7 files changed, 32 insertions, 11 deletions
diff --git a/src/input_common/motion_emu.cpp b/src/input_common/motion_emu.cpp
index a1761f184..59a035e70 100644
--- a/src/input_common/motion_emu.cpp
+++ b/src/input_common/motion_emu.cpp
@@ -74,11 +74,14 @@ private:
     bool is_tilting = false;
 
     Common::Event shutdown_event;
-    std::thread motion_emu_thread;
 
     std::tuple<Math::Vec3<float>, Math::Vec3<float>> status;
     std::mutex status_mutex;
 
+    // Note: always keep the thread declaration at the end so that other objects are initialized
+    // before this!
+    std::thread motion_emu_thread;
+
     void MotionEmuThread() {
         auto update_time = std::chrono::steady_clock::now();
         Math::Quaternion<float> q = MakeQuaternion(Math::Vec3<float>(), 0);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 1c6c15a58..aa95ef21d 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -28,6 +28,9 @@ MICROPROFILE_DEFINE(OpenGL_Blits, "OpenGL", "Blits", MP_RGB(100, 100, 255));
 MICROPROFILE_DEFINE(OpenGL_CacheManagement, "OpenGL", "Cache Mgmt", MP_RGB(100, 255, 100));
 
 RasterizerOpenGL::RasterizerOpenGL() : shader_dirty(true) {
+    // Clipping plane 0 is always enabled for PICA fixed clip plane z <= 0
+    state.clip_distance[0] = true;
+
     // Create sampler objects
     for (size_t i = 0; i < texture_samplers.size(); ++i) {
         texture_samplers[i].Create();
diff --git a/src/video_core/renderer_opengl/gl_shader_gen.cpp b/src/video_core/renderer_opengl/gl_shader_gen.cpp
index ae67aab05..015e69da9 100644
--- a/src/video_core/renderer_opengl/gl_shader_gen.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_gen.cpp
@@ -1112,7 +1112,10 @@ vec4 secondary_fragment_color = vec4(0.0);
                "gl_FragCoord.y < scissor_y2)) discard;\n";
     }
 
-    out += "float z_over_w = 1.0 - gl_FragCoord.z * 2.0;\n";
+    // After perspective divide, OpenGL transform z_over_w from [-1, 1] to [near, far]. Here we use
+    // default near = 0 and far = 1, and undo the transformation to get the original z_over_w, then
+    // do our own transformation according to PICA specification.
+    out += "float z_over_w = 2.0 * gl_FragCoord.z - 1.0;\n";
     out += "float depth = z_over_w * depth_scale + depth_offset;\n";
     if (state.depthmap_enable == RasterizerRegs::DepthBuffering::WBuffering) {
         out += "depth /= gl_FragCoord.w;\n";
@@ -1195,7 +1198,9 @@ void main() {
     texcoord0_w = vert_texcoord0_w;
     normquat = vert_normquat;
     view = vert_view;
-    gl_Position = vec4(vert_position.x, vert_position.y, -vert_position.z, vert_position.w);
+    gl_Position = vert_position;
+    gl_ClipDistance[0] = -vert_position.z; // fixed PICA clipping plane z <= 0
+    // TODO (wwylele): calculate gl_ClipDistance[1] from user-defined clipping plane
 }
 )";
 
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index bc9d34b84..06a905766 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -68,6 +68,8 @@ OpenGLState::OpenGLState() {
     draw.vertex_buffer = 0;
     draw.uniform_buffer = 0;
     draw.shader_program = 0;
+
+    clip_distance = {};
 }
 
 void OpenGLState::Apply() const {
@@ -261,6 +263,17 @@ void OpenGLState::Apply() const {
         glUseProgram(draw.shader_program);
     }
 
+    // Clip distance
+    for (size_t i = 0; i < clip_distance.size(); ++i) {
+        if (clip_distance[i] != cur_state.clip_distance[i]) {
+            if (clip_distance[i]) {
+                glEnable(GL_CLIP_DISTANCE0 + i);
+            } else {
+                glDisable(GL_CLIP_DISTANCE0 + i);
+            }
+        }
+    }
+
     cur_state = *this;
 }
 
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index 745a74479..437fe34c4 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <array>
 #include <glad/glad.h>
 
 namespace TextureUnits {
@@ -123,6 +124,8 @@ public:
         GLuint shader_program;   // GL_CURRENT_PROGRAM
     } draw;
 
+    std::array<bool, 2> clip_distance; // GL_CLIP_DISTANCE
+
     OpenGLState();
 
     /// Get the currently active OpenGL state
diff --git a/src/video_core/swrasterizer/clipper.cpp b/src/video_core/swrasterizer/clipper.cpp
index 7537689b7..cdbc71502 100644
--- a/src/video_core/swrasterizer/clipper.cpp
+++ b/src/video_core/swrasterizer/clipper.cpp
@@ -125,10 +125,6 @@ void ProcessTriangle(const OutputVertex& v0, const OutputVertex& v1, const Outpu
         {Math::MakeVec(f0, f0, f0, -f1), Math::Vec4<float24>(f0, f0, f0, EPSILON)}, // w = EPSILON
     }};
 
-    // TODO: If one vertex lies outside one of the depth clipping planes, some platforms (e.g. Wii)
-    //       drop the whole primitive instead of clipping the primitive properly. We should test if
-    //       this happens on the 3DS, too.
-
     // Simple implementation of the Sutherland-Hodgman clipping algorithm.
     // TODO: Make this less inefficient (currently lots of useless buffering overhead happens here)
     for (auto edge : clipping_edges) {
diff --git a/src/video_core/swrasterizer/rasterizer.h b/src/video_core/swrasterizer/rasterizer.h
index 2f0877581..66cd6cfd4 100644
--- a/src/video_core/swrasterizer/rasterizer.h
+++ b/src/video_core/swrasterizer/rasterizer.h
@@ -19,10 +19,9 @@ struct Vertex : Shader::OutputVertex {
 
     // Linear interpolation
     // factor: 0=this, 1=vtx
+    // Note: This function cannot be called after perspective divide
     void Lerp(float24 factor, const Vertex& vtx) {
         pos = pos * factor + vtx.pos * (float24::FromFloat32(1) - factor);
-
-        // TODO: Should perform perspective correct interpolation here...
         quat = quat * factor + vtx.quat * (float24::FromFloat32(1) - factor);
         color = color * factor + vtx.color * (float24::FromFloat32(1) - factor);
         tc0 = tc0 * factor + vtx.tc0 * (float24::FromFloat32(1) - factor);
@@ -30,12 +29,11 @@ struct Vertex : Shader::OutputVertex {
         tc0_w = tc0_w * factor + vtx.tc0_w * (float24::FromFloat32(1) - factor);
         view = view * factor + vtx.view * (float24::FromFloat32(1) - factor);
         tc2 = tc2 * factor + vtx.tc2 * (float24::FromFloat32(1) - factor);
-
-        screenpos = screenpos * factor + vtx.screenpos * (float24::FromFloat32(1) - factor);
     }
 
     // Linear interpolation
     // factor: 0=v0, 1=v1
+    // Note: This function cannot be called after perspective divide
     static Vertex Lerp(float24 factor, const Vertex& v0, const Vertex& v1) {
         Vertex ret = v0;
         ret.Lerp(factor, v1);