Paxzas CUDA Analyzer

Static analysis of NVIDIA CUDA kernels directly inside VS Code — no Python, no network, no GPU required. Point it at a PTX or SASS file and get occupancy, memory posture, bottleneck diagnosis, and instruction mix instantly, across every GPU architecture from Volta to Blackwell.

Features

Multi-Architecture What-if Analysis

Every analysis runs against all supported GPU presets in parallel. A dropdown in the panel lets you switch results between architectures instantly — no re-running needed. Presets derived from actual SASS binary targets are marked as native; all others are what-if estimates using the same instruction profile with different SM limits.

Supported architectures:

Occupancy Model

• Computes warp occupancy, blocks per SM, and the limiting factor (registers, shared memory, threads, or block limit) for any block size
• Sweep charts — occupancy, blocks/SM, and per-constraint resource limits plotted across all valid block sizes
• Waste metrics — unused threads, warps, registers, and shared bytes per SM at the current configuration
• Register what-if: how many fewer registers to reach the next occupancy tier
• Launch parameter inference from PTX hints and optional SASS register index

Bottleneck Diagnosis

• Fuses memory posture, stall profile, and pattern class into a primary and secondary bottleneck with the firing rule that triggered it
• Four stall profile flags: memory dependency, memory throttle, local memory (register spill), sync overhead
• Per-bottleneck optimization suggestions ranked by impact

Memory Model

• Classifies the kernel as memory-bound or compute-friendly
• Arithmetic intensity (ops/byte), reuse ratio, cache policy, load/store balance and vectorization score
• Distinguishes global, shared, and local (spill) traffic from SASS; falls back to PTX heuristics when SASS is absent

Pattern Model

• Classifies as tiled, streaming, reduction, compute_heavy, or mixed
• Detects archetypes: GEMM, CONV, ELEMENTWISE, STENCIL, and others
• 15 micro-flags: register spill, uncoalesced loads/stores, atomic contention, missing tensor cores, SFU-heavy, over-synchronized, FP16 scalar, warp divergence, and more
• Warp primitive detection: shuffle, vote, and reduction patterns

Instruction Mix (SASS)

• Category breakdown: arithmetic, tensor, SFU, global mem, shared mem, local mem, control/sync
• Vectorized load/store sub-counts (LDG.128 / LDG.64 / LDG.32, STG equivalents)
• Tensor core op counts (WMMA / HMMA)
• Atomic and warp-primitive counts
• Productive instruction fraction and tensor utilization fraction

Roofline Chart

• Plots the kernel's arithmetic intensity against the FP32 roof and bandwidth slope for the selected GPU
• Region classification: memory-bound or compute-bound with a ridge-point marker

Raw Feature Inspection

• Side-by-side PTX vs SASS instruction counts for every extracted feature
• Rows with differing values highlighted; column headers adapt to the available data source

Commands

Both commands are available from the Command Palette, editor title bar, editor right-click, and Explorer right-click on .ptx, .cu, and .sass files.

Settings

paxzas.gpuPreset — default GPU for the capability dropdown.

auto detects the local GPU via nvidia-smi when available; named presets force a specific architecture. Regardless of this setting, all presets are always shown in the panel dropdown.

Requirements

• VS Code ≥ 1.85
• Optional: nvidia-smi on PATH for automatic GPU detection in auto mode

Development

npm install
npm run compile   # or: npm run watch
npm test

F5 in VS Code (with this folder open) launches an Extension Development Host.

Build a .vsix

./build-vsix.sh
# or
npm run vsix

Install locally: Extensions → ··· → Install from VSIX…

Repository

github.com/CudaPaxZas/PaxZas