VectorWare enables Rust's portable SIMD on GPUs, allowing direct GPU utilization

Rust Portable SIMD on GPUs

VectorWare has achieved a milestone by enabling Rust's portable SIMD (core::simd) to run on GPUs. This integration allows developers to utilize the parallel processing capabilities of GPUs directly within Rust applications, using existing SIMD abstractions.

Understanding SIMD and GPU Parallelism

On CPUs, SIMD (Single Instruction, Multiple Data) enables a single instruction to operate on multiple data elements simultaneously within a thread, such as adding eight f32 values at once. This data parallelism occurs below the operating system's scheduling level. Traditionally, Rust SIMD involved architecture-specific intrinsics, but portable SIMD provides a generic Simd<T, N> type that abstracts these differences, allowing the compiler to generate appropriate vector instructions for the target CPU.

SIMT and SIMD Alignment

GPUs operate using a model called SIMT (Single Instruction, Multiple Thread), where a warp executes one instruction across multiple lanes, each processing its own data. This model is fundamentally similar to SIMD, as one instruction acts on many data elements. VectorWare recognized that a GPU's warp functions as a wide vector unit, making it a suitable target for Rust's portable SIMD. The portable SIMD functionality resides in core, eliminating the need for std support on the GPU.

Implications for Developers

This development means that a Simd<i16, 32> type, for example, can distribute one i16 element to each of a warp's 32 lanes. Operations like adding two such vectors compile to a single warp instruction, directly leveraging the GPU's parallel architecture. This advancement simplifies the process for Rust developers to harness GPU power for complex, high-performance applications.