Diagnose CUDA kernel bugs and bottlenecks with Nsight profiling and binary inspection.
npx skills add https://github.com/dotruru/claudemd --skill cuda-dotruru
Please help me install this Agent Skill. Skill: cuda Source: https://github.com/dotruru/claudemd/tree/main/skills/ptx-isa Command: npx skills add https://github.com/dotruru/claudemd --skill cuda-dotruru
CUDA developers often struggle to systematically develop, debug, profile, and optimize GPU kernels; this Skill provides a structured guide with workflows and practical guidance to streamline those tasks.
Start with a minimal CUDA kernel, run compute-sanitizer to surface a bug, profile with nsys/ncu to identify hotspots, and iteratively apply printf-based traces to confirm fixes.
High-intent search queries and answers about installing and using this skill.
Profile CUDA kernel performance with Nsight Systems (nsys) for non-interactive profiling to locate hotspots, then use Nsight Compute (ncu) to identify specific bottlenecks. Validate optimization improvements by comparing profiling results before and after code changes.
Yes, you can inspect PTX and SASS code using cuobjdump. This binary inspection tool verifies kernel composition and resource usage, allowing you to analyze the compiled assembly and understand how your CUDA kernel maps to the GPU hardware.
The best workflow starts with a minimal CUDA kernel, runs compute-sanitizer to surface bugs, profiles with nsys and ncu to identify hotspots, and iteratively applies traces to confirm fixes. This structured approach streamlines debugging and performance optimization across real-world workloads.
CUDA kernel crashes often stem from memory errors or thread synchronization issues. Reproduce the issue by running compute-sanitizer to surface bugs, using cuda-gdb batch execution to isolate the fault, and applying printf-based device code traces to confirm the exact failure point.
Yes, Nsight profiling tools like nsys and ncu are essential for locating hotspots and bottlenecks in GPU kernels. They provide non-interactive profiling capabilities that validate optimization improvements by quantifying performance changes across real-world CUDA workloads.
m0atQuery NVIDIA CUDA documentation for kernel development and debugging.
tamago0224Diagnose Go performance bottlenecks from pprof and fgprof profiles.
m0atAnalyze CUDA kernel performance bottlenecks with Nsight Systems and Nsight Compute.
igamenovoerConfigure CUDA build environments for Pixi projects with CMake and nvcc.
