/taichi

Productive programming language for portable, high-performance, sparse & differentiable computing

Primary LanguageC++MIT LicenseMIT

Docs | Tutorial | DiffTaichi | Examples | Contribute | Forum

Documentations Chat taichi-nightly taichi-nightly-cuda-10-0 taichi-nightly-cuda-10-1
Documentation Status Join the chat at https://gitter.im/taichi-dev/Lobby Downloads Downloads Downloads 
# Python 3.6/3.7 needed for all platforms. Python 3.8 supported only on OS X and Windows

# CPU only. No GPU/CUDA needed. (Linux, OS X and Windows)
python3 -m pip install taichi-nightly

# With GPU (CUDA 10.0) support (Linux only)
python3 -m pip install taichi-nightly-cuda-10-0

# With GPU (CUDA 10.1) support (Linux only)
python3 -m pip install taichi-nightly-cuda-10-1

# Build from source if you work in other environments

Contribution Guidelines

Linux (CUDA) OS X (10.14+) Windows
Build Build Status Build Status Build status
PyPI Build Status Build Status Build status

Updates

  • (Feb 20, 2020) v0.5.2 released
    • Gradients for ti.pow now supported (by Yubin Peng [archibate])
    • Multi-threaded unit testing (by Yubin Peng [archibate])
    • Fixed Taichi crashing when starting multiple instances simultaneously (by Yubin Peng [archibate])
    • Metal backend now supports ti.pow (by Ye Kuang [k-ye])
    • Better algebraic simplification (by Mingkuan Xu [xumingkuan])
    • ti.normalized now optionally takes a argument eps to prevent division by zero in differentiable programming
    • Improved random number generation by decorrelating PRNG streams on CUDA
    • Set environment variable TI_LOG_LEVEL to trace, debug, info, warn, error to filter out/increase verbosity. Default=info
    • [bug fix] fixed a loud failure on differentiable programming code generation due to a new optimization pass
    • Added ti.GUI.triangle example
    • Doc update: added ti.cross for 3D cross products
    • Use environment variable TI_TEST_THREADS to override testing threads
    • [For Taichi developers, bug fix] ti.init(print_processed=True) renamed to ti.init(print_preprocessed=True)
    • Various development infrastructure improvements by Yubin Peng [archibate]
  • (Feb 16, 2020) v0.5.1 released
    • Keyboard and mouse events supported in the GUI system. Check out mpm128.py for a interactive demo! (by Yubin Peng [archibate] and Ye Kuang [k-ye])
    • Basic algebraic simplification passes (by Mingkuan Xu [xumingkuan])
    • (For developers) ti (ti.exe) command supported on Windows after setting %PATH% correctly (by Mingkuan Xu [xumingkuan])
    • General power operator x ** y now supported in Taichi kernels (by Yubin Peng [archibate])
    • .dense(...).pointer() now abbreviated as .pointer(...). pointer now stands for a dense pointer array. This leads to cleaner code and better performance. (by Kenneth Lozes [KLozes])
    • (Advanced struct-fors only) for i in X now iterates all child instances of X instead of X itself. Skip this if you only use X=leaf node such as ti.f32/i32/Vector/Matrix.
    • Fixed cuda random number generator racing conditions
  • (Feb 14, 2020) v0.5.0 released with a new Apple Metal GPU backend for Mac OS X users! (by Ye Kuang [k-ye])
    • Just initialize your program with ti.init(..., arch=ti.metal) and run Taichi on your Mac GPUs!
    • A few takeaways if you do want to use the Metal backend:
      • For now, the Metal backend only supports dense SNodes and 32-bit data types. It doesn't support ti.random() or print().
      • Pre-2015 models may encounter some undefined behaviors under certain conditions (e.g. read-after-write). According to our tests, it seems like the memory order on a single GPU thread could go inconsistent on these models.
      • The [] operator in Python is slow in the current implementation. If you need to do a large number of reads, consider dumping all the data to a numpy array via to_numpy() as a workaround. For writes, consider first generating the data into a numpy array, then copying that to the Taichi variables as a whole.
      • Do NOT expect a performance boost yet, and we are still profiling and tuning the new backend. (So far we only saw a big performance improvement on a 2015 MBP 13-inch model.)
  • Full changelog

Short-term goals

  • (Done) Fully implement the LLVM backend to replace the legacy source-to-source C++/CUDA backends (By Dec 2019)
    • The only missing features compared to the old source-to-source backends:
      • Vectorization on CPUs. Given most users who want performance are using GPUs (CUDA), this is given low priority.
      • Automatic shared memory utilization. Postponed until Feb/March 2020.
  • (Done) Redesign & reimplement (GPU) memory allocator (by the end of Jan 2020)
  • (WIP) Tune the performance of the LLVM backend to match that of the legacy source-to-source backends (Hopefully by Feb, 2020. Current progress: setting up/tuning for final benchmarks)

Related papers