yfft 0.1.0

Simple fast Fourier transform library optimized for small, power-of-two, single precision transforms and x86 systems 
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

use super::super::Num;
use super::utils::{if_compatible, AlignInfo, AlignReqKernel, AlignReqKernelWrapper};
use super::{Kernel, KernelParams, SliceAccessor};

use packed_simd::{u32x4, u64x2, u64x4};

use std::{mem, ptr};

pub unsafe fn new_x86_avx_bit_reversal_kernel<T>(indices: &Vec<usize>) -> Option<Box<Kernel<T>>>
where
    T: Num,
{
    if indices.len() < 8 {
        // doesn't benefit much
        return None;
    }

    if_compatible(|| {
        Some(
            Box::new(AlignReqKernelWrapper::new(AvxDWordBitReversalKernel {
                indices: indices.clone(),
            })) as Box<Kernel<f32>>,
        )
    })
}

#[derive(Debug)]
struct AvxDWordBitReversalKernel {
    indices: Vec<usize>,
}

impl<T: Num> AlignReqKernel<T> for AvxDWordBitReversalKernel {
    fn transform<I: AlignInfo>(&self, params: &mut KernelParams<T>) {
        assert_eq!(mem::size_of::<T>(), 4);

        let indices = unsafe { SliceAccessor::new(&self.indices) };
        let size = self.indices.len();
        let mut data = unsafe { SliceAccessor::new(&mut params.coefs[0..size * 2]) };
        let mut wa = unsafe { SliceAccessor::new(&mut params.work_area[0..size * 2]) };
        wa.copy_from_slice(*data);
        let mut i = 0;
        while i + 7 < size {
            let index1 = indices[i];
            let index2 = indices[i + 1];
            let index3 = indices[i + 2];
            let index4 = indices[i + 3];
            let index5 = indices[i + 4];
            let index6 = indices[i + 5];
            let index7 = indices[i + 6];
            let index8 = indices[i + 7];

            let src1: *const u64 = &wa[index1 * 2] as *const T as *const u64;
            let src2: *const u64 = &wa[index2 * 2] as *const T as *const u64;
            let src3: *const u64 = &wa[index3 * 2] as *const T as *const u64;
            let src4: *const u64 = &wa[index4 * 2] as *const T as *const u64;
            let src5: *const u64 = &wa[index5 * 2] as *const T as *const u64;
            let src6: *const u64 = &wa[index6 * 2] as *const T as *const u64;
            let src7: *const u64 = &wa[index7 * 2] as *const T as *const u64;
            let src8: *const u64 = &wa[index8 * 2] as *const T as *const u64;
            let dest: *mut u64x4 = &mut data[i * 2] as *mut T as *mut u64x4;
            unsafe {
                I::write(dest, u64x4::new(*src1, *src2, *src3, *src4));
                I::write(dest.offset(1), u64x4::new(*src5, *src6, *src7, *src8));
            }

            i += 8;
        }
        while i < size {
            let index = indices[i];

            let src: *const u64 = &wa[index * 2] as *const T as *const u64;
            let dest: *mut u64 = &mut data[i * 2] as *mut T as *mut u64;
            unsafe {
                *dest = *src;
            }

            i += 1;
        }
    }
    fn required_work_area_size(&self) -> usize {
        self.indices.len() * 2
    }
    fn alignment_requirement(&self) -> usize {
        32
    }
}

pub unsafe fn new_x86_avx_radix2_bit_reversal_kernel<T>(
    indices: &Vec<usize>,
) -> Option<Box<Kernel<T>>>
where
    T: Num,
{
    if indices.len() < 8 || indices.len() % 8 != 0 {
        // doesn't benefit much / requires an unaligned access
        return None;
    }

    // Check legibility
    for i in 0..indices.len() / 2 {
        if indices[i] > (<u32>::max_value() / 2) as usize {
            return None;
        }
    }

    let (f1, f2, f3) = (
        indices[1] - indices[0],
        indices[2] - indices[0],
        indices[3] - indices[0],
    );
    for i in 0..indices.len() / 8 {
        let (b0, b1, b2, b3) = (
            indices[i * 4],
            indices[i * 4 + 1],
            indices[i * 4 + 2],
            indices[i * 4 + 3],
        );
        if b1 != b0 + f1 || b2 != b0 + f2 || b3 != b0 + f3 {
            return None;
        }
    }

    if_compatible(|| {
        Some(Box::new(AlignReqKernelWrapper::new(
            AvxDWordRadix2BitReversalKernel {
                indices: (0..indices.len() / 8)
                    .map(|i| (indices[i * 4] as u32) * 2)
                    .collect(),
                offs: u32x4::new(0, f1 as u32 * 2, f2 as u32 * 2, f3 as u32 * 2),
            },
        )) as Box<Kernel<f32>>)
    })
}

#[derive(Debug)]
struct AvxDWordRadix2BitReversalKernel {
    indices: Vec<u32>,
    offs: u32x4,
}

impl<T: Num> AlignReqKernel<T> for AvxDWordRadix2BitReversalKernel {
    fn transform<I: AlignInfo>(&self, params: &mut KernelParams<T>) {
        assert_eq!(mem::size_of::<T>(), 4);

        let indices = unsafe { SliceAccessor::new(&self.indices) };
        let size = self.indices.len();
        let mut data = unsafe { SliceAccessor::new(&mut params.coefs[0..size * 16]) };
        let mut wa = unsafe { SliceAccessor::new(&mut params.work_area[0..size * 16]) };
        wa.copy_from_slice(*data);

        let offs = self.offs;

        let mut i = 0;
        while i < size {
            let index1234 = offs + u32x4::splat(indices[i]);
            let index1 = index1234.extract(0) as usize;
            let index2 = index1234.extract(1) as usize;
            let index3 = index1234.extract(2) as usize;
            let index4 = index1234.extract(3) as usize;

            let src1 = unsafe { ptr::read_unaligned(&wa[index1] as *const T as *const u64x2) };
            let src2 = unsafe { ptr::read_unaligned(&wa[index2] as *const T as *const u64x2) };
            let src3 = unsafe { ptr::read_unaligned(&wa[index3] as *const T as *const u64x2) };
            let src4 = unsafe { ptr::read_unaligned(&wa[index4] as *const T as *const u64x2) };

            let t1a: u64x2 = shuffle!(src1, src2, [0, 2]); // unpcklpd
            let t2a: u64x2 = shuffle!(src3, src4, [0, 2]); // unpcklpd

            let t1b: u64x2 = shuffle!(src1, src2, [1, 3]); // unpckhpd
            let t2b: u64x2 = shuffle!(src3, src4, [1, 3]); // unpckhpd

            let out1: u64x4 = shuffle!(t1a, t2a, [0, 1, 2, 3]); // inserti128
            let out2: u64x4 = shuffle!(t1b, t2b, [0, 1, 2, 3]); // inserti128

            let dest1: *mut u64x4 = &mut data[i * 8] as *mut T as *mut u64x4;
            let dest2: *mut u64x4 = &mut data[(i + size) * 8] as *mut T as *mut u64x4;

            unsafe {
                I::write(dest1, out1);
                I::write(dest2, out2);
            }

            i += 1;
        }
    }
    fn required_work_area_size(&self) -> usize {
        self.indices.len() * 16
    }
    fn alignment_requirement(&self) -> usize {
        32
    }
}

pub unsafe fn new_x86_avx_radix4_bit_reversal_kernel<T>(
    indices: &Vec<usize>,
) -> Option<Box<Kernel<T>>>
where
    T: Num,
{
    if indices.len() < 32 || indices.len() % 32 != 0 {
        // doesn't benefit much / requires an unaligned access
        return None;
    }

    // Check legibility
    for i in 0..indices.len() / 4 {
        if indices[i] > (<u32>::max_value() / 2) as usize {
            return None;
        }
    }

    let (f1, f2, f3) = (
        indices[1] - indices[0],
        indices[2] - indices[0],
        indices[3] - indices[0],
    );
    for i in 0..indices.len() / 16 {
        let (b0, b1, b2, b3) = (
            indices[i * 4],
            indices[i * 4 + 1],
            indices[i * 4 + 2],
            indices[i * 4 + 3],
        );
        if b1 != b0 + f1 || b2 != b0 + f2 || b3 != b0 + f3 {
            return None;
        }
    }

    if_compatible(|| {
        Some(Box::new(AlignReqKernelWrapper::new(
            AvxDWordRadix4BitReversalKernel {
                indices: (0..indices.len() / 16)
                    .map(|i| (indices[i * 4] as u32) * 2)
                    .collect(),
                offs: u32x4::new(0, f1 as u32 * 2, f2 as u32 * 2, f3 as u32 * 2),
            },
        )) as Box<Kernel<f32>>)
    })
}

#[derive(Debug)]
struct AvxDWordRadix4BitReversalKernel {
    indices: Vec<u32>,
    offs: u32x4,
}

impl<T: Num> AlignReqKernel<T> for AvxDWordRadix4BitReversalKernel {
    fn transform<I: AlignInfo>(&self, params: &mut KernelParams<T>) {
        assert_eq!(mem::size_of::<T>(), 4);

        let indices = unsafe { SliceAccessor::new(&self.indices) };
        let size = self.indices.len();
        let mut data = unsafe { SliceAccessor::new(&mut params.coefs[0..size * 32]) };
        let mut wa = unsafe { SliceAccessor::new(&mut params.work_area[0..size * 32]) };
        wa.copy_from_slice(*data);

        let offs = self.offs;

        let mut i = 0;
        while i + 1 < size {
            for _ in 0..2 {
                let index1234 = offs + u32x4::splat(indices[i]);
                let index1 = index1234.extract(0) as usize;
                let index2 = index1234.extract(1) as usize;
                let index3 = index1234.extract(2) as usize;
                let index4 = index1234.extract(3) as usize;

                let src1 = unsafe { ptr::read_unaligned(&wa[index1] as *const T as *const u64x4) };
                let src2 = unsafe { ptr::read_unaligned(&wa[index2] as *const T as *const u64x4) };
                let src3 = unsafe { ptr::read_unaligned(&wa[index3] as *const T as *const u64x4) };
                let src4 = unsafe { ptr::read_unaligned(&wa[index4] as *const T as *const u64x4) };

                let t1a: u64x4 = shuffle!(src1, src2, [0, 4, 2, 6]); // unpcklpd
                let t2a: u64x4 = shuffle!(src3, src4, [0, 4, 2, 6]); // unpcklpd

                let t1b: u64x4 = shuffle!(src1, src2, [1, 5, 3, 7]); // unpckhpd
                let t2b: u64x4 = shuffle!(src3, src4, [1, 5, 3, 7]); // unpckhpd

                let out1: u64x4 = shuffle!(t1a, t2a, [0, 1, 4, 5]); // inserti128/perm2f128
                let out2: u64x4 = shuffle!(t1b, t2b, [0, 1, 4, 5]); // inserti128/perm2f128
                let out3: u64x4 = shuffle!(t1a, t2a, [2, 3, 6, 7]); // inserti128/perm2f128
                let out4: u64x4 = shuffle!(t1b, t2b, [2, 3, 6, 7]); // inserti128/perm2f128

                let dest1: *mut u64x4 = &mut data[i * 8] as *mut T as *mut u64x4;
                let dest2: *mut u64x4 = &mut data[(i + size) * 8] as *mut T as *mut u64x4;
                let dest3: *mut u64x4 = &mut data[(i + size * 2) * 8] as *mut T as *mut u64x4;
                let dest4: *mut u64x4 = &mut data[(i + size * 3) * 8] as *mut T as *mut u64x4;

                unsafe {
                    I::write(dest1, out1);
                    I::write(dest2, out2);
                    I::write(dest3, out3);
                    I::write(dest4, out4);
                }

                i += 1;
            }
        }
        assert_eq!(i, size);
    }
    fn required_work_area_size(&self) -> usize {
        self.indices.len() * 32
    }
    fn alignment_requirement(&self) -> usize {
        32
    }
}