kindhuge 32905bbf5d chore: remove repetitive words 8 months ago
..
include 56dcbca238 Move base64simd into a 3rdparty folder 10 months ago
lib 56dcbca238 Move base64simd into a 3rdparty folder 10 months ago
LICENSE 56dcbca238 Move base64simd into a 3rdparty folder 10 months ago
README.md 32905bbf5d chore: remove repetitive words 8 months ago
config.h 56dcbca238 Move base64simd into a 3rdparty folder 10 months ago

README.md

Fast Base64 stream encoder/decoder

Build Status

This is an implementation of a base64 stream encoding/decoding library in C99 with SIMD (AVX2, AVX512, NEON, AArch64/NEON, SSSE3, SSE4.1, SSE4.2, AVX) and OpenMP acceleration. It also contains wrapper functions to encode/decode simple length-delimited strings. This library aims to be:

  • FAST;
  • easy to use;
  • elegant.

On x86, the library does runtime feature detection. The first time it's called, the library will determine the appropriate encoding/decoding routines for the machine. It then remembers them for the lifetime of the program. If your processor supports AVX2, SSSE3, SSE4.1, SSE4.2 or AVX instructions, the library will pick an optimized codec that lets it encode/decode 12 or 24 bytes at a time, which gives a speedup of four or more times compared to the "plain" bytewise codec.

AVX512 support is only for encoding at present, utilizing the AVX512 VL and VBMI instructions. Decoding part reused AVX2 implementations. For CPUs later than Cannonlake (manufactured in 2018) supports these instructions.

NEON support is hardcoded to on or off at compile time, because portable runtime feature detection is unavailable on ARM.

Even if your processor does not support SIMD instructions, this is a very fast library. The fallback routine can process 32 or 64 bits of input in one round, depending on your processor's word width, which still makes it significantly faster than naive bytewise implementations. On some 64-bit machines, the 64-bit routines even outperform the SSSE3 ones.

To the author's knowledge, at the time of original release, this was the only Base64 library to offer SIMD acceleration. The author wrote an article explaining one possible SIMD approach to encoding/decoding Base64. The article can help figure out what the code is doing, and why.

Notable features:

  • Really fast on x86 and ARM systems by using SIMD vector processing;
  • Can use OpenMP for even more parallel speedups;
  • Really fast on other 32 or 64-bit platforms through optimized routines;
  • Reads/writes blocks of streaming data;
  • Does not dynamically allocate memory;
  • Valid C99 that compiles with pedantic options on;
  • Re-entrant and threadsafe;
  • Unit tested;
  • Uses Duff's Device.

Acknowledgements

The original AVX2, NEON and Aarch64/NEON codecs were generously contributed by Inkymail, who, in their fork, also implemented some additional features. Their work is slowly being backported into this project.

The SSSE3 and AVX2 codecs were substantially improved by using some very clever optimizations described by Wojciech Muła in a series of articles. His own code is here.

The AVX512 encoder is based on code from Wojciech Muła's base64simd library.

The OpenMP implementation was added by Ferry Toth (@htot) from Exalon Delft.

Building

The lib directory contains the code for the actual library. Typing make in the toplevel directory will build lib/libbase64.o and bin/base64. The first is a single, self-contained object file that you can link into your own project. The second is a standalone test binary that works similarly to the base64 system utility.

The matching header file needed to use this library is in include/libbase64.h.

To compile just the "plain" library without SIMD codecs, type:

make lib/libbase64.o

Optional SIMD codecs can be included by specifying the AVX2_CFLAGS, AVX512_CFLAGS, NEON32_CFLAGS, NEON64_CFLAGS, SSSE3_CFLAGS, SSE41_CFLAGS, SSE42_CFLAGS and/or AVX_CFLAGS environment variables. A typical build invocation on x86 looks like this:

AVX2_CFLAGS=-mavx2 SSSE3_CFLAGS=-mssse3 SSE41_CFLAGS=-msse4.1 SSE42_CFLAGS=-msse4.2 AVX_CFLAGS=-mavx make lib/libbase64.o

AVX2

To build and include the AVX2 codec, set the AVX2_CFLAGS environment variable to a value that will turn on AVX2 support in your compiler, typically -mavx2. Example:

AVX2_CFLAGS=-mavx2 make

AVX512

To build and include the AVX512 codec, set the AVX512_CFLAGS environment variable to a value that will turn on AVX512 support in your compiler, typically -mavx512vl -mavx512vbmi. Example:

AVX512_CFLAGS="-mavx512vl -mavx512vbmi" make

The codec will only be used if runtime feature detection shows that the target machine supports AVX2.

SSSE3

To build and include the SSSE3 codec, set the SSSE3_CFLAGS environment variable to a value that will turn on SSSE3 support in your compiler, typically -mssse3. Example:

SSSE3_CFLAGS=-mssse3 make

The codec will only be used if runtime feature detection shows that the target machine supports SSSE3.

NEON

This library includes two NEON codecs: one for regular 32-bit ARM and one for the 64-bit AArch64 with NEON, which has double the amount of SIMD registers and can do full 64-byte table lookups. These codecs encode in 48-byte chunks and decode in massive 64-byte chunks, so they had to be augmented with an uint32/64 codec to stay fast on smaller inputs!

Use LLVM/Clang for compiling the NEON codecs. The code generation of at least GCC 4.6 (the version shipped with Raspbian and used for testing) contains a bug when compiling vstq4_u8(), and the generated assembly code is of low quality. NEON intrinsics are a known weak area of GCC. Clang does a better job.

NEON support can unfortunately not be portably detected at runtime from userland (the mrc instruction is privileged), so the default value for using the NEON codec is determined at compile-time. But you can do your own runtime detection. You can include the NEON codec and make it the default, then do a runtime check if the CPU has NEON support, and if not, force a downgrade to non-NEON with BASE64_FORCE_PLAIN.

These are your options:

  1. Don't include NEON support;
  2. build NEON support and make it the default, but build all other code without NEON flags so that you can override the default at runtime with BASE64_FORCE_PLAIN;
  3. build everything with NEON support and make it the default;
  4. build everything with NEON support, but don't make it the default (which makes no sense).

For option 1, simply don't specify any NEON-specific compiler flags at all, like so:

CC=clang CFLAGS="-march=armv6" make

For option 2, keep your CFLAGS plain, but set the NEON32_CFLAGS environment variable to a value that will build NEON support. The line below, for instance, will build all the code at ARMv6 level, except for the NEON codec, which is built at ARMv7. It will also make the NEON codec the default. For ARMv6 platforms, override that default at runtime with the BASE64_FORCE_PLAIN flag. No ARMv7/NEON code will then be touched.

CC=clang CFLAGS="-march=armv6" NEON32_CFLAGS="-march=armv7 -mfpu=neon" make

For option 3, put everything in your CFLAGS and use a stub, but non-empty, NEON32_CFLAGS. This example works for the Raspberry Pi 2B V1.1, which has NEON support:

CC=clang CFLAGS="-march=armv7 -mtune=cortex-a7" NEON32_CFLAGS="-mfpu=neon" make

To build and include the NEON64 codec, use CFLAGS as usual to define the platform and set NEON64_CFLAGS to a nonempty stub. (The AArch64 target has mandatory NEON64 support.) Example:

CC=clang CFLAGS="--target=aarch64-linux-gnu -march=armv8-a" NEON64_CFLAGS=" " make

OpenMP

To enable OpenMP on GCC you need to build with -fopenmp. This can be by setting the OPENMP environment variable to 1.

Example:

OPENMP=1 make

This will let the compiler define _OPENMP, which in turn will include the OpenMP optimized lib_openmp.c into lib.c.

By default the number of parallel threads will be equal to the number of cores of the processor. On a quad core with hyperthreading eight cores will be detected, but hyperthreading will not increase the performance.

To get verbose information about OpenMP start the program with OMP_DISPLAY_ENV=VERBOSE, for instance

OMP_DISPLAY_ENV=VERBOSE test/benchmark

To put a limit on the number of threads, start the program with OMP_THREAD_LIMIT=n, for instance

OMP_THREAD_LIMIT=2 test/benchmark

An example of running a benchmark with OpenMP, SSSE3 and AVX2 enabled:

make clean && OPENMP=1 SSSE3_CFLAGS=-mssse3 AVX2_CFLAGS=-mavx2 make && OPENMP=1 make -C test

API reference

Strings are represented as a pointer and a length; they are not zero-terminated. This was a conscious design decision. In the decoding step, relying on zero-termination would make no sense since the output could contain legitimate zero bytes. In the encoding step, returning the length saves the overhead of calling strlen() on the output. If you insist on the trailing zero, you can easily add it yourself at the given offset.

Flags

Some API calls take a flags argument. That argument can be used to force the use of a specific codec, even if that codec is a no-op in the current build. Mainly there for testing purposes, this is also useful on ARM where the only way to do runtime NEON detection is to ask the OS if it's available. The following constants can be used:

  • BASE64_FORCE_AVX2
  • BASE64_FORCE_AVX512
  • BASE64_FORCE_NEON32
  • BASE64_FORCE_NEON64
  • BASE64_FORCE_PLAIN
  • BASE64_FORCE_SSSE3
  • BASE64_FORCE_SSE41
  • BASE64_FORCE_SSE42
  • BASE64_FORCE_AVX

Set flags to 0 for the default behavior, which is runtime feature detection on x86, a compile-time fixed codec on ARM, and the plain codec on other platforms.

Encoding

base64_encode

void base64_encode
    ( const char  *src
    , size_t       srclen
    , char        *out
    , size_t      *outlen
    , int          flags
    ) ;

Wrapper function to encode a plain string of given length. Output is written to out without trailing zero. Output length in bytes is written to outlen. The buffer in out has been allocated by the caller and is at least 4/3 the size of the input.

base64_stream_encode_init

void base64_stream_encode_init
    ( struct base64_state  *state
    , int                   flags
    ) ;

Call this before calling base64_stream_encode() to init the state.

base64_stream_encode

void base64_stream_encode
    ( struct base64_state  *state
    , const char           *src
    , size_t                srclen
    , char                 *out
    , size_t               *outlen
    ) ;

Encodes the block of data of given length at src, into the buffer at out. Caller is responsible for allocating a large enough out-buffer; it must be at least 4/3 the size of the in-buffer, but take some margin. Places the number of new bytes written into outlen (which is set to zero when the function starts). Does not zero-terminate or finalize the output.

base64_stream_encode_final

void base64_stream_encode_final
    ( struct base64_state  *state
    , char                 *out
    , size_t               *outlen
    ) ;

Finalizes the output begun by previous calls to base64_stream_encode(). Adds the required end-of-stream markers if appropriate. outlen is modified and will contain the number of new bytes written at out (which will quite often be zero).

Decoding

base64_decode

int base64_decode
    ( const char  *src
    , size_t       srclen
    , char        *out
    , size_t      *outlen
    , int          flags
    ) ;

Wrapper function to decode a plain string of given length. Output is written to out without trailing zero. Output length in bytes is written to outlen. The buffer in out has been allocated by the caller and is at least 3/4 the size of the input. Returns 1 for success, and 0 when a decode error has occured due to invalid input. Returns -1 if the chosen codec is not included in the current build.

base64_stream_decode_init

void base64_stream_decode_init
    ( struct base64_state  *state
    , int                   flags
    ) ;

Call this before calling base64_stream_decode() to init the state.

base64_stream_decode

int base64_stream_decode
    ( struct base64_state  *state
    , const char           *src
    , size_t                srclen
    , char                 *out
    , size_t               *outlen
    ) ;

Decodes the block of data of given length at src, into the buffer at out. Caller is responsible for allocating a large enough out-buffer; it must be at least 3/4 the size of the in-buffer, but take some margin. Places the number of new bytes written into outlen (which is set to zero when the function starts). Does not zero-terminate the output. Returns 1 if all is well, and 0 if a decoding error was found, such as an invalid character. Returns -1 if the chosen codec is not included in the current build. Used by the test harness to check whether a codec is available for testing.

Examples

A simple example of encoding a static string to base64 and printing the output to stdout:

#include <stdio.h>	/* fwrite */
#include "libbase64.h"

int main ()
{
	char src[] = "hello world";
	char out[20];
	size_t srclen = sizeof(src) - 1;
	size_t outlen;

	base64_encode(src, srclen, out, &outlen, 0);

	fwrite(out, outlen, 1, stdout);

	return 0;
}

A simple example (no error checking, etc) of stream encoding standard input to standard output:

#include <stdio.h>
#include "libbase64.h"

int main ()
{
	size_t nread, nout;
	char buf[12000], out[16000];
	struct base64_state state;

	// Initialize stream encoder:
	base64_stream_encode_init(&state, 0);

	// Read contents of stdin into buffer:
	while ((nread = fread(buf, 1, sizeof(buf), stdin)) > 0) {

		// Encode buffer:
		base64_stream_encode(&state, buf, nread, out, &nout);

		// If there's output, print it to stdout:
		if (nout) {
			fwrite(out, nout, 1, stdout);
		}

		// If an error occurred, exit the loop:
		if (feof(stdin)) {
			break;
		}
	}

	// Finalize encoding:
	base64_stream_encode_final(&state, out, &nout);

	// If the finalizing resulted in extra output bytes, print them:
	if (nout) {
		fwrite(out, nout, 1, stdout);
	}

	return 0;
}

Also see bin/base64.c for a simple re-implementation of the base64 utility. A file or standard input is fed through the encoder/decoder, and the output is written to standard output.

Tests

See tests/ for a small test suite. Testing is automated with GitHub Actions, which builds and tests the code across various architectures.

Benchmarks

Benchmarks can be run with the built-in benchmark program as follows:

make -C test benchmark <buildflags> && test/benchmark

It will run an encoding and decoding benchmark for all of the compiled-in codecs.

The tables below contain some results on random machines. All numbers measured with a 10MB buffer in MB/sec, rounded to the nearest integer.

*: Update needed

x86 processors

Processor Plain enc Plain dec SSSE3 enc SSSE3 dec AVX enc AVX dec AVX2 enc AVX2 dec
i7-4771 @ 3.5 GHz 833* 1111* 3333* 4444* TBD TBD 4999* 6666*
i7-4770 @ 3.4 GHz DDR1600 1790* 3038* 4899* 4043* 4796* 5709* 4681* 6386*
i7-4770 @ 3.4 GHz DDR1600 OPENMP 1 thread 1784* 3041* 4945* 4035* 4776* 5719* 4661* 6294*
i7-4770 @ 3.4 GHz DDR1600 OPENMP 2 thread 3401* 5729* 5489* 7444* 5003* 8624* 5105* 8558*
i7-4770 @ 3.4 GHz DDR1600 OPENMP 4 thread 4884* 7099* 4917* 7057* 4799* 7143* 4902* 7219*
i7-4770 @ 3.4 GHz DDR1600 OPENMP 8 thread 5212* 8849* 5284* 9099* 5289* 9220* 4849* 9200*
i7-4870HQ @ 2.5 GHz 1471* 3066* 6721* 6962* 7015* 8267* 8328* 11576*
i5-4590S @ 3.0 GHz 3356 3197 4363 6104 4243* 6233 4160* 6344
Xeon X5570 @ 2.93 GHz 2161 1508 3160 3915 - - - -
Pentium4 @ 3.4 GHz 896 740 - - - - - -
Atom N270 243 266 508 387 - - - -
AMD E-450 645 564 625 634 - - - -
Intel Edison @ 500 MHz 79* 92* 152* 172* - - - -
Intel Edison @ 500 MHz OPENMP 2 thread 158* 184* 300* 343* - - - -
Intel Edison @ 500 MHz (x86-64) 162 119 209 164 - - - -
Intel Edison @ 500 MHz (x86-64) 2 thread 319 237 412 329 - - - -

ARM processors

Processor Plain enc Plain dec NEON32 enc NEON32 dec NEON64 enc NEON64 dec
Raspberry PI B+ V1.2 46* 40* - - - -
Raspberry PI 2 B V1.1 85 141 300 225 - -
Apple iPhone SE armv7 1056* 895* 2943* 2618* - -
Apple iPhone SE arm64 1061* 1239* - - 4098* 3983*

PowerPC processors

Processor Plain enc Plain dec
PowerPC E6500 @ 1.8GHz 270* 265*

Benchmarks on i7-4770 @ 3.4 GHz DDR1600 with varrying buffer sizes: Benchmarks

Note: optimal buffer size to take advantage of the cache is in the range of 100 kB to 1 MB, leading to 12x faster AVX encoding/decoding compared to Plain, or a throughput of 24/27GB/sec. Also note the performance degradation when the buffer size is less than 10 kB due to thread creation overhead. To prevent this from happening lib_openmp.c defines OMP_THRESHOLD 20000, requiring at least a 20000 byte buffer to enable multithreading.

License

This repository is licensed under the BSD 2-clause License. See the LICENSE file.