Code generation for svod tensor operations.

This crate provides backend-agnostic code generation infrastructure for converting optimized UOp graphs into executable code.

Architecture

• Traits: Backend-agnostic interfaces (Renderer)
• LLVM: LLVM IR code generation for CPU execution
• Future: CUDA, Metal, OpenCL renderers

Usage

use svod_codegen::{llvm, program_pipeline};

let linear = svod_ir::UOp::linear(svod_schedule::linearize_with_cfg(optimized_uop_graph).into());
let kernel = llvm::text::render(&linear, Some("kernel"))?;
// Canonical staged flow: PROGRAM -> LINEAR -> SOURCE -> BINARY.
// See `program_pipeline` for the strict staged entrypoints.

Pre-render invariants

Direct callers of Renderer::render (and the per-backend render free functions) must pass a LINEAR-stage UOp produced by svod_schedule::linearize::line_rewrite_cleanups. The cleanup pass lowers gated LOADs into IF/STORE/ENDIF and provides the alt value that per-backend op handlers rely on; backends report Error::InvalidGraph if these invariants are violated. The staged entrypoints in program_pipeline run the cleanup pass automatically.

svod-codegen

Backend code generation from optimized UOp graphs.

Example

use svod_codegen::{Renderer, render};

let code = render(&kernel_graph, backend)?;

Backends

Select at runtime via SVOD_CPU_BACKEND env var (clang, llvm, mlir).

Planned:

• PTX renderer (CUDA)
• Metal renderer
• WebGPU (WGSL) renderer

Testing

cargo test -p svod-codegen