Diagnose LDS bank conflicts and lgkmcnt stalls in FlyDSL GPU kernels.
npx skills add https://github.com/ROCm/FlyDSL --skill lds-optimization-rocm
Please help me install this Agent Skill. Skill: lds-optimization Source: https://github.com/ROCm/FlyDSL/tree/main/.claude/skills/lds-optimization Command: npx skills add https://github.com/ROCm/FlyDSL --skill lds-optimization-rocm
Optimize LDS (Local Data Share / shared memory) usage in FlyDSL GPU kernels by diagnosing bank conflicts and lgkmcnt stalls, then applying swizzle or padding layouts and increasing write-read distances to hide latency. Use when trace data shows ds_read/ds_write/lgkmcnt bottlenecks.
Run /kernel-trace-analysis to identify LDS bottlenecks, then apply swizzle or padding and re-run analysis to verify improved stalls.
High-intent search queries and answers about installing and using this skill.
Bank conflicts in GPU shared memory are reduced by applying XOR-based swizzle or padding layouts to break strict bank alignment, improving memory access throughput across wavefronts.
Lgkmcnt stalls in GPU kernels are caused by data dependencies waiting on ds_read or ds_write instructions to complete, which can be mitigated by increasing write-read distance to hide latency.
LDS write latency is hidden by increasing the distance between write and read instructions, allowing the operation to overlap with independent compute work and reduce cross-wave barriers.
Apply XOR-based swizzle vs padding to local data share based on trace analysis diagnosing specific bank conflict patterns, ensuring changes maintain correctness while achieving measurable performance gains.
Diagnose LDS bottlenecks from GPU kernel trace data by running kernel-trace-analysis to identify ds_read/ds_write latency, bank conflicts, and cross-wave barriers causing lgkmcnt stalls.
Optimizing LDS usage requires re-running trace analysis to verify improved stalls and confirm that swizzle, padding, or write-read distance changes achieve measurable performance gains without breaking correctness.
fsx950223Diagnose LDS bank conflicts and latency stalls from GPU kernel trace data.
JylhisIdentify performance bottlenecks in Rust projects and recommend optimizations.
ROCmOptimizes GEMM kernels on AMD CDNA GPUs using FlyDSL tiling, memory layout, MFMA scheduling.
xchang1121Optimize GPU memory access patterns in TileLang CUDA kernels.
