A 21st Century Sea Change Taking Place in Embedded Microprocessors (cont.)
This technique will be unfamiliar to most readers because it is not an option in systems where the processors are on different chips. The reason it works is that, when both cores are on the same silicon chip, there are circuit techniques for starting and stopping core processors that can be utilized. The key is that the start / stop process has to be very fast - on the order of one instruction execution time to be really effective. But when that can be achieved, the speedup in data transfer between core processors is dramatic and improves the throughput by a factor of several times. In effect, it completely eliminates the software signaling between cores for many types of data transactions.
If the core processors are designed to use memory-mapped I/O, even more interesting types of communication can occur between cores. In this system, I/O registers are treated as memory addresses which means that the same instructions that read and write memory also perform I/O operations. But in the case of multicore chips, there is a powerful ramification of this choice for I/O structure. Not only can the core processor read and execute instructions from its local ROM and RAM, it can also read and execute instructions presented to it on I/O ports or registers.
Now the concept of tight loops transferring data without the need for reading, testing, and writing status bits becomes incredibly powerful. It allows instruction streams to be presented to the cores at I/O ports and executed directly from them. And since the shared registers between cores are essentially the same as I/O ports, that means that one core can send a code object to an adjoining core processor which can execute it directly from the shared register with no need to actually transfer the code to the other processors local memory. Code objects can now be passed among the cores, which execute them at the registers. The code objects arrive at a very high-speed since each core is essentially working entirely within its own local address space with no apparent time spent transferring code instructions.
Pages: 1 2 3 4 5 6 7 8 9
