Tensorium SIMD2GPU is a compiler module designed to lift SIMD intrinsics (such as AVX) into architecture-independent MLIR representations. It enables automatic transformation of vectorized CPU code into GPU-executable kernels through MLIR lowering pipelines.
This component is part of the Tensorium Foundation, whose goal is to simplify and accelerate tensor-based computations in numerical physics, with a focus on general relativity.
- Translation of AVX (and potentially SSE/NEON) intrinsics into MLIR
vectorandmemrefoperations - Intermediate representation suitable for targeting GPU backends (CUDA, Metal)
- Integration with Clang plugins via
#pragma tensorium target(gpu) - Support for JIT compilation and execution using MLIR's
ExecutionEngine
This module is currently under active development. The initial focus is on:
- Identifying and translating SSE/AVX2/AVX512 intrinsics from LLVM IR
- Emitting MLIR dialects (
vector,gpu,memref) - Building a lightweight infrastructure for runtime GPU execution
MIT License