What is OpenCL's select operator useful for?

user_1593077 0
Viewed 281

I've noticed that OpenCL has a select() function/builtin/operator, which seems to be similar to the ternary operator in C and C++, but not quite. What are the differences between select() and ?:, and why is the former even required if we have the latter?

opencl gpgpu conditional-operator
1 Answers

The reason the select operator is necessary/useful is for working with OpenCL vector types, like int2, float4 etc. Unlike in C++, where you could overload various operators to give them custom semantics - in C (and OpenCL C) there's only the default behavior. For the ternary operator, that means that for

x ? expression_for_true : expression_for_false

a single check will be performed, and the single appropriate value will be used - even if all three operands are of an OpenCL vector type.

Instead, with select():

  • All elements must have the same dimensions as vector types, e.g. int4, float4 and float4.
  • a different check is performed for every position within the vector type.
  • a different choice between elements of the true and false expressions is made for every position within the vector type

Also, to confuse us a bit, the order of parameters to select() is different than with the ternary operator: x ? y : z corresponds to select(z, y, x).

Thus if 

x = (int4) ( 1, 0 );
y = (float4) ( 1.2, 3.4 );
z = (float4) ( 5.6, 7.8 );

then

select(z, y, x) == (float4) ( 1.2, 7.8 );

See also the Khronos OpenCL documentation, here and here.

user_15930770
Related