Skip to content

Latest commit

 

History

History
64 lines (44 loc) · 3.3 KB

README.md

File metadata and controls

64 lines (44 loc) · 3.3 KB

Fleet-Floorplanning

Clone the following 4 repos:

https://github.com/jjthomas/Fleet

https://github.com/jjthomas/Fleet-Floorplanning

https://github.com/jjthomas/Fleet-Floorplanning-RapidWright

https://github.com/jjthomas/aws-fpga (checkout branch minimal_single_ddr)

Update your ~/.bashrc or ~/.bash_profile with the following (make sure not to include trailing slashes in directory paths):

FLEET_RW_HOME=/path/to/Fleet-Floorplanning-RapidWright
export CLASSPATH=$CLASSPATH:$FLEET_RW_HOME/rapidwright.jar:$FLEET_RW_HOME/target/fleet-rw-1.0-SNAPSHOT.jar
export AWS_FPGA=/path/to/aws-fpga
export CL_DIR=$AWS_FPGA/hdk/cl/examples/cl_dram_dma

Install Maven and make sure it is pointing to Java 1.8+. Get the latest RapidWright standalone jar for your platform from https://github.com/Xilinx/RapidWright/releases and put it at $FLEET_RW_HOME/rapidwright.jar (we've tested on RapidWright 2019.1).

Run mvn package in the Fleet-Floorplanning-RapidWright directory.

Go to the Fleet repository. Install sbt. Edit StreamingWrapperDriver to generate a shell for the desired number of cores (180 default) and core IO interface. You can also insert a core other than the default Summer (e.g. Counter). The inserted core determines the core IO interface of the shell. Notice that Counter has a different IO interface than Summer (it consumes 8-bit rather than 32-bit input words), so the two cannot share the same shell. The smaller the core IO interface in terms of total bits, the more cores the shell can support without hitting routing congestion issues in the implementation phase below.

Now run sbt "runMain edu.stanford.fleet.StreamingWrapperDriver". A file called StreamingWrapper.v should be produced that has both the shell RTL and core RTL.

Go to Fleet-Floorplanning. Make sure Vivado 2018.2 is in your PATH. Run the following commands to generate a shell with sockets for the specified number of cores and core IO interface. The socket floorplan (default for 180 cores) can be found in floorplan.txt. The aws_build_dcp_from_cl_main.sh command will take a while and should be sent to a job submission system if available. Watch for any errors in these or the below commands.

./shell_flow.sh /path/to/Fleet/StreamingWrapper.v StreamingMemoryController /path/to/Fleet/StreamingWrapper.v Summer
cd $AWS_FPGA
source hdk_setup.sh
cd $AWS_FPGA/hdk/cl/examples/cl_dram_dma/build/scripts
./aws_build_dcp_from_cl_main.sh -strategy TIMING -io_dir /path/to/Fleet-Floorplanning

Now run the following command in Fleet-Floorplanning to generate core implementations for each socket column.

./kernel_flow.sh /path/to/Fleet/StreamingWrapper.v Summer

Finally, run the following command to insert cores into the sockets and generate a final dcp (Summer_combined.dcp).

./merge_flow.sh Summer

The kernel_flow and merge_flow steps can be run again with new cores without updating the shell, as long as the core IO interface and desired number of cores remains the same.

To run the standard monolithic Vivado flow to implement StreamingWrapper.v, do the following:

cp /path/to/Fleet/StreamingWrapper.v $AWS_FPGA/hdk/cl/examples/cl_dram_dma/design/sw.sv
cd $AWS_FPGA
source hdk_setup.sh
cd $AWS_FPGA/hdk/cl/examples/cl_dram_dma/build/scripts
./aws_build_dcp_from_cl_main.sh -strategy TIMING -monolithic_flow