************************************************** |
/***************************************************************************** |
1416 |
|
* VBV compliancy check and scale |
1417 |
|
* MPEG-4 standard specifies certain restrictions for bitrate/framesize in VBR |
1418 |
|
* to enable playback on devices with limited readspeed and memory (and which |
1419 |
|
* aren't...) |
1420 |
|
* |
1421 |
|
* DivX profiles have 2 criteria: VBV as in MPEG standard |
1422 |
|
* a limit on peak bitrate for any 3 seconds |
1423 |
|
* |
1424 |
|
* But if VBV is fulfilled, peakrate is automatically fulfilled in any profile |
1425 |
|
* define so far, so we check for it (for completeness) but correct only VBV |
1426 |
|
* |
1427 |
|
*****************************************************************************/ |
1428 |
|
|
1429 |
|
#define VBV_COMPLIANT 0 |
1430 |
|
#define VBV_UNDERFLOW 1 /* video buffer runs empty */ |
1431 |
|
#define VBV_OVERFLOW 2 /* doesn't exist for VBR encoding */ |
1432 |
|
#define VBV_PEAKRATE 4 /* peak bitrate (within 3s) violated */ |
1433 |
|
|
1434 |
|
static int check_curve_for_vbv_compliancy(rc_2pass2_t * rc, const float fps) |
1435 |
|
{ |
1436 |
|
/* We do all calculations in float, for higher accuracy, |
1437 |
|
and in bytes for convenience |
1438 |
|
|
1439 |
|
typical values from DivX Home Theater profile: |
1440 |
|
vbv_size= 384*1024 (384kB), vbv_initial= 288*1024 (75% fill) |
1441 |
|
maxrate= 4000000 (4MBps), peakrate= 10000000 (10MBps) |
1442 |
|
|
1443 |
|
PAL: offset3s = 75 (3 seconds of 25fps) |
1444 |
|
NTSC: offset3s = 90 (3 seconds of 29.97fps) or 72 (3 seconds of 23.976fps) |
1445 |
|
*/ |
1446 |
|
|
1447 |
|
const float vbv_size = (float)rc->param.vbv_size/8.f; |
1448 |
|
float vbvfill = (float)rc->param.vbv_initial/8.f; |
1449 |
|
|
1450 |
|
const float maxrate = (float)rc->param.vbv_maxrate; |
1451 |
|
const float peakrate = (float)rc->param.vbv_peakrate; |
1452 |
|
const float r0 = (int)(maxrate/fps+0.5)/8.f; |
1453 |
|
|
1454 |
|
int bytes3s = 0; |
1455 |
|
int offset3s = (int)(3.f*fps+0.5); |
1456 |
|
|
1457 |
|
int i; |
1458 |
|
for (i=0; i<rc->num_frames; i++) { |
1459 |
|
/* DivX 3s peak bitrate check */ |
1460 |
|
|
1461 |
|
bytes3s += rc->stats[i].scaled_length; |
1462 |
|
if (i>=offset3s) |
1463 |
|
bytes3s -= rc->stats[i-offset3s].scaled_length; |
1464 |
|
|
1465 |
|
if (8.f*bytes3s > 3*peakrate) |
1466 |
|
return VBV_PEAKRATE; |
1467 |
|
|
1468 |
|
/* update vbv fill level */ |
1469 |
|
|
1470 |
|
vbvfill += r0 - rc->stats[i].scaled_length; |
1471 |
|
|
1472 |
|
/* this check is _NOT_ an "overflow"! only reading from disk stops then */ |
1473 |
|
if (vbvfill > vbv_size) |
1474 |
|
vbvfill = vbv_size; |
1475 |
|
|
1476 |
|
/* but THIS would be an underflow. report it! */ |
1477 |
|
if (vbvfill < 0) |
1478 |
|
return VBV_UNDERFLOW; |
1479 |
|
} |
1480 |
|
|
1481 |
|
return VBV_COMPLIANT; |
1482 |
|
} |
1483 |
|
/* TODO: store min(vbvfill) and print "minimum buffer fill" */ |
1484 |
|
|
1485 |
|
|
1486 |
|
static int scale_curve_for_vbv_compliancy(rc_2pass2_t * rc, const float fps) |
1487 |
|
{ |
1488 |
|
/* correct any VBV violations. Peak bitrate violations disappears |
1489 |
|
by this automatically |
1490 |
|
|
1491 |
|
This implementation follows |
1492 |
|
|
1493 |
|
Westerink, Rajagopalan, Gonzales "Two-pass MPEG-2 variable-bitrate encoding" |
1494 |
|
IBM J. RES. DEVELOP. VOL 43, No. 4, July 1999, p.471--488 |
1495 |
|
|
1496 |
|
Thanks, guys! This paper rocks!!! |
1497 |
|
*/ |
1498 |
|
|
1499 |
|
/* |
1500 |
|
For each scene of len N, we have to check up to N^2 possible buffer fills. |
1501 |
|
This works well with MPEG-2 where N==12 or so, but for MPEG-4 it's a |
1502 |
|
little slow... |
1503 |
|
|
1504 |
|
TODO: Better control on VBVfill between scenes |
1505 |
|
*/ |
1506 |
|
|
1507 |
|
const float vbv_size = (float)rc->param.vbv_size/8.f; |
1508 |
|
const float vbv_initial = (float)rc->param.vbv_initial/8.f; |
1509 |
|
|
1510 |
|
const float maxrate = 0.9*rc->param.vbv_maxrate; |
1511 |
|
const float vbv_low = 0.10f*vbv_size; |
1512 |
|
const float r0 = (int)(maxrate/fps+0.5)/8.f; |
1513 |
|
|
1514 |
|
int i,k,l,n,violation = 0; |
1515 |
|
float *scenefactor; |
1516 |
|
int *scenestart; |
1517 |
|
int *scenelength; |
1518 |
|
|
1519 |
|
/* first step: determine how many "scenes" there are and store their boundaries |
1520 |
|
we could get all this from existing keyframe_positions, somehow, but there we |
1521 |
|
don't have a min_scenelength, and it's no big deal to get it again. */ |
1522 |
|
|
1523 |
|
const int min_scenelength = (int)(fps+0.5); |
1524 |
|
int num_scenes = 0; |
1525 |
|
int last_scene = -999; |
1526 |
|
for (i=0; i<rc->num_frames; i++) { |
1527 |
|
if ( (rc->stats[i].type == XVID_TYPE_IVOP) && (i-last_scene>min_scenelength) ) |
1528 |
|
{ |
1529 |
|
last_scene = i; |
1530 |
|
num_scenes++; |
1531 |
|
} |
1532 |
|
} |
1533 |
|
|
1534 |
|
scenefactor = (float*)malloc( num_scenes*sizeof(float) ); |
1535 |
|
scenestart = (int*)malloc( num_scenes*sizeof(int) ); |
1536 |
|
scenelength = (int*)malloc( num_scenes*sizeof(int) ); |
1537 |
|
|
1538 |
|
if ((!scenefactor) || (!scenestart) || (!scenelength) ) |
1539 |
|
{ |
1540 |
|
free(scenefactor); |
1541 |
|
free(scenestart); |
1542 |
|
free(scenelength); |
1543 |
|
/* remember: free(0) is valid and does exactly nothing. */ |
1544 |
|
return -1; |
1545 |
|
} |
1546 |
|
|
1547 |
|
/* count again and safe the length/position */ |
1548 |
|
|
1549 |
|
num_scenes = 0; |
1550 |
|
last_scene = -999; |
1551 |
|
for (i=0; i<rc->num_frames; i++) { |
1552 |
|
if ( (rc->stats[i].type == XVID_TYPE_IVOP) && (i-last_scene>min_scenelength) ) |
1553 |
|
{ |
1554 |
|
if (num_scenes>0) |
1555 |
|
scenelength[num_scenes-1]=i-last_scene; |
1556 |
|
scenestart[num_scenes]=i; |
1557 |
|
num_scenes++; |
1558 |
|
last_scene = i; |
1559 |
|
} |
1560 |
|
} |
1561 |
|
scenelength[num_scenes-1]=i-last_scene; |
1562 |
|
|
1563 |
|
/* second step: check for each scene, how much we can scale its frames up or down |
1564 |
|
such that the VBV restriction is just fulfilled |
1565 |
|
*/ |
1566 |
|
|
1567 |
|
|
1568 |
|
#define R(k,n) (((n)+1-(k))*r0) /* how much enters the buffer between frame k and n */ |
1569 |
|
for (l=0; l<num_scenes;l++) |
1570 |
|
{ |
1571 |
|
const int start = scenestart[l]; |
1572 |
|
const int length = scenelength[l]; |
1573 |
|
twopass_stat_t * frames = &rc->stats[start]; |
1574 |
|
|
1575 |
|
float S0n,Skn; |
1576 |
|
float f,minf = 99999.f; |
1577 |
|
|
1578 |
|
S0n=0.; |
1579 |
|
for (n=0;n<=length-1;n++) |
1580 |
|
{ |
1581 |
|
S0n += frames[n].scaled_length; |
1582 |
|
|
1583 |
|
k=0; |
1584 |
|
Skn = S0n; |
1585 |
|
f = (R(k,n-1) + (vbv_initial - vbv_low)) / Skn; |
1586 |
|
if (f < minf) |
1587 |
|
minf = f; |
1588 |
|
|
1589 |
|
for (k=1;k<=n;k++) |
1590 |
|
{ |
1591 |
|
Skn -= frames[k].scaled_length; |
1592 |
|
|
1593 |
|
f = (R(k,n-1) + (vbv_size - vbv_low)) / Skn; |
1594 |
|
if (f < minf) |
1595 |
|
minf = f; |
1596 |
|
} |
1597 |
|
} |
1598 |
|
|
1599 |
|
/* special case: at the end, fill buffer up to vbv_initial again |
1600 |
|
TODO: Allow other values for buffer fill between scenes |
1601 |
|
e.g. if n=N is smallest f-value, then check for better value */ |
1602 |
|
|
1603 |
|
n=length; |
1604 |
|
k=0; |
1605 |
|
Skn = S0n; |
1606 |
|
f = R(k,n-1)/Skn; |
1607 |
|
if (f < minf) |
1608 |
|
minf = f; |
1609 |
|
|
1610 |
|
for (k=1;k<=n-1;k++) |
1611 |
|
{ |
1612 |
|
Skn -= frames[k].scaled_length; |
1613 |
|
|
1614 |
|
f = (R(k,n-1) + (vbv_initial - vbv_low)) / Skn; |
1615 |
|
if (f < minf) |
1616 |
|
minf = f; |
1617 |
|
} |
1618 |
|
|
1619 |
|
#ifdef VBV_DEBUG |
1620 |
|
printf("Scene %d (Frames %d-%d): VBVfactor %f\n", l, start, start+length-1 , minf); |
1621 |
|
#endif |
1622 |
|
|
1623 |
|
scenefactor[l] = minf; |
1624 |
|
} |
1625 |
|
#undef R |
1626 |
|
|
1627 |
|
/* last step: now we know of any scene how much it can be scaled up or down without |
1628 |
|
violating VBV. Next, distribute bits from the evil scenes to the good ones */ |
1629 |
|
|
1630 |
|
do |
1631 |
|
{ |
1632 |
|
float S_red = 0.f; /* how much to redistribute */ |
1633 |
|
float S_elig = 0.f; /* sum of bit for those scenes you can still swallow something*/ |
1634 |
|
int l; |
1635 |
|
|
1636 |
|
for (l=0;l<num_scenes;l++) /* check how much is wrong */ |
1637 |
|
{ |
1638 |
|
const int start = scenestart[l]; |
1639 |
|
const int length = scenelength[l]; |
1640 |
|
twopass_stat_t * frames = &rc->stats[start]; |
1641 |
|
|
1642 |
|
if (scenefactor[l] == 1.) /* exactly 1 means "don't touch this anymore!" */ |
1643 |
|
continue; |
1644 |
|
|
1645 |
|
if (scenefactor[l] > 1.) /* within limits */ |
1646 |
|
{ |
1647 |
|
for (n= 0; n < length; n++) |
1648 |
|
S_elig += frames[n].scaled_length; |
1649 |
|
} |
1650 |
|
else /* underflowing segment */ |
1651 |
|
{ |
1652 |
|
for (n= 0; n < length; n++) |
1653 |
|
{ |
1654 |
|
float newbytes = (float)frames[n].scaled_length * scenefactor[l]; |
1655 |
|
S_red += (float)frames[n].scaled_length - (float)newbytes; |
1656 |
|
frames[n].scaled_length =(int)newbytes; |
1657 |
|
} |
1658 |
|
scenefactor[l] = 1.f; |
1659 |
|
} |
1660 |
|
} |
1661 |
|
|
1662 |
|
if (S_red < 1.f) /* no more underflows */ |
1663 |
|
break; |
1664 |
|
|
1665 |
|
if (S_elig < 1.f) |
1666 |
|
{ |
1667 |
|
#ifdef VBV_DEBUG |
1668 |
|
fprintf(stderr,"Everything underflowing. \n"); |
1669 |
|
#endif |
1670 |
|
free(scenefactor); |
1671 |
|
free(scenestart); |
1672 |
|
free(scenelength); |
1673 |
|
return -2; |
1674 |
|
} |
1675 |
|
|
1676 |
|
const float f_red = (1.f + S_red/S_elig); |
1677 |
|
|
1678 |
|
#ifdef VBV_DEBUG |
1679 |
|
printf("Moving %.0f kB to avoid buffer underflow, correction factor: %.5f\n",S_red/1024.f,f_red); |
1680 |
|
#endif |
1681 |
|
|
1682 |
|
violation=0; |
1683 |
|
for (l=0; l<num_scenes; l++) /* scale remaining scenes up to meet total size */ |
1684 |
|
{ |
1685 |
|
const int start = scenestart[l]; |
1686 |
|
const int length = scenelength[l]; |
1687 |
|
twopass_stat_t * frames = &rc->stats[start]; |
1688 |
|
|
1689 |
|
if (scenefactor[l] == 1.) |
1690 |
|
continue; |
1691 |
|
|
1692 |
|
/* there shouldn't be any segments with factor<1 left, so all the rest is >1 */ |
1693 |
|
|
1694 |
|
for (n= 0; n < length; n++) |
1695 |
|
{ |
1696 |
|
frames[n].scaled_length = (int)(frames[n].scaled_length * f_red + 0.5); |
1697 |
|
} |
1698 |
|
|
1699 |
|
scenefactor[l] /= f_red; |
1700 |
|
if (scenefactor[l] < 1.f) |
1701 |
|
violation=1; |
1702 |
|
} |
1703 |
|
|
1704 |
|
} while (violation); |
1705 |
|
|
1706 |
|
free(scenefactor); |
1707 |
|
free(scenestart); |
1708 |
|
free(scenelength); |
1709 |
|
return 0; |
1710 |
|
} |
1711 |
|
|
1712 |
|
|
1713 |
|
/***************************************************************************** |
1714 |
* Still more low level stuff (nothing to do with stats treatment) |
* Still more low level stuff (nothing to do with stats treatment) |
1715 |
****************************************************************************/ |
****************************************************************************/ |
1716 |
|
|