Fast Matrix-Vector-Multiplication By FFT

原创已于 2023-03-03 10:55:21 修改 · 274 阅读

0 ·

本内容遵循CC 4.0 BY-SA版权协议

收录于

Python语言专栏

于 2022-11-15 10:26:46 首次发布

本文探讨了如何通过使用Toeplitz矩阵和快速傅立叶变换（FFT）技术提高矩阵乘法的效率，展示了用C++实现的示例，并比较了不同方法的时间性能。通过FFT，运算速度得到显著提升，尤其是在处理稀疏矩阵时效果明显。

在这里插入图片描述

>> l = rand(1000,1);
>> l1 = l';
>> C = toeplitz([l1(1) fliplr(l1(2:end))], l1);
>> C(1:5, 1:5)

ans =

    0.8147    0.9058    0.1270    0.9134    0.6324
    0.8667    0.8147    0.9058    0.1270    0.9134
    0.0600    0.8667    0.8147    0.9058    0.1270
    0.4440    0.0600    0.8667    0.8147    0.9058
    0.9509    0.4440    0.0600    0.8667    0.8147
>> x = rand(1000,1);
>> y=C *x;
>> for i = 1:100, y=C *x; end;
>> tic; y=C *x; toc;
Elapsed time is 0.000913 seconds.
>> tic; for i = 1:100, y=C *x; end; toc
Elapsed time is 0.074619 seconds.

在这里插入图片描述

>> y1 = ifft(fft(C(1:end,1)) .* fft(x));
>> norm(y-y1)

ans =

   4.4646e-12
>> ll = C(1:end,1);
>> tic; for i = 1:100, y1 = ifft(fft(ll) .* fft(x)); end; toc
Elapsed time is 0.011608 seconds.
误差非常小，同时运算速度得到了几倍的提升，大约 0.1 ms/loop

在这里插入图片描述

>> l = rand(1000,1);
>> T = toeplitz(l, [l(1); zeros(999,1)]);
>> T(1:5, 1:5)

ans =

    0.8147         0         0         0         0
    0.9058    0.8147         0         0         0
    0.1270    0.9058    0.8147         0         0
    0.9134    0.1270    0.9058    0.8147         0
    0.6324    0.9134    0.1270    0.9058    0.8147
>> x = rand(1000,1);
>> tic; y = T*x; toc
Elapsed time is 0.000593 seconds.
>> tic; for i = 1:100, y = T*x; end; toc
Elapsed time is 0.010883 seconds.

当然我们还可以把矩阵变成稀疏矩阵

>> T1 = sparse(T);
>> tic; y = T1*x; toc
Elapsed time is 0.826337 seconds.
>> tic; for i = 1:100, y = T1*x; end; toc
Elapsed time is 0.033503 seconds.
>> tic; y = T1*x; toc
Elapsed time is 0.000596 seconds.

在这里插入图片描述

>> y1 = T*x;
>> y2 = ifft(fft(l1).*fft([x;zeros(1000,1)]));
>> norm(y2(1:1000) - y1)

ans =

   3.4828e-12
>> tic; y2 = ifft(fft(l1).*fft([x;zeros(1000,1)])); toc
Elapsed time is 0.000527 seconds.
>> tic; for i = 1:100,  y2 = ifft(fft(l1).*fft([x;zeros(1000,1)])); end; toc
Elapsed time is 0.005841 seconds.