Merge pull request #1049 from Subv/stencil

Rasterizer: Corrected the stencil implementation.

src/video_core/pica.h

@@ -442,7 +442,13 @@ struct Regs {
 
     enum class StencilAction : u32 {
         Keep           = 0,
-        Xor  = 5,
+        Zero           = 1,
+        Replace        = 2,
+        Increment      = 3,
+        Decrement      = 4,
+        Invert         = 5,
+        IncrementWrap  = 6,
+        DecrementWrap  = 7
     };
 
     struct {
@@ -481,23 +487,29 @@ struct Regs {
 
         struct {
             union {
+                // Raw value of this register
+                u32 raw_func;
+
                 // If true, enable stencil testing
                 BitField< 0, 1, u32> enable;
 
                 // Comparison operation for stencil testing
                 BitField< 4, 3, CompareFunc> func;
 
-                // Value to calculate the new stencil value from
-                BitField< 8, 8, u32> replacement_value;
+                // Mask used to control writing to the stencil buffer
+                BitField< 8, 8, u32> write_mask;
 
                 // Value to compare against for stencil testing
                 BitField<16, 8, u32> reference_value;
 
                 // Mask to apply on stencil test inputs
-                BitField<24, 8, u32> mask;
+                BitField<24, 8, u32> input_mask;
             };
 
             union {
+                // Raw value of this register
+                u32 raw_op;
+
                 // Action to perform when the stencil test fails
                 BitField< 0, 3, StencilAction> action_stencil_fail;
 

src/video_core/rasterizer.cpp

@@ -216,14 +216,33 @@ static void SetStencil(int x, int y, u8 value) {
     }
 }
 
-// TODO: Should the stencil mask be applied to the "dest" or "ref" operands? Most likely not!
-static u8 PerformStencilAction(Regs::StencilAction action, u8 dest, u8 ref) {
+static u8 PerformStencilAction(Regs::StencilAction action, u8 old_stencil, u8 ref) {
     switch (action) {
     case Regs::StencilAction::Keep:
-        return dest;
+        return old_stencil;
 
-    case Regs::StencilAction::Xor:
-        return dest ^ ref;
+    case Regs::StencilAction::Zero:
+        return 0;
+
+    case Regs::StencilAction::Replace:
+        return ref;
+
+    case Regs::StencilAction::Increment:
+        // Saturated increment
+        return std::min<u8>(old_stencil, 254) + 1;
+
+    case Regs::StencilAction::Decrement:
+        // Saturated decrement
+        return std::max<u8>(old_stencil, 1) - 1;
+
+    case Regs::StencilAction::Invert:
+        return ~old_stencil;
+
+    case Regs::StencilAction::IncrementWrap:
+        return old_stencil + 1;
+
+    case Regs::StencilAction::DecrementWrap:
+        return old_stencil - 1;
 
     default:
         LOG_CRITICAL(HW_GPU, "Unknown stencil action %x", (int)action);
@@ -783,10 +802,16 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
             }
 
             u8 old_stencil = 0;
+
+            auto UpdateStencil = [stencil_test, x, y, &old_stencil](Pica::Regs::StencilAction action) {
+                u8 new_stencil = PerformStencilAction(action, old_stencil, stencil_test.reference_value);
+                SetStencil(x >> 4, y >> 4, (new_stencil & stencil_test.write_mask) | (old_stencil & ~stencil_test.write_mask));
+            };
+
             if (stencil_action_enable) {
                 old_stencil = GetStencil(x >> 4, y >> 4);
-                u8 dest = old_stencil & stencil_test.mask;
-                u8 ref = stencil_test.reference_value & stencil_test.mask;
+                u8 dest = old_stencil & stencil_test.input_mask;
+                u8 ref = stencil_test.reference_value & stencil_test.input_mask;
 
                 bool pass = false;
                 switch (stencil_test.func) {
@@ -824,8 +849,7 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
                 }
 
                 if (!pass) {
-                    u8 new_stencil = PerformStencilAction(stencil_test.action_stencil_fail, old_stencil, stencil_test.replacement_value);
-                    SetStencil(x >> 4, y >> 4, new_stencil);
+                    UpdateStencil(stencil_test.action_stencil_fail);
                     continue;
                 }
             }
@@ -875,22 +899,18 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
                 }
 
                 if (!pass) {
-                    if (stencil_action_enable) {
-                        u8 new_stencil = PerformStencilAction(stencil_test.action_depth_fail, old_stencil, stencil_test.replacement_value);
-                        SetStencil(x >> 4, y >> 4, new_stencil);
-                    }
+                    if (stencil_action_enable)
+                        UpdateStencil(stencil_test.action_depth_fail);
                     continue;
                 }
 
                 if (output_merger.depth_write_enable)
                     SetDepth(x >> 4, y >> 4, z);
+            }
 
-                if (stencil_action_enable) {
-                    // TODO: What happens if stencil testing is enabled, but depth testing is not? Will stencil get updated anyway?
-                    u8 new_stencil = PerformStencilAction(stencil_test.action_depth_pass, old_stencil, stencil_test.replacement_value);
-                    SetStencil(x >> 4, y >> 4, new_stencil);
-                }
-            }
+            // The stencil depth_pass action is executed even if depth testing is disabled
+            if (stencil_action_enable)
+                UpdateStencil(stencil_test.action_depth_pass);
 
             auto dest = GetPixel(x >> 4, y >> 4);
             Math::Vec4<u8> blend_output = combiner_output;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -269,7 +269,8 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
         break;
 
     // Stencil test
-    case PICA_REG_INDEX(output_merger.stencil_test):
+    case PICA_REG_INDEX(output_merger.stencil_test.raw_func):
+    case PICA_REG_INDEX(output_merger.stencil_test.raw_op):
         SyncStencilTest();
         break;
 
@@ -676,7 +677,15 @@ void RasterizerOpenGL::SyncLogicOp() {
 }
 
 void RasterizerOpenGL::SyncStencilTest() {
-    // TODO: Implement stencil test, mask, and op
+    const auto& regs = Pica::g_state.regs;
+    state.stencil.test_enabled = regs.output_merger.stencil_test.enable && regs.framebuffer.depth_format == Pica::Regs::DepthFormat::D24S8;
+    state.stencil.test_func = PicaToGL::CompareFunc(regs.output_merger.stencil_test.func);
+    state.stencil.test_ref = regs.output_merger.stencil_test.reference_value;
+    state.stencil.test_mask = regs.output_merger.stencil_test.input_mask;
+    state.stencil.write_mask = regs.output_merger.stencil_test.write_mask;
+    state.stencil.action_stencil_fail = PicaToGL::StencilOp(regs.output_merger.stencil_test.action_stencil_fail);
+    state.stencil.action_depth_fail = PicaToGL::StencilOp(regs.output_merger.stencil_test.action_depth_fail);
+    state.stencil.action_depth_pass = PicaToGL::StencilOp(regs.output_merger.stencil_test.action_depth_pass);
 }
 
 void RasterizerOpenGL::SyncDepthTest() {
@@ -867,8 +876,15 @@ void RasterizerOpenGL::ReloadDepthBuffer() {
     state.Apply();
 
     glActiveTexture(GL_TEXTURE0);
+    if (fb_depth_texture.format == Pica::Regs::DepthFormat::D24S8) {
+        // TODO(Subv): There is a bug with Intel Windows drivers that makes glTexSubImage2D not change the stencil buffer.
+        // The bug has been reported to Intel (https://communities.intel.com/message/324464)
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, fb_depth_texture.width, fb_depth_texture.height, 0,
+            GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, temp_fb_depth_buffer.get());
+    } else {
         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, fb_depth_texture.width, fb_depth_texture.height,
             fb_depth_texture.gl_format, fb_depth_texture.gl_type, temp_fb_depth_buffer.get());
+    }
 
     state.texture_units[0].texture_2d = 0;
     state.Apply();

src/video_core/renderer_opengl/gl_state.cpp

@@ -26,6 +26,9 @@ OpenGLState::OpenGLState() {
     stencil.test_ref = 0;
     stencil.test_mask = -1;
     stencil.write_mask = -1;
+    stencil.action_depth_fail = GL_KEEP;
+    stencil.action_depth_pass = GL_KEEP;
+    stencil.action_stencil_fail = GL_KEEP;
 
     blend.enabled = false;
     blend.src_rgb_func = GL_ONE;
@@ -105,6 +108,12 @@ void OpenGLState::Apply() {
         glStencilFunc(stencil.test_func, stencil.test_ref, stencil.test_mask);
     }
 
+    if (stencil.action_depth_fail != cur_state.stencil.action_depth_fail ||
+            stencil.action_depth_pass != cur_state.stencil.action_depth_pass ||
+            stencil.action_stencil_fail != cur_state.stencil.action_stencil_fail) {
+        glStencilOp(stencil.action_stencil_fail, stencil.action_depth_fail, stencil.action_depth_pass);
+    }
+
     // Stencil mask
     if (stencil.write_mask != cur_state.stencil.write_mask) {
         glStencilMask(stencil.write_mask);

src/video_core/renderer_opengl/gl_state.h

@@ -32,6 +32,9 @@ public:
         GLint test_ref; // GL_STENCIL_REF
         GLuint test_mask; // GL_STENCIL_VALUE_MASK
         GLuint write_mask; // GL_STENCIL_WRITEMASK
+        GLenum action_stencil_fail; // GL_STENCIL_FAIL
+        GLenum action_depth_fail; // GL_STENCIL_PASS_DEPTH_FAIL
+        GLenum action_depth_pass; // GL_STENCIL_PASS_DEPTH_PASS
     } stencil;
 
     struct {

src/video_core/renderer_opengl/pica_to_gl.h

@@ -152,6 +152,29 @@ inline GLenum CompareFunc(Pica::Regs::CompareFunc func) {
     return compare_func_table[(unsigned)func];
 }
 
+inline GLenum StencilOp(Pica::Regs::StencilAction action) {
+    static const GLenum stencil_op_table[] = {
+        GL_KEEP,        // StencilAction::Keep
+        GL_ZERO,        // StencilAction::Zero
+        GL_REPLACE,     // StencilAction::Replace
+        GL_INCR,        // StencilAction::Increment
+        GL_DECR,        // StencilAction::Decrement
+        GL_INVERT,      // StencilAction::Invert
+        GL_INCR_WRAP,   // StencilAction::IncrementWrap
+        GL_DECR_WRAP    // StencilAction::DecrementWrap
+    };
+
+    // Range check table for input
+    if ((unsigned)action >= ARRAY_SIZE(stencil_op_table)) {
+        LOG_CRITICAL(Render_OpenGL, "Unknown stencil op %d", action);
+        UNREACHABLE();
+
+        return GL_KEEP;
+    }
+
+    return stencil_op_table[(unsigned)action];
+}
+
 inline std::array<GLfloat, 4> ColorRGBA8(const u8* bytes) {
     return { { bytes[0] / 255.0f,
                bytes[1] / 255.0f,