rust-gpu_skill

Overview

This skill provides practical GPU memory and compute guidance for Rust engineers working with CUDA, OpenCL, compute shaders, and heterogeneous systems. It focuses on memory types, allocation patterns, zero-copy and unified memory, coalesced access, shared memory reduction, and alignment strategies. The content helps diagnose performance issues and design efficient data movement for high-throughput kernels.

How this skill works

The skill inspects typical GPU memory layers (global/VRAM, shared/SMEM, constant, local, registers) and maps Rust/CUDA patterns to each. It outlines allocation and copy idioms (cudaMalloc, cudaMallocHost, cudaMallocManaged, cudaMemcpyAsync) and recommends code motifs for coalescing, shared-memory reduction, and alignment. It also provides a checklist of performance checks and integration pointers for related Rust practices.

When to use it

• Designing kernel data layouts and access patterns for high memory bandwidth
• Implementing host↔device transfers with minimal stalls and overlapped compute
• Choosing between page-locked (zero-copy) memory, unified memory, or explicit copies
• Optimizing reductions, tiling, or other shared-memory algorithms
• Diagnosing low GPU utilization or memory-bound kernels

Best practices

• Favor coalesced global memory access: threads should read contiguous elements to maximize bandwidth
• Use shared memory to stage data and reduce global memory traffic for intra-block communication
• Prefer page-locked host memory (cudaMallocHost) only when beneficial; it increases system memory pressure
• Use asynchronous memcpy and streams to overlap data transfer with kernel execution
• Align structures to 256-byte boundaries when layouts cross cache lines or use wide loads/stores

Example use cases

• Porting a CPU vectorized loop to a GPU kernel and ensuring global accesses are coalesced
• Implementing a block-wise reduction using shared memory to reduce global traffic
• Using cudaMallocHost for real-time capture buffers to enable zero-copy access from GPU
• Switching to cudaMallocManaged to simplify pointer semantics for prototyping or irregular access patterns
• Diagnosing a memory-bound kernel by checking stride patterns, alignment, and shared-memory usage