SparseFlow v0.3-alpha — GPU Lowering Pipeline (Kernel Stub)

🚀 SparseFlow v0.3-alpha

This is an architectural alpha release that completes SparseFlow’s end-to-end GPU lowering pipeline.

This release validates the compiler flow from structured sparsity analysis to GPU IR generation using MLIR.

---

✅ What’s Included

Compiler Passes

• sparseflow-spa

  • Static N:M sparsity propagation at compile time

• sparseflow-rewrite-matmul

  • Rewrites sparse linalg.matmul into backend calls

• sparseflow-gpu-rewrite

  • Lowers sparse calls into:
    • gpu.module
    • gpu.func (kernel stub)
    • gpu.launch

Verified Pipeline

→linalg.matmul
→ sparseflow-spa
→ func.call @sparse_matmul_N_M
→ gpu.launch + gpu.func

The full pipeline executes successfully with verifier-clean MLIR.

---

⚠️ Limitations (By Design)

• GPU kernel contains no computation (gpu.return only)
• No bufferization or memory lowering
• No correctness or performance claims on GPU
• Output tensor is currently a placeholder

This release is about compiler correctness and architecture, not performance.

---

🧱 Why This Matters

SparseFlow now has:

• Stable pass registration
• Stable plugin loading
• Verified CPU → GPU lowering boundary
• A clean foundation for real GPU kernel development

This is the last structural milestone before implementing GPU kernels.

---

🔜 What’s Next

v0.3-beta

• Define GPU kernel ABI
• Memory layout and bufferization
• Replace stub kernel with real N:M computation

---

📄 License

MIT

Full Changelog: v0.7.1...v0.3-alpha

SparseFlow v0.7.1 – SPA + CPU Runtime + Python CLI (Dev Preview)

SparseFlow v0.7.1 – SPA + CPU Runtime + Python CLI (Dev Preview)

This release packages the current SparseFlow stack into a usable, reproducible demo
for researchers, compiler engineers, and hardware partners.

Highlights

• SPA v0.6 (2D Sparsity)

  • Propagates row + column sparsity through a small MLIR pipeline
  • Exports JSON metadata with row/col masks and sparsity stats
  • Proven on structured sparse matmuls with 50% row + 50% col sparsity (≈75% FLOP reduction)

• C++ CPU Runtime (OpenMP)

  • Consumes SPA JSON masks and skips zero rows/cols
  • Blocked matmul kernel with OpenMP parallelism
  • Achieves ≈3–5× speedup on larger matmuls (512–1024) at ~75% sparsity
  • Honest behavior: small sizes (<512) can be 1–3× due to OpenMP and cache overhead

• End-to-End Demo Script

  • ./spa-runner.sh:
    • Rebuilds passes/runtime if needed
    • Runs SPA on test MLIR
    • Exports spa_sparsity.json
    • Benchmarks dense vs sparse CPU matmuls

• Python CLI (Developer Preview)
  Install:

  cd sparseflow_package
  pip install -e .
  

  Tools:

  sparseflow-demo
  sparseflow-analyze tests/test_spa_v6_full_2d.mlir
  sparseflow-benchmark
  

Quick Start

git clone https://github.com/MapleSilicon/SparseFlow.git
cd SparseFlow
./spa-runner.sh

SparseFlow v0.1.1 — JSON-Driven N:M Runtime + Dimension Support

This release includes:

• Full MLIR pass pipeline (AnnotateNmPass, ExportMetadataPass, etc.)
• JSON-driven compiler → runtime integration
• Dimension-aware runtime (M, N, K correctly read from metadata)
• Sparse matmul simulator with N:M skipping logic
• End-to-end validation scripts
• Matrix + sparsity sweep tool