Programming with CUDA and Parallel Algorithms
description
Transcript of Programming with CUDA and Parallel Algorithms
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Programming with CUDA and Parallel
AlgorithmsWaqar Saleem
Jens Müller
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Organization• People
• Waqar Saleem, [email protected]
• Jens Mueller, [email protected]
• Room 3335, Ernst-Abbe-Platz 2
• The course will be conducted in English
• 6 points
• Wahl/Wahlpflicht
• Theoretical/Practical
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Organization•Meetings, before winter break
• Tue 12-14, CZ 129
• Thu 16-18, CZ 129
• Every second week
• Starting next week
• Exercises: Wed 8-10, CZ 125
• Starting tomorrow in the pool
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
The course•2 parts
• Before winter break: Lectures and assignments
• Need at least 50% in assignments to qualify for ...
• After the break: Group projects
• Project chosen by or assigned to each group
• Regular meetings
• Presentation of each project on semester end
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Assignments• Build up a minimal ray tracer on GPU
• Implement basic ray tracer on CPU
• Port to GPU
• Make ray tracer more interesting/efficient
• Utilize CUDA concepts
• Basic framework will be provided
• Scene format and scenes
• Introduction to ray tracing concepts
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Requirements
•Strong background in C programming
•Familiarity with your OS
•Modifying default settings
•Writing/understanding Makefiles
•Compiler flags and options
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Course content•Parallel programming models and
platforms
•GPGPU
•GPGPU on NVIDIA cards: CUDA
•Architecture and programming model
•OpenCL
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Today
•Organization
•Brief introduction to parallel programming and CUDA
•Short introduction to Ray tracing
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Growth of Compute Capability
•Moore’s law: the number of transistors that can be placed ... on an integrated circuit [doubles] approximately every two yearssource: wikipedia
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Growth of Compute Capability•Moore’s law
source: wikipedia
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Need for increasing compute capability
•Problems are getting more complex
•e.g. Text editing to Image editing to Video editing
•Current hardware complexity is never enough
•Impractical to stop development at current state of the art
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Barriers to growth•Natural limit on transistor size: the size
of an atom
•More transistors per unit area lead to higher power consumption and heat dissipation
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Solution: Parallel architectures
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Parallel architectures•Multiple Instructions Multiple Data
(MIMD)
•multi-threaded, multi-core architectures, clusters, grids
•Single Instruction Multiple Data (SIMD)
•Cell processor, GPUs, clusters, grids
•GPU: Graphics Processing Unit
•Parallel programming allows to program for parallel architectures
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
GPU architecture
•Simpler architecture than MIMD
•Little overhead for instruction scheduling, branch prediction etc.Subsequent figures from NVIDIA CUDA Programming Guide 2.3.1 unless mentioned otherwise
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
GPU architecture•Simpler architecture leads to higher
performance (compared to CPUs)
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
General Purpose computing on GPU, GPGPU
•Attractive because of raw GPU power
•Traditionally hard because GPU programming was closely associated to graphics
•Simplicity of GPU architecture limits the kind of problems suitable for GPGPU
•or at least requires some problems to be reformulated
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
GPGPU for the masses*•Freeing the GPU from graphics:
Nvidia CUDA, ATI Stream
•C-like programming interface to the GPU
•* - knowledge of underlying architecture required to achieve peak performance
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Freeing Parallel Programming
•OpenCL: code once, run anywhere
•single core, multi core, GPU, ...
•platform details transparent to the user
•supported by major vendors: Apple, Intel, AMD, Nvidia, ...
•OpenCL drivers made available by ATI and Nvidia for their cards
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
This course•chiefly CUDA: Nvidia specific,
mature, well documented, easily available literature
•some OpenCL: open standard, very new, limited documentation available, very similar concepts to CUDA
•no ATI Stream
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
CUDA, Compute Unified Device Architecture
•Software: C like programming interface to the GPU
•Hardware: the hardware that supports the above programming model
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
CUDA hardware model
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
CUDA programming model•CPU=host, GPU=device, work
unit=thread
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Ray tracing•A method to render a given scene
•Cast rays from a camera into the scene
•Compute ray intersections with scene geometry
•Render pixelimage source: wikipedia
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Ray tracer complexity•A ray tracer can be arbitrarily
complex
•Recursively compute intersections for reflected, refracted and shadow rays
•Account for diffuse lighting
•Consider multiple light sources
•Consider light sources other than point lights
•Account for textures: object materials
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Coding a ray tracer
•Relatively easy to code on the CPU
•Call the same intersection function recursively on secondary rays
•CPU code is not so complex
•Tricky to code on the GPU as recursion is not yet supported in GPGPU models
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
This course
•Build a trivial ray tracer on the CPU
•compute view rays only
•part of tomorrow’s exercise
•Port to GPU
•Add complexity to your GPU ray tracer
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
Reminders
•Exercise session tomorrow
•Register on CAJ
Programming with CUDA, WS09
Waqar Saleem, Jens Müller
See you next time!