FlyDSL
Author and tune high-performance AMD GPU kernels in Python
All Skills in This Repository (14)
Pure Emerald Level Indicators
ROCmport-to-layout-api
Port FlyDSL GPU kernels from raw buffer_ops to the layout API.
ROCmformat-code
Format Python and C/C++ files by removing unused imports and variables.
ROCmbuild-rocm-image
Build a ROCm development Docker image on a remote host via SSH.
ROCmflydsl-kernel-authoring
Author AMDGPU kernels with FlyDSL layout algebra and MLIR lowering.
ROCmbuild-flydsl
Automate building and installing FlyDSL from source on remote hosts or containers.
ROCmbisect-perf-regression
Automate Git bisect to locate the commit introducing GPU kernel performance regressions.
ROCmdebug-flydsl-kernel
Diagnose incorrect FlyDSL GPU kernel outputs, NaNs, infs, and crashes.
ROCmadd-target-atom-op
Add target-specific MmaOp/CopyOp to a FlyDSL backend dialect.
ROCmcapture-kernel-trace
Capture and analyze GPU kernel ATT traces using rocprofv3.
ROCmkernel-trace-analysis
Analyze ATT traces to profile GPU kernels and identify stall hotspots.
ROCmlds-optimization
Diagnose LDS bank conflicts and lgkmcnt stalls in FlyDSL GPU kernels.
ROCmprefetch-data-load
Introduce loop-carried prefetch to overlap data loads with compute in FlyDSL GPU kernel loops.
Frequently Asked Questions
FAQPage SchemaHow to install FlyDSL?▼
Run `npx skills add ROCm/FlyDSL --all -g -y` in your terminal to install all skills in this suite globally.
What is FlyDSL used for?▼
FlyDSL is a Python DSL and MLIR compiler stack for writing high-performance AMD GPU kernels with explicit control over layouts, tiling, and data movement. It ships production kernels for GEMM, attention, MoE, normalization, and convolution.
How do I debug a FlyDSL kernel that returns NaN or wrong results?▼
The included debugging skills walk you through clearing stale JIT caches, isolating faults with all-ones inputs, and checking layout and MFMA operand ordering. They also cover out-of-bounds detection using static interval analysis and runtime checks.
Can I profile and fix GPU kernel performance bottlenecks?▼
Yes. The suite captures rocprofv3 instruction-level traces, identifies stall hotspots like memory waits and bank conflicts, and produces a prioritized optimization plan. It can also bisect git history to find the exact commit that caused a regression.
Do I need GPU kernel experience to use these skills?▼
No. A step-by-step wizard guides you from a plain requirement to a tested kernel, while reference skills handle the low-level layout and hardware details for you.
Related Repositories in Software Engineering
View All in Software Engineering→
openclawopenclaw
Run a personal AI assistant across your devices and chat apps
obrasuperpowers
Gives coding agents a disciplined workflow from idea to merged code
facebookreact
AI agent skills for building, testing, and porting React core