I wrote this article mainly for the people in the demoscene. If anything is unclear or missing for the Hackernews audience, I'm happy to answer questions here.
I might write a more general article later on writing code minifiers, e.g. how it compares to writing a code formatter, how to implement the transformations, etc.
On the tech side, the code is written in F#. The web build uses Bolero (Blazor for F#). So maybe I'll write later about my experience writing an open source project with F# and evolving it.
monokai_nl 4 hours ago [-]
Thank you for making this tool. I'm calling it in a custom webpack plugin so it transforms GLSL code into a single minified string on every build / watch event. I used this setup for my latest artwork: https://monokai.com/work/origin
kristel100 20 minutes ago [-]
Feels like an entire genre of coding art. These shader size comps always blow my mind—people getting realistic lighting and animation in what amounts to glorified tweets.
pjmlp 3 hours ago [-]
The pain of GL shaders, only because they don't embrace modern programming like modules and bytecode delivery.
That is how one ends up with shader minification.
And this is still quite actual as pain point, given how shaders work in 3D Web APIs.
ryao 2 hours ago [-]
Regarding bytecode delivery, have you seen SPIR-V?
As for modules, avoiding them is a performance feature since it is easier to optimize code when you compile everything in one unit. Rather than go toward modules, modern graphics has moved away from them by removing dynamic linking from newer shader languages such as SPIR-V. I believe a limited form of dynamic linking was later introduced in the form of graphics pipeline libraries for avoiding performance issues from a combinatoric explosion of partially redundant compilations when translating shaders from Direct3D 11 and earlier, but for anything outside of translating other APIs, you are supposed to avoid using that as far as I know.
pjmlp 48 minutes ago [-]
Of course, but that is only for Vulkan, technically it is possible with OpenGL 4.6, but I doubt anyone is bothering with it.
It is also the way on most proprietary APIs.
Metal and DirectX have dynamic linking, shader libraries, and one thing slang thankfully has, is exactly being a modern modular language, which has been given to Khronos as GLSL successor, alongside the already major adoption HLSL, GLSL is done.
TazeTSchnitzel 1 hours ago [-]
Dynamic linking is returning because of ray-tracing also.
slimbuck 2 hours ago [-]
Have you tested the shader's runtime performance penalty after minification?
Also wondering how you handle named uniforms?
ryao 2 hours ago [-]
I cannot answer this for him, but I can speculate on the answer to your first question. The transformations he described his minifier applying do not change the code from the perspective of the compiler in any meaningful way, so the end result should have no performance difference. The only possible exception is the inlining, but the compiler likely would have done that anyway, so the end result should still have no difference.
To state that more precisely, everything is going to be translated into a SSA syntax in a compiler pass at some point. At that point, much of what he described should be either something the toolchain would have done (such as comment removal by the preprocessor) or effectively undone (such as the variable name reuse since by definition SSA makes each variable name be used exactly once). The rest should converge to the same result after subsequent optimization passes (such as an inlining pass). If you find a performance difference from using his tool, file a bug report with your compiler’s authors.
laurentlb 2 hours ago [-]
Most transformations don't change the code that's executed.
const float x = sin(2);
const float y = 2;
If you define these two consts, we can inline y everywhere it's used. But by default, we don't inline x as it would lead to more work at runtime (maybe it doesn't matter for sin, but other function calls can be expensive). If you notice performance issues, please file a bug.
Renaming uniforms is optional (there's a flag). If you use C++, you can generate a .h header file that contains both the minified shader and macros that tell you how they've been renamed.
So Shader Minifier will generate macros like:
# define VAR_time "f"
to tell you that "time" is now called "f" and you can use VAR_time in your C++ code.
sagacity 2 hours ago [-]
This is an awesome tool, thanks Laurent! Currently using it for my next 64k intro which will probably ship sometime in 2047 :)
zombot 4 hours ago [-]
>
#define R return
OK, that was not what I thought of when reading "shader minification".
pests 4 hours ago [-]
Thankfully that’s not what the article is about, three paragraphs later it is noticed this has no savings after compression.
keyle 3 hours ago [-]
The author is referencing the demo scene, where they often have to meet very tight budgets in terms of space, so everyone is on an equal footing and it makes for very interesting wow moments.
akomtu 3 hours ago [-]
With recursive #defines it would be possible to do LZW compression.
If anyone wants to try it, I've made a web build: https://ctrl-alt-test.fr/minifier/
I might write a more general article later on writing code minifiers, e.g. how it compares to writing a code formatter, how to implement the transformations, etc.
On the tech side, the code is written in F#. The web build uses Bolero (Blazor for F#). So maybe I'll write later about my experience writing an open source project with F# and evolving it.
That is how one ends up with shader minification.
And this is still quite actual as pain point, given how shaders work in 3D Web APIs.
As for modules, avoiding them is a performance feature since it is easier to optimize code when you compile everything in one unit. Rather than go toward modules, modern graphics has moved away from them by removing dynamic linking from newer shader languages such as SPIR-V. I believe a limited form of dynamic linking was later introduced in the form of graphics pipeline libraries for avoiding performance issues from a combinatoric explosion of partially redundant compilations when translating shaders from Direct3D 11 and earlier, but for anything outside of translating other APIs, you are supposed to avoid using that as far as I know.
It is also the way on most proprietary APIs.
Metal and DirectX have dynamic linking, shader libraries, and one thing slang thankfully has, is exactly being a modern modular language, which has been given to Khronos as GLSL successor, alongside the already major adoption HLSL, GLSL is done.
Also wondering how you handle named uniforms?
To state that more precisely, everything is going to be translated into a SSA syntax in a compiler pass at some point. At that point, much of what he described should be either something the toolchain would have done (such as comment removal by the preprocessor) or effectively undone (such as the variable name reuse since by definition SSA makes each variable name be used exactly once). The rest should converge to the same result after subsequent optimization passes (such as an inlining pass). If you find a performance difference from using his tool, file a bug report with your compiler’s authors.
Renaming uniforms is optional (there's a flag). If you use C++, you can generate a .h header file that contains both the minified shader and macros that tell you how they've been renamed.
So Shader Minifier will generate macros like:
to tell you that "time" is now called "f" and you can use VAR_time in your C++ code.