4 files changed, 17 insertions, 22 deletions

diff --git a/src/video_core/command_processor.cpp b/src/video_core/command_processor.cpp
index 694c9f169..66d19cba0 100644
--- a/src/video_core/command_processor.cpp
+++ b/src/video_core/command_processor.cpp
@@ -152,8 +152,8 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                     Shader::UnitState shader_unit;
                     shader_unit.LoadInputVertex(immediate_input, regs.vs.num_input_attributes + 1);
                     shader_engine->Run(shader_unit, regs.vs.main_offset);
-                    Shader::OutputVertex output_vertex =
-                        shader_unit.output_registers.ToVertex(regs.vs);
+                    auto output_vertex = Shader::OutputVertex::FromRegisters(
+                        shader_unit.registers.output, regs, regs.vs.output_mask);
 
                     // Send to renderer
                     using Pica::Shader::OutputVertex;
@@ -291,7 +291,8 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                 shader_engine->Run(shader_unit, regs.vs.main_offset);
 
                 // Retrieve vertex from register data
-                output_vertex = shader_unit.output_registers.ToVertex(regs.vs);
+                output_vertex = Shader::OutputVertex::FromRegisters(shader_unit.registers.output,
+                                                                    regs, regs.vs.output_mask);
 
                 if (is_indexed) {
                     vertex_cache[vertex_cache_pos] = output_vertex;
diff --git a/src/video_core/shader/shader.cpp b/src/video_core/shader/shader.cpp
index 1662b5d38..2da50bd62 100644
--- a/src/video_core/shader/shader.cpp
+++ b/src/video_core/shader/shader.cpp
@@ -19,7 +19,8 @@ namespace Pica {
 
 namespace Shader {
 
-OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
+OutputVertex OutputVertex::FromRegisters(Math::Vec4<float24> output_regs[16], const Regs& regs,
+                                         u32 output_mask) {
     // Setup output data
     OutputVertex ret;
     // TODO(neobrain): Under some circumstances, up to 16 attributes may be output. We need to
@@ -27,13 +28,13 @@ OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
     unsigned index = 0;
     for (unsigned i = 0; i < 7; ++i) {
 
-        if (index >= g_state.regs.vs_output_total)
+        if (index >= regs.vs_output_total)
             break;
 
-        if ((config.output_mask & (1 << i)) == 0)
+        if ((output_mask & (1 << i)) == 0)
             continue;
 
-        const auto& output_register_map = g_state.regs.vs_output_attributes[index];
+        const auto& output_register_map = regs.vs_output_attributes[index];
 
         u32 semantics[4] = {output_register_map.map_x, output_register_map.map_y,
                             output_register_map.map_z, output_register_map.map_w};
@@ -41,7 +42,7 @@ OutputVertex OutputRegisters::ToVertex(const Regs::ShaderConfig& config) const {
         for (unsigned comp = 0; comp < 4; ++comp) {
             float24* out = ((float24*)&ret) + semantics[comp];
             if (semantics[comp] != Regs::VSOutputAttributes::INVALID) {
-                *out = value[i][comp];
+                *out = output_regs[i][comp];
             } else {
                 // Zero output so that attributes which aren't output won't have denormals in them,
                 // which would slow us down later.
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h
index 9d2410487..7d51d0044 100644
--- a/src/video_core/shader/shader.h
+++ b/src/video_core/shader/shader.h
@@ -73,19 +73,13 @@ struct OutputVertex {
         ret.Lerp(factor, v1);
         return ret;
     }
+
+    static OutputVertex FromRegisters(Math::Vec4<float24> output_regs[16], const Regs& regs,
+                                      u32 output_mask);
 };
 static_assert(std::is_pod<OutputVertex>::value, "Structure is not POD");
 static_assert(sizeof(OutputVertex) == 32 * sizeof(float), "OutputVertex has invalid size");
 
-struct OutputRegisters {
-    OutputRegisters() = default;
-
-    alignas(16) Math::Vec4<float24> value[16];
-
-    OutputVertex ToVertex(const Regs::ShaderConfig& config) const;
-};
-static_assert(std::is_pod<OutputRegisters>::value, "Structure is not POD");
-
 /**
  * This structure contains the state information that needs to be unique for a shader unit. The 3DS
  * has four shader units that process shaders in parallel. At the present, Citra only implements a
@@ -98,11 +92,10 @@ struct UnitState {
         // required to be 16-byte aligned.
         alignas(16) Math::Vec4<float24> input[16];
         alignas(16) Math::Vec4<float24> temporary[16];
+        alignas(16) Math::Vec4<float24> output[16];
     } registers;
     static_assert(std::is_pod<Registers>::value, "Structure is not POD");
 
-    OutputRegisters output_registers;
-
     bool conditional_code[2];
 
     // Two Address registers and one loop counter
@@ -128,7 +121,7 @@ struct UnitState {
     static size_t OutputOffset(const DestRegister& reg) {
         switch (reg.GetRegisterType()) {
         case RegisterType::Output:
-            return offsetof(UnitState, output_registers.value) +
+            return offsetof(UnitState, registers.output) +
                    reg.GetIndex() * sizeof(Math::Vec4<float24>);
 
         case RegisterType::Temporary:
diff --git a/src/video_core/shader/shader_interpreter.cpp b/src/video_core/shader/shader_interpreter.cpp
index ecc227089..a6197c10a 100644
--- a/src/video_core/shader/shader_interpreter.cpp
+++ b/src/video_core/shader/shader_interpreter.cpp
@@ -175,7 +175,7 @@ static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData
 
             float24* dest =
                 (instr.common.dest.Value() < 0x10)
-                    ? &state.output_registers.value[instr.common.dest.Value().GetIndex()][0]
+                    ? &state.registers.output[instr.common.dest.Value().GetIndex()][0]
                     : (instr.common.dest.Value() < 0x20)
                           ? &state.registers.temporary[instr.common.dest.Value().GetIndex()][0]
                           : dummy_vec4_float24;
@@ -518,7 +518,7 @@ static void RunInterpreter(const ShaderSetup& setup, UnitState& state, DebugData
 
                 float24* dest =
                     (instr.mad.dest.Value() < 0x10)
-                        ? &state.output_registers.value[instr.mad.dest.Value().GetIndex()][0]
+                        ? &state.registers.output[instr.mad.dest.Value().GetIndex()][0]
                         : (instr.mad.dest.Value() < 0x20)
                               ? &state.registers.temporary[instr.mad.dest.Value().GetIndex()][0]
                               : dummy_vec4_float24;