Benchmark Research
Terminology (including terms from last week):
logical page: contains more data than can fit on the physical page
direct-mapped cache: each memory location is mapped to one specific cache entry location
fully associative cache: each memory location can be mapped to any cache entry location
two-way set associative cache: each memory location can be mapped to only one of two cache entry locations
kernel: grid of blocks of warps of scalar threads
warp: groups of threads in hardware meant to execute in lockstep
SIMT: Single Instruction, Multiple Thread, coined by NVIDIA to describe their GPU CUDA architecture
ISA: instruction set architecture
SASS: Native ISA for NVIDIA GPUs
PTX: Parallel Thread eXecution; NVIDIA's scalar, low-level, and data-parallel virtual ISA
__
This powerpoint had a lot of useful information on and good explanations of GPGPU-Sim: "GPUWattch + GPGPU-Sim: An Integrated Framework for Performance and Energy Optimizations in Manycore Architectures" by Jingwen Leng (The University of Texas at Austin) and Tayler Hetherington (University of British Columbia).
I read up on how GPUGU-Sim supports both CUDA and OpenCL.
The purpose of the presentation is to describe what GPGPU-Sim simulates, show how to setup and run CUDA (the paper focused on CUDA) applications on it, and provide a foundation for research groups.
It has some information on OpenCL supported GPUs, such as Qualcomm's Adreno 3xx GPU, ARM's Mali-T600 Series GPUs, Intel Ivy Bridge's HD 4000, the Intel Xeon Phi (Knights Corner), NVIDIA Quadro FX 5800, NVIDIA Tesla C2050 (Fermi), and NVIDIA GTX 480.
It also has useful graphs for when the RODINIA Benchmark Suite. Accuracy was measured through functional, computing time, and power models; these are what the GPGPU-Sim is supposed to simulate.
Here is the RODINIA Benchmark Suite, courtesy of the Department of Computer Science at the University of Virginia. It has descriptions on the site.
In this paper, "Analyzing CUDA Workloads Using a Detailed GPU Simulator" by Ali Bakhoda, George L. Yuan, Wilson W. L. Fung, Henry Wong and Tor M. Aamodt at the University of British Columbia, they use the following benchmarks:
-AES Cryptography/Encryption: uses the Advanced Encryption Standard algorithm in CUDA to encrypt and decrypt files.
-Graph Algorithm: Breadth First Search: breadth first search on a graph
-Coulombic Potential: unroll loops to reduce loop overheads and store point charge data in constant memory with caching
-gpuDG: calculate 3D radar scattering and analyze waves and particle accelerators in electromagnetics
-3D Laplace Solver: parallel finance application; uses shared memory and coalesced global memory accesses
-LIBOR Monte Carlo: Monte Carlo simulations based on the London Interbank Offered Rate Market Model
-MUMmerGPU: matches query strings of standard DNA nucleotides to a reference string to simulate genetic processes
-Neural Network Digit Recognition: recognize handwritten digits withthe network and pre-determined neuron weights and input digits in global memory
-N-Queens Solver: solves a chess board puzzle by placing N queens so that they cannot catch each other
-Ray Tracing: rendering graphics with near photo-realism using scalar threads in CUDA for each pixel
-StoreGPU: accelerates hashing-based primited for middleware
-Weather Prediction: uses the GPU to accelerate a portion of the Weather Research and Forecast Model
-Black-Scholes BLK option pricing: not explained
-Fast Walsh Transform: not explained
Here is another benchmark I found: