MLIR

The MLIR (Multilevel Intermediate View) project is a new approach to building a reusable and extensible compiler infrastructure. MLIR aims to address software fragmentation, improve compilation for heterogeneous hardware, significantly reduce the cost of building domain-specific compilers, and merge existing compilers. MLIR is designed for hybrid intermediate representation (IR) to support multiple different requirements in a single infrastructure.

The MLIR project defines a standard IR that brings together the infrastructure needed to run high-performance machine learning models in TensorFlow and similar ML environments. This project includes the application of HPC techniques and the integration of search algorithms such as reinforcement learning. MLIR aims to reduce the cost of new hardware implementation and improve usability for existing TensorFlow users.

MLIR brings together the infrastructure for high-performance ML models in TensorFlow. The TensorFlow ecosystem contains several compilers and optimizers that operate at multiple software and hardware stack levels. We expect the gradual adoption of MLIR to simplify every aspect of this stack. ShuraCore uses the MLIR project to design compilers for the following hardware platforms:

ShuraCore team specializes in using the following frameworks and tools:

TensorFlow
Caffe
PyTorch
LLVM
CUDA
OpenCL

Compiler Design Services

AsmJit, MIR, nj, NanoJIT, LLVM - MCJIT, Libraries, OpenMP, OpenACC, LLVM, C++, Rust, bison, flex, yacc, AntLR

Compiler Design Services

CPU, FPGA, GPU, TPU, TensorFlow, Caffe, PyTorch, LLVM, CUDA, OpenCL, WASM

ShuraCore specializes in implementing new and modern ports: GCC, GDB, GNU libraries, Binutils, LLDB, LLVM utilities, and libraries. In addition, we are engaged in the optimization and adaptation of existing compilers for any hardware platform. Finally, the ShuraCore team provides a full range of services for the development of compilers and interpreters.

We also work in the following areas: development of SDK, virtual machines, obfuscators, and code deobfuscators for our clients. We port debuggers and simulators to new hardware platforms, write high-speed optimizations. Our team also develops compilers for neural and tensor processors. ShuraCore creates developer tools based on the LLVM framework.