Vector of doubles to vector of floats type conversion

MLIR
#1

Hi,

I am currently playing around with some vectorization attempts and am a little bit stuck at vector type conversion.

My goal is to be able to convert a vector<8xf64> to a vector<8xf32>. The vector<8xf64> is obtained through a vector.load of f64 memrefs, but the rest of the program uses f32 so I need f32 in the long run. I was hoping for something similar to _mm256_cvtpd_ps or a vector version of std.fptrunc, but can’t seem to find anything like that in the dialects.

I know that I could just read scalar values from the memrefs, apply std.fptrunc and then put them into a vector, but this seems inefficient to me compared to a single vector.load.

So my question is, what is the best way to convert a vector<8xf64> to a vector<8xf32>? Is there maybe even some simple type casting I missed scanning the docs?

Thank you in advance!

#2

fptrunc should work on vectors.

#3

Oh. I didn’t try fptrunc because the docs explicitly state that only scalars are currently supported.

Thank you for the quick answer!

#4

Documentation often gets outdated. Tests are the better source of truth. Please update the documentation.