Fernando Magno Quintão Pereira

Rafael Alvarenga de Azevedo, Daniel Augusto Costa de Sa, Rodrigo Caetano Rocha et al.

Given a value computed within a program, an idempotent backward slice with respect to this value is a maximal subprogram that computes it. An informal notion of an idempotent slice has previously been used by Guimaraes et al. to transform eager into strict evaluation in the LLVM intermediate representation. However, that algorithm is insufficient to be correctly applied to general control-flow graphs. This paper addresses these omissions by formalizing the notion of idempotent backward slices and presenting a sound and efficient algorithm for extracting them from programs in Gated Static Single Assignment (GSA) form. As an example of their practical use, the paper describes how identifying and extracting idempotent backward slices enables a sparse code-size reduction optimization; that is, one capable of merging non-contiguous sequences of instructions within the control-flow graph of a single function or across functions. Experiments with the LLVM test suite show that, in specific benchmarks, this new algorithm achieves code-size reductions up to -7.24% on programs highly optimized by the -Os sequence of passes from clang 17.

Michael Canesche, Gaurav Verma, Fernando Magno Quintao Pereira

Machine-learning models consist of kernels, which are algorithms applying operations on tensors -- data indexed by a linear combination of natural numbers. Examples of kernels include convolutions, transpositions, and vectorial products. There are many ways to implement a kernel. These implementations form the kernel's optimization space. Kernel scheduling is the problem of finding the best implementation, given an objective function -- typically execution speed. Kernel optimizers such as Ansor, Halide, and AutoTVM solve this problem via search heuristics, which combine two phases: exploration and exploitation. The first step evaluates many different kernel optimization spaces. The latter tries to improve the best implementations by investigating a kernel within the same space. For example, Ansor combines kernel generation through sketches for exploration and leverages an evolutionary algorithm to exploit the best sketches. In this work, we demonstrate the potential to reduce Ansor's search time while enhancing kernel quality by incorporating Droplet Search, an AutoTVM algorithm, into Ansor's exploration phase. The approach involves limiting the number of samples explored by Ansor, selecting the best, and exploiting it with a coordinate descent algorithm. By applying this approach to the first 300 kernels that Ansor generates, we usually obtain better kernels in less time than if we let Ansor analyze 10,000 kernels. This result has been replicated in 20 well-known deep-learning models (AlexNet, ResNet, VGG, DenseNet, etc.) running on four architectures: an AMD Ryzen 7 (x86), an NVIDIA A100 tensor core, an NVIDIA RTX 3080 GPU, and an ARM A64FX. A patch with this combined approach was approved in Ansor in February 2024. As evidence of the generality of this search methodology, a similar patch, achieving equally good results, was submitted to TVM's MetaSchedule in June 2024.