OpenVX Bring Power-efficient Vision Acceleration to Mobile

OpenVX is the next open source sample specification to be launched by Khronos group, a consortium building a family of interoperating APIs for portable and power efficient vision processing. If you take a look at the OpenVX participant list, you can check that the major chip makers: Broadcom, Qualcomm, TI, Intel, Nvidia, Renesas, Samsung, ST or Xilinx, as well major IP vendors like ARM or CEVA have joined the Khronos group. Before discussing OpenVX benefits, we may comment this “Why do we NEED Standards?” slide:

These four bullets may look like obvious affirmations, but it’s worth to remind that the semiconductor industry need standard Interfaces for Interoperability, and that widely adopted standard interfaces is growing market opportunity, in such a way that it turn into a mass market, so devices can be built cheaply enough to attract even more customers. This is a kind of virtuous cycle, industries cooperate to build a market (looks naïve?)… And then compete, which is more realistic. And because today the largest market is by far the mobile industry, OpenVX has been developed to support computing vision for mobile and embedded.

Computing vision is extremely data intensive, and OpenVX bring a sophisticated image stream generation:

• Advanced, high-frequency burst control of camera and sensor operation
• Portable support more types of sensors – including depth sensors
• Tighter system integration – e.g. synch of camera and MEMS sensors

OpenVX allows developing advanced imaging & vision application, like:

• Image enhancement,
• Object tracking and detection, and
• Image manipulation

In fact, OpenVX can be implemented with CPU, GPU or DSP cores, but the goal is not only to accelerate performances but also to drastically minimize power consumption, as the primary target is mobile. Thus, implementing OpenVX on CEVA DSP core is probably the best option, allowing decreasing by 10X power consumption, when compared with CPU. Just remember that CEVA is supporting Android Multimedia Framework (AMF), a system level software solution allowing offloading of multimedia tasks from CPU/GPU to most efficient application-specific DSP platforms. The next picture illustrate OpenVX Development Flow, this example being for an Android OS using CEVA AMF:

Benefits enabled by AMF include:

• Multimedia tasks are abstracted from the CPU and are physically running on the DSP. Furthermore, tasks can be combined (“tunneled”) onto the DSP – saving data transfer, memory bandwidth and cycles overhead on the CPU
• Utilization of the PSU (Power Scaling Unit) available for CEVA DSPs to significantly lower power consumption further, when running multimedia tasks
• Easy activation of CEVA-CV computer vision (CV) software library for the development of vision-enabled applications
• Easy activation of yet to be officially released OpenVX pre-optimized software library

A chip maker developing IC for mobile has to search for differentiation, and leaders on the mobile market appear to be also the companies strongly investing to propose differentiated solutions, as we have explained in this blog for example. Is OpenVX adoption a drawback for differentiation? In fact, it’s not the case: as you can see on the above OpenCV flow, OpenVX standardized API are available in open source. Thus, it’s possible to extend OpenVX library to create differentiation, still keeping the benefits of using a standardized interface solution. This is also true for CEVA DSP customers: they will benefit from OpenVX standard interface, supported by CEVA, from the drastic reduction of the power consumption associated with DSP core usage when compared with CPU or GPU and this customer could also build a differentiated solution by extending OpenVX library with specifically developed API.

Eric Esteve from IPNEST –

lang: en_US

More Articles by Eric Esteve …..