[clang] [compiler-rt] [llvm] [X86] AVX10_V2_AUX Implementation (PR #206888)

[via llvm-commits]

================
@@ -5042,6 +5063,301 @@ let Features = "avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<512>
   def vgetmantbf16512_mask : X86Builtin<"_Vector<32, __bf16>(_Vector<32, __bf16>, _Constant int, _Vector<32, __bf16>, unsigned int)">;
 }
 
+// AVX10 V2 AUX - Convert instructions
+
+// Convert from FP32 to FP8
+
+// VCVTPS2BF8
+let Features = "avx10v2aux", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
+  def vcvtps2bf8_128 : X86Builtin<"_Vector<16, char>(_Vector<4, float>)">;
+  def vcvtps2bf8_128_mask : X86Builtin<"_Vector<16, char>(_Vector<4, float>, _Vector<16, char>, unsigned char)">;
----------------
ganeshgit wrote:

This was pending and updated now for 128/256. We use single def now. We don't use like vpmovssdb since we use select here for narrowing converts at 512.

https://github.com/llvm/llvm-project/pull/206888