GCC 13 Compiler Merges Support For Intel AVX-NE-CONVERT
Intel compiler engineers have been busy recently beginning enablement work for new instruction set capabilities for next-generation processors from Meteor Lake to Sierra Forest. This has also been happening on the LLVM side too in addition to the GNU Compiler Collection that is widely used on Linux systems.
As of this morning, Intel AVX-NE-CONVERT is the newest feature having been merged in GCC 13 by Intel engineer Lingling Kong. As part of this the "-mavxneconvert" option is added to GCC for enabling the functionality to enable AVXNECONVERT instructions while it will be turned on automatically with forthcoming "-march=" options for these next-generation Intel CPUs.
Intel's programming reference manual this month outlined AVX-NE-CONVERT alongside other new instructions for Sierra Forest and Grand Ridge. AVX-NE-CONVERT allows converting BF16 to FP32, FP16 to FP32, and single-precision floating-point values to BF16. All the AVX-NE-CONVERT details are outlined in full via the latest PRM update.
Intel's Grand Ridge platform will be up to 24-core Atom SoCs while Sierra Forest is for the first-generation E-core-only Xeon server processors. These E-core CPUs lack AVX-512 but at least add a number of new instructions to enhance their overall performance.