test/jdk/jdk/incubator/vector/benchmark/src/main/java/benchmark/jdk/incubator/vector/SumOfUnsignedBytes.java

rev 55589 : Species-phase2
rev 55594 : tests and benchmark changes

*** 55,66 ****
          data = fillByte(size, i -> (byte)(int)i);
  
          int sum = scalar();
          assertEquals(vectorInt(),   sum);
          assertEquals(vectorShort(), sum);
!         assertEquals(vectorByte(),  sum);
!         assertEquals(vectorSAD(),   sum);
      }
  
      @Benchmark
      public int scalar() {
          int sum = 0;
--- 55,66 ----
          data = fillByte(size, i -> (byte)(int)i);
  
          int sum = scalar();
          assertEquals(vectorInt(),   sum);
          assertEquals(vectorShort(), sum);
!         //assertEquals(vectorByte(),  sum);
!         //assertEquals(vectorSAD(),   sum);
      }
  
      @Benchmark
      public int scalar() {
          int sum = 0;
*** 71,81 ****
      }
  
      // 1. 32-bit accumulators
      @Benchmark
      public int vectorInt() {
!         final var lobyte_mask = I256.broadcast(0x000000FF);
  
          var acc = IntVector.zero(I256);
          for (int i = 0; i < data.length; i += B256.length()) {
              var vb = ByteVector.fromArray(B256, data, i);
              var vi = (IntVector)vb.reinterpret(I256);
--- 71,81 ----
      }
  
      // 1. 32-bit accumulators
      @Benchmark
      public int vectorInt() {
!         final var lobyte_mask = IntVector.broadcast(I256, 0x000000FF);
  
          var acc = IntVector.zero(I256);
          for (int i = 0; i < data.length; i += B256.length()) {
              var vb = ByteVector.fromArray(B256, data, i);
              var vi = (IntVector)vb.reinterpret(I256);
*** 88,98 ****
      }
  
      // 2. 16-bit accumulators
      @Benchmark
      public int vectorShort() {
!         final var lobyte_mask = S256.broadcast((short) 0x00FF);
  
          // FIXME: overflow
          var acc = ShortVector.zero(S256);
          for (int i = 0; i < data.length; i += B256.length()) {
              var vb = ByteVector.fromArray(B256, data, i);
--- 88,98 ----
      }
  
      // 2. 16-bit accumulators
      @Benchmark
      public int vectorShort() {
!         final var lobyte_mask = ShortVector.broadcast(S256, (short) 0x00FF);
  
          // FIXME: overflow
          var acc = ShortVector.zero(S256);
          for (int i = 0; i < data.length; i += B256.length()) {
              var vb = ByteVector.fromArray(B256, data, i);
*** 107,116 ****
--- 107,117 ----
          var accLo = ((IntVector)(acc             .reshape(S128).cast(I256))).and(0xFFFF); // low half as ints
          var accHi = ((IntVector)(acc.shiftEL(mid).reshape(S128).cast(I256))).and(0xFFFF); // high half as ints
          return accLo.addAll() + accHi.addAll();
      }
  
+     /*
      // 3. 8-bit halves (MISSING: _mm_adds_epu8)
      @Benchmark
      public int vectorByte() {
          int window = 256;
          var acc_hi  = IntVector.zero(I256);
*** 142,155 ****
              var v = ByteVector.fromArray(B256, data, i);
              var sad = sumOfAbsoluteDifferences(v, ByteVector.zero(B256)); // MISSING
              acc = acc.add(sad);
          }
          return acc.addAll();
!     }
  
      // Helpers
! 
      static ByteVector addSaturated(ByteVector va, ByteVector vb) {
          return ByteVectorHelper.map(va, vb, (i, a, b) -> {
              if ((a & 0xFF) + (b & 0xFF) < 0xFF) {
                  return (byte) (a + b);
              } else {
--- 143,156 ----
              var v = ByteVector.fromArray(B256, data, i);
              var sad = sumOfAbsoluteDifferences(v, ByteVector.zero(B256)); // MISSING
              acc = acc.add(sad);
          }
          return acc.addAll();
!     } */
  
      // Helpers
!     /*
      static ByteVector addSaturated(ByteVector va, ByteVector vb) {
          return ByteVectorHelper.map(va, vb, (i, a, b) -> {
              if ((a & 0xFF) + (b & 0xFF) < 0xFF) {
                  return (byte) (a + b);
              } else {
*** 165,171 ****
              } else {
                  return (byte)(b - a);
              }
          });
          return sum(vc);
!     }
  }
--- 166,172 ----
              } else {
                  return (byte)(b - a);
              }
          });
          return sum(vc);
!     } */
  }