On 01/22/21 14:14, Bruce Perens wrote:
The embedded Vision Accelerators aren't really graphics accelerators. They do some important things for DSP that we could potentially leverage, but I haven't been able to tell in short look if they are actually documented so that you can program them, or just are hardware accelerators for a proprietary thing they call Vision pack. See https://ieeexplore.ieee.org/document/6865062
The EVE is optimized to deal with pixel data, which could be input or output. From the paper you cited:
"ARP32, a 32-bit RISC core, is the master controller and programmable frontend of EVE providing the user a full ANSI-C level programming environment complete with state- of-art solution for debug and tools support for easy software development. "
"VCOP (shown in Figure 2), the Vector/Vision CO- Processor, is an 8-way SIMD engine with two parallel and identical functional units (FUs) per SIMD lane. "
You're right in that it doesn't draw circles, do fills, etc. as primitive operations.
When I speak of AI, I mean hardware Convolutional Neural Networks (CNNs), like this:
--- Zach N0ZGO