Merge pull request #1037 from aroulin/shader-ex2-lg2

Shader: Implement EX2 and LG2 in interpreter/JIT

src/video_core/shader/shader_interpreter.cpp

@@ -334,6 +334,42 @@ void RunInterpreter(UnitState<Debug>& state) {
                 Record<DebugDataRecord::CMP_RESULT>(state.debug, iteration, state.conditional_code);
                 break;
 
+            case OpCode::Id::EX2:
+            {
+                Record<DebugDataRecord::SRC1>(state.debug, iteration, src1);
+                Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest);
+
+                // EX2 only takes first component exp2 and writes it to all dest components
+                float24 ex2_res = float24::FromFloat32(std::exp2(src1[0].ToFloat32()));
+                for (int i = 0; i < 4; ++i) {
+                    if (!swizzle.DestComponentEnabled(i))
+                        continue;
+
+                    dest[i] = ex2_res;
+                }
+
+                Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest);
+                break;
+            }
+
+            case OpCode::Id::LG2:
+            {
+                Record<DebugDataRecord::SRC1>(state.debug, iteration, src1);
+                Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest);
+
+                // LG2 only takes the first component log2 and writes it to all dest components
+                float24 lg2_res = float24::FromFloat32(std::log2(src1[0].ToFloat32()));
+                for (int i = 0; i < 4; ++i) {
+                    if (!swizzle.DestComponentEnabled(i))
+                        continue;
+
+                    dest[i] = lg2_res;
+                }
+
+                Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest);
+                break;
+            }
+
             default:
                 LOG_ERROR(HW_GPU, "Unhandled arithmetic instruction: 0x%02x (%s): 0x%08x",
                           (int)instr.opcode.Value().EffectiveOpCode(), instr.opcode.Value().GetInfo().name, instr.hex);

src/video_core/shader/shader_jit_x64.cpp

@@ -25,8 +25,8 @@ const JitFunction instr_table[64] = {
     &JitCompiler::Compile_DP4,      // dp4
     nullptr,                        // dph
     nullptr,                        // unknown
-    nullptr,                        // ex2
+    &JitCompiler::Compile_EX2,      // ex2
-    nullptr,                        // lg2
+    &JitCompiler::Compile_LG2,      // lg2
     nullptr,                        // unknown
     &JitCompiler::Compile_MUL,      // mul
     nullptr,                        // lge
@@ -331,6 +331,24 @@ void JitCompiler::Compile_DP4(Instruction instr) {
     Compile_DestEnable(instr, SRC1);
 }
 
+void JitCompiler::Compile_EX2(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    MOVSS(XMM0, R(SRC1));
+    ABI_CallFunction(reinterpret_cast<const void*>(exp2f));
+    SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0, 0, 0, 0));
+    MOVAPS(SRC1, R(XMM0));
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_LG2(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    MOVSS(XMM0, R(SRC1));
+    ABI_CallFunction(reinterpret_cast<const void*>(log2f));
+    SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0, 0, 0, 0));
+    MOVAPS(SRC1, R(XMM0));
+    Compile_DestEnable(instr, SRC1);
+}
+
 void JitCompiler::Compile_MUL(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
     Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);

src/video_core/shader/shader_jit_x64.h

@@ -37,6 +37,8 @@ public:
     void Compile_ADD(Instruction instr);
     void Compile_DP3(Instruction instr);
     void Compile_DP4(Instruction instr);
+    void Compile_EX2(Instruction instr);
+    void Compile_LG2(Instruction instr);
     void Compile_MUL(Instruction instr);
     void Compile_FLR(Instruction instr);
     void Compile_MAX(Instruction instr);