mirror of https://github.com/DeviceFarmer/stf.git synced 2026-04-18 16:13:24 +02:00

Files

Gunther Brunner d381d90446 - Added benchmark comparing JPEG, PNG, WEBP, DDS DXT1, and CRN DXT1 formats.

- Pixi Canvas and WebGL render was added for quick benchmarking but it's going away soon, because WebGL is even slower than Canvas for big textures. The browsers nowadays use OpenGL in the backend for rendering Canvas.
- There is 3 different WebGL Renders but they need adjusting so they work with Non Power Of Two textures.
- For now the winner is JPEG+Canvas, however CRN+WebGL will be better almost for sure if server side supports it.

2014-02-28 16:41:49 +09:00

3.1 KiB

Raw Blame History

Benchmark results (internal data)

Versions

Canary 35
Safari 7
Firefox 27

Time to load and draw 5000 frames

Hardware	Browser	Render	Time	ms
iMac	Canary	pixi-canvas	54s
iMac	Canary	pixi-webgl	58s
iMac	Canary	canvas	54s	5.8ms
iMac	Firefox	pixi-canvas	75s	8.3ms
iMac	Firefox	pixi-webgl	82s	11.3ms
iMac	Firefox	canvas	76s	8.1ms
MacBook	Canary	pixi-canvas	68s
MacBook	Canary	pixi-webgl	83s
MacBook	Canary	canvas	76s

Time to just draw 5000 frames

Hardware	Browser	Format	Render	Time	ms
iMac	Canary	DDS DXT1	webgl	55s

Time to render 1 frame

Hardware	Browser	Format	Size	Function	ms
iMac	Canary	DDS DXT1	2.1MB	load	20-100ms
iMac	Canary	DDS DXT1	2.1MB	draw	0.04ms
iMac	Canary	DDS GZIP	271KB	load
iMac	Canary	DDS GZIP	271KB	draw
iMac	Canary	WEBP	70KB	load	9ms
iMac	Canary	WEBP	70KB	draw	24ms
iMac	Canary	DDS CRN	70KB	load	30ms
iMac	Canary	DDS CRN	70KB	draw	0.04ms
iMac	Canary	JPEG	94KB	load	25ms
iMac	Canary	JPEG	94KB	draw	6ms
iMac	Canary	PNG	590KB	load	6ms
iMac	Canary	PNG	590KB	draw	30ms

About ST3C DXT1 DDS CRN

ST3C (S3 Texture Compression)

Group of related lossy texture compression algorithms, supported by most GPUs in Mac/PCs

DXT1 (Block Compression 1)

Smallest ST3C algorithm, 1-bit Alpha, compression ratio 6:1

DDS (DirectDraw Surface)

Container file format for storing ST3C textures

CRN (DXTn Real-time Transcoding)

Container file format for transcoding crunched ST3C textures

DDS vs JPEG

Drawing a DXT1 texture is over 100 times faster than JPEG
The DXT1 texture is transferred directly to the GPU, unlike JPEG
The DXT1 uses 6 times less GPU memory
File size is big, but when gzipped it's smaller than PNG

CRN vs JPEG

It transcodes to DXT1 so has all DXT1 benefits
File size is very small, even smaller than JPEG (!)
Requires to transcode on client side, so there is a CPU penalty
However transcoding CRN->DXT1 is faster than decoding and tranfering JPEG->GPU
JPEG->GPU texture uploading blocks the main thread and is slow
Transcoding a CRN can be done async, and even offloaded to a WebWorker
Currently the Crunch library is compiled to JS with Emscripten, so enabling asm.js would make the transcoding even faster
Compressing CRN vs libjpeg-turbo benchmarks still need to be done