Paper link: https://arxiv.org/abs/1802.04799

  • Scalable frameworks, such as TensorFlow, MXNet, Caffe, and PyTorch are optimized for a narrow range of serve-class GPUs.
  • Deploying workloads to other platforms such as mobile phones, IoT, and specialized accelarators(FPGAs, ASICs) requires laborious manual effort.
  • TVM is an end-to-end optimization stack that exposes:

    • graph-level
    • operator-level optimizations

    ---> to provide performance portability to deep learning workloads across diverse hardware back-ends.


  • The number and diversity of specialized deep learning (DL) accelerators pose an adoption challenge

    • They introduce new hardware abstractions that modern compilers and frameworks are ill-equipped to deal with.
  • Providing support in various DL frameworks for diverse hardware back-ends in the present ad-hoc fashion is unsustainable.

  • Hardware targets significantly diverge in terms of memory organization, compute, etc..

  • The Goal: easily deploy DL workloads to all kinds of hardware targets, including embedded devives, GPUs, FPGAs, ASCIs (e.g, the TPU).

  • Current DL frameworks rely on a computational graph intermediate representation to implement optimizations such as:

    • auto differentiation
    • dynamic memory management
  • Graph-level optimizations are often too high-level to handle hardware back-end-specific operator transformations.

  • Current operator-level libraries that DL frameworks rely on are:

    • too rigid
    • specialized

    ---> to be easily ported across hardware devices

  • To address these weaknesses, we need a compiler framework that can expose optimization opportunities across both

    • graph-level and
    • operator-level

    ---> to deliver competitive performance across hardware back-ends.

Four fundamental challenges at the computation graph level and tensor operator level

  1. High-level dataflow rewriting:

    • Different hardware devices may have vastly different memory hierarchies.

    • Enabling strategies to fuse operators and optimize data layouts are crucial for optimizing memory access.

  2. Memory reuse across threads:

    • Modern GPUs and specialized accelerators ahve memory that can be shared across compute cores.
    • Traditional shared-nothing nested parallel model is no longer optimal.
    • Cooperation among threads on shared memory loaded is required for optimized kernels.
  3. Tensorized compute intrinsics:

    • The latest hardware provides new instructions that go beyond vector operations like the GEMM operator in TPU or the tensor core in NVIDIA's Volta.
    • Consequently, the scheduling procedure must break computation into tensor arithmetic intrinsics instead of scalar or vector code.
  4. Latency Hiding

    • Traditional architectures with simultaneous multithreading and automatically managed caches implicitly hide latency in modern CPUs/GPUs.
    • Specialized accelerator designs favor learner control and offload most of the scheduling complexity to the compiler stack.
    • Still, scheduling must be peformed carefully to hide memory access latency.

TVM: An End-to-End Optimization Stack

  • An end-to-end optimizing compiler stack to lower and fine-tune DL workloads to diverse hardware back-ends.
  • Designed to separate:
    • the algorithm description
    • schedule
    • hardware interface
  • This separation enables support for novel specialized accelerators and their corresponding new intrinsics.
  • TVM presents two optimization layers:
    • a computation graph optimization layer to address:
      • High-level dataflow rewriting
    • a tensor optimization layer with new schedule primitives to address:
      • memory reuse across threads
      • tensorized compute intrinsics
      • latency hiding

Optimizing Computational Graphs

Computational Graph

  • Computational graphs are a common way to represent programs in DL frameworks.
  • They provide a global view on computation tasks, yet avoid specifying how each computation task needs to be implemented.

Operator Fusion

  • An optimization that can greatly reduce execution time, particulary in GPUs and specialized accelerators.
  • The idea is to combine multiple operators together into a single kernel without saving the intermediate results back into global memory

Four categories of graph operators:

  • Injective (one-to-one map)
  • Reduction
  • Complex-out-fusable (can fuse element-wise map to output)
  • Opaque (cannot be fused)

Data Layout Transformation

  • Tensor operations are the basic operators of computational graphs
  • They can have divergent layout requirements across different operations
  • Optimizing data layout starts with specifying the preferred data layout of each operator given the constraints dictating their implementation in hardware.

Limitations of Graph-Level Optimizations

  • They are only as effective as what the operator library provides.
  • Currently, the few DL frameworks that support operator fusion require the operator library to provide an implementation of the fused patterns.
    • With more network operators introduced on a regular basis, this approach is no longer sustainable when targeting an increasing number of hardware back-ends.
  • It is not feasible to handcraft operator kernels for this massive space of back-end specific operators
    • TVM provides a code-generation approach that can generate tensor operators.

Optimizing Tensor Operations

Tensor Expression Language

  • TVM introduces a dataflow tensor expression language to support automatic code generation.
  • Unlike high-level computation graph languages, where the implementation of tensor operations is opaque, each operation is described in an index formula expression language.

  • TVM tensor expression language supports common arithmetic and math operations found in common language like C.
  • TVM explicitly introduces a commutative reduction operator to easily schedule commutative reductions across multiple threads.
  • TVM further introduces a high-order scan operator that can combine basic compute operators to form recurrent computations over time.

Schedule Space

  • Given a tensor expression, it is challenging to create high-performance implementations for each hardware back-end.
  • Each optimized low-level program is the result of different combinations of scheduling strategies, imposing a large burden on the kernel writer.
  • TVM adopts the principle of decoupling compute descriptions from schedule optimizations.
  • Schedules are the specific rules that lower compute descriptions down to back-end-optimized implementations.

Nested Parallelism with Cooperation

  • Parallel programming is key to improving the efficiency of compute intensive kernels in deep learning workloads.
  • Modern GPUs offer massive parallelism

    ---> Requiring TVM to bake parallel programming models into schedule transformations

  • Most existing solutions adopt a parallel programming model referred to as nested parallel programs, which is a form of fork-join parallelism.

  • TVM uses a parallel schedule primitive to parallelize a data parallel task
    • Each parallel task can be further recursively subdivided into subtasks to exploit the multi-level thread hierarchy on the target architecture (e.g, thread groups in GPU)
  • This model is called shared-nothing nested parallelism

    • One working thread cannot look at the data of its sibling within the same parallel computation stage.
    • Interactions between sibling threads happen at the join stage, when the subtasks are done and the next stage can consume the data produced by the previous stage.
    • This programming model does not enable threads to cooperate with each other in order to perform collective task within the same parallel stage.
  • A better alternative to the shared-nothing approach is to fetch data cooperatively across threads

    • This pattern is well known in GPU programming using languages like CUDA, OpenCL and Metal.
    • It has not been implemented into a schedule primitive.
  • TVM introduces the concept of memory scopes to the schedule space, so that a stage can be marked as shared.
    • Without memory scopes, automatic scope inference will mark the relevant stage as thread-local.
    • Memory scopes are useful to GPUs.
    • Memory scopes allow us to tag special memory buffers and create special lowering rules when targeting specialized deep learning accelerators.

Tensorization: Generalizing the Hardware Interface

  • Tensorization problem is analogous to the vectorization problem for SIMD architectures.
  • Tensorization differs significantly from vectorization
    • The inputs to the tensor compute primitives are multi-dimensional, with fixed or variable lengths, and dictate different data layouts.
    • Cannot resort to a fixed set of primitives, as new DL accelerators are emerging with their own flavors of tensor instructions.
  • To solve this challenge, TVM separates the hardware interface from the schedule:
    • TVM introduces a tensor intrinsic declaration mechanism
    • TVM uses the tensor expression language to declare the behavior of each new hardware intrinsic, as well as the lowering rule associated to it.
    • TVM introduces a tensorize schedule primitive to replace a unit of computation with the corresponding tensor intrinsics.
    • The compiler matches the computation pattern with a hardware declaration, and lowers it to the corresping hardware intrinsic.

Compiler Support for Latency Hiding

  • Latency Hiding: refers to the process of overlapping memory operations with computation to maximize memory and compute utilization.
  • It requires different different strategies depending on the hardware back-end that is being targeted.
  • On CPUs, memory latency hiding is achieved implicitly with simultaneous multithreading or hardware prefetching techniques.
  • GPUs rely on rapid context switching of many wraps of threads to maximize the utilization of functional units.
  • TVM provides a virtual threading schedule primitive that lets the programmer specify a high-level data parallel program that TVM automatically lowers to a low-level explicit data dependence program.

Code Generation and Runtime Support

Code Generation

  • For a specific tuple of data-flow declaration, axis relation hyper-graph, and schedule tree, TVM can generate lowered code by:
    • iteratively traversing the schedule tree
    • inferring the dependent bounds of the input tensors (using the axis relation hyergraph)
    • generating the loop nest in the low-level code
  • The code is lowered to an in-memory representation of an imperative C style loop program.
  • TVM reuses a variant of Halide's the loop program data structure in this process.
  • TVM reuses passes from Halide for common lowering primitives like storage flattening and unrolling,
    • and add GPU/accelerator-specific transformations such as:
      • synchronization point detection
      • virtual thread injection
      • module generation
  • Finally, the loop program is transformed into LLVM or CUDA/Metal/OpenCL source code.

Runtime Support

  • For GPU programs, TVM builds the host and device modules separately and provide a runtime module system that launch kernels using corresponding driver APIs.

Remote Deployment Profiling

  • TVM includes infrastructure to make profiling and autotuning easier on embedded devices.
  • Traditionally, targeting an embedded device for tuning requires:

    • cross-compiling on the host side,
    • copying to the target device,
    • and timing the execution
  • TVM provides remote function call support. Through the RPC interface:

    • TVM compiles the program on a host compiler
    • it uploads to remote embedded devices
    • it runs the funcion remotely,
    • and it accesses the results in the same script on the host.


  • TVM provides an end-to-end stack to solve fundamental optimization challenges across a diverse set of hardware back-ends.
  • TVM can encourage more studies of programming languages, compilation, and open new opportunities for hardware co-design techniques for deep learning systems.