Exo 2: A new programming language for high-performance computing, with much less code

Many firms make investments closely in hiring expertise to create the high-performance library code that underpins trendy synthetic intelligence programs. NVIDIA, for occasion, developed a few of the most superior high-performance computing (HPC) libraries, making a aggressive moat that has confirmed tough for others to breach.

However what if a few college students, inside just a few months, might compete with state-of-the-art HPC libraries with just a few hundred strains of code, as a substitute of tens or lots of of hundreds?

That is what researchers at MIT’s Laptop Science and Synthetic Intelligence Laboratory (CSAIL) have proven with a new programming language known as Exo 2.

Exo 2 belongs to a new class of programming languages that MIT Professor Jonathan Ragan-Kelley calls “user-schedulable languages” (USLs). As an alternative of hoping that an opaque compiler will auto-generate the quickest potential code, USLs put programmers within the driver’s seat, permitting them to put in writing “schedules” that explicitly management how the compiler generates code. This permits efficiency engineers to rework easy applications that specify what they need to compute into advanced applications that do the identical factor as the unique specification, however much, much quicker.

One of many limitations of current USLs (like the unique Exo) is their comparatively mounted set of scheduling operations, which makes it tough to reuse scheduling code throughout completely different “kernels” (the person parts in a high-performance library).

In distinction, Exo 2 allows customers to outline new scheduling operations externally to the compiler, facilitating the creation of reusable scheduling libraries.

Lead creator Yuka Ikarashi, an MIT Ph.D. scholar in electrical engineering and laptop science and CSAIL affiliate, says that Exo 2 can scale back whole schedule code by an element of 100 and ship efficiency aggressive with state-of-the-art implementations on a number of completely different platforms, together with Primary Linear Algebra Subprograms (BLAS) that energy many machine studying functions. This makes it a pretty possibility for engineers in HPC centered on optimizing kernels throughout completely different operations, information sorts, and goal architectures.

“It is a bottom-up strategy to automation, reasonably than doing an ML/AI search over high-performance code,” says Ikarashi. “What meaning is that efficiency engineers and {hardware} implementers can write their very own scheduling library, which is a set of optimization strategies to use on their {hardware} to achieve the height efficiency.”

One main benefit of Exo 2 is that it reduces the quantity of coding effort wanted at anybody time by reusing the scheduling code throughout functions and {hardware} targets.

The researchers carried out a scheduling library with roughly 2,000 strains of code in Exo 2, encapsulating reusable optimizations which might be linear-algebra particular and target-specific (AVX512, AVX2, Neon, and Gemmini {hardware} accelerators). This library consolidates scheduling efforts throughout greater than 80 high-performance kernels with as much as a dozen strains of code every, delivering efficiency similar to, or higher than, MKL, OpenBLAS, BLIS, and Halide.

Exo 2 features a novel mechanism known as “Cursors” that gives what they name a “steady reference” for pointing on the object code all through the scheduling course of. Ikarashi says {that a} steady reference is crucial for customers to encapsulate schedules inside a library perform, because it renders the scheduling code unbiased of object-code transformations.

“We imagine that USLs must be designed to be user-extensible, reasonably than having a set set of operations,” says Ikarashi. “On this method, a language can develop to assist massive tasks by the implementation of libraries that accommodate various optimization necessities and utility domains.”

Exo 2’s design permits efficiency engineers to concentrate on high-level optimization methods whereas making certain that the underlying object code stays functionally equal by the usage of secure primitives. Sooner or later, the staff hopes to develop Exo 2’s assist for several types of {hardware} accelerators, like GPUs. A number of ongoing tasks purpose to enhance the compiler evaluation itself, when it comes to correctness, compilation time, and expressivity.

The examine is printed on the arXiv preprint server.

This story is republished courtesy of MIT Information (internet.mit.edu/newsoffice/), a preferred website that covers information about MIT analysis, innovation and instructing.