v0.1.1 — SSA Pipeline Stabilization + 79.2 TFLOPS GEMM

🚀 T0-GPU v0.1.1 — SSA Pipeline Stabilization + 79.2 TFLOPS GEMM

A major milestone: the T0 compiler's SSA pipeline is now stable and producing 79.2 TFLOPS bf16 GEMM on RX 7900 XTX — on par with Triton on AMD, with zero external dependencies.

⚡ Performance

Config	VGPRs	4096³ TFLOPS	Status
128×128 k32	239	79.2	✅ Primary
128×64 k32	175	79.1	✅
128×128 k16	207	56.8	✅
128×64 k64	223	70.3	✅

🏗️ T0 Compiler (34 modules, ~38K LOC)

• SSA Pipeline: BlockDSL → SSA IR → 6-pass optimization (DCE, CSE, LICM, CopyProp, AlgSimp, Waitcnt) → SSA RegAlloc
• TileIR GEMM: Cooperative load + LDS double-buffer + K-loop pipeline + graduated LDS waits
• Gap Reclaim: Alignment-gap VGPR recovery, saving ~15 VGPRs (254 → 239 for k32, 64 spills → 0 for k64)
• Diagnostics: ISA verifier, GPU printf, HW probe (40+ instruction latencies), K-loop pipeline simulator

🔧 9 Critical Bug Fixes

• LICM dead code after terminator → GPU page fault
• CSE: 5 root causes (barrier-aware, inline constant key, sentinel MVal, VReg remap)
• DCE: loop-carried dependency preservation (backward branch detection)
• RegAlloc: max_vgprs 128→255, MVal interval merging
• KFD VA race: BufferPool LRU cache (zero-syscall buffer reuse)
• k48 hang: non-power-of-2 cooperative load address decomposition (assert added)

🛡️ KFD Runtime Hardening

• SIGPIPE defense (prevents queue leak on pipe close)
• KFD open 5× retry (covers MODE1 reset recovery window)
• GPU health probe (GTT write/readback verification)
• BufferPool LRU (eliminates VA remapping race)

📖 Documentation

• T0_技术手册.md: 1000+ line comprehensive technical manual
• 18 experiment records: Root cause analysis, performance optimization, correctness verification
• Tech blog: 从零到 79 TFLOPS

📊 Comparison

Engine	4096³ bf16 GEMM	Dependencies	Dispatch Latency
T0	79.2 TF	Rust only	2.26 μs
Triton	~80 TF	Python+LLVM+ROCm	~20 μs
rocBLAS	~105 TF	ROCm (WGP mode)	~20 μs

⚠️ Known Limitations

• WGP mode triggers CWSR page fault on RX 7900 XTX (Linux 6.17) — CU mode only for now
• tile_k=48 disabled (non-power-of-2 cooperative load bug, assert-protected)
• 128×128 k64 disabled (64KB LDS at CU limit causes occupancy hang)

---

Full Changelog: v0.1.0...v0.1.1