Merge pull request #2366 from FernandoS27/xmad-fix

Correct XMAD mode, psl and high_b on different encodings.

src/video_core/engines/shader_bytecode.h

@@ -1238,13 +1238,16 @@ union Instruction {
 
     union {
         BitField<20, 16, u64> imm20_16;
+        BitField<35, 1, u64> high_b_rr; // used on RR
         BitField<36, 1, u64> product_shift_left;
         BitField<37, 1, u64> merge_37;
         BitField<48, 1, u64> sign_a;
         BitField<49, 1, u64> sign_b;
+        BitField<50, 2, XmadMode> mode_cbf; // used by CR, RC
         BitField<50, 3, XmadMode> mode;
         BitField<52, 1, u64> high_b;
         BitField<53, 1, u64> high_a;
+        BitField<55, 1, u64> product_shift_left_second; // used on CR
         BitField<56, 1, u64> merge_56;
     } xmad;
 

src/video_core/shader/decode/xmad.cpp

@@ -29,39 +29,55 @@ u32 ShaderIR::DecodeXmad(NodeBlock& bb, u32 pc) {
     const bool is_signed_b = instr.xmad.sign_b == 1;
     const bool is_signed_c = is_signed_a;
 
-    auto [is_merge, op_b, op_c] = [&]() -> std::tuple<bool, Node, Node> {
+    auto [is_merge, is_psl, is_high_b, mode, op_b,
+          op_c] = [&]() -> std::tuple<bool, bool, bool, Tegra::Shader::XmadMode, Node, Node> {
         switch (opcode->get().GetId()) {
         case OpCode::Id::XMAD_CR:
             return {instr.xmad.merge_56,
+                    instr.xmad.product_shift_left_second,
+                    instr.xmad.high_b,
+                    instr.xmad.mode_cbf,
                     GetConstBuffer(instr.cbuf34.index, instr.cbuf34.GetOffset()),
                     GetRegister(instr.gpr39)};
         case OpCode::Id::XMAD_RR:
-            return {instr.xmad.merge_37, GetRegister(instr.gpr20), GetRegister(instr.gpr39)};
+            return {instr.xmad.merge_37, instr.xmad.product_shift_left, instr.xmad.high_b_rr,
+                    instr.xmad.mode,     GetRegister(instr.gpr20),      GetRegister(instr.gpr39)};
         case OpCode::Id::XMAD_RC:
-            return {false, GetRegister(instr.gpr39),
+            return {false,
+                    false,
+                    instr.xmad.high_b,
+                    instr.xmad.mode_cbf,
+                    GetRegister(instr.gpr39),
                     GetConstBuffer(instr.cbuf34.index, instr.cbuf34.GetOffset())};
         case OpCode::Id::XMAD_IMM:
-            return {instr.xmad.merge_37, Immediate(static_cast<u32>(instr.xmad.imm20_16)),
+            return {instr.xmad.merge_37,
+                    instr.xmad.product_shift_left,
+                    false,
+                    instr.xmad.mode,
+                    Immediate(static_cast<u32>(instr.xmad.imm20_16)),
                     GetRegister(instr.gpr39)};
         }
         UNIMPLEMENTED_MSG("Unhandled XMAD instruction: {}", opcode->get().GetName());
-        return {false, Immediate(0), Immediate(0)};
+        return {false, false, false, Tegra::Shader::XmadMode::None, Immediate(0), Immediate(0)};
     }();
 
     op_a = BitfieldExtract(op_a, instr.xmad.high_a ? 16 : 0, 16);
 
     const Node original_b = op_b;
-    op_b = BitfieldExtract(op_b, instr.xmad.high_b ? 16 : 0, 16);
+    op_b = BitfieldExtract(op_b, is_high_b ? 16 : 0, 16);
 
     // TODO(Rodrigo): Use an appropiate sign for this operation
     Node product = Operation(OperationCode::IMul, NO_PRECISE, op_a, op_b);
-    if (instr.xmad.product_shift_left) {
+    if (is_psl) {
         product = Operation(OperationCode::ILogicalShiftLeft, NO_PRECISE, product, Immediate(16));
     }
+    SetTemporal(bb, 0, product);
+    product = GetTemporal(0);
 
     const Node original_c = op_c;
+    const Tegra::Shader::XmadMode set_mode = mode; // Workaround to clang compile error
     op_c = [&]() {
-        switch (instr.xmad.mode) {
+        switch (set_mode) {
         case Tegra::Shader::XmadMode::None:
             return original_c;
         case Tegra::Shader::XmadMode::CLo:
@@ -80,8 +96,13 @@ u32 ShaderIR::DecodeXmad(NodeBlock& bb, u32 pc) {
         }
     }();
 
+    SetTemporal(bb, 1, op_c);
+    op_c = GetTemporal(1);
+
     // TODO(Rodrigo): Use an appropiate sign for this operation
     Node sum = Operation(OperationCode::IAdd, product, op_c);
+    SetTemporal(bb, 2, sum);
+    sum = GetTemporal(2);
     if (is_merge) {
         const Node a = BitfieldExtract(sum, 0, 16);
         const Node b =
@@ -95,4 +116,4 @@ u32 ShaderIR::DecodeXmad(NodeBlock& bb, u32 pc) {
     return pc;
 }
 
-} // namespace VideoCommon::Shader
+} // namespace VideoCommon::Shader