OpenCL/HLS/Verilog Convolution Implementation Comparison Study

OpenCL

Host: .cpp          Kernel: .cl

Host: 131

Kernel: 31

Total Time: 6hours approx.

Host: 4hours approx.

Kernel: 3hours approx.

This time includes the time taken to write and successfully compile the code.  Does not include the time to remove logical errors.

Total Time: 15 hours approx.

Host: 6hours approx.

Kernel: 4hours approx.

This time includes the time taken to remove logical errors, to understand how to create .exe and .xclbin file and how to force SW_Emulation on SDx. Also includes the time taken to perform multiple test iterations until the right output was achieved

Host(Written in SDx): 154 

Kernel(Written in HLS): 42

Total Time: 2.5hours approx.

Kernel: 20 mins. approx. Same as used in OpenCL. Added a few pragmas to change the kernel into an AXI interface 

Host: 2 hours approx.

Wrote the host code using ap_int datatypes but got errors and hence had to use standard C/C++ datatypes to resolve the issue

Time to debug: 2.5hours approx.

Kernel: 30 mins. approx.

Received a few errors in mismatch between the depth of FIFO and the array size in the kernel. Made changes and ran C/RTL co-simulation until it passed on HLS. 

Host: 2 hours approx.

Spent some time in writing the host code. Very similar to the one used in OpenCL. Got a few errors when used ap_int datatype. Hence reverted back to standard C datatypes. The majority of time spent in debugging was waiting for SDx to build the host and kernel code which took roughly a little more than 2 hours.