Frequently Asked Questions#
How to Install#
See Installing TVM.
How to add a new Hardware Backend#
If the hardware backend has LLVM support, then we can directly generate the code by setting the correct target triple as in
target.If the target hardware is a GPU, try to use the cuda, opencl or vulkan backend.
If the target hardware is a special accelerator, checkout VTA: Versatile Tensor Accelerator and Bring Your Own Codegen To TVM.
For all of the above cases, You may want to add target specific optimization templates using AutoTVM, see tutorials-autotvm-sec.
Besides using LLVM's vectorization, we can also embed micro-kernels to leverage hardware intrinsics, see tutorials-tensorize.
TVM's relation to Other IR/DSL Projects#
There are usually two levels of abstractions of IR in the deep learning systems. TensorFlow's XLA and Intel's ngraph both use a computation graph representation. This representation is high level, and can be helpful to perform generic optimizations such as memory reuse, layout transformation and automatic differentiation.
TVM adopts a low-level representation, that explicitly express the choice of memory layout, parallelization pattern, locality and hardware primitives etc. This level of IR is closer to directly target hardwares. The low-level IR adopts ideas from existing image processing languages like Halide, darkroom and loop transformation tools like loopy and polyhedra-based analysis. We specifically focus on expressing deep learning workloads (e.g. recurrence), optimization for different hardware backends and embedding with frameworks to provide end-to-end compilation stack.
TVM's relation to libDNN, cuDNN#
TVM can incorporate these libraries as external calls. One goal of TVM is to be able to generate high-performing kernels. We will evolve TVM an incremental manner as we learn from the techniques of manual kernel crafting and add these as primitives in DSL. See also top for recipes of operators in TVM.
Security#
See Security Guide