TensorFlow™ optimized open source SYCL™ library makes porting existing code easier

Imagination Technologies has announced that developers using TensorFlow will be able to target PowerVR graphics processing units (GPUs) directly, thanks to a new, optimised open source SYCL neural network library, the first version of which will be commercially available in 2019.

The SYCL version of TensorFlow supports a large number of artificial intelligence (AI) operations (Figure 1) and is easy for users to customize on demand, which means that developers can run those high-performance networks immediately available on PowerVR by using the latest neural networks or their own AI research techniques. Because TensorFlow SYCL support is both open source and based on open standards, it is an ideal solution for developers who want to accelerate the latest AI technologies on low-power devices. 

Figure 1 - TensorFlow operations powered by SYCL

SYCL is a royalty-free, open standard alternative to the CUDA architecture that breaks down barriers between ecosystems, giving developers more freedom to write code in standard C++, unleash the performance benefits of GPU hardware and ensure code portability.

In addition, Codeplay’s SYCL library enables applications to seamlessly take advantage of the PowerVR-optimized application programming interface (API) for IMGDNN, Imagination’s proprietary neural network graph compilation library that helps developers get the highest performance from PowerVR GPUs and neural network accelerators (NNAs).

Key markets where this newly expanded ecosystem can be applied include: automotive, data centers, and smart cameras.

Neal Forse, senior director of product management at Imagination Technologies, says: “Working within an open standards framework gives developers peace of mind that their code will not become obsolete or need to be rewritten. By providing access to widely used PowerVR GPUs through SYCL, developers will have easy access to powerful compute resources under TensorFlow.”

Andrew Richards, CEO of Codeplay, said: “We are seeing tremendous growth in the adoption of the SYCL standard. Renesas has enabled SYCL in its R-Car automotive AI platform and now Intel is including SYCL in its One API. The increased standardisation for programming high-performance accelerators like PowerVR GPUs will enable AI software developers to bring advanced intelligence to everything from tiny low-power battery-operated devices to large supercomputers.”

SYCL builds on the concepts and efficiency of Khronos OpenCL™. SYCL libraries optimized for PowerVR including SYCL-DNN, SYCL-BLAS and Eigen will be available on GitHub.