Conv2 GPU加速（有代码有图有真相）

conv2是matlab自带的一个二维卷积的函数，使用格式为
C = conv2(A,B,'shape')
A和B分别是二维矩阵。详细使用见matlab帮助。
卷积具有重要的物理意义，如滤波等过程。在物理光学中，卷积则可以表示瑞利-索末菲衍射的过程，在matlab变成中表达和计算都非常方便。

% 本程序计算圆形高斯光束入射到位于 束腰处的 圆盘产生的衍射及伯松亮斑
close all;clear all;
w = 0.25;                    % 入射光束腰，束腰半径为0.25mm
lamda = 0.632e-3;           % 入射激光波长
k = 2*pi/lamda;             % 波数
q = 1i*pi*w^2/lamda;        % 束腰处的q参数
N = 500;                    % 光强采样精度
x = linspace(-0.5,0.5,N);   % 观察光斑的坐标范围
z = 10;                     % 衍射传播的距离
E0 = exp(-1i*k*x.^2/2/q);   % 束腰处的光场复振幅分布

[X Y] = meshgrid(x,x);      % 坐标网格化
[Ex Ey] =meshgrid(E0,E0);   % 光场二维化
E1 = Ex.*Ey;                % 束腰处的二维光场
figure;imagesc(abs(E1).^2); % 圆形高斯光束
colormap hot;axis image;

Disk = ones(N);     % 定义半径为0.1mm的挡光圆盘
for m = 1:N
    for n = 1:N
        if sqrt(x(m)^2+x(n)^2) <= 0.1
            Disk(m,n) = 0;
        end
    end
end

Ein = E1.*Disk;         % 透过圆盘的复振幅
R = sqrt(z^2+X.^2+Y.^2);    % 衍射传播空间距离
coef = z./(2*pi*R) .*(1i*k -1./R) .*exp(-1i*k*R);   % 瑞利索末菲衍射系数
tic
Eout = conv2(coef,Ein,'same');  % 衍射的二维卷积计算，得到输出复振幅
toc

Iout = abs(Eout).^2;            % 输出光强
figure(1);imagesc(x,x,Iout);colormap hot;axis image off;
figure(2);mesh(Iout);view(0,82);colormap hot;axis off

Elapsed time is 7.304039 seconds.
修改GPU 代码：

% 本程序计算圆形高斯光束入射到位于 束腰处的 圆盘产生的衍射及伯松亮斑
close all;clear all;
w = 0.25;                    % 入射光束腰，束腰半径为0.25mm
lamda = 0.632e-3;           % 入射激光波长
k = 2*pi/lamda;             % 波数
q = 1i*pi*w^2/lamda;        % 束腰处的q参数
N = 500;                    % 光强采样精度
x = linspace(-0.5,0.5,N);   % 观察光斑的坐标范围
z = 10;                     % 衍射传播的距离
E0 = exp(-1i*k*x.^2/2/q);   % 束腰处的光场复振幅分布

x=gsingle(x);%DATA TO  GPU
E0=gsingle(E0);

[X Y] = meshgrid(x,x);      % 坐标网格化
[Ex Ey] =meshgrid(E0,E0);   % 光场二维化
E1 = Ex.*Ey;                % 束腰处的二维光场

gfigure(1);imagesc(abs(E1).^2); % 圆形高斯光束


% colormap hot;axis image;
y=x;
Disk = gones(N);     % 定义半径为0.1mm的挡光圆盘
gfor m = 1:N
     for n = 1:N
%         if sqrt(x(m)^2+x(n)^2) <= 0.1
            Disk(m,n) = Disk(m,n)*~(sqrt(x(m)^2+x(n)^2) <= 0.1);
%         end
     end
gend

Ein = E1.*Disk;         % 透过圆盘的复振幅
R = sqrt(z^2+X.^2+Y.^2);    % 衍射传播空间距离
coef = z./(2*pi*R) .*(1i*k -1./R) .*exp(-1i*k*R);   % 瑞利索末菲衍射系数

tic
gsync;
Eout = conv2(coef,Ein,'same');  % 衍射的二维卷积计算，得到输出复振幅
geval(Eout);
gsync;
toc

Iout = abs(Eout).^2;            % 输出光强

Iout=double(Iout);
x=double(x);
figure(2);imagesc(x,x,Iout);
colormap hot;
axis image off;
figure(3);mesh(Iout);
view(0,82);colormap hot;axis off

Elapsed time is 0.003671 seconds..


>> ginfo
Jacket v2.0 (build 0ae6cf5) by AccelerEyes (64-bit Windows)
A newer version of Jacket is available: http://accelereyes.com/download_jacket
License Type: Designated Computer (C:\Program Files\AccelerEyes\Jacket\engine\jlicense.dat)
Addons: MGL16, JMC, SDK, DLA, SLA
CUDA toolkit 4.0, driver 296.70
GPU1 Tesla M2090, 5376 MB, Compute 2.0 (single,double) (in use)
GPU2 Tesla M2090, 5376 MB, Compute 2.0 (single,double)
Memory Usage: 5300 MB free (5376 MB total)
>>

在JACKET两个版本上分别测试
>> ginfo
Jacket v2.1 (build e2c1fa6) by AccelerEyes (64-bit Windows)
License Type: Designated Computer (C:\Program Files\AccelerEyes\Jacket\engine\jlicense.dat)
Addons: MGL16, JMC, SDK, DLA, SLA
CUDA toolkit 4.1, driver 296.70
GPU1 Tesla M2090, 5376 MB, Compute 2.0 (single,double) (in use)
GPU2 Tesla M2090, 5376 MB, Compute 2.0 (single,double)
Memory Usage: 5105 MB free (5376 MB total)

Elapsed time is 0.003604 seconds.
>> ed
Elapsed time is 0.003579 seconds.
>>