Crate svod_runtime 

Runtime execution for svod kernels.

Provides generic kernel execution interface with backend-specific implementations (LLVM JIT, native shared libraries, CUDA, etc.).

Execution Model

execution_plan is the canonical runtime path and executes prepared operations in dependency order with hazard-aware host parallelism. The executor module remains available for explicit parallel scheduling scenarios.

Benchmarking

The benchmark module provides timing utilities for measuring kernel execution performance, used by beam search auto-tuning.

Re-exports

pub use benchmark::BenchmarkConfig;

pub use benchmark::BenchmarkResult;

pub use benchmark::benchmark_kernel;

pub use benchmark::warmup_thread_pool;

pub use custom_function::run_custom_function;

pub use device_registry::DEVICE_FACTORIES;

pub use devices::cpu::CpuBackend;

pub use devices::cpu::create_cpu_device;

pub use devices::cpu::create_cpu_device_with_backend;

pub use devices::cpu_queue::CpuQueue;

pub use execution_plan::ExecutionPlan;

pub use execution_plan::ExecutionPlanBuilder;

pub use execution_plan::PreparedBufferView;

pub use execution_plan::PreparedCopy;

pub use execution_plan::PreparedCustomFunction;

pub use execution_plan::PreparedKernel;

pub use execution_plan::PreparedOp;

pub use executor::DeviceContext;

pub use executor::ExecutionGraph;

pub use executor::ExecutionNode;

pub use executor::KernelBufferAccess;

pub use executor::SyncStrategy;

pub use executor::UnifiedExecutor;

pub use executor::global_executor;

pub use profiler::KernelProfile;

pub use error::*;

pub use kernel_cache::*;

pub use llvm::*;

Modules

benchmark

Kernel benchmarking infrastructure for auto-tuning.

clang

Clang compilation backend for C codegen.

custom_function

device_registry

Device factory registry for runtime device creation and caching.

devices

Device implementations for different backends.

error

Error types for runtime execution.

execution_plan

Pre-compiled execution plan for kernel execution.

executor

Unified parallel execution for heterogeneous devices.

jit_loader

JIT ELF loader: compiles C source via clang stdin→stdout, parses the relocatable ELF with the object crate, copies sections into an anonymous mmap, applies relocations, and returns an executable function pointer.

kernel_cache

Global kernel deduplication cache.

llvm

LLVM JIT compilation via external clang + ELF loader.

profiler

Per-kernel execution profiling.