Adding AVX512 path to Base64 encoding/Decoding (dotnet#92241)

DeepakRajendrakumaran · web-flow · commit 9ad24aec018f · 2023-10-25T09:33:26.000-07:00
* Adding AVX512 path to Base64 encoding/Decoding

* Addressing review Comments.

Signed-off-by: Deepak Rajendrakumaran &lt;deepak.rajendrakumaran@intel.com&gt;

* Removing fallback path.

* Updating Third Party Notice.

* Addressing review comments

---------

Signed-off-by: Deepak Rajendrakumaran &lt;deepak.rajendrakumaran@intel.com&gt;
diff --git a/THIRD-PARTY-NOTICES.TXT b/THIRD-PARTY-NOTICES.TXT
@@ -1297,3 +1297,37 @@ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE.
+
+License notice for Avx512Vbmi base64 encoding / decoding
+--------------------------------------------------------
+
+Copyright (c) 2015-2018, Wojciech Muła
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
+TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+--------------------------------------------------------
+
+Aspects of base64 encoding / decoding are based on algorithm described in "Base64 encoding and decoding at almost the speed of a memory
+copy", Wojciech Muła and Daniel Lemire. https://arxiv.org/pdf/1910.05109.pdf
diff --git a/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Decoder.cs b/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Decoder.cs
@@ -68,7 +68,18 @@ private static unsafe OperationStatus DecodeFromUtf8(ReadOnlySpan<byte> utf8, Sp
 
                 if (maxSrcLength >= 24)
                 {
-                    byte* end = srcMax - 45;
+                    byte* end = srcMax - 88;
+                    if (Vector512.IsHardwareAccelerated && Avx512Vbmi.IsSupported && (end >= src))
+                    {
+                        Avx512Decode(ref src, ref dest, end, maxSrcLength, destLength, srcBytes, destBytes);
+
+                        if (src == srcEnd)
+                        {
+                            goto DoneExit;
+                        }
+                    }
+
+                    end = srcMax - 45;
                     if (Avx2.IsSupported && (end >= src))
                     {
                         Avx2Decode(ref src, ref dest, end, maxSrcLength, destLength, srcBytes, destBytes);
@@ -616,6 +627,78 @@ private static OperationStatus DecodeWithWhiteSpaceFromUtf8InPlace(Span<byte> ut
             return status;
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        [CompExactlyDependsOn(typeof(Avx512BW))]
+        [CompExactlyDependsOn(typeof(Avx512Vbmi))]
+        private static unsafe void Avx512Decode(ref byte* srcBytes, ref byte* destBytes, byte* srcEnd, int sourceLength, int destLength, byte* srcStart, byte* destStart)
+        {
+            // Reference for VBMI implementation : https://github.com/WojciechMula/base64simd/tree/master/decode
+            // If we have AVX512 support, pick off 64 bytes at a time for as long as we can,
+            // but make sure that we quit before seeing any == markers at the end of the
+            // string. Also, because we write 16 zeroes at the end of the output, ensure
+            // that there are at least 22 valid bytes of input data remaining to close the
+            // gap. 64 + 2 + 22 = 88 bytes.
+            byte* src = srcBytes;
+            byte* dest = destBytes;
+
+            // The JIT won't hoist these "constants", so help it
+            Vector512<sbyte> vbmiLookup0 = Vector512.Create(
+                0x80808080, 0x80808080, 0x80808080, 0x80808080,
+                0x80808080, 0x80808080, 0x80808080, 0x80808080,
+                0x80808080, 0x80808080, 0x3e808080, 0x3f808080,
+                0x37363534, 0x3b3a3938, 0x80803d3c, 0x80808080).AsSByte();
+            Vector512<sbyte> vbmiLookup1 = Vector512.Create(
+                0x02010080, 0x06050403, 0x0a090807, 0x0e0d0c0b,
+                0x1211100f, 0x16151413, 0x80191817, 0x80808080,
+                0x1c1b1a80, 0x201f1e1d, 0x24232221, 0x28272625,
+                0x2c2b2a29, 0x302f2e2d, 0x80333231, 0x80808080).AsSByte();
+            Vector512<byte> vbmiPackedLanesControl = Vector512.Create(
+                0x06000102, 0x090a0405, 0x0c0d0e08, 0x16101112,
+                0x191a1415, 0x1c1d1e18, 0x26202122, 0x292a2425,
+                0x2c2d2e28, 0x36303132, 0x393a3435, 0x3c3d3e38,
+                0x00000000, 0x00000000, 0x00000000, 0x00000000).AsByte();
+
+            Vector512<sbyte> mergeConstant0 = Vector512.Create(0x01400140).AsSByte();
+            Vector512<short> mergeConstant1 = Vector512.Create(0x00011000).AsInt16();
+
+            // This algorithm requires AVX512VBMI support.
+            // Vbmi was first introduced in CannonLake and is avaialable from IceLake on.
+            do
+            {
+                AssertRead<Vector512<sbyte>>(src, srcStart, sourceLength);
+                Vector512<sbyte> str = Vector512.Load(src).AsSByte();
+
+                // Step 1: Translate encoded Base64 input to their original indices
+                // This step also checks for invalid inputs and exits.
+                // After this, we have indices which are verified to have upper 2 bits set to 0 in each byte.
+                // origIndex      = [...|00dddddd|00cccccc|00bbbbbb|00aaaaaa]
+                Vector512<sbyte> origIndex = Avx512Vbmi.PermuteVar64x8x2(vbmiLookup0, str, vbmiLookup1);
+                Vector512<sbyte> errorVec = (origIndex.AsInt32() | str.AsInt32()).AsSByte();
+                if (errorVec.ExtractMostSignificantBits() != 0)
+                {
+                    break;
+                }
+
+                // Step 2: Now we need to reshuffle bits to remove the 0 bits.
+                // multiAdd1: [...|0000cccc|ccdddddd|0000aaaa|aabbbbbb]
+                Vector512<short> multiAdd1 = Avx512BW.MultiplyAddAdjacent(origIndex.AsByte(), mergeConstant0);
+                // multiAdd1: [...|00000000|aaaaaabb|bbbbcccc|ccdddddd]
+                Vector512<int> multiAdd2 = Avx512BW.MultiplyAddAdjacent(multiAdd1, mergeConstant1);
+
+                // Step 3: Pack 48 bytes
+                str = Avx512Vbmi.PermuteVar64x8(multiAdd2.AsByte(), vbmiPackedLanesControl).AsSByte();
+
+                AssertWrite<Vector512<sbyte>>(dest, destStart, destLength);
+                str.Store((sbyte*)dest);
+                src += 64;
+                dest += 48;
+            }
+            while (src <= srcEnd);
+
+            srcBytes = src;
+            destBytes = dest;
+        }
+
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         [CompExactlyDependsOn(typeof(Avx2))]
         private static unsafe void Avx2Decode(ref byte* srcBytes, ref byte* destBytes, byte* srcEnd, int sourceLength, int destLength, byte* srcStart, byte* destStart)
diff --git a/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Encoder.cs b/src/libraries/System.Private.CoreLib/src/System/Buffers/Text/Base64Encoder.cs
@@ -67,7 +67,16 @@ public static unsafe OperationStatus EncodeToUtf8(ReadOnlySpan<byte> bytes, Span
 
                 if (maxSrcLength >= 16)
                 {
-                    byte* end = srcMax - 32;
+                    byte* end = srcMax - 64;
+                    if (Vector512.IsHardwareAccelerated && Avx512Vbmi.IsSupported && (end >= src))
+                    {
+                        Avx512Encode(ref src, ref dest, end, maxSrcLength, destLength, srcBytes, destBytes);
+
+                        if (src == srcEnd)
+                            goto DoneExit;
+                    }
+
+                    end = srcMax - 64;
                     if (Avx2.IsSupported && (end >= src))
                     {
                         Avx2Encode(ref src, ref dest, end, maxSrcLength, destLength, srcBytes, destBytes);
@@ -226,6 +235,80 @@ public static unsafe OperationStatus EncodeToUtf8InPlace(Span<byte> buffer, int
             }
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        [CompExactlyDependsOn(typeof(Avx512BW))]
+        [CompExactlyDependsOn(typeof(Avx512Vbmi))]
+        private static unsafe void Avx512Encode(ref byte* srcBytes, ref byte* destBytes, byte* srcEnd, int sourceLength, int destLength, byte* srcStart, byte* destStart)
+        {
+            // Reference for VBMI implementation : https://github.com/WojciechMula/base64simd/tree/master/encode
+            // If we have AVX512 support, pick off 48 bytes at a time for as long as we can.
+            // But because we read 64 bytes at a time, ensure we have enough room to do a
+            // full 64-byte read without segfaulting.
+
+            byte* src = srcBytes;
+            byte* dest = destBytes;
+
+            // The JIT won't hoist these "constants", so help it
+            Vector512<sbyte> shuffleVecVbmi = Vector512.Create(
+                0x01020001, 0x04050304, 0x07080607, 0x0a0b090a,
+                0x0d0e0c0d, 0x10110f10, 0x13141213, 0x16171516,
+                0x191a1819, 0x1c1d1b1c, 0x1f201e1f, 0x22232122,
+                0x25262425, 0x28292728, 0x2b2c2a2b, 0x2e2f2d2e).AsSByte();
+            Vector512<sbyte> vbmiLookup = Vector512.Create("ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/"u8).AsSByte();
+
+            Vector512<ushort> maskAC = Vector512.Create((uint)0x0fc0fc00).AsUInt16();
+            Vector512<uint> maskBB = Vector512.Create((uint)0x3f003f00);
+            Vector512<ushort> shiftAC = Vector512.Create((uint)0x0006000a).AsUInt16();
+            Vector512<ushort> shiftBB = Vector512.Create((uint)0x00080004).AsUInt16();
+
+            AssertRead<Vector256<sbyte>>(src, srcStart, sourceLength);
+
+            // This algorithm requires AVX512VBMI support.
+            // Vbmi was first introduced in CannonLake and is avaialable from IceLake on.
+
+            // str = [...|PONM|LKJI|HGFE|DCBA]
+            Vector512<sbyte> str = Vector512.Load(src).AsSByte();
+
+            while (true)
+            {
+                // Step 1 : Split 48 bytes into 64 bytes with each byte using 6-bits from input
+                // str = [...|KLJK|HIGH|EFDE|BCAB]
+                str = Avx512Vbmi.PermuteVar64x8(str, shuffleVecVbmi);
+
+                // TO-DO- This can be achieved faster with multishift
+                // Consider the first 4 bytes - BCAB
+                // temp1    = [...|0000cccc|cc000000|aaaaaa00|00000000]
+                Vector512<ushort> temp1 = (str.AsUInt16() & maskAC);
+
+                // temp2    = [...|00000000|00cccccc|00000000|00aaaaaa]
+                Vector512<ushort> temp2 = Avx512BW.ShiftRightLogicalVariable(temp1, shiftAC).AsUInt16();
+
+                // temp3    = [...|ccdddddd|00000000|aabbbbbb|cccc0000]
+                Vector512<ushort> temp3 = Avx512BW.ShiftLeftLogicalVariable(str.AsUInt16(), shiftBB).AsUInt16();
+
+                // str      = [...|00dddddd|00cccccc|00bbbbbb|00aaaaaa]
+                str = Vector512.ConditionalSelect(maskBB, temp3.AsUInt32(), temp2.AsUInt32()).AsSByte();
+
+                // Step 2: Now we have the indices calculated. Next step is to use these indices to translate.
+                str = Avx512Vbmi.PermuteVar64x8(vbmiLookup, str);
+
+                AssertWrite<Vector512<sbyte>>(dest, destStart, destLength);
+                str.Store((sbyte*)dest);
+
+                src += 48;
+                dest += 64;
+
+                if (src > srcEnd)
+                    break;
+
+                AssertRead<Vector512<sbyte>>(src, srcStart, sourceLength);
+                str = Vector512.Load(src).AsSByte();
+            }
+
+            srcBytes = src;
+            destBytes = dest;
+        }
+
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         [CompExactlyDependsOn(typeof(Avx2))]
         private static unsafe void Avx2Encode(ref byte* srcBytes, ref byte* destBytes, byte* srcEnd, int sourceLength, int destLength, byte* srcStart, byte* destStart)
