Struct LlvmCompiler 

pub struct LlvmCompiler { /* private fields */ }

Compiles kernels by shelling out to the custom teenyc compiler (-Zcodegen-backend=mlir) at runtime. See crate::compiler::find_teenyc for how its path is resolved, and the crate docs for the cargo-teeny setup this requires.

Implementations

impl LlvmCompiler

pub fn new( teenyc_path: impl Into<PathBuf>, cache_dir: impl Into<PathBuf>, ) -> Result<Self>

Creates a compiler that invokes the teenyc binary at teenyc_path, caching compiled kernels under cache_dir (created if it doesn’t exist).

ptx_version defaults from $TEENYC_PTX_VERSION if set (parse failures are ignored, falling back to None/teenyc’s own default). This matters beyond the compile itself: Compiler::compile’s cache-key hash folds in ptx_version, so a JIT/runtime call site that never explicitly calls Self::with_ptx_version — e.g. a deployed binary just looking up an AOT-precompiled kernel cache, with no live teenyc to fall back to — must still agree with whatever override cargo teeny package --options ptx-version=NN used at AOT time, or every lookup misses. Reading the env var here means one .env entry (see TEENYC_PTX_VERSION in the deployed package’s env) keeps both sides in sync without every call site having to thread the value through by hand.

log_level similarly defaults from $TEENYC_LOG_LEVEL if set (parse failures are ignored, falling back to None) — this lets any call site, including ones that don’t (or can’t, e.g. a fixed helper like crate::compiler::driver::cuda::compile_kernel) call Self::with_log_level directly, turn on pipeline-stage logging via the environment.

pub fn with_target_cpu(self, cpu: impl Into<String>) -> Self

Sets the target GPU architecture (e.g. sm_90) passed to teenyc as -Ctarget-cpu.

pub fn with_ptx_version(self, ptx_version: u32) -> Self

Override the PTX ISA version teenyc stamps into the generated PTX (encoded as major*10 + minor, e.g. 82 for 8.2), via TEENYC_PTX_VERSION. Without this, teenyc picks a conservative default from the target capability — set this when the deployment target’s exact CUDA version is known and needs a precise match.

pub fn with_log_level(self, log_level: LogLevel) -> Self

Sets teenyc’s diagnostic verbosity (see LogLevel). Left unset (the default), teenyc uses its own default (roughly warn) and no pipeline-stage IR is captured.

Trait Implementations

impl Clone for LlvmCompiler

fn clone(&self) -> LlvmCompiler

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Compiler for LlvmCompiler

fn compile( &self, kernel: &impl Kernel, _target: &impl Target, force: bool, ) -> Result<String>

Compiles kernel for target, returning the path to the compiled artifact. force recompiles even if a cached artifact exists.

impl Debug for LlvmCompiler

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.