Is casting vector Bool/Mask to Float valid ?

[xacond00]

Dumb question.

Sometimes it's just too convenient to write:
Float(cond) instead of dr::select(cond, 1.f, 0.f)
or:
Float(cond) * x instead of dr::select(cond, x, 0.f)

Appears to be working fine, is there any performance penalty ?

[wjakob]

* will generate an actual multiplication instruction. You will be at the mercy of the downstream compiler figuring out that this can be simplified. If in doubt, capture or the generated assembly or SASS.

Alternatively, you can use & which works for Dr.Jit arrays + masks.

[xacond00]

Alternatively, you can use & which works for Dr.Jit arrays + masks.

Wait a minute, I just now noticed, it doesn't work for Floats anyways... Which my question was about. Most likely because of scalar variant being 'float'. I cant use this in general functions, without making constexpr check first, inflating the code even more :D

[merlinND]

Could you please share an example where it doesn't output the result you expect?
The usage in diffuse.cpp that I linked above is general and works in all variants, including scalar.

[xacond00]

Could you please share an example where it doesn't output the result you expect? The usage in diffuse.cpp that I linked above is general and works in all variants, including scalar.

Doesn't work with Float type in scalar variant specifically. The examples you provided, use mask on the Spectrum type not Float.

[njroussel]

Indeed, it only applies to static arrays or structs, not to 1D types. Personally, that's always been the guiding factor between select(cond, val, 0 and val & cond.

[xacond00]

Indeed, it only applies to static arrays or structs, not to 1D types. Personally, that's always been the guiding factor between select(cond, val, 0 and val & cond.

Yeah yeah, I get it too now. Thanks.