hughperkins
commented
7 years ago Typically FPGAs work slightly differently from discrete GPUs, in that the programming time is very long, hours.
For a discrete GPU, such as AMD, or NVIDIA, the workflow for OpenCL looks something like:
- program starts
- program initializes GPU device
- program loads OpenCL source-code, which looks like C code basically
- program gives OpenCL code to GPU driver, which compiles it to GPU object code, passes it to the GPU, and passes a handle back to the program
- program gives program handle to the GPU driver, along with some data, and GPU starts processing the data, following the logic in the program
For a discrete GPU, steps 1 to 3 take ~seconds. Step 4 takes ~seconds, or less. Step 5 takes as long as it takes. minutes/hours/days/weeks, depending on what you're doing/training.
For an FPGA, step 4, takes significantly longer. Hours instead of seconds. So, the workflow would be quite different. The compilation of the OpenCL has to happen offline essentially, rather than at runtime.
It's probably not a massively blocking change, but it would need rethinking somewhat how the program runs. For example, EasyCL currently assumes that hte OpenCL will be compiled at runtime. You'd need to partition EasyCL into two parts:
- offline compilation part, that you'd run once, during compilation, and would store the FPGA program somewhere. You'd then write the FPGA program to the FPGA I believe
- runtime bit, which would simply execute the program on the already-programmed FPGA
tonyzhenyuxu
lihao2333
Would you please give me your comments on supporting OpenCL running on FPGA device instead of GPU such as Altera Arria 10? Thanks.