Introduction to GPU Programming Languages

Transcription

1 CSC 391/691: GPU Programming Fall 2011 Introduction to GPU Programming Languages Copyright 2011 Samuel S. Cho

2 research/gpu/facilities.html Maryland CPU/GPU Cluster Infrastructure

3 Intel s Response to NVIDIA GPUs

4 To Accelerate Or Not To Accelerate Pro: They make your code run faster. Cons: They re expensive. They re hard to program. Your code may not be cross-platform.

5 When is GPUs appropriate? Applications Traditional GPU Applications: Gaming, image processing i.e., manipulating image pixels, oftentimes the same operation on each pixel Scientific and Engineering Problems: physical modeling, matrix algebra, sorting, etc. Data parallel algorithms: Large data arrays Single Instruction, Multiple Data (SIMD) parallelism Floating point computations

6 Parallel Hardware Landscape: Instruction and Data Streams Flynn s Classification: Hardware dimensions of memory and control Data Streams Single Multiple Instruction Single SISD: Intel Pentium 4 SIMD: GPUs Streams Multiple MISD: No Examples Today MIMD: Intel Nalhelm

7 Single Instruction, Multiple Data (SIMD) Example Instruction: A[i][j] = A[i][j]++; Element-wise operations on vectors or matrices of data. Multiple processors: All processors execute the same set of instructions at the same time Each with data at a different address location. 0,0 0,1 0,2 0,3 1,0 1,1 1,2 1,3 Advantages: Simplifies synchronization Reduces instruction control hardware; One program, many results. Works best for highly data-parallel applications (e.g., matrix operations, monte carlo calculations) 2,0 2,1 2,2 2,3 3,0 3,1 3,2 3,3

8 CPU vs. GPU Hardware Design Philosophies Control CPU ALU ALU ALU ALU GPU DRAM DRAM

9 CUDA-capable GPU Hardware Architecture Processors execute computing threads Thread execution managers issues threads 128 thread processors grouped into 16 streaming multiprocessors (SMs) enables thread cooperation. CPU CPU GPU Host Input Assembler Thread Execution Manager GPU... Texture Texture Texture Texture Texture Texture Texture Texture Load/store Load/store Load/store Load/store Load/store Load/store Global Memory

10 CUDA-capable GPU Hardware Architecture Processors execute computing threads Thread execution managers issues threads 128 thread processors grouped into 16 streaming multiprocessors (SMs) enables thread cooperation. CPU CPU GPU Host Input Assembler Thread Execution Manager GPU... Texture Texture Texture Texture Texture Texture Texture Texture Load/store Load/store Load/store Load/store Load/store Load/store Global Memory

11 Single Instruction, Multiple Threads (SIMT) A version of SIMD used in GPUs. GPUs use a thread model to achieve a high parallel performance and hide memory latency. On a GPU, 10,000s of threads are mapped on to available processors that all execute the same set of instructions (on different data addresses). 0,0 0,1 0,2 0,3 1,0 1,1 1,2 1,3 2,0 2,1 2,2 2,3 3,0 3,1 3,2 3,3

12 Is it hard to program on a GPU? In the olden days (pre-2006) programming GPUs meant either: using a graphics standard like OpenGL (which is mostly meant for rendering), or getting fairly deep into the graphics rendering pipeline. To use a GPU to do general purpose number crunching, you had to make your number crunching pretend to be graphics. This is hard. Why bother?

13 How to Program on a GPU Today Proprietary programming language or extensions NVIDIA: CUDA (C/C++) AMD/ATI: StreamSDK/Brook+ (C/C++) OpenCL (Open Computing Language): an industry standard for doing number crunching on GPUs. Portland Group Inc (PGI) Fortran and C compilers with accelerator directives; PGI CUDA Fortran (Fortran 90 equivalent of NVIDIA s CUDA C). OpenMP version 4.0 may include directives for accelerators.