The target audience of L2 API are users who have a certain understanding of HLS and programming on FPGAs and want to make modifications on kernels including:
- Operator combinations in the kernel, such as the number of operators or operator pipeline. Most L1 APIs and glue logics inside L2 kernels are connected with streams. You could add more operators into kernels to increase its performance as long as it is not bounded by other factors like logic resource and memory bandwidth.
- Add more pre-processing and post-processing to kernel, such as compression/decompression. Certain DBMS might store its data in a compressed format to save the memory space and bandwidth to transfer. Adding a decompression module to build longer processing pipeline will save the time to decompress the data and increase system performance.
- Kernel combination to be built on single cards. You might be interested in adjusting the kernel combination to accelerate the execution plan with different weights.
- Building configs to improve frequency or migrate to other AMD FPGA devices.
- Test correctness in sw-emu, hw-emu, and on-board.
