Implement LFU sketch using arm64 intrinsics (redux)

BitFaster.Caching.Benchmarks/BitFaster.Caching.Benchmarks.csproj

@@ -3,7 +3,7 @@
   <PropertyGroup>
     <OutputType>Exe</OutputType>
     <LangVersion>latest</LangVersion>
-    <TargetFrameworks>net48;net6.0;net8.0</TargetFrameworks>
+    <TargetFrameworks>net6.0;net8.0</TargetFrameworks>
     <AllowUnsafeBlocks>True</AllowUnsafeBlocks>
     <!-- https://stackoverflow.com/a/59916801/131345 -->
     <IsWindows Condition="'$([System.Runtime.InteropServices.RuntimeInformation]::IsOSPlatform($([System.Runtime.InteropServices.OSPlatform]::Windows)))' == 'true'">true</IsWindows> 
@@ -42,4 +42,4 @@
         <DefineConstants>MacOS</DefineConstants>
     </PropertyGroup>
 
-</Project>
+</Project>

BitFaster.Caching.Benchmarks/Lfu/SketchFrequency.cs

@@ -23,7 +23,8 @@ public class SketchFrequency
 
         private CmSketchCore<int, DisableHardwareIntrinsics> blockStd;
         private CmSketchNoPin<int, DetectIsa> blockAvxNoPin;
-        private CmSketchCore<int, DetectIsa> blockAvx;
+        private CmSketchCore<int, DetectIsa> blockVector;
+        private CmSketchCore512<int, DetectIsa> blockVector512;
 
         [Params(32_768, 524_288, 8_388_608, 134_217_728)]
         public int Size { get; set; }
@@ -36,7 +37,8 @@ public void Setup()
 
             blockStd = new CmSketchCore<int, DisableHardwareIntrinsics>(Size, EqualityComparer<int>.Default);
             blockAvxNoPin = new CmSketchNoPin<int, DetectIsa>(Size, EqualityComparer<int>.Default);
-            blockAvx = new CmSketchCore<int, DetectIsa>(Size, EqualityComparer<int>.Default);
+            blockVector = new CmSketchCore<int, DetectIsa>(Size, EqualityComparer<int>.Default);
+            blockVector512 = new CmSketchCore512<int, DetectIsa>(Size, EqualityComparer<int>.Default);
         }
 
         [Benchmark(Baseline = true, OperationsPerInvoke = iterations)]
@@ -74,7 +76,7 @@ public int FrequencyBlockAvxNotPinned()
         {
             int count = 0;
             for (int i = 0; i < iterations; i++)
-                count += blockAvxNoPin.EstimateFrequency(i) > blockAvx.EstimateFrequency(i + 1) ? 1 : 0;
+                count += blockAvxNoPin.EstimateFrequency(i) > blockVector.EstimateFrequency(i + 1) ? 1 : 0;
 
             return count;
         }
@@ -84,7 +86,17 @@ public int FrequencyBlockAvxPinned()
         {
             int count = 0;
             for (int i = 0; i < iterations; i++)
-                count += blockAvx.EstimateFrequency(i) > blockAvx.EstimateFrequency(i + 1) ? 1 : 0;
+                count += blockVector.EstimateFrequency(i) > blockVector.EstimateFrequency(i + 1) ? 1 : 0;
+
+            return count;
+        }
+
+        [Benchmark(OperationsPerInvoke = iterations)]
+        public int FrequencyBlockAvxPinned512()
+        {
+            int count = 0;
+            for (int i = 0; i < iterations; i++)
+                count += blockVector512.EstimateFrequency(i) > blockVector.EstimateFrequency(i + 1) ? 1 : 0;
 
             return count;
         }

BitFaster.Caching.Benchmarks/Lfu/SketchIncrement.cs

@@ -23,6 +23,8 @@ public class SketchIncrement
         private CmSketchCore<int, DisableHardwareIntrinsics> blockStd;
         private CmSketchNoPin<int, DetectIsa> blockAvxNoPin;
         private CmSketchCore<int, DetectIsa> blockAvx;
+        private CmSketchCore512<int, DetectIsa> blockVector512;
+
 
         [Params(32_768, 524_288, 8_388_608, 134_217_728)]
         public int Size { get; set; }
@@ -36,6 +38,7 @@ public void Setup()
             blockStd = new CmSketchCore<int, DisableHardwareIntrinsics>(Size, EqualityComparer<int>.Default);
             blockAvxNoPin = new CmSketchNoPin<int, DetectIsa>(Size, EqualityComparer<int>.Default);
             blockAvx = new CmSketchCore<int, DetectIsa>(Size, EqualityComparer<int>.Default);
+            blockVector512 = new CmSketchCore512<int, DetectIsa>(Size, EqualityComparer<int>.Default);
         }
 
         [Benchmark(Baseline = true, OperationsPerInvoke = iterations)]
@@ -82,5 +85,14 @@ public void IncBlockAvxPinned()
                 blockAvx.Increment(i);
             }
         }
+
+        [Benchmark(OperationsPerInvoke = iterations)]
+        public void IncBlockAvxPinned512()
+        {
+            for (int i = 0; i < iterations; i++)
+            {
+                blockVector512.Increment(i);
+            }
+        }
     }
 }

BitFaster.Caching.UnitTests/Intrinsics.cs

@@ -1,6 +1,10 @@
 #if NETCOREAPP3_1_OR_GREATER
 using System.Runtime.Intrinsics.X86;
 #endif
+#if NET6_0_OR_GREATER
+using System.Runtime.Intrinsics.Arm;
+#endif
+
 using Xunit;
 
 namespace BitFaster.Caching.UnitTests
@@ -10,8 +14,14 @@ public static class Intrinsics
         public static void SkipAvxIfNotSupported<I>()
         {
 #if NETCOREAPP3_1_OR_GREATER
+    #if NET6_0_OR_GREATER
+            // when we are trying to test Avx2/Arm64, skip the test if it's not supported
+            Skip.If(typeof(I) == typeof(DetectIsa) && !(Avx2.IsSupported || AdvSimd.Arm64.IsSupported));
+    #else
             // when we are trying to test Avx2, skip the test if it's not supported
             Skip.If(typeof(I) == typeof(DetectIsa) && !Avx2.IsSupported);
+    #endif
+
 #else
             Skip.If(true);
 #endif

BitFaster.Caching.UnitTests/Lfu/CmSketchTests.cs

@@ -6,12 +6,12 @@
 
 namespace BitFaster.Caching.UnitTests.Lfu
 {
-    // Test with AVX2 if it is supported
-    public class CMSketchAvx2Tests : CmSketchTestBase<DetectIsa>
+    // Test with AVX2/ARM64 if it is supported
+    public class CMSketchIntrinsicsTests : CmSketchTestBase<DetectIsa>
     {
     }
 
-    // Test with AVX2 disabled
+    // Test with AVX2/ARM64 disabled
     public class CmSketchTests : CmSketchTestBase<DisableHardwareIntrinsics>
     {
     }
@@ -29,14 +29,23 @@ public CmSketchTestBase()
         public void Repro()
         {
             sketch = new CmSketchCore<int, I>(1_048_576, EqualityComparer<int>.Default);
+            var baseline = new CmSketchCore<int, DisableHardwareIntrinsics>(1_048_576, EqualityComparer<int>.Default);
 
             for (int i = 0; i < 1_048_576; i++)
             {
                 if (i % 3 == 0)
                 {
                     sketch.Increment(i);
+                    baseline.Increment(i);
                 }
             }
+
+            baseline.Size.Should().Be(sketch.Size);
+
+            for (int i = 0; i < 1_048_576; i++)
+            {
+                sketch.EstimateFrequency(i).Should().Be(baseline.EstimateFrequency(i));
+            }
         }
 
 

BitFaster.Caching/Intrinsics.cs

@@ -2,6 +2,10 @@
 using System.Runtime.Intrinsics.X86;
 #endif
 
+#if NET6_0
+using System.Runtime.Intrinsics.Arm;
+#endif
+
 namespace BitFaster.Caching
 {
     /// <summary>
@@ -12,7 +16,14 @@ public interface IsaProbe
         /// <summary>
         /// Gets a value indicating whether AVX2 is supported.
         /// </summary>
-        bool IsAvx2Supported { get; }
+        bool IsAvx2Supported { get; }
+
+#if NET6_0_OR_GREATER
+        /// <summary>
+        /// Gets a value indicating whether Arm64 is supported.
+        /// </summary>
+        bool IsArm64Supported { get => false; }
+#endif
     }
 
     /// <summary>
@@ -25,7 +36,15 @@ public interface IsaProbe
         public bool IsAvx2Supported => false;
 #else
         /// <inheritdoc/>
-        public bool IsAvx2Supported => Avx2.IsSupported;
+        public bool IsAvx2Supported => Avx2.IsSupported;
+#endif
+
+#if NET6_0_OR_GREATER
+        /// <inheritdoc/>
+        public bool IsArm64Supported => AdvSimd.Arm64.IsSupported;
+#else
+        /// <inheritdoc/>
+        public bool IsArm64Supported => false;
 #endif
     }
 
@@ -35,6 +54,9 @@ public interface IsaProbe
     public readonly struct DisableHardwareIntrinsics : IsaProbe
     {
         /// <inheritdoc/>
-        public bool IsAvx2Supported => false;
+        public bool IsAvx2Supported => false;
+
+        /// <inheritdoc/>
+        public bool IsArm64Supported => false;
     }
 }

BitFaster.Caching/Lfu/CmSketchCore.cs

@@ -10,6 +10,10 @@
 using System.Runtime.Intrinsics.X86;
 #endif
 
+#if NET6_0_OR_GREATER
+using System.Runtime.Intrinsics.Arm;
+#endif
+
 namespace BitFaster.Caching.Lfu
 {
     /// <summary>
@@ -34,13 +38,14 @@ public unsafe class CmSketchCore<T, I>
         private const long OneMask = 0x1111111111111111L;
 
         private long[] table;
-#if NET6_0_OR_GREATER
-        private long* tableAddr;
-#endif
         private int sampleSize;
         private int blockMask;
         private int size;
 
+#if NET6_0_OR_GREATER
+        private long* tableAddr;
+#endif
+
         private readonly IEqualityComparer<T> comparer;
 
         /// <summary>
@@ -81,6 +86,12 @@ public int EstimateFrequency(T value)
             {
                 return EstimateFrequencyAvx(value);
             }
+#if NET6_0_OR_GREATER
+            else if (isa.IsArm64Supported)
+            {
+                return EstimateFrequencyArm(value);
+            }
+#endif
             else
             {
                 return EstimateFrequencyStd(value);
@@ -104,6 +115,12 @@ public void Increment(T value)
             {
                 IncrementAvx(value);
             }
+#if NET6_0_OR_GREATER
+            else if (isa.IsArm64Supported)
+            {
+                IncrementArm(value);
+            }
+#endif
             else
             {
                 IncrementStd(value);
@@ -127,7 +144,7 @@ private void EnsureCapacity(long maximumSize)
 
 #if NET6_0_OR_GREATER
             I isa = default;
-            if (isa.IsAvx2Supported)
+            if (isa.IsAvx2Supported || isa.IsArm64Supported)
             {
                 // over alloc by 8 to give 64 bytes padding, tableAddr is then aligned to 64 bytes
                 const int pad = 8;
@@ -361,5 +378,96 @@ private unsafe void IncrementAvx(T value)
             }
         }
 #endif
+
+#if NET6_0_OR_GREATER
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private unsafe void IncrementArm(T value)
+        {
+            int blockHash = Spread(comparer.GetHashCode(value));
+            int counterHash = Rehash(blockHash);
+            int block = (blockHash & blockMask) << 3;
+
+            Vector128<int> h = AdvSimd.ShiftArithmetic(Vector128.Create(counterHash), Vector128.Create(0, -8, -16, -24));
+            Vector128<int> index = AdvSimd.And(AdvSimd.ShiftRightLogical(h, 1), Vector128.Create(0xf));
+            Vector128<int> blockOffset = AdvSimd.Add(AdvSimd.Add(Vector128.Create(block), AdvSimd.And(h, Vector128.Create(1))), Vector128.Create(0, 2, 4, 6));
+
+            long* tablePtr = tableAddr;
+            //fixed (long* tablePtr = table)
+            {
+                int t0 = AdvSimd.Extract(blockOffset, 0);
+                int t1 = AdvSimd.Extract(blockOffset, 1);
+                int t2 = AdvSimd.Extract(blockOffset, 2);
+                int t3 = AdvSimd.Extract(blockOffset, 3);
+
+                Vector128<long> tableVectorA = Vector128.Create(AdvSimd.LoadVector64(tablePtr + t0), AdvSimd.LoadVector64(tablePtr + t1));
+                Vector128<long> tableVectorB = Vector128.Create(AdvSimd.LoadVector64(tablePtr + t2), AdvSimd.LoadVector64(tablePtr + t3));
+
+                index = AdvSimd.ShiftLeftLogicalSaturate(index, 2);
+
+                Vector128<int> longOffA = AdvSimd.Arm64.InsertSelectedScalar(AdvSimd.Arm64.InsertSelectedScalar(Vector128<int>.Zero, 0, index, 0), 2, index, 1);
+                Vector128<int> longOffB = AdvSimd.Arm64.InsertSelectedScalar(AdvSimd.Arm64.InsertSelectedScalar(Vector128<int>.Zero, 0, index, 2), 2, index, 3);
+
+                Vector128<long> fifteen = Vector128.Create(0xfL);
+                Vector128<long> maskA = AdvSimd.ShiftArithmetic(fifteen, longOffA.AsInt64());
+                Vector128<long> maskB = AdvSimd.ShiftArithmetic(fifteen, longOffB.AsInt64());
+
+                Vector128<long> maskedA = AdvSimd.Not(AdvSimd.Arm64.CompareEqual(AdvSimd.And(tableVectorA, maskA), maskA));
+                Vector128<long> maskedB = AdvSimd.Not(AdvSimd.Arm64.CompareEqual(AdvSimd.And(tableVectorB, maskB), maskB));
+
+                var one = Vector128.Create(1L);
+                Vector128<long> incA = AdvSimd.And(maskedA, AdvSimd.ShiftArithmetic(one, longOffA.AsInt64()));
+                Vector128<long> incB = AdvSimd.And(maskedB, AdvSimd.ShiftArithmetic(one, longOffB.AsInt64()));
+
+                tablePtr[t0] += AdvSimd.Extract(incA, 0);
+                tablePtr[t1] += AdvSimd.Extract(incA, 1);
+                tablePtr[t2] += AdvSimd.Extract(incB, 0);
+                tablePtr[t3] += AdvSimd.Extract(incB, 1);
+
+                var max = AdvSimd.Arm64.MaxAcross(AdvSimd.Arm64.InsertSelectedScalar(AdvSimd.Arm64.MaxAcross(incA.AsInt32()), 1, AdvSimd.Arm64.MaxAcross(incB.AsInt32()), 0).AsInt16());
+
+                if (max.ToScalar() != 0 && (++size == sampleSize))
+                {
+                    Reset();
+                }
+            }
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private unsafe int EstimateFrequencyArm(T value)
+        {
+            int blockHash = Spread(comparer.GetHashCode(value));
+            int counterHash = Rehash(blockHash);
+            int block = (blockHash & blockMask) << 3;
+
+            Vector128<int> h = AdvSimd.ShiftArithmetic(Vector128.Create(counterHash), Vector128.Create(0, -8, -16, -24));
+            Vector128<int> index = AdvSimd.And(AdvSimd.ShiftRightLogical(h, 1), Vector128.Create(0xf));
+            Vector128<int> blockOffset = AdvSimd.Add(AdvSimd.Add(Vector128.Create(block), AdvSimd.And(h, Vector128.Create(1))), Vector128.Create(0, 2, 4, 6));
+
+            long* tablePtr = tableAddr;
+            //fixed (long* tablePtr = table)
+            {
+                Vector128<long> tableVectorA = Vector128.Create(AdvSimd.LoadVector64(tablePtr + AdvSimd.Extract(blockOffset, 0)), AdvSimd.LoadVector64(tablePtr + AdvSimd.Extract(blockOffset, 1)));
+                Vector128<long> tableVectorB = Vector128.Create(AdvSimd.LoadVector64(tablePtr + AdvSimd.Extract(blockOffset, 2)), AdvSimd.LoadVector64(tablePtr + AdvSimd.Extract(blockOffset, 3)));
+
+                index = AdvSimd.ShiftLeftLogicalSaturate(index, 2);
+
+                Vector128<int> indexA = AdvSimd.Negate(AdvSimd.Arm64.InsertSelectedScalar(AdvSimd.Arm64.InsertSelectedScalar(Vector128<int>.Zero, 0, index, 0), 2, index, 1));
+                Vector128<int> indexB = AdvSimd.Negate(AdvSimd.Arm64.InsertSelectedScalar(AdvSimd.Arm64.InsertSelectedScalar(Vector128<int>.Zero, 0, index, 2), 2, index, 3));
+
+                var fifteen = Vector128.Create(0xfL);
+                Vector128<long> a = AdvSimd.And(AdvSimd.ShiftArithmetic(tableVectorA, indexA.AsInt64()), fifteen);
+                Vector128<long> b = AdvSimd.And(AdvSimd.ShiftArithmetic(tableVectorB, indexB.AsInt64()), fifteen);
+
+                // Before: < 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, F >
+                // After:  < 0, 1, 2, 3, 8, 9, A, B, 4, 5, 6, 7, C, D, E, F >
+                var min = AdvSimd.Arm64.VectorTableLookup(a.AsByte(), Vector128.Create(0x0B0A090803020100, 0xFFFFFFFFFFFFFFFF).AsByte());
+                min = AdvSimd.Arm64.VectorTableLookupExtension(min, b.AsByte(), Vector128.Create(0xFFFFFFFFFFFFFFFF, 0x0B0A090803020100).AsByte());
+
+                var min32 = AdvSimd.Arm64.MinAcross(min.AsInt32());
+
+                return min32.ToScalar();
+            }
+        }
+#endif
     }
 }