HTVM ==== HTVM is a library which provides Haskell runtime and experimental frontend for [TVM](https://tvm.ai/about) the Machine Learning framework. **Both HTVM and TVM are under development. While TVM is somewhat stable, we don't recommend to use HTVM in applications currently** **[GitHub repository](https://github.com/grwlf/htvm) may contain newer version of HTVM** TVM in a nutshell ----------------- TVM framework extends [Halide](https://halide-lang.org) principles to Machine Learning domain. It offers (a) EDSLs for defining and hand-optimizing ML models (b) export/import facilities for translating models from other frameworks such as TensorFlow and (c) compiler to binary code for a variety of supported platforms, including LLVM (x86, arm), CUDA, OpenCL, Vulcan, ROCm, FPGAs and even WebAssembly (note: level of support may vary). DSLs for C++ and Python are best supported and also there are some support for Java, Go and Rust languages. [Watch Halide introduction video](https://youtu.be/3uiEyEKji0M) [Read more on TVM site](https://tvm.ai/about) Originally, TVM aimed at increasing speed of model's inference by providing a rich set of optimizing primitives called ['schedules'](https://docs.tvm.ai/tutorials/language/schedule_primitives.html#sphx-glr-tutorials-language-schedule-primitives-py)). At the same time it had little support for training models. Recently, training-related proposals were [added](https://sea-region.github.com/dmlc/tvm/issues/1996). TVM aims at compiling ML models in highly optimized binary code. Important parts of TVM are: * `tvm` is a core library providing `compute` interface. * `topi` is a tensor operations collection. Most of the middle-layer primitives such as `matmul`, `conv2d` and `softmax` are defined there. * `relay` is a high-level library written in Python, providing functional-style interface and its own typechecker. Currently, relay is under active development and beyond the scope of HTVM. * `nnvm` is another high-level wrapper in Python, now deprecated in favor of `relay`. Features and goals ------------------ In HTVM we are going to provide: 1. C Runtime, which makes it possible to run TVM models from Haskell. 2. Experimental EDSL for building TVM programs in Haskell. Combined TVM/HTVM-stack features are: ### FFI * Not many dependencies: TVM is much easier to build than other frameworks (hi TensorFlow). Models are compiled to binary code, no interpreters required. * Performance: HTVM uses TVM, which is designed with performace in mind. * Simplicity of code. ### EDSL * Experimental status * Simplicity again. Pure ADT-based design. * Not much type-safety yet. Expect errors in runtime. Typechecker may be implemented in future. Install ------- ### Installing dependencies 1. Make sure you have `g++` and `llvm` installed. 2. Build tvm from development repository located at https://github.com/grwlf/tvm, branch autodiff ``` $ git clone https://github.com/grwlf/tvm $ cd tvm $ git checkout origin/autodiff .. follow up with the tvm build procedure ``` ### Building HTVM We use development environment specified in [Nix](https://nixos.org/nix) language. In order to open it, please install the [Nix package manager](https://nixos.org/nix/download.html). Having Nix manager and `NIX_PATH` set, enter the environment, by running Nix development shell from the project's root folder: $ nix-shell It should get all the Haskell dependencies upon the first run. Alternatively, it should be possible to run Haskell distributions like [Haskell Platform](https://www.haskell.org/platform/). After nix-shell or Haskell distibution is ready, run `cabal` to build the project. $ cabal configure --enable-tests $ cabal build To run tests, execute the test suite. At this point you will need `g++`, `clang` and `tvm` of the correct version (see above). $ cabal test To enter the interactive shell, type $ cabal repl htvm *HTVM.EDSL.Types> :lo Demo Usage examples may be found in [Tests](./test/Main.hs) and (possibly outdated) [Demo](./src/Demo.hs). TODO: Update demo, write more examples Design notes ------------ ### TVM C Runtime FFI for TVM C Runtime library is a Haskell package, linked to `libtvm_runtime.so`. This library contains functionality, required to load and run ML code produced by TVM. 1. The module provide wrappers to `c_runtime_api.h` functions. 2. `TVMArray` is the main type describing Tensors in TVM. It is represented as ForeignPtr to internal representation and a set of accessor functions. 3. Currently, HTVM can marshal data from Haskell lists. Support for `Data.Array` is planned. 4. No backends besides LLVM are tested. Adding them should not be hard and is on the TODO list. ### TVM Haskell EDSL EDSL has a proof-of-concept status. It may be used to declare ML models in Haskell, convert them to TVM IR and finally compile. Later, compiled model may be loaded and run with Haskell FFI or with any other runtime supported by TVM. Contrary to usual practices, we don't manipulate TVM IR by calling TVM functions internally. Instead, we build AST in Haskell and print it to C++ program. After that we compile the program with common instruments. This approach has its pros and cons, which are described below. 1. `HTVM.EDSL.Types` module defines AST types which loosely corresponds to `Stmt` and `Expr` class hierarchies of TVM. 2. `HTVM.EDSL.Monad` provides monadic interface to AST builders. We favored simplicity over type-safety. We belive that overuse of Haskell type system ruined many good libraries. The interface relies on simple ADTs whenever possible. 3. `HTVM.EDSL.Print` contain functions which print AST to C++ program of Model Generator. 4. `HTVM.EDSL.Build` provides instruments to compile and run the model generator by executing `g++` and `clang` compilers: * The Model Generator program builds TVM IR and generates LLVM assembly. In HTVM, we support LLVM target, but more targets may be added later. * We execute `clang` to compile LLVM into x86 '.so' library. Resulting library may be loaded and executed by the Runtime code. The whole data transformation pipeline goes as follows: ``` Monadic --> AST --> C++ --> Model --> LLVM --> Model --> Runtime FFI Interface . . . Gen . asm . Library . . . . . . Print . Print . Run C++ g++ clang ``` Known disadvantages of C++ printing approach are: - **Compilation speed is limited by the speed of `g++`, which is slow.** Gcc is used to compile C++ to binary which may take as long as 5 seconds. Little may be done about that without changing approaches. One possible way to overcome this limitation would be to provide direct FFI to TVM IR like [Halide-hs](https://github.com/cchalmers/halide-hs) does for Halide. Unfortunately, this approach has its own downsides: * Low-level IR API is not as stable as its high-level counterpart * TVM is in its early stages and sometimes crashes. FFI to IR would provide no isolation from this. - **Calling construction-time procedures of TVM is non-trivial.** This is a consequence of previous limitation. For example, TVM may calculate Tensor shape in runtime and use it immediately to define new Tensors. In order to that in Haskell we would need to compile and run C++ program which is possible by slow. We try to avoid calling construction-time procedures. - **User may face weird C++ errors**. TVM is quite a low-level library which offers little type-checking, so user may write bad programs easily. Other high level TVM wrappers like Relay in Python, does provide their own typecheckers to catch errors earlier. HTVM offers no typechecker currently but it is certainly possible to write one. Contributions are welcome! The pros of this approach are: - C++ printer is implemented in less than 300 lines of code. Easy to maintain. - Easy to port to another TVM dialect such as Relay. - Isolation from TVM crashes. Memory problems of TVM IR will be translated to error messages in Haskell. Future plans ------------ * We aim at supporting basic `import tvm` and `import topi` functionality. * Support for Scheduling is minimal, but should be enhanced in future. * Support for TOPI is minimal, but should be enhanced in future. * No targets besides LLVM are supported. Adding them should be as simple as adding them to C++ DSL. * We plan to support [Tensor-Level AD](https://sea-region.github.com/dmlc/tvm/issues/1996) * Adding support for [Relay](https://github.com/dmlc/tvm/issues/1673) is also possible but may require some efforts like writing Python printer.