Microsoft Technology Licensing, LLC

Blaise Pascal Tine

Mikhail Lyapunov

Yuri Dotsenko

Andy Glaister

Derek Sessions

Blake Douglas Pelton

Anthony John Hodsdon

Christopher Nathaniel Raubacher

Miles Mark Cohen

Described are compiler algorithms that partition a compute shader program into maximal-size regions, called thread-loops. The algorithms may remove original barrier-based synchronization yet the thus-transformed shader program remains semantically equivalent to the original shader program (i.e., the transformed shader program is correct). Moreover, the transformed shader program is amenable to optimization via existing compiler technology, and can be executed efficiently by CPU thread(s). A Dispatch call can be load-balanced on a CPU by assigning single or multiple CPU threads to execute thread blocks. In addition, the number of concurrently executing thread blocks do not overload the CPU.

Rasterization of compute shaders

Described herein are optimizations of thread loop intermediate representation (IR) code. One embodiment involves an algorithm that, based on data-flow analysis, computes sets of temporary variables that are loaded at the beginning of a thread loop and stored upon exit from a thread loop. Another embodiment involves reducing the size of a thread loop trip for a commonly-found case where a piece of compute shader is executed by a single thread (or a compiler-analyzable range of threads). In yet another embodiment, compute shader thread indices are cached to avoid excessive divisions, further improving execution speed.

Scalar optimizations for shaders

Intermediate representation (IR) code is received as compiled from a shader in the form of shader language source code. The input IR code is first analyzed during an analysis pass, during which operations, scopes, parts of scopes, and if-statement scopes are annotated for predication, mask usage, and branch protection and predication. This analysis outputs vectorization information that is then used by various sets of vectorization transformation rules to vectorize the input IR code, thus producing vectorized output IR code.

Vectorization of shaders

Growing community of inventors

Lynnwood, WA, United States of America

Blaise Pascal Tine