User Interface

Proteus offers five ways to define JIT work. All five eventually feed the same runtime specialization, caching, and execution pipeline, but they differ in how JIT code is described and in their build-time requirements.

Interface	Best for	Specialization support	Runtime template instantiation	Compiler requirement
Code Annotations	Incremental adoption in existing host, CUDA, or HIP code	Values, arrays, objects, and launch configuration	No	Application must be compiled with Clang
C++ Frontend API	Source-string JIT, runtime template instantiation, and compiler control	Values, arrays, objects, and launch configuration	Yes	Application can be compiled with any compatible compiler
LLVM IR Frontend API	Direct LLVM IR JIT from text or bitcode	Encoded directly in the provided LLVM IR	No	Application can be compiled with any compatible compiler
MLIR Frontend API	Source-string MLIR JIT and direct access to MLIR lowering	Encoded directly in the provided MLIR source	No	Application can be compiled with any compatible compiler
DSL API	Programmatic IR construction and advanced runtime code generation	Values, arrays, and launch configuration	No	Application can be compiled with any compatible compiler

For more detail on each path:

• Code Annotations
• C++ Frontend API
• LLVM IR Frontend API
• MLIR Frontend API
• DSL API

For runtime knobs that affect specialization, caching, and compilation, see the runtime configuration guide.