Vectorized map palette conversion

build.gradle

@@ -80,7 +80,8 @@ def jvm_arguments = [
 
 def compiler_jvm_arguments = [
         '--add-exports=java.base/jdk.internal.misc=ALL-UNNAMED',
-        '--add-exports=java.base/jdk.internal.reflect=ALL-UNNAMED'
+        '--add-exports=java.base/jdk.internal.reflect=ALL-UNNAMED',
+        '--add-modules=jdk.incubator.vector' // CatRoom - SIMD support
 ]
 // Projects
 

patches/minecraft/net/minecraft/server/dedicated/DedicatedServer.java.patch

@@ -147,7 +147,19 @@
          this.field_154332_n = new ServerEula(new File("eula.txt"));
 
          if (!this.field_154332_n.func_154346_a())
-@@ -163,17 +202,17 @@
+@@ -152,6 +191,11 @@
+                 this.func_71189_e(this.field_71340_o.func_73671_a("server-ip", ""));
+             }
+
++            // CatRoom start - SIMD support
++            try {
++                gg.pufferfish.pufferfish.simd.SIMDDetection.initialize();
++            } catch (Throwable ignored) {}
++            // CatRoom end - SIMD support
+             this.func_71251_e(this.field_71340_o.func_73670_a("spawn-animals", true));
+             this.func_71257_f(this.field_71340_o.func_73670_a("spawn-npcs", true));
+             this.func_71188_g(this.field_71340_o.func_73670_a("pvp", true));
+@@ -163,17 +207,17 @@
 
              if (this.field_71340_o.func_73669_a("difficulty", 1) < 0)
              {
@@ -168,7 +180,7 @@
              InetAddress inetaddress = null;
 
              if (!this.func_71211_k().isEmpty())
-@@ -186,29 +225,42 @@
+@@ -186,29 +230,42 @@
                  this.func_71208_b(this.field_71340_o.func_73669_a("server-port", 25565));
              }
 
@@ -227,7 +239,7 @@
                  field_155771_h.warn("To change this, set \"online-mode\" to \"true\" in the server.properties file.");
              }
 
-@@ -223,7 +275,9 @@
+@@ -223,7 +280,9 @@
              }
              else
              {
@@ -238,7 +250,7 @@
                  long j = System.nanoTime();
 
                  if (this.func_71270_I() == null)
-@@ -234,7 +288,7 @@
+@@ -234,7 +293,7 @@
                  String s = this.field_71340_o.func_73671_a("level-seed", "");
                  String s1 = this.field_71340_o.func_73671_a("level-type", "DEFAULT");
                  String s2 = this.field_71340_o.func_73671_a("generator-settings", "");
@@ -247,7 +259,7 @@
 
                  if (!s.isEmpty())
                  {
-@@ -247,7 +301,7 @@
+@@ -247,7 +306,7 @@
                              k = l;
                          }
                      }
@@ -256,7 +268,7 @@
                      {
                          k = (long)s.hashCode();
                      }
-@@ -267,21 +321,21 @@
+@@ -267,21 +326,21 @@
                  this.func_71191_d(this.field_71340_o.func_73669_a("max-build-height", 256));
                  this.func_71191_d((this.func_71207_Z() + 8) / 16 * 16);
                  this.func_71191_d(MathHelper.func_76125_a(this.func_71207_Z(), 64, 256));
@@ -285,7 +297,7 @@
                      this.field_71340_o.func_187238_b("announce-player-achievements");
                      this.field_71340_o.func_73668_b();
                  }
-@@ -293,14 +347,33 @@
+@@ -293,14 +352,33 @@
                      this.field_71342_m.func_72602_a();
                  }
 
@@ -324,7 +336,7 @@
                  {
                      Thread thread1 = new Thread(new ServerHangWatchdog(this));
                      thread1.setName("Server Watchdog");
-@@ -309,7 +382,8 @@
+@@ -309,7 +387,8 @@
                  }
 
                  Items.field_190931_a.func_150895_a(CreativeTabs.field_78027_g, NonNullList.func_191196_a());
@@ -334,7 +346,7 @@
              }
          }
      }
-@@ -339,46 +413,38 @@
+@@ -339,46 +418,38 @@
 
          if (!this.field_71340_o.func_73671_a("resource-pack", "").isEmpty() && s.isEmpty())
          {
@@ -382,7 +394,7 @@
      public CrashReport func_71230_b(CrashReport p_71230_1_)
      {
          p_71230_1_ = super.func_71230_b(p_71230_1_);
-@@ -400,40 +466,34 @@
+@@ -400,40 +471,34 @@
          return p_71230_1_;
      }
 
@@ -425,7 +437,7 @@
      public boolean func_70002_Q()
      {
          return this.field_71340_o.func_73670_a("snooper-enabled", true);
-@@ -446,20 +506,39 @@
+@@ -446,20 +511,39 @@
 
      public void func_71333_ah()
      {
@@ -469,7 +481,7 @@
      public boolean func_181035_ah()
      {
          return this.field_71340_o.func_73670_a("use-native-transport", true);
-@@ -470,13 +549,11 @@
+@@ -470,13 +554,11 @@
          return (DedicatedPlayerList)super.func_184103_al();
      }
 
@@ -483,7 +495,7 @@
      public String func_71330_a(String p_71330_1_, String p_71330_2_)
      {
          return this.field_71340_o.func_73671_a(p_71330_1_, p_71330_2_);
-@@ -487,38 +564,32 @@
+@@ -487,38 +569,32 @@
          return this.field_71340_o.func_73670_a(p_71332_1_, p_71332_2_);
      }
 
@@ -522,7 +534,7 @@
      public String func_71274_v()
      {
          return this.func_71273_Y();
-@@ -530,34 +601,29 @@
+@@ -530,34 +606,29 @@
          this.field_71335_s = true;
      }
 
@@ -558,7 +570,7 @@
          {
              return false;
          }
-@@ -583,33 +649,28 @@
+@@ -583,33 +654,28 @@
          }
      }
 
@@ -593,7 +605,7 @@
      public int func_175580_aG()
      {
          int i = this.field_71340_o.func_73669_a("max-world-size", super.func_175580_aG());
-@@ -626,12 +687,14 @@
+@@ -626,12 +692,14 @@
          return i;
      }
 
@@ -609,7 +621,7 @@
      protected boolean func_152368_aE() throws IOException
      {
          boolean flag = false;
-@@ -708,8 +771,9 @@
+@@ -708,8 +776,9 @@
          {
              Thread.sleep(5000L);
          }
@@ -620,7 +632,7 @@
          }
      }
 
-@@ -718,17 +782,68 @@
+@@ -718,17 +787,68 @@
          return this.field_71340_o.func_179885_a("max-tick-time", TimeUnit.MINUTES.toMillis(1L));
      }
 

src/main/java/gg/pufferfish/pufferfish/simd/SIMDChecker.java

@@ -0,0 +1,45 @@
+package gg.pufferfish.pufferfish.simd;
+
+import jdk.incubator.vector.FloatVector;
+import jdk.incubator.vector.IntVector;
+import jdk.incubator.vector.VectorSpecies;
+
+/**
+ * Basically, java is annoying and we have to push this out to its own class.
+ */
+public class SIMDChecker {
+
+    public static void initialize() {
+        if (SIMDDetection.isInitialized()) {
+            return;
+        }
+        SIMDDetection.setInitialized();
+        try {
+            int javaVersion = SIMDDetection.getJavaVersion();
+            if (javaVersion < 17) {
+                return;
+            }
+            SIMDDetection.supportingJavaVersion = true;
+            SIMDDetection.testRunStarted = true;
+
+            VectorSpecies<Integer> ISPEC = IntVector.SPECIES_PREFERRED;
+            VectorSpecies<Float> FSPEC = FloatVector.SPECIES_PREFERRED;
+
+            SIMDDetection.intVectorBitSize = ISPEC.vectorBitSize();
+            SIMDDetection.floatVectorBitSize = FSPEC.vectorBitSize();
+
+            SIMDDetection.intElementSize = ISPEC.elementSize();
+            SIMDDetection.floatElementSize = FSPEC.elementSize();
+
+            SIMDDetection.testRunCompleted = true;
+
+            if (ISPEC.elementSize() < 2 || FSPEC.elementSize() < 2) {
+                SIMDDetection.unsupportingLaneSize = true;
+                return;
+            }
+
+            SIMDDetection.isEnabled = true;
+        } catch (Throwable ignored) {} // Basically, we don't do anything. This lets us detect if it's not functional and disable it.
+    }
+
+}

src/main/java/gg/pufferfish/pufferfish/simd/SIMDDetection.java

@@ -0,0 +1,75 @@
+package gg.pufferfish.pufferfish.simd;
+
+public class SIMDDetection {
+
+    private static boolean isInitialized = false;
+    static int intVectorBitSize;
+    static int floatVectorBitSize;
+    static int intElementSize;
+    static int floatElementSize;
+    static boolean supportingJavaVersion;
+    static boolean testRunStarted;
+    static boolean testRunCompleted;
+    static boolean unsupportingLaneSize;
+    static boolean isEnabled;
+
+    public static void initialize() {
+        try {
+            SIMDChecker.initialize();
+        } catch (Throwable ignored) {}
+    }
+
+    static void setInitialized() {
+        isInitialized = true;
+    }
+
+    public static boolean isInitialized() {
+        return isInitialized;
+    }
+
+    public static int intVectorBitSize() {
+        return intVectorBitSize;
+    }
+
+    public static int floatVectorBitSize() {
+        return floatVectorBitSize;
+    }
+
+    public static int intElementSize() {
+        return intElementSize;
+    }
+
+    public static int floatElementSize() {
+        return floatElementSize;
+    }
+
+    public static boolean supportingJavaVersion() {
+        return supportingJavaVersion;
+    }
+
+    public static boolean testRunCompleted() {
+        return testRunCompleted;
+    }
+
+    public static boolean unsupportingLaneSize() {
+        return unsupportingLaneSize;
+    }
+
+    public static boolean isEnabled() {
+        return isEnabled;
+    }
+
+    public static int getJavaVersion() {
+        // https://stackoverflow.com/a/2591122
+        String version = System.getProperty("java.version");
+        if(version.startsWith("1.")) {
+            version = version.substring(2, 3);
+        } else {
+            int dot = version.indexOf(".");
+            if(dot != -1) { version = version.substring(0, dot); }
+        }
+        version = version.split("-")[0]; // Azul is stupid
+        return Integer.parseInt(version);
+    }
+
+}

src/main/java/gg/pufferfish/pufferfish/simd/VectorMapPalette.java

@@ -0,0 +1,82 @@
+package gg.pufferfish.pufferfish.simd;
+
+import jdk.incubator.vector.FloatVector;
+import jdk.incubator.vector.IntVector;
+import jdk.incubator.vector.VectorMask;
+import jdk.incubator.vector.VectorSpecies;
+import org.bukkit.map.MapPalette;
+
+import java.awt.*;
+
+public class VectorMapPalette {
+
+    private static final VectorSpecies<Integer> I_SPEC = IntVector.SPECIES_PREFERRED;
+    private static final VectorSpecies<Float> F_SPEC = FloatVector.SPECIES_PREFERRED;
+
+    public static void matchColorVectorized(int[] in, byte[] out) {
+        int speciesLength = I_SPEC.length();
+        int i;
+        for (i = 0; i < in.length - speciesLength; i += speciesLength) {
+            float[] redsArr = new float[speciesLength];
+            float[] bluesArr = new float[speciesLength];
+            float[] greensArr = new float[speciesLength];
+            int[] alphasArr = new int[speciesLength];
+
+            for (int j = 0; j < speciesLength; j++) {
+                alphasArr[j] = (in[i + j] >> 24) & 0xFF;
+                redsArr[j] = (in[i + j] >> 16) & 0xFF;
+                greensArr[j] = (in[i + j] >> 8) & 0xFF;
+                bluesArr[j] = (in[i + j] >> 0) & 0xFF;
+            }
+
+            IntVector alphas = IntVector.fromArray(I_SPEC, alphasArr, 0);
+            FloatVector reds = FloatVector.fromArray(F_SPEC, redsArr, 0);
+            FloatVector greens = FloatVector.fromArray(F_SPEC, greensArr, 0);
+            FloatVector blues = FloatVector.fromArray(F_SPEC, bluesArr, 0);
+            IntVector resultIndex = IntVector.zero(I_SPEC);
+            VectorMask<Integer> modificationMask = VectorMask.fromLong(I_SPEC, 0xffffffff);
+
+            modificationMask = modificationMask.and(alphas.lt(128).not());
+            FloatVector bestDistances = FloatVector.broadcast(F_SPEC, Float.MAX_VALUE);
+
+            for (int c = 4; c < MapPalette.colors.length; c++) {
+                // We're using 32-bit floats here because it's 2x faster and nobody will know the difference.
+                // For correctness, the original algorithm uses 64-bit floats instead. Completely unnecessary.
+                FloatVector compReds = FloatVector.broadcast(F_SPEC, MapPalette.colors[c].getRed());
+                FloatVector compGreens = FloatVector.broadcast(F_SPEC, MapPalette.colors[c].getGreen());
+                FloatVector compBlues = FloatVector.broadcast(F_SPEC, MapPalette.colors[c].getBlue());
+
+                FloatVector rMean = reds.add(compReds).div(2.0f);
+                FloatVector rDiff = reds.sub(compReds);
+                FloatVector gDiff = greens.sub(compGreens);
+                FloatVector bDiff = blues.sub(compBlues);
+
+                FloatVector weightR = rMean.div(256.0f).add(2);
+                FloatVector weightG = FloatVector.broadcast(F_SPEC, 4.0f);
+                FloatVector weightB = FloatVector.broadcast(F_SPEC, 255.0f).sub(rMean).div(256.0f).add(2.0f);
+
+                FloatVector distance = weightR.mul(rDiff).mul(rDiff).add(weightG.mul(gDiff).mul(gDiff)).add(weightB.mul(bDiff).mul(bDiff));
+
+                // Now we compare to the best distance we've found.
+                // This mask contains a "1" if better, and a "0" otherwise.
+                VectorMask<Float> bestDistanceMask = distance.lt(bestDistances);
+                bestDistances = bestDistances.blend(distance, bestDistanceMask); // Update the best distances
+
+                // Update the result array
+                // We also AND with the modification mask because we don't want to interfere if the alpha value isn't large enough.
+                resultIndex = resultIndex.blend(c, bestDistanceMask.cast(I_SPEC).and(modificationMask)); // Update the results
+            }
+
+            for (int j = 0; j < speciesLength; j++) {
+                int index = resultIndex.lane(j);
+                out[i + j] = (byte) (index < 128 ? index : -129 + (index - 127));
+            }
+        }
+
+        // For the final ones, fall back to the regular method
+        for (; i < in.length; i++) {
+            out[i] = MapPalette.matchColor(new Color(in[i], true));
+        }
+    }
+
+}