Conditonal checks are not allowed in GPU kernels. There is no in-built function to filter out the NaN values from vectors in GPU kernels. This leads to interesting implementation of functions like SINC(x) which require filtering out NaN values in cases where supported functions like ISNAN return NaNs as logicals.
In such a scenario, filtering out the NaN values from a vector containing non-zeros and NaN's can be done using masking.
No value generated by SINC function will exceed 1 by definition which can be demonstrated by the following code:
a = -10*pi:0.01:10*pi;
plot(sinc(a));
MAX or MIN comparison of any number with NaN gives the same number which can be demonstrated by the following code:
b = NaN;
c = -1e-10;
max(b,c)
min(b,c)
Since MAX and MIN are supported on GPUs and work well on vectors, taking this into account, for the case of SINC, without using either loops or condition variables, we can create masking to filter out the NaN values.
This can be further illustrated by the following example using ONES (because the upper limit of SINC is 1)to generate the 'masking' vector:
a = [ .3 .4 NaN .5]
b = ones(size(a))
c = min(a,b)
A = gpuarray(a)
B = ones(size(A))
C = min(A,B)% this will create a GPUArray C or you can overwrite A
cc = gather(C)
isequal(c,cc) % the result was true on my machine
This approach will work as long as the MAX or MIN values of the input function are known before hand.
