AVX128: Implement AVX_NE_CONVERT

FEXCore/Source/Interface/Core/OpcodeDispatcher.h

@@ -1093,6 +1093,9 @@ public:
 
   void AVX128_VCVTPH2PS(OpcodeArgs);
   void AVX128_VCVTPS2PH(OpcodeArgs);
+  void AVX128_VBCSTNEBF162PS(OpcodeArgs);
+  void AVX128_VCVTNEEBF162PS(OpcodeArgs, bool Even);
+  void AVX128_VCVTNEPS2BF16(OpcodeArgs);
 
   // End of AVX 128-bit implementation
 

FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp

@@ -2355,4 +2355,75 @@ void OpDispatchBuilder::AVX128_VCVTPS2PH(OpcodeArgs) {
   }
 }
 
+void OpDispatchBuilder::AVX128_VBCSTNEBF162PS(OpcodeArgs) {
+  const auto DstSize = GetDstSize(Op);
+  const auto Is128Bit = DstSize == Core::CPUState::XMM_SSE_REG_SIZE;
+
+  auto Src1 = LoadSourceFPR_WithOpSize(Op, Op->Src[0], OpSize::i16Bit, Op->Flags);
+
+  // bfloat16 is converted to fp32 and broadcast through the vector.
+  // bfloat16->float32 is left-shift by 16.
+  Src1 = _VShlI(OpSize::i128Bit, OpSize::i32Bit, Src1, 16);
+  Src1 = _VDupElement(OpSize::i128Bit, OpSize::i32Bit, Src1, 0);
+
+  RefPair Result {};
+  Result.Low = Src1;
+
+  if (Is128Bit) {
+    Result.High = LoadZeroVector(OpSize::i128Bit);
+  } else {
+    Result.High = Result.Low;
+  }
+  AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
+}
+
+void OpDispatchBuilder::AVX128_VCVTNEEBF162PS(OpcodeArgs, bool Even) {
+  const auto DstSize = GetDstSize(Op);
+  const auto Is128Bit = DstSize == Core::CPUState::XMM_SSE_REG_SIZE;
+
+  auto Src = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, !Is128Bit);
+
+  RefPair Result {};
+
+  auto Convert = [&](auto Src, bool Even) -> Ref {
+    // Even or odd elements of each 128-bit register is converted from bfloat16 to float32.
+    // For even, just left shift the elements.
+    if (Even) {
+      return _VShlI(OpSize::i128Bit, OpSize::i32Bit, Src, 16);
+    }
+
+    // For Odds, we just need to mask off the lower 16-bits of each 32-bit element.
+    // Easiest way to do this is with a right then left shift.
+    auto Tmp = _VUShrI(OpSize::i128Bit, OpSize::i32Bit, Src, 16);
+    return _VShlI(OpSize::i128Bit, OpSize::i32Bit, Tmp, 16);
+  };
+
+  Result.Low = Convert(Src.Low, Even);
+
+  if (Is128Bit) {
+    Result.High = LoadZeroVector(OpSize::i128Bit);
+  } else {
+    Result.High = Convert(Src.High, Even);
+  }
+  AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
+}
+
+void OpDispatchBuilder::AVX128_VCVTNEPS2BF16(OpcodeArgs) {
+  const auto SrcSize = OpSizeFromSrc(Op);
+  const auto Is128BitSrc = SrcSize == OpSize::i128Bit;
+  auto Src = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, !Is128BitSrc);
+
+  // Convert 4 or 8 32-bits floats in to bfloat16.
+  // This instruction only operates with rounding mode set to `Rounds to Nearest (even)`.
+
+  RefPair Result {};
+  Result.Low = _VBFCVTN(OpSize::i128Bit, Src.Low);
+  if (!Is128BitSrc) {
+    Result.Low = _VBFCVTN2(OpSize::i128Bit, Result.Low, Src.High);
+  }
+
+  Result.High = LoadZeroVector(OpSize::i128Bit);
+  AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
+}
+
 } // namespace FEXCore::IR

FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp

@@ -321,6 +321,7 @@ namespace AVX128 {
     {OPD(2, 0b01, 0x59), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VBROADCAST, OpSize::i64Bit>},
     {OPD(2, 0b01, 0x5A), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VBROADCAST, OpSize::i128Bit>},
 
+    {OPD(2, 0b10, 0x72), 1, &OpDispatchBuilder::AVX128_VCVTNEPS2BF16},
     {OPD(2, 0b01, 0x78), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VBROADCAST, OpSize::i8Bit>},
     {OPD(2, 0b01, 0x79), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VBROADCAST, OpSize::i16Bit>},
 
@@ -365,6 +366,13 @@ namespace AVX128 {
     {OPD(2, 0b01, 0xA6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, true, 2, 1, 3>},  // VFMADDSUB
     {OPD(2, 0b01, 0xA7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, false, 2, 1, 3>}, // VFMSUBADD
 
+    {OPD(2, 0b00, 0xB0), 1, &OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::AVX128_VCVTNEEBF162PS, false>},
+    {OPD(2, 0b01, 0xB0), 1, &OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::AVX128_VCVTNEEBF162PS, true>},
+    {OPD(2, 0b10, 0xB0), 1, &OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::AVX128_VCVTNEEBF162PS, true>},
+    {OPD(2, 0b11, 0xB0), 1, &OpDispatchBuilder::Bind<&IR::OpDispatchBuilder::AVX128_VCVTNEEBF162PS, false>},
+    {OPD(2, 0b01, 0xB1), 1, &IR::OpDispatchBuilder::AVX128_VBCSTNEBF162PS},
+    {OPD(2, 0b10, 0xB1), 1, &IR::OpDispatchBuilder::AVX128_VBCSTNEBF162PS},
+
     {OPD(2, 0b01, 0xB6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, true, 2, 3, 1>},  // VFMADDSUB
     {OPD(2, 0b01, 0xB7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::AVX128_VFMAddSubImpl, false, 2, 3, 1>}, // VFMSUBADD
 
@@ -771,6 +779,7 @@ namespace AVX256 {
     {OPD(2, 0b01, 0x59), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VBROADCASTOp, OpSize::i64Bit>},
     {OPD(2, 0b01, 0x5A), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VBROADCASTOp, OpSize::i128Bit>},
 
+    {OPD(2, 0b10, 0x72), 1, &OpDispatchBuilder::UnimplementedOp},
     {OPD(2, 0b01, 0x78), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VBROADCASTOp, OpSize::i8Bit>},
     {OPD(2, 0b01, 0x79), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VBROADCASTOp, OpSize::i16Bit>},
 
@@ -815,6 +824,13 @@ namespace AVX256 {
     {OPD(2, 0b01, 0xA6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VFMAddSubImpl, true, 2, 1, 3>},  // VFMADDSUB
     {OPD(2, 0b01, 0xA7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VFMAddSubImpl, false, 2, 1, 3>}, // VFMSUBADD
 
+    {OPD(2, 0b00, 0xB0), 1, &IR::OpDispatchBuilder::UnimplementedOp},
+    {OPD(2, 0b01, 0xB0), 1, &IR::OpDispatchBuilder::UnimplementedOp},
+    {OPD(2, 0b10, 0xB0), 1, &IR::OpDispatchBuilder::UnimplementedOp},
+    {OPD(2, 0b11, 0xB0), 1, &IR::OpDispatchBuilder::UnimplementedOp},
+    {OPD(2, 0b01, 0xB1), 1, &IR::OpDispatchBuilder::UnimplementedOp},
+    {OPD(2, 0b10, 0xB1), 1, &IR::OpDispatchBuilder::UnimplementedOp},
+
     {OPD(2, 0b01, 0xB6), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VFMAddSubImpl, true, 2, 3, 1>},  // VFMADDSUB
     {OPD(2, 0b01, 0xB7), 1, &OpDispatchBuilder::Bind<&OpDispatchBuilder::VFMAddSubImpl, false, 2, 3, 1>}, // VFMSUBADD