Merge pull request #1062 from aroulin/shader-rcp-rsq

Shader: RCP and RSQ computes only the 1st component

src/common/x64/emitter.cpp

@@ -1535,6 +1535,7 @@ void XEmitter::MAXSS(X64Reg regOp, const OpArg& arg)   {WriteSSEOp(0xF3, sseMAX,
 void XEmitter::MAXSD(X64Reg regOp, const OpArg& arg)   {WriteSSEOp(0xF2, sseMAX, regOp, arg);}
 void XEmitter::SQRTSS(X64Reg regOp, const OpArg& arg)  {WriteSSEOp(0xF3, sseSQRT, regOp, arg);}
 void XEmitter::SQRTSD(X64Reg regOp, const OpArg& arg)  {WriteSSEOp(0xF2, sseSQRT, regOp, arg);}
+void XEmitter::RCPSS(X64Reg regOp, const OpArg& arg)   {WriteSSEOp(0xF3, sseRCP, regOp, arg);}
 void XEmitter::RSQRTSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseRSQRT, regOp, arg);}
 
 void XEmitter::ADDPS(X64Reg regOp, const OpArg& arg)   {WriteSSEOp(0x00, sseADD, regOp, arg);}

src/common/x64/emitter.h

@@ -586,6 +586,7 @@ public:
     void MAXSD(X64Reg regOp, const OpArg& arg);
     void SQRTSS(X64Reg regOp, const OpArg& arg);
     void SQRTSD(X64Reg regOp, const OpArg& arg);
+    void RCPSS(X64Reg regOp, const OpArg& arg);
     void RSQRTSS(X64Reg regOp, const OpArg& arg);
 
     // SSE/SSE2: Floating point bitwise (yes)

src/video_core/shader/shader_interpreter.cpp

@@ -228,13 +228,12 @@ void RunInterpreter(UnitState<Debug>& state) {
             {
                 Record<DebugDataRecord::SRC1>(state.debug, iteration, src1);
                 Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest);
+                float24 rcp_res = float24::FromFloat32(1.0f / src1[0].ToFloat32());
                 for (int i = 0; i < 4; ++i) {
                     if (!swizzle.DestComponentEnabled(i))
                         continue;
 
-                    // TODO: Be stable against division by zero!
-                    // TODO: I think this might be wrong... we should only use one component here
-                    dest[i] = float24::FromFloat32(1.0f / src1[i].ToFloat32());
+                    dest[i] = rcp_res;
                 }
                 Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest);
                 break;
@@ -245,13 +244,12 @@ void RunInterpreter(UnitState<Debug>& state) {
             {
                 Record<DebugDataRecord::SRC1>(state.debug, iteration, src1);
                 Record<DebugDataRecord::DEST_IN>(state.debug, iteration, dest);
+                float24 rsq_res = float24::FromFloat32(1.0f / std::sqrt(src1[0].ToFloat32()));
                 for (int i = 0; i < 4; ++i) {
                     if (!swizzle.DestComponentEnabled(i))
                         continue;
 
-                    // TODO: Be stable against division by zero!
-                    // TODO: I think this might be wrong... we should only use one component here
-                    dest[i] = float24::FromFloat32(1.0f / sqrt(src1[i].ToFloat32()));
+                    dest[i] = rsq_res;
                 }
                 Record<DebugDataRecord::DEST_OUT>(state.debug, iteration, dest);
                 break;

src/video_core/shader/shader_jit_x64.cpp

@@ -529,9 +529,10 @@ void JitCompiler::Compile_MOV(Instruction instr) {
 void JitCompiler::Compile_RCP(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
-    // TODO(bunnei): RCPPS is a pretty rough approximation, this might cause problems if Pica
+    // TODO(bunnei): RCPSS is a pretty rough approximation, this might cause problems if Pica
     // performs this operation more accurately. This should be checked on hardware.
-    RCPPS(SRC1, R(SRC1));
+    RCPSS(SRC1, R(SRC1));
+    SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
 
     Compile_DestEnable(instr, SRC1);
 }
@@ -539,9 +540,10 @@ void JitCompiler::Compile_RCP(Instruction instr) {
 void JitCompiler::Compile_RSQ(Instruction instr) {
     Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
 
-    // TODO(bunnei): RSQRTPS is a pretty rough approximation, this might cause problems if Pica
+    // TODO(bunnei): RSQRTSS is a pretty rough approximation, this might cause problems if Pica
     // performs this operation more accurately. This should be checked on hardware.
-    RSQRTPS(SRC1, R(SRC1));
+    RSQRTSS(SRC1, R(SRC1));
+    SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(0, 0, 0, 0)); // XYWZ -> XXXX
 
     Compile_DestEnable(instr, SRC1);
 }