[KongchangAI]
Concept

Dynamic Shape Compilation

A compilation technique in deep learning compilers for handling dynamic tensor shapes, generating kernels that can process variable-length inputs through shape specialization or symbolic inference.

Timeline (last 90 days)

Nov 4

Dynamic shape compilation can reduce cumulative shape-switching cost by about 70%, at the cost of a slight decrease in steady-state performance.

Unverified50%
Nov 4

The core challenge of dynamic shape compilation is that the compiler cannot determine the specific dimensions of tensors at compile time, making it difficult to perform optimizations that depend on static sizes, such as loop unrolling, memory pre-allocation, and fixed-tile-based scheduling.

Unverified50%
Nov 4

The roughly 70% reduction in cumulative shape-switching cost achieved by dynamic-shape compilation comes from reducing the number of recompilations; the slight drop in steady-state performance stems from the general kernel being unable to fully exploit the optimization space afforded by static sizes.

Unverified50%

All Facts (3)

Source Articles