WebGPU Shading Language

[SPIR-V]: We should define a safe subset of SPIR-V and accept it directly in WebGPU implementations.
<SPIR-V is too low level>: SPIR-V is meant to be fed into drivers, while WebGPU needs to be able to generate HLSL and MSL, which are higher level. This makes some translation paths inefficient and awkward to implement.
+
it has a lot of instructions, some with undefined behavior.
+
it represents control flow as a graph (CFG), which forces us to recover the structure when generating HLSL and MSL.
[Text formats are better on the Web]: Using a text human readable format is more idiomatic to the Web.
+
JavaScript is the language of the Web.
+
Being able to see the web source and tinker with it was one of the principles of the Web as it was emerging, and it's still important today.
-
WASM is the new form of the language of the Web, and it's binary.
-
Parsing a text format is more complex and takes more time than parsing a SPIR-V binary.
-
shader parsing is only a couple of % for the time total spent in a pipeline creation. Most time is still spent within the driver.
-
Text is more verbose and redundant, from the point of view of the data transferred from the client.
-
there is no indication that text-based shaders are going to be larger than SPIR-V, which is known for its big size. Compressed, text should be close in size, if not smaller.
-
Text format is more complex than binary
-
complexity is independent of the form. There can be complex binary languages and simple text ones.
<SPIR-V can't be used in the tests and examples>: If we can't write it, we'd be writing tests and examples in another language, which would make WebGPU less strong API.
<SPIR-V is developed indepenently>: SPIR-V is developed by a working group inside Khronos, which may bring tension to using it.
(1)
Khronos working group has different constraints, it doesn't have any obligation to make SPIR-V changes that WebGPU would need.
(2)
It is likely that WebGPU would need to have different versions of the same functionality present in core SPIR-V, like the instructions for ray tracing.
(3)
That will effectively lead to the forking SPIR-V
<Nobody wants yet another shading language>: we already have a diversity of shading languages, as well as huge shader code bases. Adding support for another one is something nobody is excited about.
<Parsing text is prone to errors>: there is a large class of parsing/processing errors that doesn't exist in a binary format like SPIR-V/
+
Angle team has experience with WebGL shading language, there is a impressive bug list
<SPIR-V is well-specified>: It's an existing modern shading language that has a very detailed specification and a good conformance test suite.
-
SPIR-V isn't tested directly, but rather via GLSL
-
It's specified in 2 different documents: language itself and the execution environment. Sometimes it's not easy to figure out which of them is responsible for describing a particular behavior.
+
It has a memory model
-
WebGPU can borrow the memory model without taking the full SPIR-V
<SPIR-V has a large ecosystem>: Leveraging this ecosystem can help WebGPU to succeed.
+
There are libraries for reflecting SPIR-V in popular languages: C++, Rust, Go, Python etc.
+
There is both an optimizer and a validator in SPIRV-Tools.
+
SPIRV-Cross exists, it can translate SPIR-V into MSL, HLSL, and GLSL
-
After inspecting it and experimenting with it, both Google and Mozilla decided to not use SPIRV-Cross. It's hard to bring it into shape that will be suitable for shipping in a browser.