SPIR-🇹 is a research project aimed at exploring shader-oriented IR designs derived from SPIR-V, and producing a framework around such an IR to facilitate advanced compilation pipelines, beyond what existing SPIR-V tooling allows for.
Such a need arose in the Rust-GPU project, which requires a variety of legalization passes to turn general-purpose (Rust1) code operating on untyped memory, into GPU-friendly direct data-flow.
Our goal is to replace the existing Rust-GPU SPIR-V legalizations passes with SPIR-🇹 equivalents - but even more imporantly, SPIR-🇹 should allow writing much more powerful legalization/optimization passes, that would've been unfathomable2 for direct SPIR-V manipulation.
1 Rust is not unique in its needs here, and more languages (or IRs) could eventually make use of such a framework, but the initial design and implementation work has focused on Rust-GPU
2 not outright impossible, but requiring excessive development/maintenance cost, having to constantly balance correctness and power (more conservative passes are easier to trust), etc.
This project is not affiliated, associated, authorized, endorsed by, or in any way officially connected with Khronos Group Inc., or any of its subsidiaries or its affiliates. The official Khronos Group Inc. website can be found at https://www.khronos.org.
The names SPIR, SPIR-V, as well as related names, marks, emblems and images are trademarks of their respective owners.
Additional context: the name of this project is a pun on SPIR-V, and entirely unrelated to SPIR (the older IR standard).
🚧 This project is in active design and development, many details can and will change 🚧
If you're interested in using SPIR-🇹 yourself, you may want to first take a look at the issue tracker for relevant issues, and even open new ones describing your usecase.
With the initial focus being on Rust-GPU's usecase, various (otherwise desirable) functionality/APIs/docs may be lacking, or rapidly changing - at the same time, discussions around widening the scope and usability of SPIR-🇹 in the long term are still welcome.
- supporting the ("OpenCL")
Kerneldialect of SPIR-VKernelSPIR-V is much closer to LLVM IR, thanShaderSPIR-V, and as such tooling oriented around LLVM is more likely to be a better fit
- textual syntax that can be parsed back
- i.e. the pretty-printer output is purely a visualization
|
IR data types:
|
Framework utilities:
Passes (to/from/on SPIR-🇹):
|
|
GLSL ( #version 450
out int output0;
void main() {
int o = 1;
for(int i = 1; i < 10; i++)
o *= i;
output0 = o;
}WGSL ( @vertex
fn main() -> @location(0) i32 {
var o: i32 = 1;
for(var i: i32 = 1; i < 10; i++) {
o *= i;
}
return o;
} |
SPIR-🇹 #[spv.Decoration.Flat]
#[spv.Decoration.Location(Location: 0)]
global_var GV0 in spv.StorageClass.Output: s32
func F0() -> spv.OpTypeVoid {
loop(v0: s32 <- 1s32, v1: s32 <- 1s32) {
v2 = spv.OpSLessThan(v1, 10s32): bool
(v3: s32, v4: s32) = if v2 {
v5 = spv.OpIMul(v0, v1): s32
v6 = spv.OpIAdd(v1, 1s32): s32
(v5, v6)
} else {
(spv.OpUndef: s32, spv.OpUndef: s32)
}
(v3, v4) -> (v0, v1)
} while v2
spv.OpStore(Pointer: &GV0, Object: v0)
} |
SPIR-V ( %typeof_output0 = OpTypePointer Output %i32
%output0 = OpVariable %typeof_output0 Output
%typeof_main = OpTypeFunction %void
%main = OpFunction %void None %typeof_main
%entry = OpLabel
OpBranch %bb0
%bb0 = OpLabel
OpBranch %bb1
%bb1 = OpLabel
%o = OpPhi %i32 %1_i32 %bb0 %o_next %bb5
%i = OpPhi %i32 %0_i32 %bb0 %i_next %bb5
OpLoopMerge %bb6 %bb5 None
OpBranch %bb2
%bb2 = OpLabel
%cond = OpSLessThan %bool %i %10_i32
OpSelectionMerge %bb4 None
OpBranchConditional %cond %bb4 %bb3
%bb3 = OpLabel
OpBranch %bb6
%bb4 = OpLabel
%o_next = OpIMul %i32 %o %i
OpBranch %bb5
%bb5 = OpLabel
%i_next = OpIAdd %i32 %i %1_i32
OpBranch %bb1
%bb6 = OpLabel
OpStore %output0 %o
OpReturn
OpFunctionEnd |
(and the vision of how SPIR-🇹 fits into it)
The distinction being made here is between:
- Interchange IRs (standards that many tools can use to interoperate)
- SPIR-V was very much intended as such a standard (outside of the GPU space, wasm is also a great example)
- they only need to encode the right concepts, not straying too far away from what tools understand, but the design effort is often oriented around being a "serialization" format
- Compiler IRs (non-standard implementation details of compilers)
- LLVM is quite well-known, but Mesa's NIR is even closer to SPIR-🇹 (both being shader-oriented, and having similar specialized choices of e.g. handling control-flow)
- these have to handle legalization/optimization passes quite well, and in general a lot of on-the-fly transformations - as their main purpose is to expedite such operations
- this is where SPIR-🇹 sits, as a kind of "relative"/dialect of SPIR-V, but making trade-offs in favor of the "intra-compiler" usage
We welcome community contributions to this project.
Please read our Contributor Guide for more information on how to get started. Please also read our Contributor Terms before you make any contributions.
Any contribution intentionally submitted for inclusion shall comply with the Rust standard licensing model (MIT OR Apache 2.0) and therefore be dual licensed as described below, without any additional terms or conditions:
This contribution is dual licensed under EITHER OF
- Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.
For clarity, "your" refers to any licensee/user of the contribution.