In part A, we write three simple assembly programs to implement three functions in example.c.
Based on ensuring correctness, we especially focus on the functional equivalence with the example C functions.
By selecting and placing labels in the assembly code appropriately, the code is also very readable.
In part B, we modify the HCL file of the Y86's sequential design to add a new instruction --- iaddl.
The following is the roadmap to finish this part:
- Clarify the computation process of
iaddand write it down at the beginning inseq-full.hcl. - Add dependency relations of
iaddlto all boolean signals. - Design the datapath for
iaddl, i.e., generate control signals forsrcanddst
We achieve full scores in the benchmark testing in just 2 hours, but we spent 2 more days researching all the potential methods to optimize the performance even further. The following is our roadmap:
- Change the order of the instruction sequence to avoid data hazard and structure hazards as much as possible, which leaves CPE = 12.96.
- Beyond the changes on instruction order, we apply loop unrolling to reduce the number of conditional check and registers updating, which leaves CPE = 9.83.
- Use a binary search tree to find the precise remaining number of elements after several rounds of unrolling to achieve complete unrolling, which leaves CPE = 8.95.
- Modify the pipeline design in HCL file to achieve 100% accuracy in branch prediction for certain code pattern, which brings CPE down to 7.79.