AVX128: Implement support for vpsubsw

FEX-Emu · Jun 18, 2024 · b7f8202 · b7f8202
1 parent 1849c79
commit b7f8202
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 1 deletion.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -1079,6 +1079,9 @@ class OpDispatchBuilder final : public IREmitter {
   template<size_t ElementSize>
   void AVX128_VPHSUB(OpcodeArgs);
 
+  Ref AVX128_PHSUBSWImpl(Ref Src1, Ref Src2);
+  void AVX128_VPHSUBSW(OpcodeArgs);
+
   // End of AVX 128-bit implementation
 
   void InvalidOp(OpcodeArgs);

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp
@@ -257,7 +257,7 @@ void OpDispatchBuilder::InstallAVX128Handlers() {
 
     {OPD(2, 0b01, 0x05), 1, &OpDispatchBuilder::AVX128_VPHSUB<2>},
     {OPD(2, 0b01, 0x06), 1, &OpDispatchBuilder::AVX128_VPHSUB<4>},
-    // TODO: {OPD(2, 0b01, 0x07), 1, &OpDispatchBuilder::VPHSUBSWOp},
+    {OPD(2, 0b01, 0x07), 1, &OpDispatchBuilder::AVX128_VPHSUBSW},
 
     {OPD(2, 0b01, 0x08), 1, &OpDispatchBuilder::AVX128_VPSIGN<1>},
     {OPD(2, 0b01, 0x09), 1, &OpDispatchBuilder::AVX128_VPSIGN<2>},
@@ -1791,4 +1791,18 @@ void OpDispatchBuilder::AVX128_VPHSUB(OpcodeArgs) {
   AVX128_VectorBinaryImpl(Op, GetDstSize(Op), ElementSize, [this](size_t _ElementSize, Ref Src1, Ref Src2) { return AVX128_PHSUBImpl(Src1, Src2, _ElementSize); });
 }
 
+Ref OpDispatchBuilder::AVX128_PHSUBSWImpl(Ref Src1, Ref Src2) {
+  const uint8_t ElementSize = 2;
+
+  auto Even = _VUnZip(OpSize::i128Bit, ElementSize, Src1, Src2);
+  auto Odd = _VUnZip2(OpSize::i128Bit, ElementSize, Src1, Src2);
+
+  // Saturate back down to the result
+  return _VSQSub(OpSize::i128Bit, ElementSize, Even, Odd);
+}
+
+void OpDispatchBuilder::AVX128_VPHSUBSW(OpcodeArgs) {
+  AVX128_VectorBinaryImpl(Op, GetDstSize(Op), OpSize::i16Bit, [this](size_t _ElementSize, Ref Src1, Ref Src2) { return AVX128_PHSUBSWImpl(Src1, Src2); });
+}
+
 } // namespace FEXCore::IR