Skip to main content

Introduction

The Daisytuner Optimizing Compiler Collection (docc) compiles ordinary C/C++, Python/NumPy, and PyTorch code for multi-core CPUs, GPUs, and accelerators. You keep your code as it is and choose a target; docc parallelizes, offloads, and optimizes it for that target.

What docc offers​

  • Drop-in C/C++ compiler: use docc instead of clang or gcc.
  • Python JIT: compile NumPy functions with the @native decorator.
  • PyTorch backend: run models through torch.compile(..., backend="docc").
  • Automatic parallelization and offloading: OpenMP, CUDA, ROCm, and Tenstorrent from the same source code.
  • Data-transfer optimization: avoids unnecessary copies between host and device, even across library boundaries.
  • Runtime instrumentation: measure performance counters and capture kernel arguments of generated code.
  • Transfer tuning (optional): reuse optimizations from the Daisytuner cloud database.

Same idea, three frontends​

Each example parallelizes the same computation across CPU cores with OpenMP. Change the target to cuda or rocm to run it on a GPU.

C/C++

docc -O3 -docc-tune=openmp app.c -o app

Python

from docc.python import native

@native(target="openmp")
def matmul(A, B):
    return A @ B

PyTorch

import docc.pytorch

compiled = torch.compile(model, backend="docc", options={"target": "openmp"})

How it works​

All frontends translate code into one shared intermediate representation, the stateful dataflow multigraph (SDFG). docc analyzes and optimizes the SDFG, then generates code for the selected hardware:

Frontends          Intermediate representation       Targets

┌───────────┐
│ C/C++     │──┐                                     ┌──────────────────────────────────┐
└───────────┘  │   ┌─────────────────────────┐   ┌──►│ CPU          sequential, OpenMP  │
┌───────────┐  │   │ SDFG IR                 │   ├──►│ NVIDIA GPU   CUDA                │
│ Python    │──┼──►│ analysis & optimization │───┼──►│ AMD GPU      ROCm                │
└───────────┘  │   │ code generation         │   ├──►│ Tenstorrent  Wormhole, Blackhole │
┌───────────┐  │   └─────────────────────────┘   └──►│ Apple GPU    Metal (in progress) │
│ PyTorch   │──┘                                     └──────────────────────────────────┘
└───────────┘

Architecture & Development describes the components in detail.

Compatibility​

OpenMPCUDAROCmTenstorrentMetal
C/C++ (Linux)✅✅✅✅—
Python (Linux)✅✅✅——
PyTorch (Linux)✅✅✅——
Python (macOS)✅———🚧
PyTorch (macOS)🚧———🚧

✅ Supported | 🚧 Work in progress

Where to go next​