Introducing Triton: Open-Source GPU Programming for Neural Networks

We're releasing Triton 1.0, an open-source Python-like programming language which enables researchers with no CUDA experience to write highly efficient GPU codemost of the time on par with what an expert would be able to produce. Triton makes it possible to reach peak hardware performance with relatively little effort; for example, it can be used to write FP16 matrix multiplication kernels that match the performance of cuBLASsomething that many GPU programmers can't doin under 25 lines of code. Our researchers have already used it to produce kernels that are up to 2x more efficient than equivalent Torch implementations, and we're excited to work with the community to make GPU programming more accessible to everyone.

Novel research ideas in the field of Deep Learning are generally implemented using a combination of native framework operators. While convenient, this approach often requires the creation (and/or movement) of many temporary tensors, which can hurt the performance of neural networks at scale. These issues can be mitigated by writing specialized GPU kernels, but doing so can be surprisingly difficult due to the many intricacies of GPU programming. And, although a variety of systems have recently emerged to make this process easier, we have found them to be either too verbose, lack flexibility or generate code noticeably slower than our hand-tuned baselines. This has led us to extend and improve Triton, a recent language and compiler whose original creator now works at OpenAI.

The architecture of modern GPUs can be roughly divided into three major componentsDRAM, SRAM and ALUseach of which must be considered when optimizing CUDA code:

Basic architecture of a GPU.

Reasoning about all these factors can be challenging, even for seasoned CUDA programmers with many years of experience. The purpose of Triton is to fully automate these optimizations, so that developers can better focus on the high-level logic of their parallel code. Triton aims to be broadly applicable, and therefore does not automatically schedule work across SMs -- leaving some important algorithmic considerations (e.g. tiling, inter-SM synchronization) to the discretion of developers.

Compiler optimizations in CUDA vs Triton.

Out of all the Domain Specific Languages and JIT-compilers available, Triton is perhaps most similar to Numba: kernels are defined as decorated Python functions, and launched concurrently with different program_ids on a grid of so-called instances. However, as shown in the code snippet below, the resemblance stops there: Triton exposes intra-instance parallelism via operations on blockssmall arrays whose dimensions are powers of tworather than a Single Instruction, Multiple Thread (SIMT) execution model. In doing so, Triton effectively abstracts away all the issues related to concurrency within CUDA thread blocks (e.g., memory coalescing, shared memory synchronization/conflicts, tensor core scheduling).

Vector addition in Triton.

While this may not be particularly helpful for embarrassingly parallel (i.e., element-wise) computations, it can greatly simplify the development of more complex GPU programs.

Consider for example the case of a fused softmax kernel (below) in which each instance normalizes a different row of the given input tensor $X in mathbb{R}^{M times N}$. Standard CUDA implementations of this parallelization strategy can be challenging to write, requiring explicit synchronization between threads as they concurrently reduce the same row of $X$. Most of this complexity goes away with Triton, where each kernel instance loads the row of interest and normalizes it sequentially using NumPy-like primitives.

Fused softmax in Triton.

Note that the Triton JIT treats X and Y as pointers rather than tensors; we felt like retaining low-level control of memory accesses was important to address more complex data structures (e.g., block-sparse tensors).

Importantly, this particular implementation of softmax keeps the rows of $X$ in SRAM throughout the entire normalization process, which maximizes data reuse when applicable (~<32K columns). This differs from PyTorchs internal CUDA code, whose use of temporary memory makes it more general but significantly slower (below). The bottom line here is not that Triton is inherently better, but that it simplifies the development of specialized kernels that can be much faster than those found in general-purpose libraries.

A100 performance of fused softmax for M=4096.

The lower performance of the Torch (v1.9) JIT highlights the difficulty of automatic CUDA code generation from sequences of high-level tensor operations.

Fused softmax with the Torch JIT.

Being able to write fused kernels for element-wise operations and reductions is important, but not sufficient given the prominence of matrix multiplication tasks in neural networks. As it turns out, Triton also works very well for those, achieving peak performance with just ~25 lines of Python code. On the other hand, implementing something similar in CUDA would take a lot more effort and would even be likely to achieve lower performance.

Matrix multiplication in Triton.

One important advantage of handwritten matrix multiplication kernels is that they can be customized as desired to accommodate fused transformations of their inputs (e.g., slicing) and outputs (e.g., Leaky ReLU). Without a system like Triton, non-trivial modifications of matrix multiplication kernels would be out-of-reach for developers without exceptional GPU programming expertise.

V100 tensor-core performance of matrix multiplication with appropriately tuned values for BLOCK$_M$, BLOCK$_N$, BLOCK$_K$, GROUP$_M$.

The good performance of Triton comes from a modular system architecture centered around Triton-IR, an LLVM-based intermediate representation in which multi-dimensional blocks of values are first-class citizens.

High-level architecture of Triton.

The @triton.jit decorator works by walking the Abstract Syntax Tree (AST) of the provided Python function so as to generate Triton-IR on-the-fly using a common SSA construction algorithm. The resulting IR code is then simplified, optimized and automatically parallelized by our compiler backend, before being converted into high-quality LLVM-IRand eventually PTXfor execution on recent NVIDIA GPUs. CPUs and AMD GPUs are not supported at the moment, but we welcome community contributions aimed at addressing this limitation.

We have found that the use of blocked program representations via Triton-IR allows our compiler to automatically perform a wide variety of important program optimizations. For example, data can be automatically stashed to shared memory by looking at the operands of computationally intensive block-level operations (e.g., tl.dot)and allocated/synchronized using standard liveness analysis techniques.

The Triton compiler allocates shared memory by analyzing the live range of block variables used in computationally intensive operations.

On the other hand, Triton programs can be efficiently and automatically parallelized both (1) across SMs by executing different kernel instances concurrently, and (2) within SMs by analyzing the iteration space of each block-level operation and partitioning it adequately across different SIMD units, as shown below.

Element-wise

FP16 matrix multiplication

Vectorized

Tensorized

SM

GPU

Element-wise

FP16 matrix mult.multiplication

Vectorized

Tensorized

SM

GPU

Automatic parallelization in Triton. Each block-level operation defines a blocked iteration space that is automatically parallelized to make use of the resources available on a Streaming Multiprocessor (SM).

We intend for Triton to become a community-driven project. Feel free to fork our repository on GitHub!

If youre interested in joining our team and working on Triton & GPU kernels, were hiring!

Go here to see the original:
Introducing Triton: Open-Source GPU Programming for Neural Networks