CUDA optimization

FABRIC DRAPING

CS 750 HIGH PERFORMANCE COMPUTING – COURSE PROJECT

​

Contents

• Draping Introduction
• Blossom Polynomials
• Problem definition and Inputs
• Parallelized Reduction Algorithm
• CUDA basics
• Results
• References

​

Draping

• Draping of virtual characters is done using cloth simulation models.
• The Fabric needs to be rendered and simulated in a time efficient manner.
• CUDA optimization can help lower the time required to achieve this.
• For maximum performance and process control, we choose programming language C++.

​

Rendering process

CAD model of the target

Target feature points

Grid of points for fabric

Physical simulation

Fabric geometry

Fully Rendered Fabric on Target

Our scope

• Converting a grid of control points into a fabric geometry

Scope of project

B-Spline surface

Blossoming polynomials

Bi-variate Blossom : Quadratic Bspline

Tri-variate Blossom : Cubic Bspline

B-spline Construction

B-spline Construction

Blossom Construction

Surface blossoms

Grid of Control Points + u,v grid

Inputs

u, v coordinates

CUDA Intro

CUDA virtualizes the physical hardware into threads and blocks

Threads

• Thread is a virtualized Scalar Processor

Blocks

• Thread blocks is a virtualized Streaming multiprocessor.
• Thread blocks need to be independent
• They run to completion.
• Order of blocks is undecided.

B-spline basis Construction

Thread(i,j): Across domain

Block(i_b,j_b) = 16 threads/ Per Block = 1 grid point (u,v)

Parallel Reduction: Sequential Addressing

Warps (Scheduling unit)

Each warp runs threads in a lock step fashion

Data transfer rates

Memory hierarchy

Code walk through

• __Global__ function is called from the host and runs on the device.

​

• * variables are for input and output.

​

• __Shared variables are shared across the threads within a block.

Thread and Block IDs obtained

• Thread ID i,j select the control points for each reduction operation.
• Block ID ib,jb select the u,v grid co-ordinate.

Barriers an Thread Synchronization

• __syncthreads ensures all threads within a block have reached here.
• Used to prevent memory conflicts and race conditions from occurring.
• Use atomic operations key words like and volatile, to prevent race conditions.

​

CUDA memory management.

• Memory Allocation

​

​

• Data transfer from host to device.

​

​

• Data transfer from device to host.

​

Nsight : Visual Studio

NVVP

NVVP

Results

Sample duration: (without yarn info)

• Cuda time (ms)= 1.227840
• Sequential time (ms) = 4.081682

​

• Speed up factor :3.33
• Tolerance<1e-4

CUDA streams for Asynchronous data transfer

Profiler

References

CUDA

• NVIDIA, GPU Programming Guide, Version 8.0.
• http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html
• CS 759 Slides: High Performance Computing applications for Engineering
• Jason Sanders and Edward Kandrot: CUDA by Example: An Introduction to General-Purpose GPU Programming, Addison-Wesley Professional, 2010
• Lecture1.pdf (bu.edu)

Graphics

• ME 535 slides : CAGD
• Yuksel, C., Kaldor, J. M., James, D. L., & Marschner, S. (2012). Stitch meshes for modeling knitted clothing with yarn-level detail. ACM Transactions on Graphics (TOG), 31(4), 1-12.