Your SlideShare is downloading. ×
GPGPU and Stream Computing
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×
Saving this for later? Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime – even offline.
Text the download link to your phone
Standard text messaging rates apply

GPGPU and Stream Computing

850

Published on

This talk was held in the seminar for the course Parallel Programming and dealt with general purpose computation on graphics hardware and fundamentals of stream computing. Building on previous …

This talk was held in the seminar for the course Parallel Programming and dealt with general purpose computation on graphics hardware and fundamentals of stream computing. Building on previous knowledge about computer architecture and parallelization strategies, I contextualized GPGPU and introduced stream computing as its background. I then demonstrated a few modern languages and technologies (CUDA, OpenCL) and briefly touched upon compilation processes (NVIDIA PTX, AMD IL). The talk ended with perspectives on programmability and efficiency of the technologies and a short overview of the latest trends.

Published in: Technology
0 Comments
1 Like
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total Views
850
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
40
Comments
0
Likes
1
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. GPGPU and Stream Computing Julian Fietkau University of Hamburg June 30th, 2011
  • 2. Julian FietkauThings to clear up beforehand. . . These slides are published under the CC-BY-SA 3.0 license. Sources for the numbered figures are in the →list of figures. Non- numbered pictures and illustrations are from the OpenClipArt Project or are based on content from there. Download these slides and give feedback: http://www.julian-fietkau.de/gpgpu_and_stream_computing 2 / 21
  • 3. Agenda Julian Fietkau Agenda Introduction General Idea of GPGPU Stream Computing Languages Common Ideas OpenCL CUDA Others Compilation to Intermediary Languages Properties Programmability Efficiency Prospects and Conclusions Future Developments Conclusion 3 / 21
  • 4. Introduction: General Idea of GPGPU Julian Fietkau Flynn’s Taxonomy SISD MISD SIMD MIMD 4 / 21
  • 5. Introduction: General Idea of GPGPU Julian Fietkau Why Does It Exist? How long can Moore’s law hold true? → parallelism as a possible answer to computational demands “swiss army knife” (generally optimal solution) for parallel programming has not been found idea: exploit consumer-grade graphics Figure 1: Moore’s law – 2011 hardware 5 / 21
  • 6. Introduction: General Idea of GPGPU Julian Fietkau About Graphics Hardware games need to display increasingly realistic objects/scenes in real time need to calculate a lot of vertices and a lot of pixels very quickly → Pixel/Vertex Shaders, later Unified Shader Model consumer market ensures that graphics adapters remain (relatively) cheap General Purpose computation on Graphics Processing Units 6 / 21
  • 7. Introduction: Stream Computing Julian Fietkau Stream Computing idea: operate on a “stream” of data passing through different “kernels” related to SIMD mitigates some of the difficulties of parallelism on von Neumann architectures as well as simple SIMD implementations like SSE or AltiVec first came up in the 70ies, didn’t gain much traction as “pure” implementations, but hybrid architectures survived 7 / 21
  • 8. Introduction: Stream Computing Julian Fietkau Stream Computing Example Input: u, v, w; x = u - (v + w); y = u * (v + w); Output: x, y; Figure 2: Stream Computing Example 8 / 21
  • 9. Languages: Common Ideas Julian Fietkau Common Ideas modern streaming programming languages. . . . . . are verbose about different usage scenarios for memory . . . help with partitioning problem spaces in a multitude of ways . . . are not afraid to introduce limitations to faciliate optimization 9 / 21
  • 10. Languages: OpenCL Julian Fietkau OpenCL™ Open Computing Language, free standard by Khronos™ Group Application Kernel Command Queue Context Device Figure 3: OpenCL™ Application Model 10 / 21
  • 11. Languages: OpenCL Julian Fietkau OpenCL™ in Detail Device NDRange Host Work Work Work group group group Application (0,0) (1,0) (2,0) Work Work Work group group group (0,1) (1,1) (1,2) Work Work Work Work item item item item (2,0,1) (0,0,0) (1,0,0) (2,0,0) Work Work Work Work item item item item (2,1,1) (0,1,0) (1,1,0) (2,1,0) Figure 4: OpenCL™ Problem Partitioning 11 / 21
  • 12. Languages: CUDA Julian Fietkau CUDA NVIDIA’s custom framework for high-level GPGPU (it’s actually older than OpenCL though) same basic idea, but specific to NVIDIA GPUs conceptually only minor differences between CUDA and OpenCL biggest one: CUDA is compiled at application compile time while OpenCL is (typically) compiled at application run time also, annoying nomenclature differences (e.g. shared vs. local vs. private memory) 12 / 21
  • 13. Languages: Others Julian Fietkau Others There are several more stream processing languages, some of them long in development. Notable: Brook (and Brook+) Cilk, compare also Intel Array Building Blocks 13 / 21
  • 14. Languages: Compilation to Intermediary Languages Julian Fietkau Intermediary Languages Problem The actual binary code that runs on devices needs to “know” about exact numbers for cores, memory, registers etc., information that is generally not known at compile time. → compilation to an intermediary language like NVIDIA’s PTX and AMD’s IL, low-level and assembly-like yet abstracting some hardware limitations 14 / 21
  • 15. Languages: Compilation to Intermediary Languages Julian Fietkau PTX and AMD IL PTX example .reg .b32 r1, r2; .global .f32 array[N]; start: mov.b32 r1, %tid.x; shl.b32 r1, r1, 2; // shift thread id by 2 bits ld.global.b32 r2, array[r1]; // thread[tid] gets array[tid] add.f32 r2, r2, 0.5; // add 1/2 AMD IL example sample_resource(0)_sampler(0) r0.x, v0.xy00 mov r2.x, r0.xxxx dcl_output_generic o0 ret 15 / 21
  • 16. Properties: Programmability Julian Fietkau Programmability as they’re mostly custom versions of C, GPGPU languages are rather simple to pick up for someone with C experience OpenCL™ and CUDA both look slightly boilerplate-y for small tasks hypothesis: they might not be designed for small tasks disadvantage of the cutting edge: toolchain maturity might be lacking watch out for vendor dependencies! 16 / 21
  • 17. Properties: Efficiency Julian Fietkau Efficiency hard to find actual data optimizations and proficiency might skew the results conceptual similarities indicate that implementations would also be similar CUDA can get a (constant) head start vs. OpenCL™ due to being precompiled CUDA might generally perform faster, sometimes significantly, than OpenCL (but take this with a grain of salt) 17 / 21
  • 18. Prospects and Conclusions: Future Developments Julian Fietkau Things to Come The future remains notoriously hard to predict. at the moment, we see increased interest in specialized GPGPU boards (cf. NVIDIA Tesla and AMD FireStream) OpenCL promotes device flexibility at the cost of efficiency – no way to know if this strategy will win Intel pushes for integrated solutions with more processing power (cf. Sandy Bridge, Ivy Bridge) 18 / 21
  • 19. Prospects and Conclusions: Conclusion Julian Fietkau Conclusion GPGPU is a viable way to to massively parallel work even on a home PC will be further developed and refined, knowledge may be valuable 19 / 21
  • 20. External Links: Weblinks Julian Fietkau Weblinks AMD Developer Central: Introduction to OpenCL™ Programming http://developer.amd.com/zones/openclzone/...-may-2010.aspx GPGPU: OpenCL™ (Università di Catania) http://www.dmi.unict.it/~bilotta/gpgpu/notes/11-opencl.html NVIDIA: PTX ISA Version 2.1 http://developer.download.nvidia.com/compute/.../ptx_isa_2.1.pdf AMD: High Level Programming for GPGPU http://coachk.cs.ucf.edu/courses/CDA6938/s08/AMD_IL.pdf 20 / 21
  • 21. External Links: List of figures Julian Fietkau List of figures 1 Moore’s Law – 2011, by Wgsimon via Wikimedia Commons, CC-BY-SA 2 Stream Computing Example, by Kallistratos via German Wikipedia, public domain 3 OpenCL – Simple Kernel Exec, by Joachim Weging, CC-BY-SA 4 OpenCL – Problem Partitioning, by Joachim Weging, CC-BY-SA 21 / 21

×