pica_core: Propogate vertex uniforms to geometry setup when not in exclusive mode

src/video_core/pica/pica_core.cpp

@@ -256,6 +256,10 @@ void PicaCore::WriteInternalReg(u32 id, u32 value, u32 mask) {
 
     case PICA_REG_INDEX(vs.bool_uniforms):
         vs_setup.WriteUniformBoolReg(regs.internal.vs.bool_uniforms.Value());
+        if (!regs.internal.pipeline.gs_unit_exclusive_configuration &&
+            regs.internal.pipeline.use_gs == PipelineRegs::UseGS::No) {
+            gs_setup.WriteUniformBoolReg(regs.internal.vs.bool_uniforms.Value());
+        }
         break;
 
     case PICA_REG_INDEX(vs.int_uniforms[0]):
@@ -264,6 +268,10 @@ void PicaCore::WriteInternalReg(u32 id, u32 value, u32 mask) {
     case PICA_REG_INDEX(vs.int_uniforms[3]): {
         const u32 index = (id - PICA_REG_INDEX(vs.int_uniforms[0]));
         vs_setup.WriteUniformIntReg(index, regs.internal.vs.GetIntUniform(index));
+        if (!regs.internal.pipeline.gs_unit_exclusive_configuration &&
+            regs.internal.pipeline.use_gs == PipelineRegs::UseGS::No) {
+            gs_setup.WriteUniformIntReg(index, regs.internal.vs.GetIntUniform(index));
+        }
         break;
     }
 
@@ -275,7 +283,11 @@ void PicaCore::WriteInternalReg(u32 id, u32 value, u32 mask) {
     case PICA_REG_INDEX(vs.uniform_setup.set_value[5]):
     case PICA_REG_INDEX(vs.uniform_setup.set_value[6]):
     case PICA_REG_INDEX(vs.uniform_setup.set_value[7]): {
-        vs_setup.WriteUniformFloatReg(regs.internal.vs, value);
+        const auto index = vs_setup.WriteUniformFloatReg(regs.internal.vs, value);
+        if (!regs.internal.pipeline.gs_unit_exclusive_configuration &&
+            regs.internal.pipeline.use_gs == PipelineRegs::UseGS::No && index) {
+            gs_setup.uniforms.f[index.value()] = vs_setup.uniforms.f[index.value()];
+        }
         break;
     }
 

src/video_core/pica/shader_setup.cpp

@@ -27,21 +27,23 @@ void ShaderSetup::WriteUniformIntReg(u32 index, const Common::Vec4<u8> values) {
     uniforms.i[index] = values;
 }
 
-void ShaderSetup::WriteUniformFloatReg(ShaderRegs& config, u32 value) {
+std::optional<u32> ShaderSetup::WriteUniformFloatReg(ShaderRegs& config, u32 value) {
     auto& uniform_setup = config.uniform_setup;
     const bool is_float32 = uniform_setup.IsFloat32();
     if (!uniform_queue.Push(value, is_float32)) {
-        return;
+        return std::nullopt;
     }
 
     const auto uniform = uniform_queue.Get(is_float32);
     if (uniform_setup.index >= uniforms.f.size()) {
         LOG_ERROR(HW_GPU, "Invalid float uniform index {}", uniform_setup.index.Value());
-        return;
+        return std::nullopt;
     }
 
-    uniforms.f[uniform_setup.index] = uniform;
+    const u32 index = uniform_setup.index.Value();
-    uniform_setup.index.Assign(uniform_setup.index + 1);
+    uniforms.f[index] = uniform;
+    uniform_setup.index.Assign(index + 1);
+    return index;
 }
 
 u64 ShaderSetup::GetProgramCodeHash() {

src/video_core/pica/shader_setup.h

@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <optional>
 #include "common/vector_math.h"
 #include "video_core/pica/packed_attribute.h"
 #include "video_core/pica_types.h"
@@ -58,7 +59,7 @@ public:
 
     void WriteUniformIntReg(u32 index, const Common::Vec4<u8> values);
 
-    void WriteUniformFloatReg(ShaderRegs& config, u32 value);
+    std::optional<u32> WriteUniformFloatReg(ShaderRegs& config, u32 value);
 
     u64 GetProgramCodeHash();
 

src/video_core/shader/generator/glsl_shader_gen.cpp

@@ -39,19 +39,6 @@ layout (binding = 1, std140) uniform vs_data {
     bool enable_clip1;
     vec4 clip_coef;
 };
-
-const vec2 EPSILON_Z = vec2(0.00000001f, -1.00001f);
-
-vec4 SanitizeVertex(vec4 vtx_pos) {
-    float ndc_z = vtx_pos.z / vtx_pos.w;
-    if (ndc_z > 0.f && ndc_z < EPSILON_Z[0]) {
-        vtx_pos.z = 0.f;
-    }
-    if (ndc_z < -1.f && ndc_z > EPSILON_Z[1]) {
-        vtx_pos.z = -vtx_pos.w;
-    }
-    return vtx_pos;
-}
 )";
 
 static std::string GetVertexInterfaceDeclaration(bool is_output, bool use_clip_planes,
@@ -107,7 +94,13 @@ std::string GenerateTrivialVertexShader(bool use_clip_planes, bool separable_sha
     out += GetVertexInterfaceDeclaration(true, use_clip_planes, separable_shader);
     out += VSUniformBlockDef;
 
+    // Certain games render 2D elements very close to clip plane 0 resulting in very tiny
+    // negative/positive z values when computing with f32 precision,
+    // causing some vertices to get erroneously clipped. To workaround this problem,
+    // we can use a very small epsilon value for clip plane comparison.
     out += R"(
+const float EPSILON_Z = 0.00000001f;
+
 void main() {
     primary_color = vert_color;
     texcoord0 = vert_texcoord0;
@@ -116,7 +109,10 @@ void main() {
     texcoord0_w = vert_texcoord0_w;
     normquat = vert_normquat;
     view = vert_view;
-    vec4 vtx_pos = SanitizeVertex(vert_position);
+    vec4 vtx_pos = vert_position;
+    if (abs(vtx_pos.z) < EPSILON_Z) {
+        vtx_pos.z = 0.f;
+    }
     gl_Position = vec4(vtx_pos.x, vtx_pos.y, -vtx_pos.z, vtx_pos.w);
 )";
     if (use_clip_planes) {
@@ -219,6 +215,7 @@ std::string GenerateVertexShader(const ShaderSetup& setup, const PicaVSConfig& c
             return "1.0";
         };
 
+        out += "const float EPSILON_Z = 0.00000001f;\n\n";
         out += "vec4 GetVertexQuaternion() {\n";
         out += "    return vec4(" + semantic(VSOutputAttributes::QUATERNION_X) + ", " +
                semantic(VSOutputAttributes::QUATERNION_Y) + ", " +
@@ -231,7 +228,9 @@ std::string GenerateVertexShader(const ShaderSetup& setup, const PicaVSConfig& c
                semantic(VSOutputAttributes::POSITION_Y) + ", " +
                semantic(VSOutputAttributes::POSITION_Z) + ", " +
                semantic(VSOutputAttributes::POSITION_W) + ");\n";
-        out += "    vtx_pos = SanitizeVertex(vtx_pos);\n";
+        out += "    if (abs(vtx_pos.z) < EPSILON_Z) {\n";
+        out += "        vtx_pos.z = 0.f;\n";
+        out += "    }\n";
         out += "    gl_Position = vec4(vtx_pos.x, vtx_pos.y, -vtx_pos.z, vtx_pos.w);\n";
         if (config.state.use_clip_planes) {
             out += "    gl_ClipDistance[0] = -vtx_pos.z;\n"; // fixed PICA clipping plane z <= 0
@@ -313,6 +312,7 @@ struct Vertex {
         return "1.0";
     };
 
+    out += "const float EPSILON_Z = 0.00000001f;\n\n";
     out += "vec4 GetVertexQuaternion(Vertex vtx) {\n";
     out += "    return vec4(" + semantic(VSOutputAttributes::QUATERNION_X) + ", " +
            semantic(VSOutputAttributes::QUATERNION_Y) + ", " +
@@ -325,7 +325,9 @@ struct Vertex {
            semantic(VSOutputAttributes::POSITION_Y) + ", " +
            semantic(VSOutputAttributes::POSITION_Z) + ", " +
            semantic(VSOutputAttributes::POSITION_W) + ");\n";
-    out += "    vtx_pos = SanitizeVertex(vtx_pos);\n";
+    out += "    if (abs(vtx_pos.z) < EPSILON_Z) {\n";
+    out += "        vtx_pos.z = 0.f;\n";
+    out += "    }\n";
     out += "    gl_Position = vec4(vtx_pos.x, vtx_pos.y, -vtx_pos.z, vtx_pos.w);\n";
     if (state.use_clip_planes) {
         out += "    gl_ClipDistance[0] = -vtx_pos.z;\n"; // fixed PICA clipping plane z <= 0