lsp-dsp-lib

DSP library for digital signal processing (and more)

This library provides set of functions that perform SIMD-optimized computing on several hardware architectures.

Currently supported set of SIMD extensions:

• i586 architecture (32-bit): SSE, SSE2, SSE3, AVX, AVX2 and FMA3;
• x86_64 architecture (64-bit): SSE, SSE2, SSE3, AVX, AVX2 and FMA3;
• armv7 architecture (32-bit): NEON;
• AArch64 architecture (64-bit): ASIMD.

All functions currently operate on IEEE-754 single-precision floating-point numbers.

Current set of functions provided:

• Functions that gather system information and optimize CPU for better computing;
• Cooley-Tukey 1-dimensional FFT algorithms with unpacked complex numbers;
• Cooley-Tukey 1-dimensional FFT algorithms with packed complex numbers;
• Direct convolution algorithm;
• Fast convolution functions that enhance performance of FFT-based convolution algorithms;
• Biquad static filter transform and processing algorithms;
• Biquad dynamic filter transform and processing algorithms;
• Floating-point operations: copying, moving, protection from NaNs and denormals;
• Parallel arithmetics functions on long vectors including fused multiply operations;
• Basic unpacked complex number arithmetics;
• Basic packed complex number arithmetics;
• Some functions that operate on RGB and HSL colors and their conversions;
• Mid/Side matrix functions for converting Stereo channel to Mid/Side and back;
• Functions for searching minimums and maximums;
• Resampling functions based on Lanczos filter;
• Interpolation functions;
• Some set of function to work with 3D mathematics.

Supported platforms

The build and correct unit test execution has been confirmed for following platforms:

• FreeBSD
• GNU/Linux
• OpenBSD
• Windows 32-bit
• Windows 64-bit

Supported architectures

The support of following list of hardware architectures has been implemented:

• i386 (32-bit) - full support.
• x86_64 (64-bit) - full support.
• ARMv6A - full support.
• ARMv7A - full support.
• AArch64 - most functions.

For all other architectures the generic implementation of algorithms is used, without any architecture-specific optimizations.

Requirements

The following packages need to be installed for building:

• gcc >= 4.9
• make >= 4.0

Building

To build the library, perform the following commands:

make config # Configure the build
make fetch # Fetch dependencies from Git repository
make
sudo make install

To get more build options, run:

make help

To uninstall library, simply issue:

make uninstall

To clean all binary files, run:

make clean

To clean the whole project tree including configuration files, run:

make prune

To fetch all possible dependencies and make the source code tree portable between different architectures and platforms, run:

make tree

To build source code archive with all possible dependencies, run:

make distsrc

Usage

Here's the code snippet of how the library can be initialized and used in C++:

#include <lsp-plug.in/dsp/dsp.h>
#include <stdio.h>
#include <stdlib.h>

int main(int argc, const char **argv)
{
    // Initialize DSP
    lsp::dsp::init();

    // Optionally: output information about the system
    lsp::dsp::info_t *info = lsp::dsp::info();
    if (info != NULL)
    {
        printf("Architecture:   %s\n", info->arch);
        printf("Processor:      %s\n", info->cpu);
        printf("Model:          %s\n", info->model);
        printf("Features:       %s\n", info->features);

        ::free(info);
    }
    
    // For faster computing we can tune CPU by updating thread context.
    // This will enable Flush-to-Zero and Denormals-are-Zero flags on
    // CPUs that support them. This is thread-local change and should
    // be called in each individual processing thread
    lsp::dsp::context_t ctx;
    lsp::dsp::start(&ctx);
    
    // Here we call some dsp functions, for example dsp::fill_zero
    float v[0x1000];
    lsp::dsp::fill_zero(v, sizeof(v)/sizeof(float));
    
    // At the end, we need to restore the context and reset CPU settings to defaults
    lsp::dsp::finish(&ctx);
    
    return 0;
}

Also all functions can be accessed from pure C with lsp_dsp_ prefix in the funcion and type names:

#include <lsp-plug.in/dsp/dsp.h>
#include <stdio.h>
#include <stdlib.h>

int main(int argc, const char **argv)
{
    // Initialize DSP
    lsp_dsp_init();

    // Optionally: output information about the system
    lsp_dsp_info_t *info = lsp_dsp_info();
    if (info != NULL)
    {
        printf("Architecture:   %s\n", info->arch);
        printf("Processor:      %s\n", info->cpu);
        printf("Model:          %s\n", info->model);
        printf("Features:       %s\n", info->features);

        free(info);
    }
    
    // For faster computing we can tune CPU by updating thread context.
    // This will enable Flush-to-Zero and Denormals-are-Zero flags on
    // CPUs that support them. This is thread-local change and should
    // be called in each individual processing thread
    lsp_dsp_context_t ctx;
    lsp_dsp_start(&ctx);
    
    // Here we call some dsp functions, for example lsp_dsp_fill_zero
    float v[0x1000];
    lsp_dsp_fill_zero(v, sizeof(v)/sizeof(float));
    
    // At the end, we need to restore the context and reset CPU settings to defaults
    lsp_dsp_finish(&ctx);
    
    return 0;
}