r8brain-free-src - High-Quality, Fast Resampler

Introduction

Open source (under the MIT license) high-quality professional audio sample rate converter (SRC) / resampler C++ library. Features routines for SRC, both up- and downsampling, to/from any sample rate, including non-integer sample rates: it can be also used for conversion to/from SACD/DSD sample rates, and even go beyond that. SRC routines were implemented in a multi-platform C++ code, and have a high level of optimality. Also suitable for fast general-purpose 1D time-series resampling / interpolation (with relaxed filter parameters).

The structure of this library's objects is such that they can be frequently created and destroyed in large applications with a minimal performance impact due to a high level of reusability of its most "initialization-expensive" objects: the fast Fourier transform and FIR filter objects.

The SRC algorithm at first produces 2X oversampled (relative to the source sample rate, or the destination sample rate if the downsampling is performed) signal, then performs interpolation using a bank of short (8 to 30 taps, depending on the required precision) polynomial-interpolated sinc function-based fractional delay filters. This puts the algorithm into the league of the fastest among the most precise SRC algorithms. The more precise alternative being only the whole number-factored SRC, which can be slower.

P.S. Please credit the creator of this library in your documentation in the following way: "Sample rate converter designed by Aleksey Vaneev of Voxengo".

Requirements

C++ compiler and system with the "double" floating point type (53-bit mantissa) support. No explicit code for the "float" type is present in this library, because as practice has shown the "float"-based code performs considerably slower on a modern processor, at least in this library. This library does not have dependencies beside the standard C library, the "windows.h" on Windows and the "pthread.h" on macOS and Linux.

Usage Information

The sample rate converter (resampler) is represented by the r8b::CDSPResampler class, which is a single front-end class for the whole library. You do not basically need to use nor understand any other classes beside this class. Several derived classes that have varying levels of precision are also available (for full-resolution 16-bit and 24-bit resampling).

The code of the library resides in the "r8b" C++ namespace, effectively isolating it from all other code. The code is thread-safe. A separate resampler object should be created for each audio channel or stream being processed concurrently.

Note that you will need to compile the "r8bbase.cpp" source file and include the resulting object file into your application build. This source file includes definitions of several global static objects used by the library. You may also need to include to your project: the "Kernel32" library (on Windows) and the "pthread" library on macOS and Linux.

The library is able to process signal of any scale and loudness: it is not limited to just a "usual" -1.0 to 1.0 range.

By defining the R8B_IPP configuration macro it is possible to enable Intel IPP back-end for FFT functions, instead of the default Ooura FFT. IPP FFT makes sample rate conversion faster by 23% on average.

#define R8B_IPP 1

If a larger initial processing delay and a very minor sample-timing error are not an issue, for the most efficiency you can define these macros at the beginning of the r8bconf.h file, or during compilation:

#define R8B_IPP 1
#define R8B_FASTTIMING 1
#define R8B_EXTFFT 1

If you do not have access to the Intel IPP then you may consider enabling the PFFFT which is only slightly slower than Intel IPP FFT in performance. There are two macros available: R8B_PFFFT and R8B_PFFFT_DOUBLE. The first macro enables PFFFT that works in single-precision resolution, thus limiting the overall resampler's precision to 24-bit sample rate conversions (for mission-critical professional audio applications, using the R8B_PFFFT macro is not recommended as its peak error is quite large). The second macro enables PFFFT implementation that works in double-precision resolution, making use of SSE2, AVX, and NEON intrinsics, yielding precision that is equal to both Intel IPP and Ooura FFT implementations.

To use the PFFFT, define the R8B_PFFFT or R8B_PFFFT_DOUBLE macro, compile and include the supplied pffft.cpp or pffft_double/pffft_double.c file to your project build.

#define R8B_PFFFT 1

or

#define R8B_PFFFT_DOUBLE 1

The code of this library was commented in the Doxygen style. To generate the documentation locally you may run the doxygen ./other/r8bdoxy.txt command from the library's folder.

Preliminary tests show that the r8b::CDSPResampler24 resampler class achieves 38*n_cores Mrops (56*n_cores for Intel IPP FFT) when converting 1 channel of 24-bit audio from 44100 to 96000 sample rate (2% transition band), on a Ryzen 3700X processor-based 64-bit system. This approximately translates to a real-time resampling of 860*n_cores (1270*n_cores) audio streams, at 100% CPU load. Performance when converting to other sample rates may vary greatly. When comparing performance of this resampler library to another library make sure that the competing library is also tuned to produce a fully linear-phase response, has similar stop-band characteristics and similar sample-timing precision.

Dynamic Link Library

The functions of this SRC library are also accessible in simplified form via the DLL file on Windows, requiring a processor with SSE2 support (Win64 version includes AVX2 auto-dispatch code). Delphi Pascal interface unit file for the DLL file is available. DLL and C LIB files are distributed in the DLL folder on the project's homepage. On non-Windows systems it is preferrable to use the C++ library directly. Note that the DLL was compiled with the Intel IPP enabled.

Real-Time Applications

The resampler class of this library was designed as an asynchronous processor: it may produce any number of output samples, depending on the input sample data length and the resampling parameters. The resampler must be fed with the input sample data until enough output sample data were produced, with any excess output samples used before feeding the resampler with more input data. A "relief" factor here is that the resampler removes the initial processing latency automatically, and that after initial moments of processing the output becomes steady, with only minor output sample data length fluctuations.

So, while for an off-line resampling a "push" method can be used, demonstrated in the example.cpp file, for a real-time resampling a "pull" method should be used which calls the resampling process until the output buffer is filled.

Notes

When using the r8b::CDSPResampler class directly, you may select the transition band/steepness of the low-pass (reconstruction) filter, expressed as a percentage of the full spectral bandwidth of the input signal (or the output signal if the downsampling is performed), and the desired stop-band attenuation in decibel.

The transition band is specified as the normalized spectral space of the input signal (or the output signal if the downsampling is performed) between the low-pass filter's -3 dB point and the Nyquist frequency, and ranges from 0.5% to 45%. Stop-band attenuation can be specified in the range from 49 to 218 decibel. Both the transition band and stop-band attenuation affect resampler's overall performance and initial output delay. For your information, transition frequency range spans 175% of the specified transition band, which means that for 2% transition band, frequency response below 0.965*Nyquist is linear.

This SRC library also implements a much faster "power of 2" resampling (e.g. 2X, 4X, 8X, 16X, 3X, 3*2X, 3*4X, 3*8X, etc. upsampling and downsampling), which is engaged automatically if the resampling parameters permit.

This library was tested for compatibility with GNU C++, Microsoft Visual C++, LLVM and Intel C++ compilers, on 32- and 64-bit Windows, macOS, and CentOS Linux.

Most code is "inline", without the need to compile many source files. The memory footprint is quite modest.

Acknowledgements

r8brain-free-src is bundled with the following code:

FFT routines Copyright (c) 1996-2001 Takuya OOURA. Homepage
PFFFT Copyright (c) 2013 Julien Pommier. Homepage
PFFFT DOUBLE Copyright (c) 2020 Hayati Ayguen, Dario Mambro. Homepage

Users

This library is used by:

Please send me a note via [email protected] and I will include a link to your software product to this list of users. This list is important in maintaining confidence in this library among the interested parties. The inclusion into this list is not mandatory.

Change Log

Version 6.3:

Improved the findGCD() function, to cover a wider range of sample rate ratios.
Added the R8B_DSPBASECLASS macro, to redefine the base class of non-cached objects.

Version 6.2:

Fixed miscalculation in the recently introduced getInLenBeforeOutPos() function for minimum-phase filters.
Fixed a mistake in the getInputRequiredForOutput() function.
Fixed a long-standing mistake in LatencyFrac value of whole-stepping interpolation. However, this mistake gave no practical issues before (absent for linear-phase filters, and minor for minimum-phase filters).

Version 6.1:

Made a micro-optimization of the "whole stepping" interpolation yielding 18% performance increase in some conversions (e.g., 44100 to 96000).
Implemented the getInLenBeforeOutPos() function which is an ultra-fast and flexible replacement for the getInLenBeforeOutStart() function (that became a legacy function now). Also added the getInputRequiredForOutput() helper function.
Updated comment sections across the codebase, to match the latest Doxygen version.
Reintroduced the r8b_inlen() function in the DLL.

Version 6.0:

Added SSE and NEON implementations to CDSPHBDownsampler yielding 5-16% performance improvement of power-of-2 downsampling.
Further optimization of filter calculation making it 15% faster.
Upped "SpinCount" in Windows mutex to 2000, to be on a safer side when the filter cache is fully filled.
Made the latest used "static" filter bank pop to the top of the list, for cases when multiple "ReqAtten" values are in use in an application.

Version 5.9:

Optimized filter calculation (Kaiser window function) with negligible change in filtering results.
Optimized min-phase filter's group delay calculation.
Reduced "SpinCount" in Windows mutex to 1000.
Made non-essential changes across the codebase and comments.

Version 5.8:

Rearranged FFT macros, added R8B_PFFFT and R8B_PFFFT_DOUBLE collision check.

Version 5.7:

Removed the defined( __ARM_NEON ) macro detection so that the code compiles on non-ARM64 platforms.

Version 5.6:

Added SSE and NEON implementations to CDSPHBUpsampler yielding 15% performance improvement of power-of-2 upsampling.
Added SSE and NEON implementations to the CDSPRealFFT::multiplyBlocksZP function, for 2-3% performance improvement.
Added intermediate interpolator's transition band limitation, for logical hardness (not practically needed).
Added the aDoConsumeLatency parameter to CDSPHBUpsampler constructor, for "inline" DSP uses of the class.
Made various minor changes across the codebase.

Version 5.5:

Hardened positional logic of fractional filter calculation, removed redundant multiplications.
Removed unnecessary function templating from the CDSPSincFilterGen class.
Added the __ARM_NEON macro to NEON availability detection.

Version 5.4:

Added compiler specializations to previously optimized inner loops. "Shuffled" SIMD interpolation code is not efficient on Apple M1. Intel C++ Compiler vectorizes "whole stepping" interpolation as good as a manually-written SSE.
Reorganized SIMD instructions for a slightly better performance.
Changed internal buffer sizes of half-band resamplers (1-2% performance boost).
Fixed compiler warnings in PFFFT code.
Added several asserts to the code.

Version 5.3:

Optimized inner loops of the fractional interpolator, added SSE2 and NEON intrinsics, resulting in a measurable (8-25%) performance gain.
Optimized filter calculation functions: changed some divisions by a constant to multiplications.
Renamed M_PI macros to R8B_PI, to avoid macro collisions.
Removed redundant code and macros.

Version 5.2:

Modified PFFFT and PFFFT DOUBLE conditional pre-processor directives to always enable NEON on aarch64/arm64 (this includes code built for Apple M1).

Version 5.1:

Changed alignment in the CFixedBuffer class to 64 bytes. This improves AVX performance of the PFFFT DOUBLE implementation by a few percent.
Removed redundant files from the pffft_double folder, integrated the pffft_common.c file into the pffft_double.c file.

Version 5.0:

Removed a long-outdated macros from the r8bconf.h file.
Changed a conditional pre-processor directive in the pf_sse2_double.h file as per PFFFT DOUBLE author's suggestion, to allow SSE2 intrinsics in most compilers.
Fixed "License.txt" misnaming in the source files to "LICENSE".

Version 4.10:

Added the PFFFT DOUBLE implementation support. Now available via the R8B_PFFFT_DOUBLE definition macro.

Version 4.9:

Reoptimized half-band and fractional interpolation filters with a stricter frequency response linearity constraints. This did not impact the average speed performance.

Version 4.8:

Added a limit to the intermediate filter's transition band, to keep the latency under control at any resampling ratio.

Version 4.7:

Added #ifndef _USE_MATH_DEFINES to pffft.cpp.
Moved #include "pffft.h" to CDSPRealFFT.h.

Version 4.6:

Removed the MaxInLen parameter from the oneshot() function.
Decreased intermediate low-pass filter's transition band slightly, for more stable quality.

Version 4.5:

Fixed VS2017 compiler warnings.

Version 4.4:

Fixed the "Declaration hides a member" Intel C++ compiler warnings.

Version 4.3:

Added //$ markers for internal debugging purposes.

Version 4.2:

Backed-off max transition band to 45 and MinAtten to 49.
Implemented Wave64 and AIFF file input in the r8bfreesrc bench tool. The tool is now compiled with the R8B_IPP 1 and R8B_EXTFFT 1 macros to demonstrate the maximal achievable performance.

Version 4.1:

Updated allowed ReqAtten range to 52-218, ReqTransBand 0.5-56. It is possible to specify filter parameters slightly beyond these values, but the resulting filter will be slightly out of specification as well.
Optimized static filter banks allocation.

Version 4.0:

A major overhaul of interpolation classes: now templated parameters are not used, all required parameters are calculated at runtime. Static filter bank object is not used, it is created when necessary, and then cached.
Implemented one-third interpolation filters, however, this did not measurably increase resampler's speed.

Version 3.7:

Used ippsMul_64f_I() in the CDSPRealFFT::multiplyBlockZ() function for a minor conversion speed increase in Intel IPP mode.

Version 3.6:

Added memory alignment to allocated buffers which boosts performance by 1.5% when Intel IPP FFT is in use.
Implemented PFFFT support.

Version 3.5:

Improved resampling speed very slightly.
Updated the r8bfreesrc benchmark tool to support RF64 WAV files.

Version 3.4:

Added a more efficient half-band filters for >= 256 resampling ratios.

Version 3.3:

Made minor fix to downsampling for some use cases of CDSPBlockConvolver, did not affect resampler.
Converted CDSPHBUpsampler and CDSPHBDownsampler's inner functions to static functions, which boosted high-ratio resampling performance measurably.

Version 3.2:

Minor fix to the latency consumption mechanism.

Version 3.1:

Reoptimized fractional delay filter's windowing function.

Version 3.0:

Implemented a new variant of the getInLenBeforeOutStart() function.
Reimplemented oneshot() function to support float buffer types.
Considerably improved downsampling performance at high resampling ratios.
Implemented intermediate interpolation technique which boosted upsampling performance for most resampling ratios considerably.
Removed the ConvCount constant - now resampler supports virtually any resampling ratios.
Removed the UsePower2 parameter from the resampler constructor.
Now resampler's process() function always returns pointer to the internal buffer, input buffer is returned only if no resampling happens.
Resampler's getMaxOutLen() function can now be used to obtain the maximal output length that can be produced by the resampler in a single call.
Added a more efficient "one-third" filters to half-band upsampler and downsampler.

Version 2.1:

Optimized 2X half-band downsampler.

Version 2.0:

Optimized power-of-2 upsampling.

Version 1.9:

Optimized half-band downsampling filter.
Implemented whole-number stepping resampling.
Added R8B_EXTFFT configuration option.
Fixed initial sub-sample offseting on downsampling.

Version 1.8:

Added R8B_FASTTIMING configuration option.

Version 1.7:

Improved sample timing precision.
Increased CDSPResampler :: ConvCountMax to 28 to support a lot wider resampling ratios.
Added bench tools.
Removed getInLenBeforeOutStart() due to incorrect calculation.

avaneev/r8brain-free-src

avaneev

Reviews

Repository Details