Skip to content

architecture.md

Latest commit

 

History

History
48 lines (34 loc) · 2.95 KB

architecture.md

File metadata and controls

48 lines (34 loc) · 2.95 KB

Architecture

MIND is structured as a modular compiler-runtime stack. This document captures the top-level components and how data flows between them.

High-Level Overview

  1. Frontend – Lexer, parser, and surface-level validations produce a typed abstract syntax tree (AST).
  2. Type & Shape System – A constraint solver assigns concrete ranks, shapes, and element types while validating effect capabilities.
  3. Intermediate Representation (IR) – The typed AST lowers into a static single assignment (SSA) IR purpose-built for tensor programs.
  4. Lowering Pipelines – Dedicated passes perform canonicalization, fusion, layout selection, and eventually emit MLIR.
  5. Execution Runtimes – Backends convert MLIR to CPU or GPU executables, or interpret the IR directly for debugging.
  6. Tooling – Packaging, FFI, benchmarking, and developer tooling live alongside the compiler in feature-gated crates.

The architecture diagram in ../assets/diagrams/architecture.svg mirrors this flow.

Crate Layout

Crate/Module Responsibility
mind-syntax Lexer, parser, AST definitions, surface diagnostics
mind-types Type lattice, constraint solver, effect tracking
mind-ir Core SSA structures, pattern-matching utilities, graph rewrites
mind-lowering High-level → mid-level lowering, canonicalization passes
mind-mlir Emission of MLIR dialects, translation to LLVM
mind-runtime Tensor buffer management, host/device executors
mind-cli Command-line interface, REPL, and package tooling

The root crate enables features to bring specific components into the final binary. For example --features cpu-exec,mlir-exec compiles both the native interpreter and MLIR JIT.

Data Flow

Source → AST → Typed AST → MIND IR → MLIR → LLVM IR / Runtime Calls → Execution

Key invariants:

  • Types and shapes must be fully resolved before lowering to MLIR.
  • Autodiff annotations expand into explicit IR functions prior to optimization.
  • Runtime backends operate on host-device descriptors generated during lowering.

Extensibility Hooks

  • Dialect Extensions – New operators are introduced via pattern definitions in mind-ir and mirrored in MLIR dialect extensions.
  • Backend Plugins – Trait-based executors allow embedding custom accelerators via the Runtime trait.
  • Pass Pipelines – Pass managers can be configured per target, making it safe to ship experimental transformations behind feature flags.

For detailed discussions of the intermediate representation see ir-mlir.md; for runtime integration details refer to ffi-runtime.md.