Skip to content

Latest commit

 

History

History
83 lines (59 loc) · 4.13 KB

compile_arch.md

File metadata and controls

83 lines (59 loc) · 4.13 KB
Raw

Compiling ThunderGP

This page provides details of quickly deploying build-in graph analytic algorithms of ThunderGP.
The compilation of ThunderGP is managed by GNU make.

Compilation Arguments

Currently, ThunderGP supports four graph analytic applications, namely PR, SpMV, BFS and SSSP. The wanted application can be implemented by passing argument app=[the wanted application] to the make command. The below table are details of this argument.

Argument Accelerated algorithm
app=pr PageRank
app=spmv Sparse matrix-vector multiplication (SpMV)
app=bfs Breadth first search
app=sssp Single Source Shortest Path

Other arguments the developers may use.

Argument Description
all Compile host + cccelerator programs, it is time-costly (10+ hours)
exe Compile host program only, it is very fast
cleanexe Clean the host program
clean Clean the accelerator program
cleanall Clean all of the host + accelerator programs

For example, if you want to implement the PageRank, the command is:

$ make app=pr all

Makefile Tree

The Makefile file structure is shown in the bellowing tree:

├── application
│   ├── common.mk    # general configuration for both host and accelerator programs
│   ├── gs_kernel.mk # accelerator configuration for gather-scatter kernel 
│   ├── pr                    # application specific folder
│   │   ├── apply_kernel.mk   # accelerator configuration for apply kernel
│   │   ├── build.mk          # application specific build configuration
│   │   ├── config.mk         # application specific design parameters
│   └── ...
├── docs
├── libfpga
├── libgraph
├── Makefile  # Main entrance for make
├── README.md
├── utils
│   ├── help.mk      # Help information
│   ├── opencl.mk    # OpenCL library
│   └── utils.mk     # Misc.
└── xcl
    └── xcl.mk       # Xilinx OpenCL library

Fast Debugging Compilation

Compiling the entire accelerator can be very time-costly (14+ hours), and it is very unfriendly for iteratively debugging. ThunderGP provides a simplified and fast mode for compilation by only utilizing one SLR, it also means that in this mode, only have one Scatter-Gather CU and the compilation time can be significantly reduced to about 4 hours.

To enable fast compilation for debugging, you need to change the HAVE_FULL_SLR to false in the build.mk located at the application-specific directory, then start the build. The host program do not need any manual change, as it will automatically adapt partitioning according to this configuration.

Waveform-based Debugging

Xilinx provide real-time waveform display in hw_emu mode, but it need many setting-up steps in command-line. ThunderGP simplified the details to use this feature.

  • Change the TARGETS to hw_emu in main Makefile, or directly pass this argument from the make command.
  • After the application has been built, run make hwemuprepare
  • Run env setup script:source ./utils/hw_emu.sh
  • Start the program

Notes: We found that the waveform debugging in SDAccel 2019.2 have one problem: all of the stream interfaces are not displayed correctly. Therefore we only use SDAccel 2018.3 to perform the waveform-based debugging.

Timing Problems

As ThunderGP has a very high resources utilization, the timing issues can be very difficult to be fixed from high-level-synthesis code. our framework used the stream slicing technique and multi-level data duplication technique to achieve a higher frequency(250MHz). the details can be found in our paper. However, there are still many randomized factors in the placement and routine stage which can not be controlled by our HLS code. From our experience, we always start 3 same compilation, which could result with a better frequency finally.