ovid 0.1.2

fast, bidirectional PDF and Image converter
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145

# Linux Profiling & Cross-Platform Findings

Test system: 2-core AMD EPYC-Milan @ 2.0GHz, 7.6GB RAM, Ubuntu 24.04, kernel 6.8.0
Test PDF: 150-page lecture document at 300 DPI

## Build Notes

### libjpeg symbol collision (resolved)

MuPDF vendors libjpeg-9 and turbojpeg vendors libjpeg-turbo. Both define identical symbols.
This caused duplicate symbol errors or runtime version mismatches.

**Fix:** `Cargo.toml` enables the `sys-lib-libjpeg` feature on `mupdf`, which tells MuPDF to
use the system libjpeg (libjpeg-turbo) instead of its vendored libjpeg-9. This eliminates
the symbol conflict entirely.

Build dependencies:
```
apt install cmake nasm libclang-dev libfontconfig1-dev libjpeg-turbo8-dev pkg-config
```

## JPEG Encoding — 4:4:4 vs 4:2:0 Subsampling

### Discovery

Initial benchmarks showed JPG encoding was **2x slower than PNG** on both platforms.
This was traced to the `turbojpeg` Rust crate defaulting to `Subsamp::None` (4:4:4) —
no chroma subsampling. Every other JPEG tool (pdftoppm, ghostscript, cjpeg) defaults
to 4:2:0.

4:4:4 performs DCT + quantization + Huffman on all three color channels at full resolution.
4:2:0 subsamples chrominance by 2x in both dimensions, reducing encoding work by ~40%
with no perceptible quality loss for document/photo content.

### Measured encoding speed (tjbench, 2550x3300 page, Q75)

| Subsampling | FPS | ms/page | vs 4:4:4 |
|-------------|-----|---------|----------|
| 4:4:4       | 53  | 18.9    | baseline |
| 4:2:2       | 71  | 14.1    | 1.34x    |
| 4:2:0       | 88  | 11.4    | 1.66x    |
| Grayscale   | 118 | 8.5     | 2.22x    |

### Impact of fix (`set_subsamp(Sub2x2)`)

**Linux (150pg, 300 DPI, 2 cores):**

| Config | 4:4:4 (before) | 4:2:0 (after) | Improvement |
|--------|----------------|---------------|-------------|
| j=1    | 3.49s          | 2.29s         | 34%         |
| j=2    | 2.74s          | 1.84s         | 33%         |

**macOS (150pg, 300 DPI, 8 cores, Apple Silicon):**

| Config | 4:4:4 (before) | 4:2:0 (after) | Improvement |
|--------|----------------|---------------|-------------|
| j=1    | 2.91s          | 1.75s         | 40%         |
| j=8    | 0.56s          | 0.40s         | 29%         |

## Competitive Benchmarks — Linux

### Split: 150-page PDF, 300 DPI

**PNG:**

| Tool     | Wall    | User   | Sys   | RSS   | Speedup |
|----------|---------|--------|-------|-------|---------|
| ovid     | 1.40s   | 2.49s  | 0.26s | 104MB | —       |
| mutool   | 15.66s  | 13.67s | 1.98s | 81MB  | 11.2x   |
| gs       | 25.17s  | 24.87s | 0.29s | 37MB  | 18.0x   |
| pdftoppm | 85.79s  | 85.60s | 0.17s | 80MB  | 61.3x   |

**JPG:**

| Tool     | Wall    | User   | Sys   | RSS   | Speedup |
|----------|---------|--------|-------|-------|---------|
| ovid     | 1.84s   | 3.39s  | 0.22s | 101MB | —       |
| gs       | 2.89s   | 2.59s  | 0.29s | 37MB  | 1.6x    |
| pdftoppm | 4.12s   | 3.97s  | 0.14s | 80MB  | 2.2x    |

### Split: 50-page PDF, 300 DPI

**JPG:**

| Tool     | Wall    | Speedup |
|----------|---------|---------|
| ovid     | 0.67s   | —       |
| gs       | 1.08s   | 1.6x    |
| pdftoppm | 2.37s   | 3.5x    |

### Thread scaling (15pg, 300 DPI, 2 cores)

| Format | j=1   | j=2   | Speedup |
|--------|-------|-------|---------|
| PNG    | 0.79s | 0.60s | 1.32x   |
| JPG    | 0.95s | 0.69s | 1.38x   |

Parallel efficiency is lower on 2 cores than 8 — expected, since coordination overhead
is a larger fraction of total work.

### Binary size

| Tool     | Size  | Notes                         |
|----------|-------|-------------------------------|
| ovid     | 11MB  | standalone (static MuPDF + turbojpeg) |
| mutool   | ~1MB  | + shared libmupdf             |
| gs       | ~1MB  | + shared libs                 |
| pdftoppm | ~50KB | + shared libpoppler           |

Linux binary is larger than macOS (8.7MB) due to x86_64 SIMD assembly and
different LTO behavior across platforms.

## Time Breakdown — Where Cycles Go

Measured by comparing single-threaded PNG vs JPG and isolating via tjbench:

**150pg, 300 DPI, single-threaded (Linux):**

| Phase                 | PNG path | JPG path (4:2:0) |
|-----------------------|----------|-------------------|
| MuPDF rendering       | ~0.66s   | ~0.66s            |
| Image encoding        | ~1.14s   | ~1.63s            |
| **Total**             | 1.80s    | 2.29s             |

MuPDF rendering at 4.4ms/page is the floor — no tool can go faster than its renderer.
Encoding overhead is the only lever, and PNG with `Compression::Fast` (zlib level 1,
Paeth filter) is cheaper than JPEG at 4:2:0 on this workload because deflate on
document-like content (large flat regions, sharp edges) compresses very efficiently
with minimal CPU work.

## Why JPEG Margin Is Lower Than PNG Margin

1. **PNG competitors are single-threaded and slow.** pdftoppm takes 85s for PNG —
   its PNG encoder (libpng) is far slower than ovid's (png crate + zlib-rs). This
   gives ovid a 61x advantage. For JPEG, pdftoppm uses libjpeg-turbo internally
   and takes only 4.1s — the encoding gap is much smaller.

2. **MuPDF rendering dominates JPEG total time.** For JPG, encoding is ~71% of
   single-threaded time (1.63s of 2.29s). For PNG, encoding is ~63% (1.14s of 1.80s).
   The rendering floor is the same for both, so the encoding is where ovid
   differentiates itself, and the PNG encoding gap vs competitors is wider.

3. **2-core parallelism has diminishing returns.** ovid's primary advantage is
   parallel rendering across cores. On 2 cores, the speedup ceiling is 2x. On 8 cores
   (macOS), ovid achieves 4-5x speedup, making the margin much wider.