Implement w-buffer (#13)

* Pica: Implement W-Buffer in SW rasterizer

* OpenGL: Implement W-Buffers and fix depth-mapping

src/video_core/clipper.cpp

@@ -64,8 +64,6 @@ static void InitScreenCoordinates(OutputVertex& vtx)
     viewport.halfsize_y = float24::FromRaw(regs.viewport_size_y);
     viewport.offset_x   = float24::FromFloat32(static_cast<float>(regs.viewport_corner.x));
     viewport.offset_y   = float24::FromFloat32(static_cast<float>(regs.viewport_corner.y));
-    viewport.zscale     = float24::FromRaw(regs.viewport_depth_range);
-    viewport.offset_z   = float24::FromRaw(regs.viewport_depth_far_plane);
 
     float24 inv_w = float24::FromFloat32(1.f) / vtx.pos.w;
     vtx.color *= inv_w;
@@ -78,7 +76,7 @@ static void InitScreenCoordinates(OutputVertex& vtx)
 
     vtx.screenpos[0] = (vtx.pos.x * inv_w + float24::FromFloat32(1.0)) * viewport.halfsize_x + viewport.offset_x;
     vtx.screenpos[1] = (vtx.pos.y * inv_w + float24::FromFloat32(1.0)) * viewport.halfsize_y + viewport.offset_y;
-    vtx.screenpos[2] = viewport.offset_z + vtx.pos.z * inv_w * viewport.zscale;
+    vtx.screenpos[2] = vtx.pos.z * inv_w;
 }
 
 void ProcessTriangle(const OutputVertex &v0, const OutputVertex &v1, const OutputVertex &v2) {

src/video_core/pica.h

@@ -69,7 +69,7 @@ struct Regs {
     INSERT_PADDING_WORDS(0x9);
 
     BitField<0, 24, u32> viewport_depth_range; // float24
-    BitField<0, 24, u32> viewport_depth_far_plane; // float24
+    BitField<0, 24, u32> viewport_depth_near_plane; // float24
 
     BitField<0, 3, u32> vs_output_total;
 
@@ -121,7 +121,20 @@ struct Regs {
         BitField<16, 10, s32> y;
     } viewport_corner;
 
-    INSERT_PADDING_WORDS(0x17);
+    INSERT_PADDING_WORDS(0x1);
+
+    //TODO: early depth
+    INSERT_PADDING_WORDS(0x1);
+
+    INSERT_PADDING_WORDS(0x2);
+
+    enum DepthBuffering : u32 {
+        WBuffering  = 0,
+        ZBuffering  = 1,
+    };
+    BitField< 0, 1, DepthBuffering> depthmap_enable;
+
+    INSERT_PADDING_WORDS(0x12);
 
     struct TextureConfig {
         enum WrapMode : u32 {
@@ -1273,10 +1286,11 @@ ASSERT_REG_POSITION(cull_mode, 0x40);
 ASSERT_REG_POSITION(viewport_size_x, 0x41);
 ASSERT_REG_POSITION(viewport_size_y, 0x43);
 ASSERT_REG_POSITION(viewport_depth_range, 0x4d);
-ASSERT_REG_POSITION(viewport_depth_far_plane, 0x4e);
+ASSERT_REG_POSITION(viewport_depth_near_plane, 0x4e);
 ASSERT_REG_POSITION(vs_output_attributes[0], 0x50);
 ASSERT_REG_POSITION(vs_output_attributes[1], 0x51);
 ASSERT_REG_POSITION(viewport_corner, 0x68);
+ASSERT_REG_POSITION(depthmap_enable, 0x6D);
 ASSERT_REG_POSITION(texture0_enable, 0x80);
 ASSERT_REG_POSITION(texture0, 0x81);
 ASSERT_REG_POSITION(texture0_format, 0x8e);

src/video_core/rasterizer.cpp

@@ -859,10 +859,30 @@ static void ProcessTriangleInternal(const Shader::OutputVertex& v0,
                 }
             }
 
-            unsigned num_bits = Regs::DepthBitsPerPixel(regs.framebuffer.depth_format);
+            // interpolated_z = z / w
-            u32 z = (u32)((v0.screenpos[2].ToFloat32() * w0 +
+            float interpolated_z_over_w = (v0.screenpos[2].ToFloat32() * w0 +
                                            v1.screenpos[2].ToFloat32() * w1 +
-                           v2.screenpos[2].ToFloat32() * w2) * ((1 << num_bits) - 1) / wsum);
+                                           v2.screenpos[2].ToFloat32() * w2) / wsum;
+
+            // Not fully accurate. About 3 bits in precision are missing.
+            // Z-Buffer (z / w * scale + offset)
+            float depth_scale = float24::FromRaw(regs.viewport_depth_range).ToFloat32();
+            float depth_offset = float24::FromRaw(regs.viewport_depth_near_plane).ToFloat32();
+            float depth = interpolated_z_over_w * depth_scale + depth_offset;
+
+            // Potentially switch to W-Buffer
+            if (regs.depthmap_enable == Pica::Regs::DepthBuffering::WBuffering) {
+
+                // W-Buffer (z * scale + w * offset = (z / w * scale + offset) * w)
+                depth *= interpolated_w_inverse.ToFloat32() * wsum;
+            }
+
+            // Clamp the result
+            depth = MathUtil::Clamp(depth, 0.0f, 1.0f);
+
+            // Convert float to integer
+            unsigned num_bits = Regs::DepthBitsPerPixel(regs.framebuffer.depth_format);
+            u32 z = (u32)(depth * ((1 << num_bits) - 1));
 
             if (output_merger.depth_test_enable) {
                 u32 ref_z = GetDepth(x >> 4, y >> 4);

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -250,10 +250,15 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
 
     // Depth modifiers
     case PICA_REG_INDEX(viewport_depth_range):
-    case PICA_REG_INDEX(viewport_depth_far_plane):
+    case PICA_REG_INDEX(viewport_depth_near_plane):
         SyncDepthModifiers();
         break;
 
+    // Depth buffering
+    case PICA_REG_INDEX(depthmap_enable):
+        state.draw.shader_dirty = true;
+        break;
+
     // Blending
     case PICA_REG_INDEX(output_merger.alphablend_enable):
         SyncBlendEnabled();
@@ -865,9 +870,9 @@ void RasterizerOpenGL::SyncCullMode() {
 
 void RasterizerOpenGL::SyncDepthModifiers() {
     float depth_scale = -Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_range).ToFloat32();
-    float depth_offset = Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_far_plane).ToFloat32() / 2.0f;
+    float depth_offset = Pica::float24::FromRaw(Pica::g_state.regs.viewport_depth_near_plane).ToFloat32();
 
-    // TODO: Implement scale modifier
+    uniform_block_data.data.depth_scale = depth_scale;
     uniform_block_data.data.depth_offset = depth_offset;
     uniform_block_data.dirty = true;
 }

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -35,6 +35,8 @@ struct PicaShaderConfig {
         PicaShaderConfig res;
         const auto& regs = Pica::g_state.regs;
 
+        res.depthmap_enable = regs.depthmap_enable;
+
         res.alpha_test_func = regs.output_merger.alpha_test.enable ?
             regs.output_merger.alpha_test.func.Value() : Pica::Regs::CompareFunc::Always;
 
@@ -141,6 +143,8 @@ struct PicaShaderConfig {
         return std::memcmp(this, &o, sizeof(PicaShaderConfig)) == 0;
     };
 
+    Pica::Regs::DepthBuffering depthmap_enable = Pica::Regs::DepthBuffering::WBuffering;
+
     Pica::Regs::CompareFunc alpha_test_func = Pica::Regs::CompareFunc::Never;
     std::array<Pica::Regs::TevStageConfig, 6> tev_stages = {};
     u8 combiner_buffer_input = 0;
@@ -303,6 +307,7 @@ private:
         GLvec4 const_color[6];
         GLvec4 tev_combiner_buffer_color;
         GLint alphatest_ref;
+        GLfloat depth_scale;
         GLfloat depth_offset;
         alignas(16) GLvec3 lighting_global_ambient;
         LightSrc light_src[8];

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -525,6 +525,7 @@ layout (std140) uniform shader_data {
     vec4 const_color[NUM_TEV_STAGES];
     vec4 tev_combiner_buffer_color;
     int alphatest_ref;
+    float depth_scale;
     float depth_offset;
     vec3 lighting_global_ambient;
     LightSrc light_src[NUM_LIGHTS];
@@ -566,7 +567,15 @@ vec4 secondary_fragment_color = vec4(0.0);
     }
 
     out += "color = last_tex_env_out;\n";
-    out += "gl_FragDepth = gl_FragCoord.z + depth_offset;\n}";
+
+    out += "float z_over_w = gl_FragCoord.z * 2.0 - 1.0;\n";
+    out += "float depth = z_over_w * depth_scale + depth_offset;\n";
+    if (config.depthmap_enable == Pica::Regs::DepthBuffering::WBuffering) {
+        out += "depth /= gl_FragCoord.w;\n";
+    }
+    out += "gl_FragDepth = depth;\n";
+
+    out += "}";
 
     return out;
 }