Maxeler Open Stream Processing (OSP)

Key observation (from DJ Greaves): Much high-performance computing on FPGA is limited by I/O bandwidth, so let's make sure we use 100 percent of that and factor everything else around it.

No program counter - Stream Processing processes the stream in stream order.

Data flows as fast as possible to/from DRAM bank or fileserver - good. Programs need to be linearised to wrap around the stream.

• Transmit (T) the entire input to the program,
• Compute (C) sophisticated functions on large pieces of input at the rate it is presented,
• and Store (S), capture temporarily or archive all of it long term generator and reducer conver from and to scalars (or small vectors) respectively.

Trivial example : blurrer: compute the local 3-average:

HWType flt = hwFloat(8,24);
HWVar x = io.input(”x”, flt ) ;
HWVar x_prev = stream.offset(x, − 1);
HWVar x_next = stream.offset(x, +1);
HWVar cnt = control.count.simpleCounter(32, N);
HWVar sel_nl = cnt > 0;
HWVar sel_nh = cnt < (N − 1);
HWVar sel_m = sel_nl & sel_nh;
HWVar prev = sel_nl ? x_prev : 0;
HWVar next = sel_nh ? x_next : 0;
HWVar divisor = sel_m ? 3.0 : 2.0;
HWVar y = (prev+x+next)/divisor;
io.output(”y” , y, flt ) ;