KUPL
KUPL provides comprehensive support for multi-threading, data management, and matrix programming. It optimizes scheduling and synchronization algorithms and enables asynchronous data transfers, achieving peak performance with superior ease of use.
Many-Core Parallelism
- Inter-operator parallelism leverages graph- and queue-based dependencies to achieve dynamic load balancing across operators.
- Intra-operator parallelism enables automated task decomposition and concurrent execution for multi-dimensional data.
- Inter-core synchronization provides low-latency synchronization and reduction interfaces.
Data Management
- Asynchronous memory copy enables efficient data movement across tiered memory.
- Shared memory accelerates data movement between NUMA nodes.
- Tiered memory allocation enhances efficiency.
Matrix Programming
- Streamlined programming for matrix tiling, multiplication, and I/O (read/save)
Key Features
parallel_for
Enables multi-threading. It functions like the OpenMP parallel for directive, accelerating for loop operations using multiple threads. It offers both static and dynamic scheduling policies.
It defines operator parallel computing logic and can work with inter-operator parallelism of computational graph and multi-queue programming.
Multi-Queue Programming
Allows developers to define queues and events to orchestrate operator submission, multi-threaded execution, and dependency management.
Computational Graph Programming
Defines static graphs, which are submitted to dynamic execution engines to achieve multi-threaded parallelism and load balancing.