University:
Boston University
Course:
MET CS 201 | Introduction to Programming
Academic year:

2020
Views:

349

Pages:

6
Author:

Clint W.

GPU Computing with CUDA Lab 2 - FD in global and texture memory Objectives ‣ Implement a simple finite difference problem in CUDA ‣ Learn how to use texture memory Heat transfer with FD ‣ Heat diffusion on square 2D flat plate with Dirichlet boundary conditions and 0 initial condition T=0 T = 200 T=0 3.5 ∂u 2 = α∇ u ∂t α = 0.645 k = 1e-5 N = 128 T = 200 3.5 n+1 ui,j = uni,j αk n + 2 (ui,j+1 + uni,j−1 + uni+1,j + uni−1,j − 4uni,j ) h Heat transfer with FD ‣ Each thread will do one mesh point ‣ Global memory data lives for whole execution! ‣ Experiment with other mesh sizes (N) that are not a multiple of block size Heat transfer with FD ‣ Changes to global memory case: (start from pervious code!) - Initialize texture (beginning of code) texture tex_u; - After allocation, bind texture to global memory cudaChannelFormatDesc desc = cudaCreateChannelDesc(); cudaBindTexture2D(NULL, tex_u, u_d, desc, N_max, N_max, sizeof(float)*N_max); - In the kernel, fetch values from texture memory int x = threadIdx.x + blockIdx.x*BSZ; int y = threadIdx.y + blockIdx.y*BSZ; c = tex2D(tex_u_prev, x, y); - Unbind texture at the end cudaUnbindTexture(tex_u); Heat transfer with FD ‣ Using texture memory: - Should give good results as it is designed for 2D data locality ‣ Textures map multiples of 32 or 64: - Need to pad or use cudaMallocPitch ‣ Texture memory is read only within a kernel - Need to write separate kernel to update values in texture

GPU Computing with CUDA Lab 2 - FD in Global and Texture Memory

GPU Computing with CUDA Lab 2 - FD in Global and Texture Memory - Page 3

GPU Computing with CUDA Lab 2 - FD in Global and Texture Memory - Page 4

GPU Computing with CUDA Lab 2 - FD in Global and Texture Memory - Page 5

GPU Computing with CUDA Lab 2 - FD in Global and Texture Memory - Page 6

of 6

Description

Free up your schedule!

Take 5 seconds to unlock