graphics: support signed 16-bit integer packing

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.cpp

@@ -664,6 +664,7 @@ std::string OpcodeToString(Opcode opcode) {
 		case Opcode::VCvtPknormI16F32: return "v_cvt_pknorm_i16_f32";
 		case Opcode::VCvtPknormU16F32: return "v_cvt_pknorm_u16_f32";
 		case Opcode::VCvtPkU16U32: return "v_cvt_pk_u16_u32";
+		case Opcode::VCvtPkI16I32: return "v_cvt_pk_i16_i32";
 		case Opcode::VPkMadI16: return "v_pk_mad_i16";
 		case Opcode::VPkMulLoU16: return "v_pk_mul_lo_u16";
 		case Opcode::VPkAddI16: return "v_pk_add_i16";

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h

@@ -249,6 +249,7 @@ enum class Opcode {
 	VCvtPknormI16F32,
 	VCvtPknormU16F32,
 	VCvtPkU16U32,
+	VCvtPkI16I32,
 	VPkMadI16,
 	VPkMulLoU16,
 	VPkAddI16,

src/graphics/shader/recompiler/frontend/decode/VectorAluOps.cpp

@@ -272,11 +272,11 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x363u, Opcode::VBfmB32},          {0x364u, Opcode::VBcntU32B32},
     {0x365u, Opcode::VMbcntLoU32B32},   {0x366u, Opcode::VMbcntHiU32B32},
     {0x368u, Opcode::VCvtPknormI16F32}, {0x369u, Opcode::VCvtPknormU16F32},
-    {0x36au, Opcode::VCvtPkU16U32},     {0x36fu, Opcode::VLshlOrB32},
+    {0x36au, Opcode::VCvtPkU16U32},     {0x36bu, Opcode::VCvtPkI16I32},
-    {0x371u, Opcode::VAndOrB32},        {0x372u, Opcode::VOr3B32},
+    {0x36fu, Opcode::VLshlOrB32},       {0x371u, Opcode::VAndOrB32},
-    {0x377u, Opcode::VPermlane16B32},   {0x378u, Opcode::VPermlanex16B32},
+    {0x372u, Opcode::VOr3B32},          {0x377u, Opcode::VPermlane16B32},
-    {0x34bu, Opcode::VFmaF16},          {0x36du, Opcode::VAdd3U32},
+    {0x378u, Opcode::VPermlanex16B32},  {0x34bu, Opcode::VFmaF16},
-    {0x14fu, Opcode::VAlignbyteB32},
+    {0x36du, Opcode::VAdd3U32},         {0x14fu, Opcode::VAlignbyteB32},
 };
 
 constexpr auto VOP3_OPS = Detail::MakeOpcodeTable<0x400>(VOP3_OPCODE_LIST);
@@ -1208,7 +1208,8 @@ uint32_t NativeVop3SourceCount(Opcode opcode) {
 		case Opcode::VBcntU32B32:
 		case Opcode::VCvtPknormI16F32:
 		case Opcode::VCvtPknormU16F32:
-		case Opcode::VCvtPkU16U32: return 2;
+		case Opcode::VCvtPkU16U32:
+		case Opcode::VCvtPkI16I32: return 2;
 		default: return 3;
 	}
 }

src/graphics/shader/recompiler/ir/opcodes/ShaderIROpcodes.cpp

@@ -201,6 +201,7 @@ constexpr LowerMap LOWER_OPS[] = {
     {Decoder::Opcode::VCvtPknormI16F32, Opcode::PackSnorm2x16F32},
     {Decoder::Opcode::VCvtPknormU16F32, Opcode::PackUnorm2x16F32},
     {Decoder::Opcode::VCvtPkU16U32, Opcode::PackU16U32},
+    {Decoder::Opcode::VCvtPkI16I32, Opcode::PackU16U32},
     {Decoder::Opcode::VCvtPkU8F32, Opcode::PackU8F32},
     {Decoder::Opcode::VPackB32F16, Opcode::PackB32F16},
     {Decoder::Opcode::VPkMadI16, Opcode::PackedMadI16},

tests/ShaderRecompilerComputeTests.cpp

@@ -11809,6 +11809,34 @@ TestCase Vop2SdwaSubNcPreservesByteAndWordDestinations() {
           {O::VMovB32, O::VSubNcU32, O::BufferStoreDword, O::SEndpgm}};
 }
 
+TestCase Vop3CvtPkI16I32Captured() {
+  using O = ShaderOpcode;
+
+  std::vector<u32> code;
+  AppendBufferLoadDword(&code, 7, 30);
+  AppendVMovU32(&code, 30, 4);
+  AppendBufferLoadDword(&code, 16, 30);
+  code.push_back(0xd76b0005u);
+  code.push_back(0x00022107u); // v_cvt_pk_i16_i32 v5, v7, v16
+  AppendStoreVgpr(&code, 5, 2);
+  AppendEnd(&code);
+
+  TestCase test;
+  test.name = "Vop3CvtPkI16I32Captured";
+  test.code = std::move(code);
+  test.initial = {0x00018001u, 0xfffe7ffeu, 0};
+  test.expected = {0x00018001u, 0xfffe7ffeu, 0x7ffe8001u};
+  test.opcodes = {O::BufferLoadDword, O::VMovB32, O::VCvtPkI16I32,
+                  O::BufferStoreDword, O::SEndpgm};
+  test.decoded_counts = {
+      {"0x00000014: v_cvt_pk_i16_i32 v5, v7, v16\n", 1}};
+  test.ir_counts = {{" = BitwiseAnd32 ", 2},
+                    {" = ShiftLeftLogical32 ", 1},
+                    {" = BitwiseOr32 ", 1}};
+  test.required_spirv = {"OpBitwiseAnd", "OpShiftLeftLogical", "OpBitwiseOr"};
+  return test;
+}
+
 TestCase Vop2SdwaMinU32PreservesWordDestination() {
   using O = ShaderOpcode;
 
@@ -17567,6 +17595,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(Vop2SdwaAddNcCapturedHighWordDestination);
   AddCase(Vop2SdwaAshrrevCapturedWord0SignExtends);
   AddCase(Vop2SdwaSubNcPreservesByteAndWordDestinations);
+  AddCase(Vop3CvtPkI16I32Captured);
   AddCase(Vop2SdwaMinU32PreservesWordDestination);
   AddCase(VectorShiftCountsMaskLowBits);
   AddCase(VectorVop3IntegerOps);
@@ -21012,6 +21041,11 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, Vop2SdwaAshrrevCapturedWord0SignExtends());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--cvt-pk-i16-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, Vop3CvtPkI16I32Captured());
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--sff1-b64-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, ScalarBitfieldPack());