| 1 |
use crate::{Fft, FftDirection, FftNum}; |
| 2 |
use std::arch::x86_64::{__m256, __m256d}; |
| 3 |
use std::sync::Arc; |
| 4 |
|
| 5 |
pub trait AvxNum: FftNum { |
| 6 |
type VectorType: AvxVector256<ScalarType = Self>; |
| 7 |
} |
| 8 |
|
| 9 |
impl AvxNum for f32 { |
| 10 |
type VectorType = __m256; |
| 11 |
} |
| 12 |
impl AvxNum for f64 { |
| 13 |
type VectorType = __m256d; |
| 14 |
} |
| 15 |
|
| 16 |
// Data that most (non-butterfly) SIMD FFT algorithms share |
| 17 |
// Algorithms aren't required to use this struct, but it allows for a lot of reduction in code duplication |
| 18 |
struct CommonSimdData<T, V> { |
| 19 |
inner_fft: Arc<dyn Fft<T>>, |
| 20 |
twiddles: Box<[V]>, |
| 21 |
|
| 22 |
len: usize, |
| 23 |
|
| 24 |
inplace_scratch_len: usize, |
| 25 |
outofplace_scratch_len: usize, |
| 26 |
|
| 27 |
direction: FftDirection, |
| 28 |
} |
| 29 |
|
| 30 |
macro_rules! boilerplate_avx_fft { |
| 31 |
($struct_name:ident, $len_fn:expr, $inplace_scratch_len_fn:expr, $out_of_place_scratch_len_fn:expr) => { |
| 32 |
impl<A: AvxNum, T: FftNum> Fft<T> for $struct_name<A, T> { |
| 33 |
fn process_outofplace_with_scratch( |
| 34 |
&self, |
| 35 |
input: &mut [Complex<T>], |
| 36 |
output: &mut [Complex<T>], |
| 37 |
scratch: &mut [Complex<T>], |
| 38 |
) { |
| 39 |
let required_scratch = self.get_outofplace_scratch_len(); |
| 40 |
if scratch.len() < required_scratch |
| 41 |
|| input.len() < self.len() |
| 42 |
|| output.len() != input.len() |
| 43 |
{ |
| 44 |
// We want to trigger a panic, but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 45 |
fft_error_outofplace( |
| 46 |
self.len(), |
| 47 |
input.len(), |
| 48 |
output.len(), |
| 49 |
self.get_outofplace_scratch_len(), |
| 50 |
scratch.len(), |
| 51 |
); |
| 52 |
return; // Unreachable, because fft_error_outofplace asserts, but it helps codegen to put it here |
| 53 |
} |
| 54 |
|
| 55 |
let scratch = &mut scratch[..required_scratch]; |
| 56 |
let result = array_utils::iter_chunks_zipped( |
| 57 |
input, |
| 58 |
output, |
| 59 |
self.len(), |
| 60 |
|in_chunk, out_chunk| { |
| 61 |
self.perform_fft_out_of_place(in_chunk, out_chunk, scratch) |
| 62 |
}, |
| 63 |
); |
| 64 |
|
| 65 |
if result.is_err() { |
| 66 |
// We want to trigger a panic, because the buffer sizes weren't cleanly divisible by the FFT size, |
| 67 |
// but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 68 |
fft_error_outofplace( |
| 69 |
self.len(), |
| 70 |
input.len(), |
| 71 |
output.len(), |
| 72 |
self.get_outofplace_scratch_len(), |
| 73 |
scratch.len(), |
| 74 |
) |
| 75 |
} |
| 76 |
} |
| 77 |
fn process_with_scratch(&self, buffer: &mut [Complex<T>], scratch: &mut [Complex<T>]) { |
| 78 |
let required_scratch = self.get_inplace_scratch_len(); |
| 79 |
if scratch.len() < required_scratch || buffer.len() < self.len() { |
| 80 |
// We want to trigger a panic, but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 81 |
fft_error_inplace( |
| 82 |
self.len(), |
| 83 |
buffer.len(), |
| 84 |
self.get_inplace_scratch_len(), |
| 85 |
scratch.len(), |
| 86 |
); |
| 87 |
return; // Unreachable, because fft_error_inplace asserts, but it helps codegen to put it here |
| 88 |
} |
| 89 |
|
| 90 |
let scratch = &mut scratch[..required_scratch]; |
| 91 |
let result = array_utils::iter_chunks(buffer, self.len(), |chunk| { |
| 92 |
self.perform_fft_inplace(chunk, scratch) |
| 93 |
}); |
| 94 |
|
| 95 |
if result.is_err() { |
| 96 |
// We want to trigger a panic, because the buffer sizes weren't cleanly divisible by the FFT size, |
| 97 |
// but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 98 |
fft_error_inplace( |
| 99 |
self.len(), |
| 100 |
buffer.len(), |
| 101 |
self.get_inplace_scratch_len(), |
| 102 |
scratch.len(), |
| 103 |
) |
| 104 |
} |
| 105 |
} |
| 106 |
#[inline(always)] |
| 107 |
fn get_inplace_scratch_len(&self) -> usize { |
| 108 |
$inplace_scratch_len_fn(self) |
| 109 |
} |
| 110 |
#[inline(always)] |
| 111 |
fn get_outofplace_scratch_len(&self) -> usize { |
| 112 |
$out_of_place_scratch_len_fn(self) |
| 113 |
} |
| 114 |
} |
| 115 |
impl<A: AvxNum, T> Length for $struct_name<A, T> { |
| 116 |
#[inline(always)] |
| 117 |
fn len(&self) -> usize { |
| 118 |
$len_fn(self) |
| 119 |
} |
| 120 |
} |
| 121 |
impl<A: AvxNum, T> Direction for $struct_name<A, T> { |
| 122 |
#[inline(always)] |
| 123 |
fn fft_direction(&self) -> FftDirection { |
| 124 |
self.direction |
| 125 |
} |
| 126 |
} |
| 127 |
}; |
| 128 |
} |
| 129 |
|
| 130 |
macro_rules! boilerplate_avx_fft_commondata { |
| 131 |
($struct_name:ident) => { |
| 132 |
impl<A: AvxNum, T: FftNum> Fft<T> for $struct_name<A, T> { |
| 133 |
fn process_outofplace_with_scratch( |
| 134 |
&self, |
| 135 |
input: &mut [Complex<T>], |
| 136 |
output: &mut [Complex<T>], |
| 137 |
scratch: &mut [Complex<T>], |
| 138 |
) { |
| 139 |
if self.len() == 0 { |
| 140 |
return; |
| 141 |
} |
| 142 |
|
| 143 |
let required_scratch = self.get_outofplace_scratch_len(); |
| 144 |
if scratch.len() < required_scratch |
| 145 |
|| input.len() < self.len() |
| 146 |
|| output.len() != input.len() |
| 147 |
{ |
| 148 |
// We want to trigger a panic, but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 149 |
fft_error_outofplace( |
| 150 |
self.len(), |
| 151 |
input.len(), |
| 152 |
output.len(), |
| 153 |
self.get_outofplace_scratch_len(), |
| 154 |
scratch.len(), |
| 155 |
); |
| 156 |
return; // Unreachable, because fft_error_outofplace asserts, but it helps codegen to put it here |
| 157 |
} |
| 158 |
|
| 159 |
let scratch = &mut scratch[..required_scratch]; |
| 160 |
let result = array_utils::iter_chunks_zipped( |
| 161 |
input, |
| 162 |
output, |
| 163 |
self.len(), |
| 164 |
|in_chunk, out_chunk| { |
| 165 |
self.perform_fft_out_of_place(in_chunk, out_chunk, scratch) |
| 166 |
}, |
| 167 |
); |
| 168 |
|
| 169 |
if result.is_err() { |
| 170 |
// We want to trigger a panic, because the buffer sizes weren't cleanly divisible by the FFT size, |
| 171 |
// but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 172 |
fft_error_outofplace( |
| 173 |
self.len(), |
| 174 |
input.len(), |
| 175 |
output.len(), |
| 176 |
self.get_outofplace_scratch_len(), |
| 177 |
scratch.len(), |
| 178 |
); |
| 179 |
} |
| 180 |
} |
| 181 |
fn process_with_scratch(&self, buffer: &mut [Complex<T>], scratch: &mut [Complex<T>]) { |
| 182 |
if self.len() == 0 { |
| 183 |
return; |
| 184 |
} |
| 185 |
|
| 186 |
let required_scratch = self.get_inplace_scratch_len(); |
| 187 |
if scratch.len() < required_scratch || buffer.len() < self.len() { |
| 188 |
// We want to trigger a panic, but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 189 |
fft_error_inplace( |
| 190 |
self.len(), |
| 191 |
buffer.len(), |
| 192 |
self.get_inplace_scratch_len(), |
| 193 |
scratch.len(), |
| 194 |
); |
| 195 |
return; // Unreachable, because fft_error_inplace asserts, but it helps codegen to put it here |
| 196 |
} |
| 197 |
|
| 198 |
let scratch = &mut scratch[..required_scratch]; |
| 199 |
let result = array_utils::iter_chunks(buffer, self.len(), |chunk| { |
| 200 |
self.perform_fft_inplace(chunk, scratch) |
| 201 |
}); |
| 202 |
|
| 203 |
if result.is_err() { |
| 204 |
// We want to trigger a panic, because the buffer sizes weren't cleanly divisible by the FFT size, |
| 205 |
// but we want to avoid doing it in this function to reduce code size, so call a function marked cold and inline(never) that will do it for us |
| 206 |
fft_error_inplace( |
| 207 |
self.len(), |
| 208 |
buffer.len(), |
| 209 |
self.get_inplace_scratch_len(), |
| 210 |
scratch.len(), |
| 211 |
); |
| 212 |
} |
| 213 |
} |
| 214 |
#[inline(always)] |
| 215 |
fn get_inplace_scratch_len(&self) -> usize { |
| 216 |
self.common_data.inplace_scratch_len |
| 217 |
} |
| 218 |
#[inline(always)] |
| 219 |
fn get_outofplace_scratch_len(&self) -> usize { |
| 220 |
self.common_data.outofplace_scratch_len |
| 221 |
} |
| 222 |
} |
| 223 |
impl<A: AvxNum, T> Length for $struct_name<A, T> { |
| 224 |
#[inline(always)] |
| 225 |
fn len(&self) -> usize { |
| 226 |
self.common_data.len |
| 227 |
} |
| 228 |
} |
| 229 |
impl<A: AvxNum, T> Direction for $struct_name<A, T> { |
| 230 |
#[inline(always)] |
| 231 |
fn fft_direction(&self) -> FftDirection { |
| 232 |
self.common_data.direction |
| 233 |
} |
| 234 |
} |
| 235 |
}; |
| 236 |
} |
| 237 |
|
| 238 |
#[macro_use] |
| 239 |
mod avx_vector; |
| 240 |
|
| 241 |
mod avx32_butterflies; |
| 242 |
mod avx32_utils; |
| 243 |
|
| 244 |
mod avx64_butterflies; |
| 245 |
mod avx64_utils; |
| 246 |
|
| 247 |
mod avx_bluesteins; |
| 248 |
mod avx_mixed_radix; |
| 249 |
mod avx_raders; |
| 250 |
|
| 251 |
pub mod avx_planner; |
| 252 |
|
| 253 |
pub use self::avx32_butterflies::{ |
| 254 |
Butterfly11Avx, Butterfly128Avx, Butterfly12Avx, Butterfly16Avx, Butterfly24Avx, |
| 255 |
Butterfly256Avx, Butterfly27Avx, Butterfly32Avx, Butterfly36Avx, Butterfly48Avx, |
| 256 |
Butterfly512Avx, Butterfly54Avx, Butterfly5Avx, Butterfly64Avx, Butterfly72Avx, Butterfly7Avx, |
| 257 |
Butterfly8Avx, Butterfly9Avx, |
| 258 |
}; |
| 259 |
|
| 260 |
pub use self::avx64_butterflies::{ |
| 261 |
Butterfly11Avx64, Butterfly128Avx64, Butterfly12Avx64, Butterfly16Avx64, Butterfly18Avx64, |
| 262 |
Butterfly24Avx64, Butterfly256Avx64, Butterfly27Avx64, Butterfly32Avx64, Butterfly36Avx64, |
| 263 |
Butterfly512Avx64, Butterfly5Avx64, Butterfly64Avx64, Butterfly7Avx64, Butterfly8Avx64, |
| 264 |
Butterfly9Avx64, |
| 265 |
}; |
| 266 |
|
| 267 |
pub use self::avx_bluesteins::BluesteinsAvx; |
| 268 |
pub use self::avx_mixed_radix::{ |
| 269 |
MixedRadix11xnAvx, MixedRadix12xnAvx, MixedRadix16xnAvx, MixedRadix2xnAvx, MixedRadix3xnAvx, |
| 270 |
MixedRadix4xnAvx, MixedRadix5xnAvx, MixedRadix6xnAvx, MixedRadix7xnAvx, MixedRadix8xnAvx, |
| 271 |
MixedRadix9xnAvx, |
| 272 |
}; |
| 273 |
pub use self::avx_raders::RadersAvx2; |
| 274 |
use self::avx_vector::AvxVector256; |