From d46e2a6e7ac22661d6debe090c8c6b25d565613a Mon Sep 17 00:00:00 2001
From: ReinUsesLisp <reinuseslisp@airmail.cc>
Date: Mon, 15 Oct 2018 02:04:31 -0300
Subject: gl_shader_decompiler: Implement non-immediate HADD2 and HMUL2
 instructions

---
 src/video_core/engines/shader_bytecode.h           | 25 +++++++++++
 .../renderer_opengl/gl_shader_decompiler.cpp       | 50 ++++++++++++++++++++++
 2 files changed, 75 insertions(+)

diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index d6d46d277..f84b9883c 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -573,6 +573,22 @@ union Instruction {
         BitField<49, 1, u64> negate_a;
     } alu_integer;
 
+    union {
+        BitField<39, 1, u64> ftz;
+        BitField<32, 1, u64> saturate;
+        BitField<49, 2, HalfMerge> merge;
+
+        BitField<43, 1, u64> negate_a;
+        BitField<44, 1, u64> abs_a;
+        BitField<47, 2, HalfType> type_a;
+
+        BitField<31, 1, u64> negate_b;
+        BitField<30, 1, u64> abs_b;
+        BitField<47, 2, HalfType> type_b;
+
+        BitField<35, 2, HalfType> type_c;
+    } alu_half;
+
     union {
         BitField<40, 1, u64> invert;
     } popc;
@@ -1165,6 +1181,10 @@ public:
         LEA_RZ,
         LEA_IMM,
         LEA_HI,
+        HADD2_C,
+        HADD2_R,
+        HMUL2_C,
+        HMUL2_R,
         POPC_C,
         POPC_R,
         POPC_IMM,
@@ -1238,6 +1258,7 @@ public:
         ArithmeticImmediate,
         ArithmeticInteger,
         ArithmeticIntegerImmediate,
+        ArithmeticHalf,
         Bfe,
         Shift,
         Ffma,
@@ -1409,6 +1430,10 @@ private:
             INST("001101101101----", Id::LEA_IMM, Type::ArithmeticInteger, "LEA_IMM"),
             INST("010010111101----", Id::LEA_RZ, Type::ArithmeticInteger, "LEA_RZ"),
             INST("00011000--------", Id::LEA_HI, Type::ArithmeticInteger, "LEA_HI"),
+            INST("0111101-1-------", Id::HADD2_C, Type::ArithmeticHalf, "HADD2_C"),
+            INST("0101110100010---", Id::HADD2_R, Type::ArithmeticHalf, "HADD2_R"),
+            INST("0111100-1-------", Id::HMUL2_C, Type::ArithmeticHalf, "HMUL2_C"),
+            INST("0101110100001---", Id::HMUL2_R, Type::ArithmeticHalf, "HMUL2_R"),
             INST("0101000010000---", Id::MUFU, Type::Arithmetic, "MUFU"),
             INST("0100110010010---", Id::RRO_C, Type::Arithmetic, "RRO_C"),
             INST("0101110010010---", Id::RRO_R, Type::Arithmetic, "RRO_R"),
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index c6ae8c3b4..a1a0babe8 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -1827,6 +1827,56 @@ private:
 
             break;
         }
+        case OpCode::Type::ArithmeticHalf: {
+            if (opcode->GetId() == OpCode::Id::HADD2_C || opcode->GetId() == OpCode::Id::HADD2_R) {
+                ASSERT_MSG(instr.alu_half.ftz == 0, "Unimplemented");
+            }
+            const bool negate_a =
+                opcode->GetId() != OpCode::Id::HMUL2_R && instr.alu_half.negate_a != 0;
+            const bool negate_b =
+                opcode->GetId() != OpCode::Id::HMUL2_C && instr.alu_half.negate_b != 0;
+
+            const std::string op_a =
+                GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr8, 0, false), instr.alu_half.type_a,
+                             instr.alu_half.abs_a != 0, negate_a);
+
+            std::string op_b;
+            switch (opcode->GetId()) {
+            case OpCode::Id::HADD2_C:
+            case OpCode::Id::HMUL2_C:
+                op_b = regs.GetUniform(instr.cbuf34.index, instr.cbuf34.offset,
+                                       GLSLRegister::Type::UnsignedInteger);
+                break;
+            case OpCode::Id::HADD2_R:
+            case OpCode::Id::HMUL2_R:
+                op_b = regs.GetRegisterAsInteger(instr.gpr20, 0, false);
+                break;
+            default:
+                UNREACHABLE();
+                op_b = "0";
+                break;
+            }
+            op_b = GetHalfFloat(op_b, instr.alu_half.type_b, instr.alu_half.abs_b != 0, negate_b);
+
+            const std::string result = [&]() {
+                switch (opcode->GetId()) {
+                case OpCode::Id::HADD2_C:
+                case OpCode::Id::HADD2_R:
+                    return '(' + op_a + " + " + op_b + ')';
+                case OpCode::Id::HMUL2_C:
+                case OpCode::Id::HMUL2_R:
+                    return '(' + op_a + " * " + op_b + ')';
+                default:
+                    LOG_CRITICAL(HW_GPU, "Unhandled half float instruction: {}", opcode->GetName());
+                    UNREACHABLE();
+                    return std::string("0");
+                }
+            }();
+
+            regs.SetRegisterToHalfFloat(instr.gpr0, 0, result, instr.alu_half.merge, 1, 1,
+                                        instr.alu_half.saturate != 0);
+            break;
+        }
         case OpCode::Type::Ffma: {
             const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
             std::string op_b = instr.ffma.negate_b ? "-" : "";
-- 
cgit v1.2.3