Library for C++ masochists. It does short vector math, SIMD, encodes and decodes images, virtual filesystem and other tricks.
"We wasted our time so that you don't have to waste yours."
CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz
Image: 2560 x 1600 (3531 KB)
---------------------------------------------------
decode(ms) encode(ms) size(KB)
---------------------------------------------------
libpng: 57.3 820.5 2843
lodepng: 105.0 722.6 2624
stb: 48.6 562.4 3864
spng: 37.7 215.9 3232
fpng: N/A 46.0 3715
wuffs: 23.9 N/A 0
mango: 3.2 57.2 3531
CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz
image: 4833 x 5875 (33598 KB)
---------------------------------------------------
decode(ms) encode(ms) size(KB)
---------------------------------------------------
libpng: 379.5 14051.3 28180
lodepng: 900.9 4953.5 26318
stb: 407.2 5249.0 45144
spng: 296.6 2424.7 34083
fpng: N/A 341.0 36958
wuffs: 165.9 N/A 0
mango: 53.7 287.0 33598
CPU: Intel Core i9 8950HK @ 2.90GHz
image: 5184 x 3456 (12652 KB)
-----------------------------------------------------
decode(ms) encode(ms) size(KB)
-----------------------------------------------------
libjpeg: 197.8 70.1 6677
stb: 300.7 1090.9 16784
toojpeg: N/A 292.6 5370
wuffs: 185.9 N/A 0
mango: 23.3 18.5 ms 6925
CPU: Intel Core i9 8950HK @ 2.90GHz
image: 2560 x 1600 (683 KB)
-----------------------------------------------------
decode(ms) encode(ms) size(KB)
-----------------------------------------------------
libjpeg: 32.9 16.3 678
stb: 54.8 118.9 1607
toojpeg: N/A 64.2 618
wuffs: 29.4 N/A 0
mango: 3.2 4.3 699
Portable vector programming abstraction that works seamlessly with native intrinsics.
float32x4 test(float32x4 a, float32x4 b, float32x4 c)
{
return a.wwww * b.xxyy + (c.xxzz - a).zzzz * b.w;
}
ARM64 GCC 15.2.0 generates the following instructions:
trn1 v2.4s, v2.4s, v2.4s
dup v3.4s, v1.s[3]
zip1 v1.4s, v1.4s, v1.4s
fsub v2.4s, v2.4s, v0.4s
fmul v2.4s, v3.4s, v2.s[2]
fmla v2.4s, v1.4s, v0.4s[3]
orr v0.16b, v2.16b, v2.16b
x86_64 clang 18.1.0 (SSE2):
shufps xmm2, xmm2, 160
subps xmm2, xmm0
shufps xmm0, xmm0, 255
movaps xmm3, xmm1
unpcklps xmm3, xmm3
mulps xmm0, xmm3
shufps xmm2, xmm2, 170
shufps xmm1, xmm1, 255
mulps xmm1, xmm2
addps xmm0, xmm1
ConcurrentQueue queue;
queue.enqueue([]
{
// your work here
});
The filesystem abstraction hides away details such as container format, compression and encryption. THere is a custom container format which is designed for modern multicore systems where large files are broken into blocks that can be processed in parallel, and small files are combined into blocks for higher compression ratio. The access to small files is through LRU cache so decompression cost is per block not per file (accessing two small files would otherwise cause decopression twice).
// accessing file from .iso container
File file("test.iso/data/image.png");
ConstMemory memory = file;
// same as above but using parent path
Path path("test.iso/data/");
File file(path, "image.png");
The parent Path is the virtual filesystem mechanism where the path provides filesystem. If no path is given the default is native filesystem. It is also possible to provide memory as path which allows to use Windows Resource Files using the same interface as any other file (this is just example use case for the path backed by raw memory).
Unified interface for large number of supported compression methods.
Unified encryption API with hardware acceleration for Intel AES-NI and AMD Crypto instructions.
"I'm skeptical of those benchmarks." -- randy408