Add SOFA+ Architecture with fracturable 18x18 multiplier

[tangxifan]

Checklist

• [X] Added the VPR architecture XML for the SOFA+ architecture, which is a heterogeneous FPGA with integrated fracturable 18x18 multipliers.
  • [X] Each fracturable 18x18 multiplier can either operate as a 18x18 multiplier or two independent 9x9 multipliers.
  • [x] Each flip-flop can operate in multiple modes, e.g., latch, d-type, d-type with active-high reset, d-type with active-low reset.
  • [x] Removed shift register mode since it is not used most of the time.
  • [x] Use scan-chain flip-flop as configurable memory.
  • [x] Use custom LUT netlist to implement LUT with carry logic
  • [x] Use timing values in SOFA architecture
  • [x] Build timing variable files
• [X] Added the associated OpenFPGA architecture XML for the SOFA+ architecture.
• [X] Added technology library files required to synthesis these DSP blocks using Yosys.

Screenshots on technical details

• [X] A detailed view on the SOFA+ FPGA architecture (routing architecture not shown), where the green blocks are the DSP blocks
• [X] A detailed view on the SOFA+ FPGA architecture (routing architecture is shown). There are routing tracks passing through the DSP blocks to enhance routability. Each block has only pins on the top and right sides, to keep a granularity on the tile-based architecture.

[tangxifan]

@ganeshgore I plan to merge this PR and then start creating a SOFA+ directory like other SOFA project. Still a few work to do:

• Think about how many clocks we want to support
• Provide all the custom netlists, LUT, CCFF and custom data FF
• Upload all the benchmarks we have to test the functionality of the architecture, DSP, adder etc.
• Create a regression test for the benchmarks so that we can run sign-off easily

I also plan to increase the routing channel width to 80 or even 100, depending on the benchmarks we want to implement. We may reduce the array size to 10x10 to fit the area budget.

[tpagarani]

@tangxifan Channel width of 100 may be an overkill for 10x10 layout if you have crossbar. If you don't have crossbar, you may need to push up to 100.

[tangxifan]

@tangxifan Channel width of 100 may be an overkill for 10x10 layout if you have crossbar. If you don't have crossbar, you may need to push up to 100.

@tpagarani Indeed. In this version, we do not have crossbar. Considering the large number of inputs and outputs of a multiplier, increasing the routing channel width to 100 looks very necessary.

Another concern is that even when we have crossbar in SLC, the nets connected to multipliers may cause congestions (think about the number of pins of a SLC is far more than a SLC). We may have to increase the routing channel width anyhow but will definitely so significant as the one without crossbar.