如何使用GCC或LLVM编译Intel VNNI指令（VPDPBUSD、VPDPWSSD）？

Hey there! I’ve looked into this topic in detail, so here’s a breakdown of how to get those VNNI instructions compiling for your simulation work:

GCC Setup #

• Stick to recent versions: GCC 12 and later offer reliable support for AVX512 VNNI intrinsics. If you’re on an older stable release, you’ll need to grab the trunk (development) build to access the latest VNNI-related features.
• Enable with compile flags: Add -mavx512vnni (pair with -mavx512f if your code relies on base AVX512 functionality) to your compile command. Example:

  gcc -mavx512vnni -O2 your_vnni_code.c -o vnni_binary
  

• If intrinsics like _mm512_dpbusd_epi32 (which maps to VPDPBUSD) aren’t recognized, double-check your GCC version—trunk builds have the most up-to-date intrinsic definitions.

LLVM/Clang Setup #

• Clang version requirements: Clang 13 and above include full support for AVX512 VNNI. For experimental or newly added VNNI instructions, using the latest trunk build is your best bet.
• Compile flags: Use -mavx512vnni to enable the extension. Example command:

  clang -mavx512vnni -O2 your_vnni_code.c -o vnni_binary
  

• In rare cases for cutting-edge features, you might need to add -mllvm -enable-vnni, but this is usually unnecessary in recent stable releases.

Fallback: Inline Assembly #

If compiler intrinsics aren’t working for specific instructions (like VPDPBUSD or VPDPWSSD), you can directly embed assembly code to bypass compiler support gaps—perfect for your simulation needs. Here’s an example for VPDPBUSD:

#include <immintrin.h>

__m512i vpdpbusd_simulate(__m512i src1, __m512i src2, __m512i dest) {
    // VPDPBUSD: Multiply and accumulate packed unsigned bytes into dwords
    asm volatile ("vpdpbusd %[dest], %[src2], %[src1]"
                  : [src1] "+x" (src1)
                  : [src2] "x" (src2), [dest] "x" (dest));
    return src1;
}

You can adapt this pattern directly for VPDPWSSD—just use the same assembly mnemonic as the instruction name.

Simulation Tip #

If you’re running code on hardware that doesn’t natively support AVX512 VNNI, use QEMU to simulate a compatible CPU. Launch it with a flag that enables VNNI, like:

qemu-x86_64 -cpu Skylake-Server,avx512vnni your_vnni_binary

This lets you test your compiled code even without physical VNNI-supporting hardware.

内容的提问来源于stack exchange，提问作者user9477432