name old speed new speed delta wuffs_bmp_decode_40k/clang11 4.36GB/s ± 0% 4.48GB/s ± 0% +2.74% (p=0.008 n=5+5) wuffs_bmp_decode_40k/gcc10 4.49GB/s ± 0% 4.50GB/s ± 0% +0.15% (p=0.008 n=5+5) wuffs_gif_decode_1k_bw/clang11 723MB/s ± 0% 742MB/s ± 0% +2.64% (p=0.008 n=5+5) wuffs_gif_decode_1k_color_full_init/clang11 177MB/s ± 0% 177MB/s ± 0% ~ (p=0.841 n=5+5) wuffs_gif_decode_1k_color_part_init/clang11 216MB/s ± 0% 217MB/s ± 0% +0.66% (p=0.008 n=5+5) wuffs_gif_decode_10k_bgra/clang11 793MB/s ± 0% 785MB/s ± 0% -1.04% (p=0.016 n=4+5) wuffs_gif_decode_10k_indexed/clang11 210MB/s ± 0% 209MB/s ± 0% -0.38% (p=0.008 n=5+5) wuffs_gif_decode_20k/clang11 254MB/s ± 0% 255MB/s ± 0% +0.37% (p=0.008 n=5+5) wuffs_gif_decode_100k_artificial/clang11 568MB/s ± 0% 569MB/s ± 0% +0.14% (p=0.008 n=5+5) wuffs_gif_decode_100k_realistic/clang11 221MB/s ± 0% 221MB/s ± 0% ~ (p=0.222 n=5+5) wuffs_gif_decode_1000k_full_init/clang11 224MB/s ± 0% 224MB/s ± 0% ~ (p=0.690 n=5+5) wuffs_gif_decode_1000k_part_init/clang11 224MB/s ± 0% 224MB/s ± 0% ~ (p=0.841 n=5+5) wuffs_gif_decode_anim_screencap/clang11 1.30GB/s ± 0% 1.30GB/s ± 0% ~ (p=0.111 n=5+5) wuffs_gif_decode_1k_bw/gcc10 638MB/s ± 0% 672MB/s ± 0% +5.23% (p=0.008 n=5+5) wuffs_gif_decode_1k_color_full_init/gcc10 175MB/s ± 0% 181MB/s ± 0% +3.03% (p=0.008 n=5+5) wuffs_gif_decode_1k_color_part_init/gcc10 203MB/s ± 0% 209MB/s ± 0% +3.38% (p=0.008 n=5+5) wuffs_gif_decode_10k_bgra/gcc10 798MB/s ± 0% 804MB/s ± 0% +0.70% (p=0.008 n=5+5) wuffs_gif_decode_10k_indexed/gcc10 210MB/s ± 0% 212MB/s ± 0% +0.84% (p=0.008 n=5+5) wuffs_gif_decode_20k/gcc10 265MB/s ± 0% 265MB/s ± 0% ~ (p=0.222 n=5+5) wuffs_gif_decode_100k_artificial/gcc10 576MB/s ± 0% 577MB/s ± 0% ~ (p=0.690 n=5+5) wuffs_gif_decode_100k_realistic/gcc10 226MB/s ± 0% 227MB/s ± 0% +0.21% (p=0.008 n=5+5) wuffs_gif_decode_1000k_full_init/gcc10 230MB/s ± 0% 231MB/s ± 0% +0.32% (p=0.008 n=5+5) wuffs_gif_decode_1000k_part_init/gcc10 230MB/s ± 0% 231MB/s ± 0% +0.43% (p=0.008 n=5+5) wuffs_gif_decode_anim_screencap/gcc10 1.32GB/s ± 0% 1.32GB/s ± 0% +0.18% (p=0.008 n=5+5) wuffs_png_decode_image_19k_8bpp/clang11 267MB/s ± 0% 273MB/s ± 0% +2.47% (p=0.008 n=5+5) wuffs_png_decode_image_40k_24bpp/clang11 326MB/s ± 0% 327MB/s ± 0% +0.51% (p=0.008 n=5+5) wuffs_png_decode_image_77k_8bpp/clang11 954MB/s ± 0% 969MB/s ± 0% +1.58% (p=0.008 n=5+5) wuffs_png_decode_image_552k_32bpp_ignore_checksum/clang11 865MB/s ± 0% 864MB/s ± 0% -0.18% (p=0.016 n=5+5) wuffs_png_decode_image_552k_32bpp_verify_checksum/clang11 832MB/s ± 0% 831MB/s ± 0% -0.11% (p=0.016 n=5+5) wuffs_png_decode_image_4002k_24bpp/clang11 331MB/s ± 0% 332MB/s ± 0% +0.18% (p=0.016 n=5+5) wuffs_png_decode_image_19k_8bpp/gcc10 298MB/s ± 0% 292MB/s ± 0% -2.17% (p=0.008 n=5+5) wuffs_png_decode_image_40k_24bpp/gcc10 352MB/s ± 0% 349MB/s ± 0% -0.75% (p=0.008 n=5+5) wuffs_png_decode_image_77k_8bpp/gcc10 1.05GB/s ± 0% 1.03GB/s ± 0% -1.56% (p=0.008 n=5+5) wuffs_png_decode_image_552k_32bpp_ignore_checksum/gcc10 950MB/s ± 0% 949MB/s ± 0% ~ (p=0.310 n=5+5) wuffs_png_decode_image_552k_32bpp_verify_checksum/gcc10 905MB/s ± 0% 902MB/s ± 0% -0.35% (p=0.008 n=5+5) wuffs_png_decode_image_4002k_24bpp/gcc10 360MB/s ± 0% 358MB/s ± 0% -0.40% (p=0.008 n=5+5) wuffs_tga_decode_19k_8bpp/clang11 7.81GB/s ± 0% 7.90GB/s ± 0% +1.22% (p=0.008 n=5+5) wuffs_tga_decode_77k_24bpp/clang11 4.34GB/s ± 0% 4.36GB/s ± 0% +0.59% (p=0.008 n=5+5) wuffs_tga_decode_19k_8bpp/gcc10 7.10GB/s ± 0% 7.13GB/s ± 0% +0.33% (p=0.008 n=5+5) wuffs_tga_decode_77k_24bpp/gcc10 4.61GB/s ± 0% 4.60GB/s ± 0% -0.35% (p=0.008 n=5+5) Updates #96
PNG
PNG (Portable Network Graphics) is a lossless image compression format for still images. APNG (Animated PNG) is an unofficial extension for animated images.
Wuffs does not currently implement the APNG extension.
File Structure
A PNG file consists of an 8-byte magic identifier and then a series of chunks. Each chunk is:
- a 4-byte uint32 payload length
N. - a 4-byte chunk type (e.g.
gAMAfor gamma correction metadata). - an
N-byte payload. - a 4-byte uint32 CRC-32 checksum of the previous
(N + 4)bytes, including the chunk type but excluding the payload length.
All multi-byte numbers (including 16-bit depth RGBA colors) are stored
big-endian. The first chunk has an IHDR type (whose 13-byte payload contains
the uint32 width and height). The last chunk has an IEND type (and a 0-byte
payload).
For example, this 36 × 24 (0x24 × 0x1C), 8 bits-per-pixel PNG file's chunk
sequence starts: IHDR, gAMA, cHRM, bKGD, IDAT...
$ hd test/data/hippopotamus.interlaced.png | head
00000000 89 50 4e 47 0d 0a 1a 0a 00 00 00 0d 49 48 44 52 |.PNG........IHDR|
00000010 00 00 00 24 00 00 00 1c 08 02 00 00 01 f1 4c ba |...$..........L.|
00000020 99 00 00 00 04 67 41 4d 41 00 00 b1 8f 0b fc 61 |.....gAMA......a|
00000030 05 00 00 00 20 63 48 52 4d 00 00 7a 26 00 00 80 |.... cHRM..z&...|
00000040 84 00 00 fa 00 00 00 80 e8 00 00 75 30 00 00 ea |...........u0...|
00000050 60 00 00 3a 98 00 00 17 70 9c ba 51 3c 00 00 00 |`..:....p..Q<...|
00000060 06 62 4b 47 44 00 ff 00 ff 00 ff a0 bd a7 93 00 |.bKGD...........|
00000070 00 09 7a 49 44 41 54 48 c7 65 56 69 90 54 d5 19 |..zIDATH.eVi.T..|
00000080 bd eb 5b fb 75 bf de a6 a7 e9 9e b5 67 06 74 80 |..[.u.......g.t.|
00000090 08 a8 50 a8 10 b4 2c 2c b5 a2 21 d1 44 8d 26 24 |..P...,,..!.D.&$|
The upper / lower case bit of a chunk type's first of four letters denote
critical / ancillary chunks. There are four critical chunk types, which must
occur in this order (although a PLTE chunk is optional and there can be more
than one IDAT chunk):
IHDRcontains what Wuffs calls the image config (width, height and pixel format (e.g. 4 bits per pixel gray, RGB, RGBA_NONPREMUL_4X16BE)) and the interlacing bit.PLTEcontains the color palette.IDATcontains the zlib-compressed filtered pixel data. If there are multipleIDATchunks, their payloads are treated as concatenated.IENDcontains an empty payload.
The PNG specification allows decoders to ignore all ancillary chunks, but when
converting a PNG file to pixels on a screen, high quality decoders should still
process transparency related (tRNS) and color space related (cHRM, gAMA,
iCCP, sBIT and sRGB) chunks.
Filtering
Encoding a PNG involves subtracting (using modular uint8 arithmetic) a
predicted value from each pixel (before zlib compressing the residuals). The
predicted value of pixel x depends on the pixel to the left (a), the pixel
above (b) and the pixel above-left (c). Decoding a PNG involves reversing
that process (after zlib decompression).
Each row of pixels uses one of five prediction algorithms, also called filters: 0=None, 1=Sub, 2=Up, 3=Average and 4=Paeth.
- Filter 0:
Prediction = 0. - Filter 1:
Prediction = a. - Filter 2:
Prediction = b. - Filter 3:
Prediction = floor((a + b) / 2). - Filter 4:
Prediction = paeth(a, b, c).
The Paeth prediction function is around 10 lines of code and is described in the PNG spec.
Prediction conceptually involves pixels but in practice works on bytes. 8-bit depth RGB images have 3 bytes per pixel, so the 'pixel to the left' means 3 bytes prior in the zlib-decompressed stream. Low depth images (e.g. bi-level images have 0.125 bytes per pixel) use a filter distance of 1 byte.
The bytes of the 'pixel to the left' of the first column is implicitly zero. Likewise for the 'pixel above' the first row.
Per-Row Filter Byte
Just after zlib decompression when decoding (or just before zlib compression
when encoding), the image is represented as ((ceil(width * bpp) + 1) * height) bytes, where bpp is the possibly-fractional number of bytes per
pixel (e.g. 3 for RGB, 4 for RGBA). Each of the height rows of pixel data
start with an additional byte that denotes the per-row filter.
For example, a 4 pixel wide RGB (3 bpp) image would have 13 bytes per row.
Undoing the filter on this row (whose initial 0x01 byte denotes the Sub
filter) of residuals:
0x01, 0x57, 0x68, 0x61, 0x74, 0x73, 0x49, 0x6E, 0x41, 0x4E, 0x61, 0x6D, 0x65
reconstitutes the original RGB pixel data for that row:
____ 0x57, 0x68, 0x61, 0xCB, 0xDB, 0xAA, 0x39, 0x1C, 0xF8, 0x9A, 0x89, 0x5D
Interlacing
TODO.
Further Reading
See the PNG Wikipedia article.