Skip to content

Latest commit

 

History

History
101 lines (81 loc) · 5.89 KB

README.md

File metadata and controls

101 lines (81 loc) · 5.89 KB
Raw

64px vkpolybench

If used for your research, please cite this work as: Nicola Capodieci, Roberto Cavicchioli, vkpolybench: A crossplatform Vulkan Compute port of the PolyBench/GPU benchmark suite, SoftwareX, Volume 15, 2021, 100793, ISSN 2352-7110, Link to paper

vkpolybench is a Vulkan port of the original version of PolyBench/GPU.
All the host code and device code has been translated into Vulkan compute shaders. For the host code, an updated version of vkcomp has been used: vkcomp is a wrapper of the compute pipeline of the Vulkan API. vkcomp was previously used for a 'CUDA vs Vulkan compute' comparison with respect to CPU activity during CPU to GPU command submission (paper here).
For device (kernel/compute shader) code, GLSL shaders are provided in the respective benchmark-specific subfolders; vkcomp will generate SPIR-V binary files starting from the provided GLSL code (.comp files).
See license information here

Depedencies and pre-requisites

  • A Vulkan capable GPU/device.
  • Properly installed GPU/device driver(s).
  • LunarG Vulkan SDK (available here).

For Linux

vkpolybech has been tested on Ubuntu 18.04

  • glslc executable from LunarG SDK must be reachable from the command line.
    • type glslc --version in a terminal to test if it is working.
  • The Makefile script assumes Vulkan headers are located in /usr/include/vulkan

For Windows

vkpolybench has been tested on Windows 10 Pro, 1903

  • Download and install cygwin. Be sure to install the following packages when asked (they are NOT installed by default):
    • gcc-g++
    • gcc-core
    • make
  • The Makefile script assumes cygwin bin folder is appended to the path environment variable
  • The Makefile script assumes the Vulkan SDK path is appended to the path environment variable

For Android

vkpolybench has been tested on a smartphone running Android 9. Host PC was Windows 10 with Android NDK and SDK properly installed and set to path.
Also, Gradle 6.1.1 must be installed and set to path.
Developer and USB debug mode on your smartphone device must be enabled.

Compile and test

Once this repository is cloned, move to the project root folder.
For building

  • In Linux:
    • Open a terminal and type chmod a+x compileCodes.sh
    • To build all the benchmarks type: ./compileCodes
  • In Windows:
    • Open a powershell terminal and type compileCodes.bat
  • In Android:
    • Each benchmark has to be invidually compiled. On a shell, navigate through vkpolybench/android/[BENCHMARK_NAME]
    • type gradle assembleDebug

For running all the benchmarks (after build)

  • In Linux:

    • Open a terminal and type ./compileCodes test

  • In Windows:

    • Open a powershell terminal and type compileCodes.bat test

  • In case more than one Vulkan capables device are present in the tested system:

    • in compileCodes.sh or compileCodes.bat change DEVICE_SELECTION variable from 0 to:
      • [deviceId], as an integer indicating the device identifier of your preferred GPU, or
      • [deviceVendorString], as a string identifier of the device vendor (e.g. nvidia or intel)

  • For Android:

    • Once a benchmark app is built and your phone is connected to the host PC, type on a shell:
      [path to ANDROID_SDK]\sdk\platform-tools\adb.exe install [path to the repo]\vkpolybench\android\2DCONV\build\outputs\apk\debug\2DCONV-debug.apk
      Your app should now be installed in your smartphone. Substitute 2DCONV with any other compiled benchmark.
    • type: [path to ANDROID_SDK]\sdk\platform-tools\adb.exe logcat -s "vulkanandroid". This will capture the logcat output related to vkpolybench
    • On your phone tap the newly installed app and check the output on the logcat console

For cleaning (obj files and generated binary spv related to all the benchmarks)

  • In Linux:
    • Open a terminal and type ./compileCodes clean
  • In Windows:
    • Open a powershell terminal and type compileCodes.bat clean
  • In Android:
    • Open a shell on your host PC and type gradle clean

Tests

alt text
alt text
alt text

Tested GPU devices

  • NVIDIA GeForce RTX 2070
  • NVIDIA TITAN V
  • NVIDIA TITAN RTX
  • NVIDIA Xavier SoC
  • INTEL HD Graphics 530
  • ARM MALI G72

Known issues

  • For each benchmark, data types, sizes and other constants are defined in the HDcommon.h header file, that is present in each benchmark folder and relates to both device and host code. If GLSL_TO_SPV_UPDATE_COMPILATION_CHECK in VulkanCompute.h is defined, vkcomp will check the last modification date of the glsl source and compares it with last modification of the generated binary spv file (if present). If no mismatches are present, the spv binary will not be generated; hence, the last generated .spv file will be read by the application. This mechanism is unable to track modification to the HDcommon.h file, so that if a modification is performed in the header file, the previously generated .spv file must be manually deleted.