blob: dbb4e2901ecc83bfaa2f2557f1e6e474e605a26a
1 | The official guide to swscale for confused developers. |
2 | ======================================================== |
3 | |
4 | Current (simplified) Architecture: |
5 | --------------------------------- |
6 | Input |
7 | v |
8 | _______OR_________ |
9 | / \ |
10 | / \ |
11 | special converter [Input to YUV converter] |
12 | | | |
13 | | (8-bit YUV 4:4:4 / 4:2:2 / 4:2:0 / 4:0:0 ) |
14 | | | |
15 | | v |
16 | | Horizontal scaler |
17 | | | |
18 | | (15-bit YUV 4:4:4 / 4:2:2 / 4:2:0 / 4:1:1 / 4:0:0 ) |
19 | | | |
20 | | v |
21 | | Vertical scaler and output converter |
22 | | | |
23 | v v |
24 | output |
25 | |
26 | |
27 | Swscale has 2 scaler paths. Each side must be capable of handling |
28 | slices, that is, consecutive non-overlapping rectangles of dimension |
29 | (0,slice_top) - (picture_width, slice_bottom). |
30 | |
31 | special converter |
32 | These generally are unscaled converters of common |
33 | formats, like YUV 4:2:0/4:2:2 -> RGB12/15/16/24/32. Though it could also |
34 | in principle contain scalers optimized for specific common cases. |
35 | |
36 | Main path |
37 | The main path is used when no special converter can be used. The code |
38 | is designed as a destination line pull architecture. That is, for each |
39 | output line the vertical scaler pulls lines from a ring buffer. When |
40 | the ring buffer does not contain the wanted line, then it is pulled from |
41 | the input slice through the input converter and horizontal scaler. |
42 | The result is also stored in the ring buffer to serve future vertical |
43 | scaler requests. |
44 | When no more output can be generated because lines from a future slice |
45 | would be needed, then all remaining lines in the current slice are |
46 | converted, horizontally scaled and put in the ring buffer. |
47 | [This is done for luma and chroma, each with possibly different numbers |
48 | of lines per picture.] |
49 | |
50 | Input to YUV Converter |
51 | When the input to the main path is not planar 8 bits per component YUV or |
52 | 8-bit gray, it is converted to planar 8-bit YUV. Two sets of converters |
53 | exist for this currently: One performs horizontal downscaling by 2 |
54 | before the conversion, the other leaves the full chroma resolution, |
55 | but is slightly slower. The scaler will try to preserve full chroma |
56 | when the output uses it. It is possible to force full chroma with |
57 | SWS_FULL_CHR_H_INP even for cases where the scaler thinks it is useless. |
58 | |
59 | Horizontal scaler |
60 | There are several horizontal scalers. A special case worth mentioning is |
61 | the fast bilinear scaler that is made of runtime-generated MMXEXT code |
62 | using specially tuned pshufw instructions. |
63 | The remaining scalers are specially-tuned for various filter lengths. |
64 | They scale 8-bit unsigned planar data to 16-bit signed planar data. |
65 | Future >8 bits per component inputs will need to add a new horizontal |
66 | scaler that preserves the input precision. |
67 | |
68 | Vertical scaler and output converter |
69 | There is a large number of combined vertical scalers + output converters. |
70 | Some are: |
71 | * unscaled output converters |
72 | * unscaled output converters that average 2 chroma lines |
73 | * bilinear converters (C, MMX and accurate MMX) |
74 | * arbitrary filter length converters (C, MMX and accurate MMX) |
75 | And |
76 | * Plain C 8-bit 4:2:2 YUV -> RGB converters using LUTs |
77 | * Plain C 17-bit 4:4:4 YUV -> RGB converters using multiplies |
78 | * MMX 11-bit 4:2:2 YUV -> RGB converters |
79 | * Plain C 16-bit Y -> 16-bit gray |
80 | ... |
81 | |
82 | RGB with less than 8 bits per component uses dither to improve the |
83 | subjective quality and low-frequency accuracy. |
84 | |
85 | |
86 | Filter coefficients: |
87 | -------------------- |
88 | There are several different scalers (bilinear, bicubic, lanczos, area, |
89 | sinc, ...). Their coefficients are calculated in initFilter(). |
90 | Horizontal filter coefficients have a 1.0 point at 1 << 14, vertical ones at |
91 | 1 << 12. The 1.0 points have been chosen to maximize precision while leaving |
92 | a little headroom for convolutional filters like sharpening filters and |
93 | minimizing SIMD instructions needed to apply them. |
94 | It would be trivial to use a different 1.0 point if some specific scaler |
95 | would benefit from it. |
96 | Also, as already hinted at, initFilter() accepts an optional convolutional |
97 | filter as input that can be used for contrast, saturation, blur, sharpening |
98 | shift, chroma vs. luma shift, ... |
99 |