faster rect command creation

include/pd/lithium/drawlist.hpp

@@ -45,6 +45,8 @@ class PD_API Drawlist {
 
   /** Command Allocation */
   Command& NewCommand();
+  bool HasCommands() { return pCommands.size() > 0; }
+  Command& GetLastCommand() { return pCommands[pCommands.size() - 1]; }
 
   /** Path API */
   void PathAdd(const fvec2& point) { pPath.Push(point); }

include/pd/lithium/font.hpp

@@ -92,6 +92,8 @@ class PD_API Font {
    * **Now using unordered map**
    */
   std::unordered_map<u32, Codepoint> CodeMap;
+  /** need some performance tweaks */
+  std::array<Codepoint, 128> pAsciiCache;
   /** TMS */
   struct TMELEM {
     PD::u32 ID;

include/pd/lithium/vertex.hpp

@@ -9,6 +9,8 @@ class Vertex {
   Vertex() {}
   Vertex(const fvec2& pos, const fvec2& uv, u32 color)
       : pos(pos), uv(uv), color(color) {}
+  Vertex(float x, float y, float u, float v, u32 color)
+      : pos(x, y), uv(u, v), color(color) {}
   ~Vertex() {}
 
   void Reset() {

source/lithium/font.cpp

@@ -136,11 +136,26 @@ PD_API void Font::LoadTTF(const std::vector<u8>& data, int px_height) {
   if (!empty) {
     BakeAndPush(true, font_tex, texszs);
   }
+
+  for (u32 i = 0; i < 128; i++) {
+    auto r = CodeMap.find(i);
+    if (r == CodeMap.end()) {
+      static Codepoint invalid;
+      invalid.pInvalid = true;
+      pAsciiCache[i] = invalid;
+    } else {
+      pAsciiCache[i] = CodeMap[i];
+    }
+  }
 }
 
 PD_API void Font::LoadDefaultFont(int id, int pixel_height) {}
 
 PD_API Font::Codepoint& Font::GetCodepoint(u32 c) {
+  if (c < 128) {
+    // Direct Access (~11% improvement)
+    return pAsciiCache[c];
+  }
   // Check if codepoijt exist or return a static invalid one
   auto res = CodeMap.find(c);
   if (res == CodeMap.end()) {
@@ -192,6 +207,17 @@ PD_API fvec2 Font::GetTextBounds(const char* text, float scale) {
   return res;
 }
 
+PD_API void PrimTextQuad(Command& cmd, float x, float y, float w, float h,
+                         const fvec4& uv, const PD::Color& color) {
+  cmd.Reserve(4, 6);
+  cmd.Add(2, 1, 0);
+  cmd.Add(3, 2, 0);
+  cmd.Add(Vertex(x, y, uv.x, uv.y, color));
+  cmd.Add(Vertex(x + w, y, uv.z, uv.y, color));
+  cmd.Add(Vertex(x + w, y + h, uv.z, uv.w, color));
+  cmd.Add(Vertex(x, y + h, uv.x, uv.w, color));
+}
+
 PD_API void Font::CmdTextEx(Drawlist& dl, const fvec2& pos, u32 color,
                             float scale, const char* text, LiTextFlags flags,
                             const fvec2& box) {
@@ -210,7 +236,7 @@ PD_API void Font::CmdTextEx(Drawlist& dl, const fvec2& pos, u32 color,
 
   U8Iterator it(text);
   u32 c;
-  Command* cmd = nullptr;
+  Command* cmd = dl.HasCommands() ? &dl.GetLastCommand() : nullptr;
   while (it.Decode32(c)) {
     auto cp = GetCodepoint(c);
     if ((cp.pInvalid && c != L'\n' && c != L'\t' && c != L' ') && c != L'\r')
@@ -237,15 +263,20 @@ PD_API void Font::CmdTextEx(Drawlist& dl, const fvec2& pos, u32 color,
       cmd->Tex = Textures[cp.Tex];
     }
 
-    if (flags & LiTextFlags_Shaddow) {
-      Rect rec = Math::PrimRect(rpos + off + fvec2(1, cp.Offset * cfs + 1),
-                                cp.Size * cfs, 0.f);
-      dl.PrimQuad(*cmd, rec, cp.SimpleUV, 0xff111111);
-    }
+    // calculating once and using PrimTextQuad to directly push
+    // saves ~42% on raw multiline text draw time 6.3 -> 3.7 ms
+    // and ~25% on the whole frametime 20.3 -> 15.2 ms
+    // tested with Craftus-Next 0.8.0 commit:
+    // 33298ccc276bf996d341710e69ecf05f99c57961
+    float cw = cp.Size.x * cfs;
+    float ch = cp.Size.y * cfs;
+    float cx = rpos.x + off.x;
+    float cy = rpos.y + off.y + (cp.Offset * cfs);
 
-    Rect rec = Math::PrimRect(rpos + off + fvec2(0, cp.Offset * cfs),
-                              cp.Size * cfs, 0.f);
-    dl.PrimQuad(*cmd, rec, cp.SimpleUV, color);
+    if (flags & LiTextFlags_Shaddow) {
+      PrimTextQuad(*cmd, cx + 1.f, cy + 1.f, cw, ch, cp.SimpleUV, 0xff111111);
+    }
+    PrimTextQuad(*cmd, cx, cy, cw, ch, cp.SimpleUV, color);
 
     off.x += cp.Size.x * cfs + 2 * cfs;
   }