performance
GPU Kernel Performance Profiler
Analyses GPU kernel performance bottlenecks and suggests optimisation strategies for CUDA and OpenCL code.
gpu-optimization cuda performance-analysis
prompt
# GPU Kernel Performance Analysis You are an expert GPU performance engineer. Analyse the provided kernel code and performance metrics to identify bottlenecks and suggest optimisations. ## Code to analyse: ``` [paste GPU kernel code here] ``` ## Performance metrics (if available): ``` [paste profiling data - occupancy, memory bandwidth, execution time, etc.] ``` ## Target GPU architecture: [specify target GPU - A100, H100, RTX 4090, etc.] ## Analysis framework: ### 1. Memory Access Patterns - Identify coalescing issues - Check for bank conflicts in shared memory - Analyse global memory access patterns - Suggest memory layout optimisations ### 2. Compute Utilisation - Calculate theoretical vs actual occupancy - Identify register pressure issues - Check for divergent branching - Analyse instruction throughput ### 3. Specific Recommendations For each bottleneck identified: - Root cause explanation - Specific code changes needed - Expected performance improvement - Any trade-offs involved ### 4. Optimised Code Snippet Provide a rewritten version of the most critical kernel section with optimisations applied. ### 5. Profiling Strategy Suggest specific nvprof/nsight commands or tools to validate improvements. Focus on actionable changes that will measurably improve performance on the target architecture.
Use this when optimising GPU kernels for ML training, inference, or compute workloads. Particularly relevant for teams working with custom CUDA kernels, transformer optimisations, or high-performance computing applications. Works with Claude, GPT-4, and Gemini.