drjit 1.4.0

Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering

About this project

Dr.Jit is a just-in-time (JIT) compiler for ordinary and differentiable computation. It was originally created as the numerical foundation of Mitsuba 3, a differentiable Monte Carlo renderer. However, Dr.Jit is a general-purpose tool that can also help with various other types of embarrassingly parallel computation.

Dr.Jit helps with three steps:

• Tracing and vectorization: Dr.Jit executes arithmetic (e.g. a + b) by recording it into a computation graph instead of performing the operation right away. It then just-in-time (JIT) compiles this graph into fused kernels targeting GPUs, using either Metal on macOS or CUDA on other platforms. It can also target the host’s CPU using vector instruction sets like AVX512 or NEON via LLVM. The generated kernels are very efficient. Dr.Jit can also be used without JIT-compilation, which turns the project into a header-only vector library without external dependencies.

• Differentiation: If desired, Dr.Jit can compute derivatives using automatic differentiation (AD), using either forward or reverse-mode accumulation. Differentiation and tracing go hand-in-hand to produce specialized derivative evaluation code.

• Python: Dr.Jit types are accessible within C++17 and Python. Code can be developed in either language, or even both at once. Combinations of Python and C++ code can be jointly traced and differentiated.

Dr.Jit handles large programs with custom data structures, side effects, and polymorphism. It includes a mathematical support library including transcendental functions and types like vectors, matrices, complex numbers, quaternions, etc.

Difference to machine learning frameworks

Why did we create Dr.Jit, when dynamic derivative compilation is already possible using Python-based ML frameworks like JAX, Tensorflow, and PyTorch along with backends like XLA and TorchScript?

The reason is related to the typical workloads: machine learning involves smallish computation graphs that are, however, made of arithmetically intense operations like convolutions, matrix multiplications, etc. The application motivating Dr.Jit (differentiable rendering) creates giant and messy computation graphs consisting of 100K to millions of “trivial” nodes (elementary arithmetic operations). In our experience, ML compilation backends use internal representations and optimization passes that are too rich for this type of input, causing them to crash or time out during compilation. If you have encountered such issues, you may find Dr.Jit useful.

Cloning

Dr.Jit recursively depends on two other repositories: nanobind for Python bindings, and drjit-core providing core components of the JIT-compiler.

To fetch the entire project including these dependencies, clone the project using the --recursive flag as follows:

$ git clone --recursive https://github.com/mitsuba-renderer/drjit

Documentation

Please see Dr.Jit’s page on readthedocs.io for example code and reference documentation.

References, citations

Please see the paper Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering for the nitty-gritty details and details on the problem motivating this project. There is also a video presentation explaining the design decisions at a higher level.

If you use Dr.Jit in your own research, please cite it using the following BibTeX entry:

@article{Jakob2022DrJit,
  author = {Wenzel Jakob and S{\'e}bastien Speierer and Nicolas Roussel and Delio Vicini},
  title = {Dr.Jit: A Just-In-Time Compiler for Differentiable Rendering},
  journal = {Transactions on Graphics (Proceedings of SIGGRAPH)},
  volume = {41},
  number = {4},
  year = {2022},
  month = jul,
  doi = {10.1145/3528223.3530099}
}

Logo and history

The Dr.Jit logo was generously created by Otto Jakob. The “Dr.” prefix simultaneously abbreviates differentiable rendering with the stylized partial derivative D, while also conveying a medical connotation that is emphasized by the Rod of Asclepius. Differentiable rendering algorithms are growing beyond our control in terms of conceptual and implementation-level complexity. A doctor is a person, who can offer help in such a time of great need. Dr.Jit tries to fill this role to improve the well-being of differentiable rendering researchers.

Dr.Jit is the successor of the Enoki project, and its high-level API still somewhat resembles that of Enoki. The system evolved towards a different approach and has an all-new implementation, hence the decision to switch to a different project name.