Supported DSLs

Agent skill: /gwatch_cuda_xtrace

Xtrace inserts probes at one of two levels of the compilation stack.

Tip: SASS-level instrumentation offers the highest-fidelity tracing.

Where Xtrace inserts probes

┌──────────────┐
│ Source code  │ ◀── PTX tier: scope markers
└──────┬───────┘
       ▼
█ DSL compiler █
       ▼
┌──────────────┐
│     PTX      │ ◀── PTX tier: markers become probes
└──────┬───────┘
       ▼
█    ptxas     █     compiles the probes on the PTX tier
       ▼
┌──────────────┐
│     SASS     │ ◀── SASS tier: probes spliced into the cubin
└──────┬───────┘
       ▼
   execution

PTX tier   instrumentation_tier="ptx"    CUDA, TileLang, CuTeDSL, Triton
SASS tier  instrumentation_tier="sass"   SASS (no markers, no rebuild)