Compiler-Native Sparse Optimization and Automatic Design Selection
MLIR-integrated analysis and selection.
OSU STAR Lab · NSF CROSS Project · 2027
From measured prototype
to compiler capability.
Generalize MLIR-native sparse optimization and automatic architecture selection across workloads, formats, toolchains, and parallel systems.
Programme goals
Three directions define 2027
Generalization Across More Contractions, Formats, and Toolchains
More contractions, formats, and backends.
Scaling Parallel Sparse Execution and Memory Systems
Parallel execution and scalable memory.
Toolchain contract
The model specifies semantics; the compiler selects the architecture
01Tensor contraction
→
02Common sparse MLIR
→
03Profile + legality
→
04Design selection
→
05Backend lowering
→
06Accelerator + QoR
Success criteria
Evidence required for a reusable toolchain
Target
Profile-guided selection consistently outperforms fixed schedules across contraction families and sparsity distributions.
Target
HIDA, the STAR Lab framework, and Stream-HLS consume common program, profile, and design artifacts through isolated adapters.
Target
Maintain the matched CPU/GPU/FPGA evaluation contract while scaling parallel designs and memory systems; physical comparison is a 2026 evaluation prerequisite.
Target
A versioned manifest reproduces each published design, generated source, tool invocation, and QoR report with one command.