Column-wise dot-products of two gpuArrays

Question

Ahmed 2013-9-5

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/86418-column-wise-dot-products-of-two-gpuarrays

Is there a good way to compute column-wise dot-products of two matrices of same size when stored on the GPU? Specifically, say A and B are n (rows) times m (columns) matrices, I seek to find the 1 (row) times m (columns) row-vector of dot-products of the columns of A and B (i.e. the diagonal elements of the dot-product matrix). Following minimum working example demonstrates the computations:

n = 10;
m = 100;
A = randn(n,m);
B = randn(n,m);
C = bsxfun(@dot,A,B);

However, if A and B are gpuArrays, then this is not possible:

C = bsxfun(@dot,gpuArray(A),gpuArray(B));
% Error using gpuArray/bsxfun
% Use of 'dot' is not supported

Similarly, using anonymous function '@(a,b) a'*b' instead of '@dot' does not work. In my application n is roughly 100 and m roughly 500000. Hope that someone has an idea how C can be computed directly on the GPU and without for loops, as this was very slow for the given problem size.

Ahmed --