Howard Mao

Creek Vector Co-processor

Creek is a floating-point vector co-processor I developed in order to learn more about computer architecture and the Chisel hardware description language.

The basic functional blocks of the co-processor are streaming floating-point adder and multiplier units. Each of these units have four pipelined vector lanes and operate on single-precision floating point numbers. These functional blocks, along with a memory controller, are connected by a crossbar switch to seven vector registers (effectively small sequential-access memories).

Each vector register can support simultaneous reads and writes. This allows a higher degree of parallelism, but it becomes the programmer’s responsibility to schedule operations in such a way that a fast writer doesn’t overwrite the values a slow reader is trying to read (or, vice-versa, that a fast reader does not try to read values that a slow writer has not yet updated).

The co-processor also does not have any control flow instructions. Instructions are issued one after another until a special instruction which pauses the co-processor. This instruction allows the host processor to take back control, copy data between its memory and co-processor memory, write new instructions, and restart the co-processor.

The Chisel source code, Verilog FPGA glue code, and test software for this project is available on Github.