flare is a high-performance tensor processing basic component library of EA. In the field of vector search technology,
search, recommendation and advertising will eventually settle on the calculation of tensors. One aspect of flare is to
abstract tensor operations to an operation plane, such as tensor distance calculation, delay expression (to facilitate
operator fusion in the future), matrix multiplication and other basic operations. On the other hand, the tensor operations
of different devices are abstracted to an operation plane, such as CPU and CUDA. When the business layer is used, it
reduces the consumption of the system design layer and focuses on the design of algorithms and the improvement of
algorithm quality.
- High performance
- Support cpu and gpu
- Standard math
- Vector algorithms
- Machine learning
- Linear algebra
- unified api
flare also can work as a plugin to hercules.
flare provides software developers with a high-level abstraction of data that resides on the
accelerator, the flare::tensorobject. Developers write code that performs operations on flare
arrays, which, in turn, are automatically translated into near-optimal kernels that execute on
the computational device.
When designing flare, we fully considered the convenience of the deployment in the production environment and the R&D environment. Considering different dimensions, we provide corresponding solutions, especially in the R&D environment based on large-scale servers and the consistency of deployment. For detailed compilation and installation process, please refer to build and install
- tann knn search application.