Hetero-Mark

A Benchmark Suite for collaborative CPU-GPU computing.

Prerequisite

OpenCL Environment

• OpenCL - The OpenCL driver

We use to use the FGLRX driver on Ubuntu 14.04, which supports OpenCL 2.0. Since AMD stopped the support for the FGLRX driver on ubuntu 16.04, we switched to the AMDGPU-pro driver. Currently, OpenCL 2.0 benchmarks are not supported on this platform.

HSA Environment

• ROCm - Radeon Open Compute

Although ROCm 1.6 platform does not fully support OpenCL 2.0, all the features we have been using in the benchmark suite is supported by ROCm 1.6.

OpenCV Library

The Background Extraction benchmark will use OpenCV for video decoding and encoding. The benchmark suite will detect if your system has OpenCV installed or not. If OpenCV libraries are not found, CMAKE will skip compiling the BE benchmarks.

We use the following command to install OpenCV libraries.

sudo apt install libopencv-dev

Applications

Hetero-Mark is designed to model the workloads that is similar to real world applications, where the major part of the application is written in general purpose programming languages, while only a small, performance critical portion is written using GPU-accelerated libraries. So for each benchmark, we provide a base class that provides platform independent functionalities, such as input data loading, result verification. For each GPU programming method (such as CUDA, HC, HIP), we extend the base class with a sub-class and impletment the "Run" method.

Since the base classes are platform independent, we use plain pointers for input and output data. Each benchmark will have to read from plain pointers and finally write the result into other plain pointers. We believe this behavior is closer to real world senario, since most programmers do not carry a platform specific memory management system to the whole application and usually will only use GPU program as a library. This also suggests that the benchmarking time considers the data copy time between the CPU and the GPU memory.

All the benchmarks has a verification process where the GPU result is compared with the CPU result. Although we report the execution time of the verification process, the time is not meant to compare the CPU performance to GPU performance. The verification process can be very useful if the benchmark runs in simulators or if the validity of the platform is under evaluation.

• Advanced Encryption Standard (AES) - The program takes plaintext as input and encrypts it using a given encryption key. Our implementation uses a key size of 256 bits. The AES algorithm is comprised of many rounds, that ultimately turnplaintext into cipher-text. Each round has multiple processing stepsthat include AddRoundKey, SubBytes, ShiftRows and MixColumns. Key bits must be expanded using a precise key expansion

• Background Extraction (BE) - An useful algorithm in video and image processing, background extraction algorithms usually create a background model based on static components of the frame. Our implementation uses a Running Gaussian Average, and takes an input video file and extracts the background of that video.

• Black Scholes (BS) - The Black–Scholes or Black–Scholes–Merton model is a mathematical model of a financial market containing derivative investment instruments. From the model, one can deduce the Black–Scholes formula, which gives a theoretical estimate of the price of European-style options.

• Binary Search Tree Insertion (BSTI) - Binary Search Tree is a useful data structure for its balanced insertion and in-order accessing performance, but rearranging an array into a binary search tree is usually time consuming. The GPU can help with inserting the nodes of a binary search tree in parallel, using one thread to insert one node. However, as the output is a irregular tree structure, we need to let the CPU and the GPU collaborate under the Co-Contributing pattern.

• Color Histogramming (CH) - Color Histogramming is a popular method in image processing to divide the color space into groups, and counts the number of pixels in a picture that fall into each group. The implementation of Color Histogramming is divided into two phases. In the first phase the GPU kernel scans the whole image and each GPU thread covers a small portion of the image. Each thread stores the histogram information of the pixels it has scanned in a region of the private memory that is dedicated to that thread. In the second phase, each GPU thread takes the histogram produced in the first phase and adds it to an output histogram using atomic operations.

• Force Directed Edge Bundling (FDEB) - Force Directed Edge Bundling is a graph-based data visualization algorithm that helps readers identify patterns in a complex graph. The algorithm models a spring between each pair of the edges and calculates the forces applied to points on each edge. Then each point moves a certain distance towards the direction of the combined force.

• Evolutionary Programming (EP) - Evolutionary Programming solves optimization problems using an approach that mimics the natural selection process. In our benchmark implementation, we use Evolutionary Programming to solve a non-convex optimization problem.

• Finite Impulse Response (FIR) - FIR filter produces an impulse response of finite duration. The impulse response is the response to any finite length input. The FIR filtering program is designed to have the host send array data to the FIR kernel on the OpenCL device. Then the FIR filter is calculated on the device, and the result is transferred back to the host.

• Gene Alignment (GA) - Gene Alignment algorithms are used to answer questions about specific gene sequences (e.g., “CATGCATG”) that occur in the human gene sequence. Our implementation uses a modified version of the Basic Local Alignment Search Tool (BLAST).

• K-Nearest Neighbors (KNN) - Given a large number of labeled training samples in a multi-dimensional feature space, the K- Nearest Neighbors (KNN) algorithm takes a query point and searches for the K training samples that are close to that point. Using a majority vote approach, the KNN algorithm can categorize the query point with the label that appears the most number of times in the selected K training samples.

• KMeans (KM) - k-means clustering is a method of vector quantization, originally from signal processing, that is popular for cluster analysis in data mining. k-means clustering aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean, serving as a prototype of the cluster. In this implementation, we have varied the number of objects of 34 features and put them into 5 clusters. The input file contains features and attributes.

• Page Rank (PR) - PageRank is an algorithm used by Google Search to rank websites in their search engine results. It is a link analysis algorithm and it assigns a numerical weighting to each element of a hyperlinked set of documents, such as the World Wide Web, with the purpose of "measuring" its relative importance within the set. So the computations are representatives of graph based applications.

Compiling the code

OpenCL

Use the following commands to compile the OpenCL benchmarks.

mkdir build
cd build
cmake -DCOMPILE_OPENCL12=On ../
make

If OpenCL is properly configured in your system, the command above will use the system default compiler to compile OpenCL benchmarks.

HCC Compilation

Use the following commands to compile the HCC benchmarks

mkdir build
cd build
CXX=hcc cmake ../
make

This command will also use the HCC compiler to compile the OpenCL benchmarks.

CUDA Compilation

Use the following commands to compile CUDA benchmarks. Make sure your system has NVCC compiler installed.

mkdir build
cd build
cmake -DCOMPILE_CUDA=On ../
make

You need to modify the file cmake/CUDA.cmake to adjust your GPU's capability.

If you want to compile the system-level atomic benchmarks, you need to make sure to use compute capacity greater than 60. And you need the following commands to include those benchmarks into the compilation list:

mkdir build
cd build
cmake -DCOMPILE_CUDA=On -DCOMPILE_SYSTEM_ATOMIC_CUDA=On ../
make

HIP Compilation

Use the following commands to compile HIP benchmarks.

mkdir build
cd build
cmake -DCOMPILE_HIP=On ../
make

HIP works for both CUDA platform and the ROCm platform.

Run the code

The executables are in the build folder under Hetero-Mark/build/src/<application name>/<environment> if you follow the default compile guide, where <application name> is the name of the application, such as, fir, be, bs etc and replace <environment> for cl12, cl20, cuda or hc.

The executables support the following arguments:

• -t is for timing information
• -v is for cpu verification
• -q is for suppressing the output

All benchmark executables has a -h option. The help documentation of each benchmark explains how to use the benchmark and what parameter is needed.

Input data

Download standard input data

You can download the standard data from the following url https://drive.google.com/file/d/1IItjFFUIfANgrUUI7jebNS9rfSEe32lZ/view?usp=sharing".

Generate your own input data

• To generate custom data in data folder

  • AES - Generates the input file and keys for AES. For keys, only 16-byte is allowed.

    ./datagen <num_bytes> > file.data

  • Gene-alignment - Generates the input file for Gene Alignment. The target sequence length should be much shorter than the query sequence length.

    python data_gen.py <target_sequence_len> <query_sequence_len>

  • KMeans - It generates the input file for KMeans. Usage:

    g++ datagen.cpp -o datagen
    ./datagen <numObjects> [ <numFeatures> ] [-f]

  • PageRank - It generates the input matrix for PageRank. Usage:

    python datagen.py

Known Issue

• The HIP version of Background Extraction (BE) benchmark cannot compile on CUDA platform currently. This is due to that fact that the HIP cmake configuration uses NVCC as program linker which NVCC cannot handle the linking to OpenCV properly. We have been actively resolving the problem with ROCm developers. See #120.

Development guide

Please raise issues on the Github page if you have any questions or problems using the benchmark suite.

We accept pull requests on github if you want to contribute to the benchmark suite. If you have any question or problem with HeteroMark, please file an issue in our github repo.

Hetero-mark follows google c++ coding style in header files and source files. We also use make check to lint the source code using the clang-format tool and cpplint tool.

Citation

• Yifan Sun, Saoni Mukherjee, Trinayan Baruah, Shi Dong, Julian Gutierrez, Prannoy Mohan and David Kaeli, "Evaluating Performance Tradeoffs on the Radeon Open Compute Platform" 2018 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), Belfast, 2018.

• Yifan Sun, Xiang Gong, Amir Kavyan Ziabari, Leiming Yu, Xiangyu Li, Saoni Mukherjee, Carter McCardwell, Alejandro Villegas, David Kaeli, "Hetero-mark, a benchmark suite for CPU-GPU collaborative computing" 2016 IEEE International Symposium on Workload Characterization (IISWC), Providence, RI, USA, 2016, pp. 1-10.

• Saoni Mukherjee, Yifan Sun, Paul Blinzer, Amir Kavyan Ziabari and David Kaeli, "A comprehensive performance analysis of HSA and OpenCL 2.0" 2016 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), Uppsala, 2016, pp. 183-193.

• Saoni Mukherjee, Xiang Gong, Leiming Yu, Carter McCardwell, Yash Ukidave, Tuan Dao, Fanny Nina Paravecino, and David Kaeli. Exploring the Features of OpenCL 2.0 The International Workshop on OpenCL (IWOCL), 2015.