There is a broadcast instruction for float 32 values but there does not seem to be one listed for int32 values.
As far as I understand it can be done using VPBROADCASTD instruction.
Detailed information can be found in Intel Xeon Phi Coprocessor Instruction Set Architecture Reference Manual (http://download-software.intel.com/sites/default/files/forum/278102/327364001en.pdf), page 459.
Also you can use Intel C++ Compiler or Intel Fortran Compiler with automatic vectorization. In that case the compiler will care what SIMD instructions to use. Just use -O2 key in the compiler command line and #pragma simd before the loop. Of course, we need to care about simple enough loop structure and avoid data dependencies. Additional information can be found here: http://software.intel.com/mic-developer, training, advanced videos, vectorization.
I am writing a compiler myself so I dont have the option of using their one. I will try your suggestion
Experiment seems to show that you can just issue an instruction to load it as 32 bit floats without conversion and it works using vbroadcastss
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators