Docs | Tutorial | DiffTaichi | Examples | Contribute | Forum

Documentations	Chat	taichi-nightly	taichi-nightly-cuda-10-0	taichi-nightly-cuda-10-1

# 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

Note:

• On Ubuntu 19.04+, please sudo apt install libtinfo5.
• On Windows, please install Microsoft Visual C++ Redistributable if you haven't.

Contribution Guidelines

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

Updates

• (Mar 4, 2020) v0.5.7 released
  • Deprecated ti.classfunc. Please simply use ti.func, even if you are decorating a class member function
  • Upgrade spdlog from 0.14.0 to 1.5.0 with git submodule (#565) (by Mingkuan Xu [xumingkuan])
  • Metal backend refactored (by Ye Kuang [k-ye])
  • Fixed infinitely looping signal handlers
  • Fixed ti test on release mode
  • Doc updated
• (Mar 3, 2020) v0.5.6 released
  • Fixed runtime LLVM bitcode loading failure on Linux
  • Fixed a GUI bug in ti.GUI.line (by Mingkuan Xu [xumingkuan])
  • Fixed frontend syntax error false positive (static range-fors) (by Mingkuan Xu [xumingkuan])
  • arch=ti.arm64 is now supported. (Please build from source)
  • CUDA supported on NVIDIA Jetson. (Please build from source)
• (Mar 2, 2020) v0.5.5 released: Experimental CUDA 10.0/10.1 support on Windows. Feedbacks are welcome!
• (Mar 1, 2020) v0.5.4 released
  • Metal backend now supports < 32bit args (#530) (by Ye Kuang [k-ye])
  • Added ti.imread/imwrite/imshow for convenient image IO (by Yubin Peng [archibate])
  • ti.GUI.set_image now takes all numpy unsigned integer types (by Yubin Peng [archibate])
  • Bug fix: Make sure KernelTemplateMapper extractors's size is the same as the number of args (by Ye Kuang [k-ye])
  • Avoid duplicate evaluations in chaining comparison (such as 1 < ti.append(...) < 3 < 4) (by Mingkuan Xu [xumingkuan])
  • Frontend kernel/function structure checking (#544) (by Mingkuan Xu [xumingkuan])
  • Throw exception instead of SIGABRT to obtain RuntimeError in Python-scope (by Yubin Peng [archibate])
  • Mark sync bit only after running a kernel on GPU (by Ye Kuang [k-ye])
  • @ti.classkernel is deprecated. Always use ti.kernel, no matter you are decorating a class member function or not (by Ye Kuang [k-ye])
  • Fix ti.func AST transform (due to locals() not saving compile result) #538, #539 (by Yubin Peng [archibate])
  • Add a KernelSimplicityASTChecker to ensure grad kernel is compliant (#553) (by Ye Kuang [k-ye])
  • Fixed MSVC C++ mangling which leads to unsupported characters in LLVM NVPTX ASM printer
  • CUDA unified memory dependency is now removed. Set TI_USE_UNIFIED_MEMORY=0 to disable unified memory usage
  • Improved ti.GUI.line performance
  • (For developers) compiler significantly refactored and folder structure reorganized
• (Feb 25, 2020) v0.5.3 released
  • Better error message when try to declare tensors after kernel invocation (by Yubin Peng [archibate])
  • Logging: ti.warning renamed to ti.warn
  • Arch: ti.x86_64 renamed to ti.x64. ti.x86_64 is deprecated and will be removed in a future release
  • (For developers) Improved runtime bit code compilation thread safety (by Yubin Peng [archibate])
  • Improved OS X GUI performance (by Ye Kuang [k-ye])
  • Experimental support for new integer types u8, i8, u16, i16, u32 (by Yubin Peng [archibate])
  • Update doc (by Ye Kuang [k-ye])
• (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]
  • Official Python3.6 - Python3.8 packages on OS X (by wYw [Detavern])
• (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

• (ICLR 2020) Differentiable Programming for Physical Simulation [Video] [BibTex] [Code]
  • by Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand
• (SIGGRAPH Asia 2019) High-Performance Computation on Sparse Data Structures [Video] [BibTex] [Code]
  • by Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand