Micro-optimization: reducing function calls in hot code.

hamsham · Jan 2, 2019 · 1856066 · 1856066
1 parent 3930a12
commit 1856066
Showing 1 changed file with 11 additions and 2 deletions.
diff --git a/soft_render/src/SR_VertexProcessor.cpp b/soft_render/src/SR_VertexProcessor.cpp
@@ -194,9 +194,18 @@ void SR_VertexProcessor::push_fragments(
     // Copy all per-vertex coordinates and varyings to the fragment bins which
     // will need the data for interpolation
     SR_FragmentBin bin;
-    ls::utils::fast_memcpy(bin.mScreenCoords, screenCoords, sizeof(bin.mScreenCoords));
+    //ls::utils::fast_memcpy(bin.mScreenCoords, screenCoords, sizeof(bin.mScreenCoords));
+    bin.mScreenCoords[0] = screenCoords[0];
+    bin.mScreenCoords[1] = screenCoords[1];
+    bin.mScreenCoords[2] = screenCoords[2];
+
     bin.mPerspDivide = math::rcp(math::vec4{worldCoords[0][3], worldCoords[1][3], worldCoords[2][3], 1.f});
-    ls::utils::fast_memcpy(bin.mVaryings, varyings, sizeof(bin.mVaryings));
+
+    //ls::utils::fast_memcpy(bin.mVaryings, varyings, sizeof(bin.mVaryings));
+    for (unsigned i = 0; i < LS_ARRAY_SIZE(bin.mVaryings); ++i)
+    {
+        bin.mVaryings[i] = varyings[i];
+    }
 
     const math::vec4& p0 = screenCoords[0];
     const math::vec4& p1 = screenCoords[1];