1. CUDA Interoperability with Graphical Environments
Martin Kruliš
by Martin Kruliš (v1.1)

2. Graphics Interoperability
Allows CUDA code to read/write graphical buffers
Works with OpenGL and Direct3D libraries
Motivation
Direct visualization of complex simulations
Augmenting 3D rendering with visualization routines which are difficult to implement in shaders
How it works
The graphics resource is registered and represented by struct cudaGraphicResource
The resource may be mapped to CUDA memory space cudaGraphicsMapResources(), …
by Martin Kruliš (v1.1)

3. OpenGL Initialization Resources
Device must be selected by cudaGLSetGLDevice()
Resources
cudaGraphicsGLRegisterBuffer() for buffers
The mapped buffers can be accessed in the same way as CUDA allocated memory
cudaGraphicsGLRegisterImage() for images and render buffers
The image buffers can be also accessed through texture and surface mechanisms
Code example at CUDA Programming Gude (page 54-56)
CUDA Samples:
2_Graphics\simpleGL – the same example as in the Guide (vertex buffer filling)
2_Graphics\Mandelbrot – fractal rendering
3_Imaging\postProcessGL – render buffer post processing (i.e., reading and writing)
5_Simulations\fluidsGL
5_Simulations\nbody
5_Simulations\particles
5_Simulations\smokeParticles
5_Simulations\oceanFFT
Examples
by Martin Kruliš (v1.1)

4. Direct3D Direct3D Support Versions 9, 10, and 11 are supported
Each version has its own API
CUDA context may operate with one Direct3D device at a time
And special HW mode must be set on the device
Initialization is similar to OpenGL
cudaD3D[9|10|11]SetDirect3DDevice()
Available Direct3D resources
Buffers, textures, and surfaces
All using cudaGraphicsD3DXXRegisterResource()
by Martin Kruliš (v1.1)

5. SLI Interoperability GPU SLI Mode
Multiple GPUs are interconnected (physically) and cooperating in rendering the scene
AFR mode – different GPUs render subsequent frames
CUDA interoperability issues
Any CUDA allocation on one GPU is automatically performed on all SLI-connected GPUs
CUDA has to use separate contexts for each GPU
cudaGLGetDevices() – identify, which devices are in SLI
cudaGLDeviceListAll
cudaGLDeviceListCurrentFrame
cudaGLDeviceListNextFrame
by Martin Kruliš (v1.1)

6. Discussion
by Martin Kruliš (v1.1)

7. Libraries Using CUDA
Martin Kruliš
by Martin Kruliš (v1.1)

8. CUBLAS CUDA Basic Linear Algebra Subroutines
CUDA implementation of standard BLAS library
Complete support of all 152 functions on vectors/matrices
copy, move, rotate, swap
maximum, minimum, multiply by scalar
sum, dot products, Euclidean norms
matrix multiplications, inverses, linear combinations
Some operations have batch versions
Supports floats, doubles, and complex numbers
by Martin Kruliš (v1.1)

9. CUSP CUDA Sparse Linear Algebra
Open source C++ library for sparse linear structures (matrices, linear systems, …)
Key features
Sparse matrix operations (add, substraction, max independent set, polynomial relaxation, …)
Supports various matrix formats
COO, CSR, DIA, ELL, and HYB
Require CUDA CC 2.0 or higher
by Martin Kruliš (v1.1)

10. CUFFT CUDA Fast Fourier Transform
Decompose signal to frequency spectrum
1-3D transforms (up to 128M elements)
Many variations (precision, complex/real types, …)
API similar to FFTW library
Create plan (cufftHandle) which holds the configuration
Associate/allocate work space (buffers)
cufftExecC2C() (or R2C, C2R) starts execution
FFT plan can be associated with CUDA stream
For synchronization and overlapping
by Martin Kruliš (v1.1)

11. Thrust CUDA Thrust C++ template library based on STL API
Basic idea is to develop C++ parallel applications with minimal overhead
STL like vectors (for devices) and vector operations
copy, fill, create sequences, reordering, sorting, …
Algorithms
Transformations
Reductions
Prefix-sums
by Martin Kruliš (v1.1)

12. GPU AI GPU AI for Board Games
Specific AI library designed for games with large, but well-defined configuration space
Requires CUDA CC 2.0
Currently supports
Game Tree Split – alpha/beta pruning
Single and multiple recursion (with large depths)
Zero-sum games (3D Tic-Tac-Toe, Reversi, …)
Sudoku backtracking generator and solver
Statistical simulations (Monte Carlo for Go)
by Martin Kruliš (v1.1)

13. PhysX PhysX APEX Realtime physics engine
Originally developed by Ageia for PPU card
NVIDIA bought it and re-implemented it for CUDA
Most important features
Simulation of rigid bodies (collisions, destruction)
Cloths and fluid particle systems
APEX
Framework built on top of PhysX
Designed for easy usage (artists, games, …)
by Martin Kruliš (v1.1)

14. Discussion
by Martin Kruliš (v1.1)