Vitis Hardware AccelerationSee Vitis™ Development Environment on xilinx.com |
In this section…
Perform C synthesis to create an RTL representation of the design and discuss the compiler hints (pragmas) used…
Run Vitis HLS C Synthesis¶
In the previous phase, the C simulation, the code execution was purely C based and although special libraries are used for streams, no RTL was yet involved.
In C synthesis, the top function tsp
is analyzed and compiled based on the hints (called pragma or directives) passed to the HLS compiler. Once the operations are scheduled and mapped onto hardware constructs, the final code is generated in RTL (with both Verilog and VHDL).
The code uses 3 of these “hints”:
PIPELINE
: Requests execution of the main loop (labeledloop_compute
) at each clock cycle as specified by theII=1
optionINLINE
: Dissolves a sub-function for better optimization results. This is used for thecompute
functionINTERFACE
: Specifies a protocol for a given top function argument. This is optional and only to demonstrate how AXI-Stream can be added to an HLS streamBIND_STORAGE
: Assigns an array to a specific type of on-chip memory. Here thedistance
array is mapped onto a RAM with 1 write port and multiple read ports to allow simultaneous access to multiple city to city distances and calculate the full route quicker.
Before running synthesis open the tsp.h
file and set the number to 13 (N=13
). You can open the tsp.h
file in the Vitis HLS GUI from the tsp.cpp
file by holding the mouse over the #include "tsp.h"
line and press CTRL
key while clicking the mouse.
To run HLS synthesis from the GUI:
Use the same shortcut as we used for C simulation earlier and select ‘Run C Synthesis’ or…
Via the main menu, Project
-> Run C Synthesis
- > C Synthesis
Once synthesis has completed, the main window shows “Performance and Resource Estimates”:
(collapse the “General Information” and “Timing Estimates” sections by clicking on their title to make more room if necessary)
This “Performance and Resource Estimates” section shows a table in which we see the main function tsp
and the main loops and since we gave them a label in the source code it’s easy to know which they are.
The full latency for the
tsp
function is 479,001,957 (close to half a billion clock cycles) mainly contributed byloop_compute
That latency of 479,001,600 is exactly factorial 12 (12!) which corresponds to the the scenario with 13 cities (N=13) given that the first city (the route starting point) is fixed, so the permutations are applied on the 12 remaining cities.