unsafe fn round4(value: &[f64; 4]) -> [i64; 4]Examples found in repository?
crates/competitive/src/math/mint_fft_convolve.rs (line 233)
199pub unsafe fn convolve_u64_avx2(a: Vec<u64>, b: Vec<u64>) -> Vec<u64> {
200 let len = a.len() + b.len() - 1;
201 let n = len.next_power_of_two() / 2;
202 let a_parts = if a.iter().all(|&value| value <= u32::MAX as u64) {
203 3
204 } else {
205 5
206 };
207 let mut fa = split_u64_coefficients(&a, n);
208 drop(a);
209 let b_parts = if b.iter().all(|&value| value <= u32::MAX as u64) {
210 3
211 } else {
212 5
213 };
214 let mut fb = split_u64_coefficients(&b, n);
215 drop(b);
216 for part in 0..3 {
217 fft_soa(&mut fa[part]);
218 fft_soa(&mut fb[part]);
219 }
220 for part in &mut fa[3..a_parts] {
221 fft_soa(part);
222 }
223 for part in &mut fb[3..b_parts] {
224 fft_soa(part);
225 }
226 dot_u64_soa(&mut fa, &fb);
227 drop(fb);
228 for part in &mut fa {
229 ifft_soa(part);
230 }
231 let mut result = vec![0; len];
232 for (block, _) in fa[0].iter().enumerate() {
233 let real: [[i64; 4]; 5] = std::array::from_fn(|part| round4(&fa[part][block].re));
234 let imag: [[i64; 4]; 5] = std::array::from_fn(|part| round4(&fa[part][block].im));
235 for lane in 0..4 {
236 let i = block * 4 + lane;
237 if i < len {
238 result[i] = (real[0][lane] as u64)
239 .wrapping_add((real[1][lane] as u64) << 13)
240 .wrapping_add((real[2][lane] as u64) << 26)
241 .wrapping_add((real[3][lane] as u64) << 39)
242 .wrapping_add((real[4][lane] as u64) << 52);
243 }
244 if i + n < len {
245 result[i + n] = (imag[0][lane] as u64)
246 .wrapping_add((imag[1][lane] as u64) << 13)
247 .wrapping_add((imag[2][lane] as u64) << 26)
248 .wrapping_add((imag[3][lane] as u64) << 39)
249 .wrapping_add((imag[4][lane] as u64) << 52);
250 }
251 }
252 }
253 result
254}