Is it possible to transform from one type to another using thrust::transform/thrust::transform_reduce?

[SagaciousFish]

I'm recently trying to GPU-accelerate some function that finds maximum of some_property hidden in a struct. This requires transforming the vector of the struct into the vector of some_property, then run reduce to find the maximum.

I tried to implement that, but it's not compiling. Then I realised that all examples are showing the case where the input vector and output vector are of the same type. Also, the Unary Functor is taking one and only type in the template.

Is it true that thrust is just not capable of doing such an operation? Or am I being stupid and having something missing?

[miscco]

I believe you should have a look at the minmax example.

That should get you started

[SagaciousFish]

Thank you for your suggestion, and yes, it's a perfect example.

[pauleonix]

As the example cited by miscco shows, transforming from one type to another is certainly possible. Generally it is desirable to use Structs-of-Arrays (SoA) instead of Arrays-of-Structs (AoS) because for good performance (especially in memory-bound kernels like transform_reduce) you want coalesced global memory access. The arrays can then be combined with thrust::zip_iterator to provide access to multiple properties in algorithms that need them.

Also, the Unary Functor is taking one and only type in the template.

Which unary functor are you referring to? Please be more specific.

[SagaciousFish]

Thank you for your proposal of using Structs-of-Arrays, and you are absolutely correct.

My attempt was an acceleration of parallel data preparation on an array of cuda::std::bitsets (the struct). Hence, it's not quite defying your idea, but thank you for your suggestions anyway.