__m512i _bswap endianness in C/C++?

__m512i _bswap endianness in C/C++?

This might be a noob question, but is there a way to perform a byte swap between big and little endian on a __m512i vector without having to iterate through each value? I really want to avoid using unions, as they can be slow. I can't find any info in documentation or googling around, so hopefully somebody here can help me.

8 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.

I'm checking with our Developer. Stay tuned.

Developer's guidance is that vector byte swap is not available as there is no hardware support for it.

Let me know if you have additional questions.

Look at using _mm256_shuffle_epi8

This will work on AVX2, I haven't checked to see if this is supported on Phi. As long as the byte swaps do not cross 128-bit lanes, this will work for you. You will also have to check to see if this disturbs the high 256-bits. This will affect your strategy to perform 512-bit byte endian shuffle.

Jim Dempsey

>>...I really want to avoid using unions, as they can be slow...

Almost all Intel's intrinsic data types are union based ( there are just a couple of exceptions when they declared as structs ) and if you have some performance issues more technical details need to be provided.

you can read in the data to vector as byte, then use swizzle to swap the elements (bytes),  then write back the vector.  This only do 16 bytes at a time, though.  Reading in full full 64 bytes and swap the bytes by shifts and bit or/and may actually be faster.



@Jim:  Xeon Phi (Knight's Corner) doesn't support any MMX, SSE*, or AVX* flavors.  So __mm256 instructions can't be used.   Sorry


You could just write your own byte swap routine involving the bitwise logical and shift __m512i operations.

The relevant functions are listed under AVX-512... I didn't test it, but i guess they should also work on KNC. There are a lot of functions not listed under AVX-512, but still work on KNC...


Leave a Comment

Please sign in to add a comment. Not a member? Join today