SPMDWhat a Reduction Loop Reveals About SPMD vs Per-Op Intrinsics
A side-by-side disassembly of the same AVX2 reduction reveals a structural advantage of whole-loop vectorization over per-operation intrinsics
Read More
SPMDWe Built Cross-Lane SIMD Primitives. None of Them Helped.
The most important negative result from our SPMD-for-Go proof of concept: explicit shuffles and rotations lost to compiler pattern detection on idiomatic Go
Read More
SPMDHow the Compiler Knows Your Load Is Contiguous
The most important backend optimization in SPMD: recognizing contiguous memory access through ChangeType and BinOp chains
Read More
SPMD16 Bytes That Saved a Thousand Branches
The cheapest optimization in our SPMD proof of concept: a WASM linear memory guard zone for safe vector overreads
Read More
SPMDByte Iteration at 32 Lanes: The Decomposed Index Path
How to iterate a []byte on AVX2 without drowning in index-register pressure
Read More
SPMDPattern Matching Outperformed Hand-Written SIMD
How compiler pattern detection on idiomatic Go outperformed explicit cross-lane SIMD builtins in our SPMD proof of concept
Read More
SPMDLoop Peeling: Where Most of the Speed Comes From
How SSA-level loop peeling enables the all-ones mask fast path that delivers ~2x of SPMD benchmark wins
Read More
SPMDHow SPMD Lives in the Compiler: Lessons from Building It
The mask-stack detour, predicated SSA, and why SPMD has to live at the heart of the compiler
Read More
SPMDWriting SPMD Go: A Practical Guide
How to think about uniform vs varying, write go for loops, use reductions, and avoid the common pitfalls
Read More