Modern neural networks are increasingly bottlenecked by the limited capacity of on-device
GPU memory. Prior work explores drop** activations as a strategy to scale to larger neural …

Engineering a Compiler, Third Edition covers the latest developments in compiler
technology, with new chapters focusing on semantic elaboration (the problems that arise in …

Chaitin and his colleagues at IBM in Yorktown Heights built the first global register allocator
based on graph coloring. This thesis describes a series of improvements and extensions to …

Compilers commonly translate an input program into an intermediate representation (IR)
before optimizing it and generating code. Over time there have been a number of different …

The widespread use of object-oriented languages and Internet security concerns are just the
beginning. Add embedded systems, multiple memory banks, highly pipelined units …

Modern electronics is driven by the explosive growth of digital communications and multi-
media technology. A basic challenge is to design first-time-right complex digital systems, that …

The value dependence graph (VDG) is a sparse dataflow-like representation that simplifies
program analysis and transformation. It is a functional representation that represents control …

Sets are a fundamental abstraction widely used in programming. Many representations are
possible, each offering different advantages. We describe a representation that supports …

When training complex neural networks, memory usage can be an important bottleneck. The
question of when to rematerialize, ie, to recompute intermediate values rather than retaining …

We consider the elimination of false paths in combinational circuits. We give the single
generic algorithm that is used to solve this problem, and demonstrate that it is parameterized …