Update Nvptx backend for Zig 0.10 #12878
Conversation
| if (mod.comp.bin_file.options.target.cpu.arch.isNvptx()) { | ||
| for (buffer.items) |*byte| switch (byte.*) { | ||
| '{', '}', '*', '[', ']', '(', ')', ',', ' ', '\'' => byte.* = '_', | ||
| else => {}, | ||
| }; | ||
| } |
There was a problem hiding this comment.
Is this not gonna cause linker errors when two symbols map to the same value? Say functions @"A()" and @"A{}".
There was a problem hiding this comment.
yes, it's true. I don't know how robust this should be. I could sanitize and append some hash. Would that be better ?
src/Sema.zig
Outdated
| .Fn => switch (ty.fnCallingConvention()) { | ||
| // For now we want to authorize PTX kernel to use zig objects, even if we end up exposing the ABI. | ||
| // The goal is to experiment with more integrated CPU/GPU code. | ||
| .PtxKernel => return true, | ||
| else => return !Type.fnCallingConventionAllowsZigTypes(ty.fnCallingConvention()), | ||
| }, |
There was a problem hiding this comment.
Why not just change fnCallingConventionAllowsZigTypes to return true for .PtxKernel?
There was a problem hiding this comment.
so fnCallingConventionAllowsZigTypes is already returning true for .PtxKernel. But that's the issue. Because I still want to be able to export the kernels. So we allow exporting functions that don't have Zig types in their CC, and we also allow exporting PtxKernel even if they have Zig types.
gwenzek
force-pushed
the
ptx
branch
2 times, most recently
from
a325183 to
15fbf59
Compare
|
Nice work. Would you be willing to write up some corresponding release notes that I can put into the upcoming 0.10.0 release notes? |
Due to the progress on 0.10 a few things broke Nvptx backend.
Some of the changes I'm pushing here are outside of
link/NvPtx.zig, and are modifying shared code.Let me know if you can think of a better way to structure the changes. From least controversial to more controversial.
link/NvPtx.zig -> I'm modifying in place the "Compilation" object. Now I restore the changes before returning. A bit hacky but it allows to reuse the rest of the LLVM backend code without modification there.
Target.zig -> PTX only support debug information starting from PTX 7.5. Not sure how to test "ptx >= 7.5".
Sema.zig -> allow to use Zig objects in exported .PtxKernel functions. This allow tight integration between CPU/GPU code and was approved by Andrew given that we can change it later if we realize it was a bad idea to expose the ABI like this. Handling of heterogeneous computing #10397 (comment)
Module.zig -> sanitize the name of variables. This is required because the names can be like "io.fixed_buffer_stream.FixedBufferStream([]u8).writer" and don't match the C-like identifiers expected by PTX.
Tagging @Snektron who expressed interest in this workstream and @Vexu who kindly merge previous PR (#10189)
Thanks to the progress of self hosted, we can now generate debug information in the PTX files. Also the assembly syntax now works as documented, so it's simpler to use special ptx registers. And I can use the same zig binary for building the device and host code, so it's pretty exciting.
See https://github.com/gwenzek/cudaz/tree/e8895596009c689300fe7c7193fa2dbf7db07629 for user code using this Zig branch.