1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314
//! Basic digital signal processing (DSP) operations
//!
//! Digital signal processing based on real or complex vectors in time or frequency domain.
//! Vectors are expected to typically have a size which is at least in the order
//! of magnitude of a couple of thousand elements. This crate tries to balance between a clear
//! API and performance in terms of processing speed.
//!
//! Take this example:
//!
//! ```
//! # extern crate num_complex;
//! # extern crate basic_dsp_vector;
//! # use basic_dsp_vector::*;
//! # fn main() {
//! let mut vector1 = vec!(1.0, 2.0).to_real_time_vec();
//! let vector2 = vec!(10.0, 11.0).to_real_time_vec();
//! vector1.add(&vector2).expect("Ignoring error handling in examples");
//! # }
//!
//! ```
//! If `vector2` would be a complex or frequency vector then this won't compile. The type mismatch
//! indicates that a conversation is missing and that this might be a programming mistake. This lib uses
//! the Rust type system to catch such errors.
//!
//! DSP vectors are meant to integrate well with other types and so they can for example be converted from and to a Rust standard vector:
//!
//! ```
//! # use std::f32;
//! # use basic_dsp_vector::*;
//! # use basic_dsp_vector::conv_types::*;
//! let mut dsp_vec = vec![0.0; 1000].to_real_time_vec();
//! let mut buffer = SingleBuffer::new();
//! dsp_vec.interpolatei(&mut buffer, &RaisedCosineFunction::new(0.35), 2).unwrap();
//! let vec: Vec<f64> = dsp_vec.into();
//! assert_eq!(vec.len(), 2000);
//! ````
//!
//! DSP algorithms are often executed in loops. If you work with large vectors you typically try to avoid
//! allocating buffers in every iteration. Preallocating buffers is a common practice to safe a little time
//! with every iteration later on, but also to avoid heap fragmentation. At the same time it's a tedious task
//! to calculate the right buffer sizes for all operations. As an attempt to provide a more convenient solution
//! buffer types exist which don't preallocate, but store temporary memory segments so that they can be reused in the
//! next iteration. Here is an example:
//!
//! ```
//! # use std::f32;
//! # use basic_dsp_vector::*;
//! let vector = vec!(1.0, 0.0, -0.5, 0.8660254, -0.5, -0.8660254).to_complex_time_vec();
//! let mut buffer = SingleBuffer::new();
//! let _ = vector.fft(&mut buffer);
//! ```
//!
//! The vector types don't distinguish between the shapes 1xN or Nx1. This is a difference to other
//! conventions such as in MATLAB or GNU Octave.
//! The reason for this decision is that most operations are only defined if the shape of the
//! vector matches. So it appears to be more practical and clearer to implement the few operations
//! where the arguments can be of different shapes as seperate methods. The methods `mul` and `dot_product`
//! are one example for this.
//!
//! The trait definitions in this lib can look complex and might be overwhelming at the beginning.
//! There is a wide range of DSP vectors, e.g. a slice can be DSP vector, a boxed array can be a DSP vector,
//! a standard vector can be a DSP vector and so on. This lib tries to work with all of that and tries
//! to allow all those different DSP vector types to work together. The price for this flexibility is a more complex
//! trait definition. As a mental model, this is what the traits are specifiying:
//! Whenever you have a complex vector in time domain, it's binary operations will work with all other
//! complex vectors in time domain, but not with real valued vectors or frequency domain vectors.
//! And the type `GenDspVec` serves as wild card at compile time since it defers all checks to run time.
#![cfg_attr(feature = "use_simd", feature(portable_simd))]
#![cfg_attr(feature = "use_simd", feature(stdarch_x86_avx512))]
extern crate arrayvec;
#[cfg(feature = "use_gpu")]
extern crate clfft;
#[cfg(feature = "std")]
extern crate crossbeam;
#[cfg(feature = "std")]
#[macro_use]
extern crate lazy_static;
#[cfg(feature = "std")]
extern crate linreg;
extern crate num_complex;
#[cfg(feature = "std")]
extern crate num_cpus;
extern crate num_traits;
#[cfg(feature = "use_gpu")]
extern crate ocl;
extern crate rustfft;
#[macro_use]
mod simd_extensions;
pub mod conv_types;
pub mod meta;
mod multicore_support;
mod vector_types;
pub mod window_functions;
pub use crate::multicore_support::print_calibration;
pub use crate::multicore_support::MultiCoreSettings;
pub use crate::vector_types::*;
mod gpu_support;
mod inline_vector;
use crate::numbers::*;
use std::ops::Range;
pub mod numbers {
//! Traits from the `num` crate which are used inside `basic_dsp` and extensions to those traits.
use crate::gpu_support::{Gpu32, Gpu64, GpuFloat, GpuRegTrait};
use crate::simd_extensions;
use crate::simd_extensions::*;
pub use num_complex::Complex;
use num_traits;
pub use num_traits::Float;
pub use num_traits::Num;
pub use num_traits::One;
use rustfft;
use std::fmt::Debug;
/// A trait for a numeric value which at least supports a subset of the operations defined in this crate.
/// Can be an integer or a floating point number. In order to have support for all operations in this crate
/// a must implement the `RealNumber`.
pub trait DspNumber:
Num
+ Copy
+ Clone
+ Send
+ Sync
+ ToSimd
+ Debug
+ num_traits::Signed
+ num_traits::FromPrimitive
+ rustfft::FftNum
+ 'static
{
}
impl<T> DspNumber for T where
T: Num
+ Copy
+ Clone
+ Send
+ Sync
+ ToSimd
+ Debug
+ num_traits::Signed
+ num_traits::FromPrimitive
+ rustfft::FftNum
+ 'static
{
}
/// Associates a number type with a SIMD register type.
pub trait ToSimd: Sized + Sync + Send {
/// Type for the SIMD register on the CPU.
type RegFallback: SimdGeneric<Self>;
type RegSse: SimdGeneric<Self>;
type RegAvx: SimdGeneric<Self>;
type RegAvx512: SimdGeneric<Self>;
/// Type for the SIMD register on the GPU. Defaults to an arbitrary type if GPU support is not
/// compiled in.
type GpuReg: GpuRegTrait;
}
impl ToSimd for f32 {
type RegFallback = simd_extensions::fallback::f32x4;
#[cfg(all(feature = "use_sse2", target_feature = "sse2"))]
type RegSse = simd_extensions::sse::f32x4;
#[cfg(not(all(feature = "use_sse2", target_feature = "sse2")))]
type RegSse = simd_extensions::fallback::f32x4;
#[cfg(all(feature = "use_avx2"))]
type RegAvx = simd_extensions::avx::f32x8;
#[cfg(not(all(feature = "use_avx2")))]
type RegAvx = simd_extensions::fallback::f32x4;
#[cfg(feature = "use_avx512")]
type RegAvx512 = simd_extensions::avx512::f32x16;
#[cfg(not(feature = "use_avx512"))]
type RegAvx512 = simd_extensions::fallback::f32x4;
type GpuReg = Gpu32;
}
impl ToSimd for f64 {
type RegFallback = simd_extensions::fallback::f64x2;
#[cfg(all(feature = "use_sse2", target_feature = "sse2"))]
type RegSse = simd_extensions::sse::f64x2;
#[cfg(not(all(feature = "use_sse2", target_feature = "sse2")))]
type RegSse = simd_extensions::fallback::f64x2;
#[cfg(all(feature = "use_avx2", target_feature = "avx2"))]
type RegAvx = simd_extensions::avx::f64x4;
#[cfg(not(all(feature = "use_avx2", target_feature = "avx2")))]
type RegAvx = simd_extensions::fallback::f64x2;
#[cfg(feature = "use_avx512")]
type RegAvx512 = simd_extensions::avx512::f64x8;
#[cfg(not(feature = "use_avx512"))]
type RegAvx512 = simd_extensions::fallback::f64x2;
type GpuReg = Gpu64;
}
/// A real floating pointer number intended to abstract over `f32` and `f64`.
pub trait RealNumber: Float + DspNumber + GpuFloat + num_traits::FloatConst {}
impl<T> RealNumber for T where T: Float + DspNumber + GpuFloat + num_traits::FloatConst {}
/// This trait is necessary so that we can define zero for types outside this crate.
/// It calls the `num_traits::Zero` trait where possible.
pub trait Zero {
fn zero() -> Self;
}
impl<T> Zero for T
where
T: DspNumber,
{
fn zero() -> Self {
<Self as num_traits::Zero>::zero()
}
}
impl<T> Zero for Complex<T>
where
T: DspNumber,
{
fn zero() -> Self {
<Self as num_traits::Zero>::zero()
}
}
}
/// Transmutes a slice. Both S and D must be `#[repr(C)]`.
/// The code panics if the slice has a length which doesn't allow conversion.
fn transmute_slice<S, D>(source: &[S]) -> &[D] {
let len = get_target_slice_len::<S, D>(source);
unsafe {
let trans: &[D] = &*(source as *const [S] as *const [D]);
std::slice::from_raw_parts(trans.as_ptr(), len)
}
}
/// Transmutes a mutable slice. Both S and D must be `#[repr(C)]`.
/// The code panics if the slice has a length which doesn't allow conversion.
fn transmute_slice_mut<S, D>(source: &mut [S]) -> &mut [D] {
let len = get_target_slice_len::<S, D>(source);
unsafe {
let trans: &mut [D] = &mut *(source as *mut [S] as *mut [D]);
std::slice::from_raw_parts_mut(trans.as_mut_ptr(), len)
}
}
/// Helper method which finds the length of a target slice and also perform checks on the length
fn get_target_slice_len<S, D>(source: &[S]) -> usize {
let to_larger_type = std::mem::size_of::<D>() >= std::mem::size_of::<S>();
if to_larger_type {
assert_eq!(std::mem::size_of::<D>() % std::mem::size_of::<S>(), 0);
let ratio = std::mem::size_of::<D>() / std::mem::size_of::<S>();
assert_eq!(source.len() % ratio, 0);
source.len() / ratio
} else {
assert_eq!(std::mem::size_of::<S>() % std::mem::size_of::<D>(), 0);
let ratio = std::mem::size_of::<S>() / std::mem::size_of::<D>();
source.len() * ratio
}
}
// Returns a complex slice from a real slice
fn array_to_complex<T>(array: &[T]) -> &[Complex<T>] {
transmute_slice(array)
}
// Returns a complex slice from a real slice
fn array_to_complex_mut<T>(array: &mut [T]) -> &mut [Complex<T>] {
transmute_slice_mut(array)
}
/// Copies memory inside a slice
fn memcpy<T: Copy>(data: &mut [T], from: Range<usize>, to: usize) {
use std::ptr::copy;
assert!(from.start <= from.end);
assert!(from.end <= data.len());
assert!(to <= data.len() - (from.end - from.start));
unsafe {
let ptr = data.as_mut_ptr();
copy(ptr.add(from.start), ptr.add(to), from.end - from.start)
}
}
// Zeros a range within the slice
fn memzero<T: Copy>(data: &mut [T], range: Range<usize>) {
use std::ptr::write_bytes;
assert!(range.start <= range.end);
assert!(range.end <= data.len());
unsafe {
let ptr = data.as_mut_ptr();
write_bytes(ptr.add(range.start), 0, range.end - range.start);
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::simd_extensions::Simd;
#[test]
fn to_simd_test() {
// This is more a check for syntax. So if it compiles
// then the test already passes. The assert is then only
// a sanity check.
let reg = <f32 as ToSimd>::RegFallback::splat(1.0);
let sum = reg.sum_real();
assert!(sum > 0.0);
}
}