Based on Vivado HLS 2019.2, implement High-Performance Karatsuba Multiplier in High-Level Synthesis (HLS) for FPGA Based on Recursive Template. The design provides parameters which can help user to customize the multiplier according to different requirements of operand bitwidth, performance and resource cost.
The utilization of this open source project should follow the GPL-3.0 License.
(^_^) if this project helps you, a STAR will be encouragement.
When handling 4096x4096=>8192, the performance and resource cost are shown below (left is for non-dataflow design, 133333x(4096x4096=>8192 multiplication) per second ; right is for dataflow design, throughput increased by 27.0%, 166666x(4096x4096=>8192 multiplication) per second ):
non-dataflow design located in the directory "HLS_Implementation"
dataflow design located in the directory "HLS_Implementation_dataflow"
Use template in C++ to generate recursive multiplication tree during compilation which can make full use of parallelism of Karatsuba multiplier
each loop are pipelined.
One of the versions provides dataflow processing to increase throughput. However, although the design is validated by me on VivadoHLS 2019.2, dataflow implementation in VivadoHLS is not very reliable currently. If you use another version of VivadoHLS, please run C/RTL co-simulation before deploying the design in your project.
Detailed comments can be found in the source code.
the outline of the design is shown below:
located in the directory "Software_Implementation"
the same as HLS version
but generate random input and use GMP library to verify the results