TensorSuite
A real and synthetic sparse tensor suite: curated real-world sparse tensors with element-wise and block-wise patterns, plus synthetic tensor generators for benchmarking under configurable size, density, and memory constraints.
What We Build
A real and synthetic sparse tensor suite: curated real-world sparse tensors with element-wise and block-wise patterns, plus synthetic tensor generators for benchmarking under configurable size, density, and memory constraints.
Cross-layer coordination and optimization for scalable and sparse tensor networks. TensorWorld leads NSF-supported work on abstractions, systems, and performance portability for tensor network workloads.
Hierarchical optimization of ragged (irregularly shaped) tensor operators for deep learning workloads, spanning representations, scheduling, and code generation across the software stack.
Formally verified and performance-optimized tensor contraction sequences for quantum many-body computations, combining formal methods with high-performance code generation.
We design algorithms, data structures, and runtime support for sparse tensor contraction, decomposition, and related kernels on CPUs, GPUs, and emerging architectures.
Our group builds profiling and analysis tools for modern ML workloads, with an emphasis on data types, memory behavior, GPU execution, and framework-level performance bottlenecks.
We study performance optimization for irregular algorithms in graph, hypergraph, tensor, and scientific computing workloads, including auto-tuning and architecture-aware implementation strategies.