8 files changed, 103 insertions, 7 deletions

diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 988a6433e..cc1f90de6 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -478,7 +478,9 @@ public:
 
                 u32 depth_write_enabled;
 
-                INSERT_PADDING_WORDS(0x8);
+                INSERT_PADDING_WORDS(0x7);
+
+                u32 d3d_cull_mode;
 
                 BitField<0, 3, ComparisonOp> depth_test_func;
 
@@ -498,7 +500,13 @@ public:
                     u32 enable[NumRenderTargets];
                 } blend;
 
-                INSERT_PADDING_WORDS(0x2D);
+                INSERT_PADDING_WORDS(0xB);
+
+                union {
+                    BitField<4, 1, u32> triangle_rast_flip;
+                } screen_y_control;
+
+                INSERT_PADDING_WORDS(0x21);
 
                 u32 vb_element_base;
 
@@ -528,7 +536,12 @@ public:
                     }
                 } tic;
 
-                INSERT_PADDING_WORDS(0x22);
+                INSERT_PADDING_WORDS(0x21);
+
+                union {
+                    BitField<2, 1, u32> coord_origin;
+                    BitField<3, 10, u32> enable;
+                } point_coord_replace;
 
                 struct {
                     u32 code_address_high;
@@ -818,11 +831,14 @@ ASSERT_REG_POSITION(rt_control, 0x487);
 ASSERT_REG_POSITION(depth_test_enable, 0x4B3);
 ASSERT_REG_POSITION(independent_blend_enable, 0x4B9);
 ASSERT_REG_POSITION(depth_write_enabled, 0x4BA);
+ASSERT_REG_POSITION(d3d_cull_mode, 0x4C2);
 ASSERT_REG_POSITION(depth_test_func, 0x4C3);
 ASSERT_REG_POSITION(blend, 0x4CF);
+ASSERT_REG_POSITION(screen_y_control, 0x4EB);
 ASSERT_REG_POSITION(vb_element_base, 0x50D);
 ASSERT_REG_POSITION(tsc, 0x557);
 ASSERT_REG_POSITION(tic, 0x55D);
+ASSERT_REG_POSITION(point_coord_replace, 0x581);
 ASSERT_REG_POSITION(code_address, 0x582);
 ASSERT_REG_POSITION(draw, 0x585);
 ASSERT_REG_POSITION(index_array, 0x5F2);
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index ec9050d3d..3b70efeec 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -329,6 +329,19 @@ union Instruction {
     } isetp;
 
     union {
+        BitField<0, 3, u64> pred0;
+        BitField<3, 3, u64> pred3;
+        BitField<12, 3, u64> pred12;
+        BitField<15, 1, u64> neg_pred12;
+        BitField<24, 2, PredOperation> cond;
+        BitField<29, 3, u64> pred29;
+        BitField<32, 1, u64> neg_pred29;
+        BitField<39, 3, u64> pred39;
+        BitField<42, 1, u64> neg_pred39;
+        BitField<45, 2, PredOperation> op;
+    } psetp;
+
+    union {
         BitField<39, 3, u64> pred39;
         BitField<42, 1, u64> neg_pred;
         BitField<43, 1, u64> neg_a;
@@ -646,7 +659,7 @@ private:
             INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
             INST("1110111110010---", Id::LD_C, Type::Memory, "LD_C"),
             INST("1110111111110---", Id::ST_A, Type::Memory, "ST_A"),
-            INST("1100000000111---", Id::TEX, Type::Memory, "TEX"),
+            INST("110000----111---", Id::TEX, Type::Memory, "TEX"),
             INST("1101111101001---", Id::TEXQ, Type::Memory, "TEXQ"),
             INST("1101100---------", Id::TEXS, Type::Memory, "TEXS"),
             INST("1101101---------", Id::TLDS, Type::Memory, "TLDS"),
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index e516eb1ad..bacb389e1 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -686,7 +686,10 @@ u32 RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, GLuint program,
 
         // Bind the uniform to the sampler.
         GLint uniform = glGetUniformLocation(program, entry.GetName().c_str());
-        ASSERT(uniform != -1);
+        if (uniform == -1) {
+            continue;
+        }
+
         glProgramUniform1i(program, uniform, current_bindpoint);
 
         const auto texture = maxwell3d.GetStageTexture(entry.GetStage(), entry.GetOffset());
@@ -771,6 +774,16 @@ void RasterizerOpenGL::SyncCullMode() {
     if (state.cull.enabled) {
         state.cull.front_face = MaxwellToGL::FrontFace(regs.cull.front_face);
         state.cull.mode = MaxwellToGL::CullFace(regs.cull.cull_face);
+
+        // If the GPU is configured to flip the rasterized triangles, then we need to flip the
+        // notion of front and back. Note: We flip the triangles when the value of the register is 0
+        // because OpenGL already does it for us.
+        if (regs.screen_y_control.triangle_rast_flip == 0) {
+            if (state.cull.front_face == GL_CCW)
+                state.cull.front_face = GL_CW;
+            else if (state.cull.front_face == GL_CW)
+                state.cull.front_face = GL_CCW;
+        }
     }
 }
 
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index 50469c05c..57d7763ff 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -108,7 +108,8 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
     {GL_DEPTH24_STENCIL8, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, ComponentType::UNorm,
      false}, // Z24S8
     {GL_DEPTH24_STENCIL8, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, ComponentType::UNorm,
-     false}, // S8Z24
+     false},                                                                            // S8Z24
+    {GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F
 }};
 
 static const FormatTuple& GetFormatTuple(PixelFormat pixel_format, ComponentType component_type) {
@@ -191,7 +192,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
         MortonCopy<true, PixelFormat::DXT1>,         MortonCopy<true, PixelFormat::DXT23>,
         MortonCopy<true, PixelFormat::DXT45>,        MortonCopy<true, PixelFormat::DXN1>,
         MortonCopy<true, PixelFormat::ASTC_2D_4X4>,  MortonCopy<true, PixelFormat::Z24S8>,
-        MortonCopy<true, PixelFormat::S8Z24>,
+        MortonCopy<true, PixelFormat::S8Z24>,        MortonCopy<true, PixelFormat::Z32F>,
 };
 
 static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
@@ -213,6 +214,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, Tegra::GPUVAddr),
         MortonCopy<false, PixelFormat::ABGR8>,
         MortonCopy<false, PixelFormat::Z24S8>,
         MortonCopy<false, PixelFormat::S8Z24>,
+        MortonCopy<false, PixelFormat::Z32F>,
 };
 
 // Allocate an uninitialized texture of appropriate size and format for the surface
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index 8005a81b8..b4d7f8ebe 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -42,6 +42,7 @@ struct SurfaceParams {
         // DepthStencil formats
         Z24S8 = 13,
         S8Z24 = 14,
+        Z32F = 15,
 
         MaxDepthStencilFormat,
 
@@ -94,6 +95,7 @@ struct SurfaceParams {
             4, // ASTC_2D_4X4
             1, // Z24S8
             1, // S8Z24
+            1, // Z32F
         }};
 
         ASSERT(static_cast<size_t>(format) < compression_factor_table.size());
@@ -120,6 +122,7 @@ struct SurfaceParams {
             32,  // ASTC_2D_4X4
             32,  // Z24S8
             32,  // S8Z24
+            32,  // Z32F
         }};
 
         ASSERT(static_cast<size_t>(format) < bpp_table.size());
@@ -135,6 +138,8 @@ struct SurfaceParams {
             return PixelFormat::S8Z24;
         case Tegra::DepthFormat::Z24_S8_UNORM:
             return PixelFormat::Z24S8;
+        case Tegra::DepthFormat::Z32_FLOAT:
+            return PixelFormat::Z32F;
         default:
             LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
             UNREACHABLE();
@@ -235,6 +240,8 @@ struct SurfaceParams {
             return Tegra::DepthFormat::S8_Z24_UNORM;
         case PixelFormat::Z24S8:
             return Tegra::DepthFormat::Z24_S8_UNORM;
+        case PixelFormat::Z32F:
+            return Tegra::DepthFormat::Z32_FLOAT;
         default:
             UNREACHABLE();
         }
@@ -284,6 +291,8 @@ struct SurfaceParams {
         case Tegra::DepthFormat::S8_Z24_UNORM:
         case Tegra::DepthFormat::Z24_S8_UNORM:
             return ComponentType::UNorm;
+        case Tegra::DepthFormat::Z32_FLOAT:
+            return ComponentType::Float;
         default:
             LOG_CRITICAL(HW_GPU, "Unimplemented format={}", static_cast<u32>(format));
             UNREACHABLE();
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index d5259e0b1..e817aca5a 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -1213,6 +1213,9 @@ private:
                 switch (instr.conversion.f2f.rounding) {
                 case Tegra::Shader::F2fRoundingOp::None:
                     break;
+                case Tegra::Shader::F2fRoundingOp::Round:
+                    op_a = "roundEven(" + op_a + ')';
+                    break;
                 case Tegra::Shader::F2fRoundingOp::Floor:
                     op_a = "floor(" + op_a + ')';
                     break;
@@ -1477,6 +1480,36 @@ private:
             }
             break;
         }
+        case OpCode::Type::PredicateSetPredicate: {
+            std::string op_a =
+                GetPredicateCondition(instr.psetp.pred12, instr.psetp.neg_pred12 != 0);
+            std::string op_b =
+                GetPredicateCondition(instr.psetp.pred29, instr.psetp.neg_pred29 != 0);
+
+            using Tegra::Shader::Pred;
+            // We can't use the constant predicate as destination.
+            ASSERT(instr.psetp.pred3 != static_cast<u64>(Pred::UnusedIndex));
+
+            std::string second_pred =
+                GetPredicateCondition(instr.psetp.pred39, instr.psetp.neg_pred39 != 0);
+
+            std::string combiner = GetPredicateCombiner(instr.psetp.op);
+
+            std::string predicate =
+                '(' + op_a + ") " + GetPredicateCombiner(instr.psetp.cond) + " (" + op_b + ')';
+
+            // Set the primary predicate to the result of Predicate OP SecondPredicate
+            SetPredicate(instr.psetp.pred3,
+                         '(' + predicate + ") " + combiner + " (" + second_pred + ')');
+
+            if (instr.psetp.pred0 != static_cast<u64>(Pred::UnusedIndex)) {
+                // Set the secondary predicate to the result of !Predicate OP SecondPredicate,
+                // if enabled
+                SetPredicate(instr.psetp.pred0,
+                             "!(" + predicate + ") " + combiner + " (" + second_pred + ')');
+            }
+            break;
+        }
         case OpCode::Type::FloatSet: {
             std::string op_a = instr.fset.neg_a ? "-" : "";
             op_a += regs.GetRegisterAsFloat(instr.gpr8);
diff --git a/src/video_core/renderer_opengl/maxwell_to_gl.h b/src/video_core/renderer_opengl/maxwell_to_gl.h
index 6b9bb3df1..6ce53bbd9 100644
--- a/src/video_core/renderer_opengl/maxwell_to_gl.h
+++ b/src/video_core/renderer_opengl/maxwell_to_gl.h
@@ -29,6 +29,10 @@ inline GLenum VertexType(Maxwell::VertexAttribute attrib) {
         switch (attrib.size) {
         case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
             return GL_UNSIGNED_BYTE;
+        case Maxwell::VertexAttribute::Size::Size_16_16:
+            return GL_UNSIGNED_SHORT;
+        case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
+            return GL_UNSIGNED_INT_2_10_10_10_REV;
         }
 
         LOG_CRITICAL(Render_OpenGL, "Unimplemented vertex size={}", attrib.SizeString());
@@ -41,6 +45,10 @@ inline GLenum VertexType(Maxwell::VertexAttribute attrib) {
         switch (attrib.size) {
         case Maxwell::VertexAttribute::Size::Size_8_8_8_8:
             return GL_BYTE;
+        case Maxwell::VertexAttribute::Size::Size_16_16:
+            return GL_SHORT;
+        case Maxwell::VertexAttribute::Size::Size_10_10_10_2:
+            return GL_INT_2_10_10_10_REV;
         }
 
         LOG_CRITICAL(Render_OpenGL, "Unimplemented vertex size={}", attrib.SizeString());
diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp
index 7b06fea3e..d5ab4e4f9 100644
--- a/src/video_core/textures/decoders.cpp
+++ b/src/video_core/textures/decoders.cpp
@@ -78,6 +78,7 @@ static u32 DepthBytesPerPixel(DepthFormat format) {
     switch (format) {
     case DepthFormat::S8_Z24_UNORM:
     case DepthFormat::Z24_S8_UNORM:
+    case DepthFormat::Z32_FLOAT:
         return 4;
     default:
         UNIMPLEMENTED_MSG("Format not implemented");
@@ -132,6 +133,7 @@ std::vector<u8> UnswizzleDepthTexture(VAddr address, DepthFormat format, u32 wid
     switch (format) {
     case DepthFormat::S8_Z24_UNORM:
     case DepthFormat::Z24_S8_UNORM:
+    case DepthFormat::Z32_FLOAT:
         CopySwizzledData(width, height, bytes_per_pixel, bytes_per_pixel, data,
                          unswizzled_data.data(), true, block_height);
         break;