sw_rasterizer: truncate depth operations to f24 precision

vitor-k · Jan 16, 2024 · 19c30fc · 19c30fc
1 parent 6a7841d
commit 19c30fc
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 10 deletions.
diff --git a/src/video_core/pica_types.h b/src/video_core/pica_types.h
@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <bit>
 #include <cmath>
 #include <cstring>
 #include <boost/serialization/access.hpp>
@@ -31,6 +32,17 @@ struct Float {
         return ret;
     }
 
+    static constexpr Float<M, E> Trunc(float val) {
+        Float<M, E> ret;
+        u32 hex = std::bit_cast<u32>(val) & (0xffffffff ^ ((1 << (23 - M)) - 1));
+        ret.value = std::bit_cast<float>(hex);
+        return ret;
+    }
+
+    static constexpr Float<M, E> Trunc(const Float<M, E>& val) {
+        return Trunc(val.ToFloat32());
+    }
+
     static constexpr Float<M, E> FromRaw(u32 hex) {
         Float<M, E> res;
 
@@ -50,7 +62,7 @@ struct Float {
             hex = sign;
         }
 
-        std::memcpy(&res.value, &hex, sizeof(float));
+        res.value = std::bit_cast<float>(hex);
 
         return res;
     }

diff --git a/src/video_core/renderer_software/sw_rasterizer.cpp b/src/video_core/renderer_software/sw_rasterizer.cpp
@@ -19,6 +19,7 @@
 
 namespace SwRenderer {
 
+using Pica::f20;
 using Pica::f24;
 using Pica::FramebufferRegs;
 using Pica::RasterizerRegs;
@@ -331,18 +332,20 @@ void RasterizerSoftware::ProcessTriangle(const Vertex& v0, const Vertex& v1, con
                     f24::One() / Common::Dot(w_inverse, baricentric_coordinates);
 
                 // interpolated_z = z / w
-                const float interpolated_z_over_w =
-                    (v0.screenpos[2].ToFloat32() * w0 + v1.screenpos[2].ToFloat32() * w1 +
-                     v2.screenpos[2].ToFloat32() * w2) /
-                    wsum;
+                const f24 interpolated_z_over_w =
+                    f24::Trunc(f24::Trunc(f24::Trunc(v0.screenpos[2] * baricentric_coordinates[0]) +
+                                          f24::Trunc(v1.screenpos[2] * baricentric_coordinates[1]) +
+                                          f24::Trunc(v2.screenpos[2] * baricentric_coordinates[2]))
+                                   .ToFloat32() /
+                               wsum);
 
                 // Not fully accurate. About 3 bits in precision are missing.
                 // Z-Buffer (z / w * scale + offset)
-                const float depth_scale =
-                    f24::FromRaw(regs.rasterizer.viewport_depth_range).ToFloat32();
-                const float depth_offset =
-                    f24::FromRaw(regs.rasterizer.viewport_depth_near_plane).ToFloat32();
-                float depth = interpolated_z_over_w * depth_scale + depth_offset;
+                const f24 depth_scale = f24::FromRaw(regs.rasterizer.viewport_depth_range);
+                const f24 depth_offset = f24::FromRaw(regs.rasterizer.viewport_depth_near_plane);
+                float depth =
+                    f24::Trunc(f24::Trunc(interpolated_z_over_w * depth_scale) + depth_offset)
+                        .ToFloat32();
 
                 // Potentially switch to W-Buffer
                 if (regs.rasterizer.depthmap_enable ==