glsl-load-speed

glsl-load-speed copied to clipboard

A small Windows / Mac app to test GLSL shader loading speed.

TL;DR: even simple offline GLSL "pre-optimization" improves shader loading speed, often 2x faster to load on Mac! Loading D3D9 assembly shaders is several times faster than loading GLSL ones.

Measuring process:

1. load shaders from files
2. compile super small empty shader (to load/"warmup" the compiler)
3. set 1x1 pixel viewport
4. start timer
5. compile & link the shaders
6. set the program, draw two triangles (some drivers defer final compilation until first use)
7. delete shader
8. glFinish()
9. stop the timer

Several shaders included:

• Tree Leaf shader with per pixel lighting and some translucency calculations
• FXAA 3.11 PC-39 preset
• Deferred Lighting light application shader
• Parallax normal mapping with self-illumination
• Naive "port" of part of Mental Ray's architectural shader (not optimized)
• SSAO

For each shader, it's "raw" form is there as produced by hlsl2glslfork1 and the same shader, still in GLSL, but pre-optimized offline by glsl-optimizer2. For some shaders, D3D9 shader assembly is provided and performance of loading that is measured when on Windows.

See results.txt. Even loading "still a GLSL" shaders, but pre-optimized a bit, is faster. On Mac it's faster quite a lot!

A big game might use 1000 to 10000 shaders. According to these results, even simple offline pre-optimization can save 3-10ms from each shader loading time. That is 3 to 100 seconds saved from loading time. I guess loading preoptimized GLSL bytecode would be even faster, but there's no way to measure that until GLSL has it. This test compares loading speed of D3D9 bytecode but of course that is not a fully apples-to-apples comparison.

Source code includes parts of Ryan Gordon's Mojoshader 3 to assemble D3D9 shaders into bytecode. It felt like a saner choice than depending on a D3DX DLL from DirectX SDK.

TODO:

• Check on Mac OS X 10.7.x with Core profile.
• Check on Windows with latest AMD drivers.
• Check on Windows on NVIDIA hardware.
• Check equivalent shaders (where possible) using ARB vertex/fragment program assembly.