https://community.intel.com/t5/Intel-Fortran-Compiler/auto-vectorization/m-p/777126#M25665 <DIV></DIV><P><FONT face="Times New Roman" size="3">(1) This is generally true, although transformations such as loop interchanging and collapsing can involve outer loops too. Conversely, loop materialization enables vectorization of straightline code.</FONT></P><P><FONT face="Times New Roman" size="3">(2) For double-precision data, loops with just two iterations can still be vectorized, although the compiler may, by default, reject such short running loops for efficiency reasons, like misaligned or too much setup/cleanup overhead. Use the vectorization diagnostics (-vec_report2) to inspect the reason why loop are not vectorized, and experiment with pragmas and other compiler hints to see if efficient vectorization is possible. Perhaps you can post an example?</FONT></P><P><FONT face="Times New Roman" size="3">Hope this helps.</FONT></P><P><FONT face="Times New Roman" size="3">Aart Bik<BR /></FONT><A href="http://www.aartbik.com/" target="_blank"><FONT face="Times New Roman" size="3"></FONT></A><A href="http://www.aartbik.com/" target="_blank">http://www.aartbik.com/</A></P><P><FONT face="Times New Roman" size="3"></FONT></P> Mon, 24 Apr 2006 09:40:48 GMT Intel_C_Intel 2006-04-24T09:40:48Z https://community.intel.com/t5/Intel-Fortran-Compiler/auto-vectorization/m-p/777125#M25664 <DIV><DIV>1. IFort vectorizes only inner loops.</DIV><DIV>2. IFort does not vectorize low trip loops (2 steps)</DIV><DIV>Can I conclude that small (2x2) matrix operations (sum, multiply) will not be auto-vectorized althoug there are sse instructions to perform similar operations ?</DIV></DIV> Mon, 24 Apr 2006 08:36:33 GMT https://community.intel.com/t5/Intel-Fortran-Compiler/auto-vectorization/m-p/777125#M25664 manoel 2006-04-24T08:36:33Z https://community.intel.com/t5/Intel-Fortran-Compiler/auto-vectorization/m-p/777126#M25665 <DIV></DIV><P><FONT face="Times New Roman" size="3">(1) This is generally true, although transformations such as loop interchanging and collapsing can involve outer loops too. Conversely, loop materialization enables vectorization of straightline code.</FONT></P><P><FONT face="Times New Roman" size="3">(2) For double-precision data, loops with just two iterations can still be vectorized, although the compiler may, by default, reject such short running loops for efficiency reasons, like misaligned or too much setup/cleanup overhead. Use the vectorization diagnostics (-vec_report2) to inspect the reason why loop are not vectorized, and experiment with pragmas and other compiler hints to see if efficient vectorization is possible. Perhaps you can post an example?</FONT></P><P><FONT face="Times New Roman" size="3">Hope this helps.</FONT></P><P><FONT face="Times New Roman" size="3">Aart Bik<BR /></FONT><A href="http://www.aartbik.com/" target="_blank"><FONT face="Times New Roman" size="3"></FONT></A><A href="http://www.aartbik.com/" target="_blank">http://www.aartbik.com/</A></P><P><FONT face="Times New Roman" size="3"></FONT></P> Mon, 24 Apr 2006 09:40:48 GMT https://community.intel.com/t5/Intel-Fortran-Compiler/auto-vectorization/m-p/777126#M25665 Intel_C_Intel 2006-04-24T09:40:48Z