PantheonSim v0.2.0

Getting started

• make installs what is missing (it asks first) and builds; ./run starts a simulated GPU machine from a
  settings file. The README is those two commands; the rest is in docs/getting-started.md, docs/config.md
  and docs/cli.md. The simulator uses half your CPUs unless you say otherwise.
• scripts/ci-setup.sh does for GitLab CI, CircleCI, Forgejo, Jenkins and Buildkite what the GitHub Action does
  (docs/ci-other-platforms.md).
• Every document was checked against the code, and the profile counts, command lists, ROCm versions and
  limitations were corrected. 48 profiles (38 NVIDIA, 10 AMD); 14 NVIDIA and 3 AMD are read from physical cards.
• CONTRIBUTING.md, CODE_OF_CONDUCT.md, issue templates (bug, something unsupported, a wrong GPU profile) and a
  pull request template say how to send a change and what it needs.

Faster

• The SASS engine's pre-decoded tier is on by default; the AMD engine has a tier of its own, and matrix and
  packed-math fast paths. Atomics, kernel launch and the first launch of a fatbin cost less. A second tier that
  compiles hot straight-line code (VGPU_JIT=1) is available and off by default. Every fast path produces
  the same bits as the path it replaces, and the old path stays available through the documented VGPU_*
  switches so the two can be compared (docs/cli.md).
• Blocks that wait on one another (grid-wide spins, decoupled look-back) now run together, taking turns when
  they cannot all be resident. This applies to SASS, PTX and AMD code.

NVIDIA

• New profile: RTX PRO 4500 (read from a card).
• CUPTI traces were compared with the real thing on 19 rented GPUs; torch.profiler traces are checked against a
  card in CI. VGPU_PLATFORM=linux switches CUPTI to the answers a Linux driver gives (the default is wsl,
  nvidia/docs/cupti.md).
• SASS: a cloned device function runs its own copy; a yielding warp rotates through its lane groups.
• LuisaCompute's CUDA backend runs (scalar suld/sust, txq, the device runtime library in cuLinkAddData);
  cuTensorNet and cuda-samples fixes.

AMD

• rocprofv3 traces of small HIP programs match what it recorded on a real RX 6800 and RX 6700 XT.
• gfx942 and gfx950 were swept against AMD's instruction guides; gfx90a images, kernel-to-host calls (hostcall),
  PC sampling, the ROCm Validation Suite and vLLM graphs run (amd/README.md).

Known limits

• Functional only: there is no timing model, so a run on a physical GPU before shipping still counts.
• Programs must link the shared CUDA runtime (nvcc -cudart shared); a statically linked runtime needs NVIDIA's
  undocumented driver export tables (this is why CuPy does not work).
• Nsight Systems and Nsight Compute as NVIDIA ships them, and OptiX, are not supported. See docs/status.md.

PantheonSim v0.1.0

First tagged release of PantheonSim: simulated NVIDIA and AMD GPUs, so CUDA and HIP programs run unmodified on a machine with no GPU (development, teaching and CI).

Ways to use it

• GitHub Actions: pantheongpu/setup-pantheonsim (@v0).
• Container: docker run --rm ghcr.io/pantheongpu/pantheonsim:0.1.0 vgpu list-gpus (CUDA 13.0 build; 0.1.0-cuda12.8 for CUDA 12.8). See docs/docker.md.
• From source: cmake -S . -B build -G Ninja && cmake --build build.

The simulator is functional, not cycle- or bit-accurate, so it complements testing on real hardware and does not replace it. NVIDIA's cuDNN headers are not part of this repository; the build downloads them from NVIDIA's nvidia-cudnn-cu12 wheel.