/SimdBase64

Fast WHATWG forgiving-base64 decoding in C#

Primary LanguageC#MIT LicenseMIT

SimdBase64

Fast WHATWG forgiving base64 decoding in C#

Base64 is a standard approach to represent any binary data as ASCII. It is part of the email standard (MIME) and is commonly used to embed data in XML, HTML or JSON files. For example, images can be encoded as text using base64. Base64 is also used to represent cryptographic keys.

Our processors have fast instructions (SIMD) that can process blocks of data at once. They are ideally suited to encode and decode base64. The C# .NET runtime library has fast (SIMD-based) base64 functions1 when the input is UTF-8.

Encoding is somewhat easier than decoding. Decoding is a more challenging problem than base64 encoding because of the presence of allowable white space characters and the need to validate the input. Indeed, all inputs are valid for encoding, but only some inputs are valid for decoding. Having to skip white space characters makes accelerated decoding somewhat difficult. We refer to this decoding as WHATWG forgiving-base64 decoding.

To handle spaces and validation, we recently designed faster base64 decoding algorithm. It has been deployed in the simdutf C++ library and used in production systems (e.g., the JavaScript runtime systems Node.js and Bun). With this new algorithm, we beat the C# .NET runtime functions by 1.7 x to 2.3 x on realistic inputs of a few kilobytes.

The algorithm is unpatented (free) and we make our C# code available under a liberal open-source licence (MIT).

Results (SimdBase64 vs. fast .NET functions)

We use the enron base64 data for benchmarking, see benchmark/data/email. We process the data as UTF-8 (ASCII) using the .NET accelerated functions as a reference (System.Buffers.Text.Base64.DecodeFromUtf8). Our benchmark results are fully reproducible.

processor and base freq. SimdBase64 (GB/s) .NET speed (GB/s) speed up
Apple M2 processor (ARM, 3.5 Ghz) 6.5 3.8 1.7 x
AWS Graviton 3 (ARM, 2.6 GHz) 3.6 2.0 1.8 x
Intel Ice Lake (2.0 GHz) 6.5 3.4 1.9 x
AMD EPYC 7R32 (Zen 2, 2.8 GHz) 6.8 2.9 2.3 x 

Results (SimdBase64 vs. string .NET functions)

The .NET runtime did not accelerate the Convert.FromBase64String(mystring) functions. We can multiply the decoding speed compared to the .NET standard library.

Replace the following code based on the standard library...

byte[] newBytes = Convert.FromBase64String(s);

with our version...

byte[] newBytes = SimdBase64.Base64.FromBase64String(s);
processor and base freq. SimdBase64 (GB/s) .NET speed (GB/s) speed up
Apple M2 processor (ARM, 3.5 Ghz) 4.0 1.1 3.6 x
Intel Ice Lake (2.0 GHz) 2.5 0.65 3.8 x

AVX-512

As for .NET 9, the support for AVX-512 remains incomplete in C#. In particular, important VBMI2 instructions are missing. Hence, we are not using AVX-512 under x64 systems at this time. However, as soon as .NET offers the necessary support, we will update our results.

Requirements

We require .NET 9 or better: https://dotnet.microsoft.com/en-us/download/dotnet/9.0

Usage

The library only provides Base64 decoding functions, because the .NET library already has fast Base64 encoding functions. We support both Span<byte> (ASCII or UTF-8) and Span<char> (UTF-16) as input. If you have C# string, you can get its Span<char> with the AsSpan() method.

        string base64 = "SGVsbG8sIFdvcmxkIQ=="; // could be span<byte> in UTF-8 as well
        byte[] buffer = new byte[SimdBase64.Base64.MaximalBinaryLengthFromBase64(base64.AsSpan())];
        int bytesConsumed; // gives you the number of characters consumed
        int bytesWritten;
        var result = SimdBase64.Base64.DecodeFromBase64(base64.AsSpan(), buffer, out bytesConsumed, out bytesWritten, false); // false is for regular base64, true for base64url
        // result == OperationStatus.Done
        var answer = buffer.AsSpan().Slice(0, bytesWritten); // decoded result
        // Encoding.UTF8.GetString(answer) == "Hello, World!"

Running tests

dotnet test

To get a list of available tests, enter the command:

dotnet test --list-tests

To run specific tests, it is helpful to use the filter parameter:

dotnet test -c Release --filter DecodeBase64CasesScalar

Running Benchmarks

To run the benchmarks, run the following command:

cd benchmark
dotnet run -c Release

To run just one benchmark, use a filter:

cd benchmark
dotnet run -c Release --filter "SimdUnicodeBenchmarks.RealDataBenchmark.AVX2DecodingRealDataUTF8(FileName: \"data/email/\")"

If you are under macOS or Linux, you may want to run the benchmarks in privileged mode:

cd benchmark
sudo dotnet run -c Release

For UTF-16 benchmarks, you need to pass a flag as they are not enabled by default:

cd benchmark
dotnet run -c Release --anyCategories UTF16

Building the library

cd src
dotnet build

Code format

We recommend you use dotnet format. E.g.,

cd test
dotnet format

Programming tips

You can print the content of a vector register like so:

        public static void ToString(Vector256<byte> v)
        {
            Span<byte> b = stackalloc byte[32];
            v.CopyTo(b);
            Console.WriteLine(Convert.ToHexString(b));
        }
        public static void ToString(Vector128<byte> v)
        {
            Span<byte> b = stackalloc byte[16];
            v.CopyTo(b);
            Console.WriteLine(Convert.ToHexString(b));
        }

You can convert an integer to a hex string like so: $"0x{MyVariable:X}".

Performance tips

  • Be careful: Vector128.Shuffle is not the same as Ssse3.Shuffle nor is Vector256.Shuffle the same as Avx2.Shuffle. Prefer the latter.
  • Similarly Vector128.Shuffle is not the same as AdvSimd.Arm64.VectorTableLookup, use the latter.
  • stackalloc arrays should probably not be used in class instances.
  • In C#, struct might be preferable to class instances as it makes it clear that the data is thread local.
  • You can ask for an asm dump: DOTNET_JitDisasm=NEON64HTMLScan dotnet run -c Release. See Viewing JIT disassembly and dumps.

Scientific References

References

More reading

Footnotes

  1. The .NET runtime appear to have received some of its fast SIMD base64 functions from gfoidl.Base64 who built on earlier work by Klomp, Muła and others. See Faster Base64 Encoding and Decoding using AVX2 Instructions for a review. ↩