SMOL-V: like Vulkan/Khronos SPIR-V, but smaller.
SMOL-V encodes Vulkan/Khronos SPIR-V format programs into a form that is smoller, and is more compressible. Normally no changes to the programs are done; they decode into exactly same program as what was encoded. Optionally, debug information can be removed too.
SPIR-V is a very verbose format, several times larger than same programs expressed in other shader formats (e.g. DX11 bytecode, GLSL, DX9 bytecode etc.). The SSA-form with ever increasing IDs is not very appreciated by regular data compressors either. SMOL-V does several things to improve this:
- Many words, especially ones that most often have small values, are encoded using "varint" scheme (1-5 bytes per word, with just one byte for values in 0..127 range).
- Some IDs used in the program are delta-encoded, relative to previously seen IDs (e.g. Result IDs). Often instructions reference things that were computed just before, so this results in small deltas. These values are also encoded using "varint" scheme.
- Reordering instruction opcodes so that the most common ones are the smallest values, for smaller varint encoding.
- Encoding several instructions in a more compact form, e.g. the "typical <=4 component swizzle" shape of a VectorShuffle instruction, or sequences of MemberDecorate instructions.
A somewhat similar utility is spirv-remap from glslang.
See this blog post for more information about how I did SMOL-V.
smolv::Decode is the basic functionality. See smolv.h.
Other functions are for development/statistics purposes, to figure out frequencies and distributions of the instructions.
There's a test + compression benchmarking suite in
testing/testmain.cpp, using that needs adding
other files under testing/external to the build too (3rd party code: glslang remapper, Zstd, LZ4, miniz).
Limitations / TODO
- SPIR-V where the words got stored in big-endian layout is not supported yet.
- The whole thing might not work on Big-Endian CPUs. It might, but I'm not 100% sure.
- Not much prevention is done against malformed/corrupted inputs, TODO.
- Out of memory cases are not handled. The code will either throw exception or crash, depending on your compilation flags.
Code itself: Public Domain.
There is 3rd party code under the testing framework (
testing/external); it is not required for
using SMOL-V. Most of that code (glslang,
LZ4, Zstd) is BSD-licensed,
and taken from github repositories of the respective projects. miniz
is public domain.
There are SPIR-V binary shader dumps under
tests/spirv-dumps for compression testing;
these are not required for using SMOL-V. Not sure how to appropriately
"license" them (but hey they are kinda useless by themselves out of context),
so I'll go with this: "Binary shader dumps under 'tests' folder are only to be
used for SMOL-V testing". Details on them:
tests/spirv-dumps/dota2- some shaders from DOTA2, Copyright Valve Corporation, all rights reserved.
tests/spirv-dumps/shadertoy- several most popular shaders from Shadertoy, converted to Vulkan SPIR-V via glslang. Copyrights by their individual authors (filename matches last component of shadertoy URL).
tests/spirv-dumps/talos- some shaders from The Talos Principle, Copyright (c) 2002-2016 Croteam All rights reserved.
tests/spirv-dumps/unity- various Unity shaders, produced through a HLSL -> DX11 bytecode -> HLSLcc -> glslang toolchain.
As of 2016 September 1, results on 323 shaders (under
0 Remap 4540.2KB 93.3% 0 SMOL-V 1629.1KB 33.5% 1 zlib 1212.7KB 24.9% 1 re+zlib 1079.4KB 22.2% 1 sm+zlib 601.8KB 12.4% 2 LZ4HC 1342.6KB 27.6% 2 re+LZ4HC 1147.9KB 23.6% 2 sm+LZ4HC 606.0KB 12.4% 3 Zstd 898.7KB 18.5% 3 re+Zstd 742.6KB 15.3% 3 sm+Zstd 445.3KB 9.1% 4 Zstd20 589.3KB 12.1% 4 re+Zstd20 508.8KB 10.5% 4 sm+Zstd20 347.8KB 7.1%
- "Remap" is spirv-remap from glslang, with debug info stripping.
- SMOL-V is what you're looking at, with debug info stripping too.
- zlib, LZ4HC and Zstd are general compression algorithms at default settings (Zstd20 is Zstd compression with almost max setting of 20).
- "re+" is "remapper + compression", "sm+" is "SMOL-V + compression".
- Compression is done on the whole blob of all the test programs (not individually for each program).