LMMP 4.3.0
Lamina High-Precision Arithmetic Library
载入中...
搜索中...
未找到
fft_fuse.c
浏览该文件的文档.
1/**
2 * Copyright (C) 2026 HJimmyK (Jericho Knox)
3 *
4 * This file is part of LMMP.
5 *
6 * LMMP is free software: you can redistribute it and/or modify it
7 * under the terms of the GNU Lesser General Public License as
8 * published by the Free Software Foundation; either version 3 of
9 * the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library. If not, see
18 * <https://www.gnu.org/licenses/>.
19 *
20 * SPDX-License-Identifier: LGPL-3.0-or-later
21 */
22
23/*
24 fft/ifft 蝶形数据面融合内核 lmmp_fft_bfy_fused_ / lmmp_ifft_bfy_fused_ 的 C 参考实现。
25
26 x64 汇编构建下符号由 arch/x64/fft_fuse.S 提供(本文件被构建系统排除);
27 GENERIC 与 arm64 汇编构建走此 C 实现(arm64 NEON 版待移植,由 CMake 补入)。
28
29 顺序流式访问(每 limb 恰好读写一次),不做 PART_SIZE 分块。
30 差链补码化同 add_n_sub_n_: d = a + ~t + k, k0 = 1 - b, 借位 = 1 - k。
31*/
32
33#include "lmmpn/fft_ssa.h"
34
36 mp_size_t l, mp_size_t w, mp_size_t shr) {
37 const unsigned rshr = LIMB_BITS - shr;
38 mp_limb_t acy = 0, bcy = 0; // acy=numa 链, bcy=numc 链(各段进出解包/重包)
39 mp_limb_t cur;
40
41 // 扫描1: numc[l-w+i] = numa[l-w+i] + t, numa[l-w+i] -= t(和→numc 即 bcy, 差→numa 即 acy)
42 cur = numb[0];
43 {
44 mp_limb_t ca = bcy, k = 1 - acy;
45 for (mp_size_t i = 0; i < w; i++) {
46 mp_limb_t nx = (i + 1 < w) ? numb[i + 1] : numb[w]; // 顶 limb 的 hi 来自 numb[w]
47 mp_limb_t t = (cur >> shr) | (nx << rshr);
48 cur = nx;
49 mp_limb_t a = numa[l - w + i];
50
51 __uint128_t s = (__uint128_t)a + t + ca;
52 __uint128_t d = (__uint128_t)a + ~t + k;
53 numc[l - w + i] = (mp_limb_t)s;
54 numa[l - w + i] = (mp_limb_t)d;
55 ca = (mp_limb_t)(s >> 64);
56 k = (mp_limb_t)(d >> 64);
57 }
58 bcy = ca;
59 acy = 1 - k;
60 }
61
62 // 扫描2: numc[i] = numa[i] - t', numa[i] += t'(和→numa 即 acy, 差→numc 即 bcy)
63 cur = numb[w];
64 {
65 mp_limb_t ca = acy, k = 1 - bcy;
66 for (mp_size_t i = 0; i < l - w; i++) {
67 mp_limb_t nx = (i + 1 < l - w) ? numb[w + i + 1] : numb[l]; // 顶 limb 的 hi 来自 numb[l]
68 mp_limb_t t = (cur >> shr) | (nx << rshr);
69 cur = nx;
70 mp_limb_t a = numa[i];
71
72 __uint128_t s = (__uint128_t)a + t + ca;
73 __uint128_t d = (__uint128_t)a + ~t + k;
74 numa[i] = (mp_limb_t)s;
75 numc[i] = (mp_limb_t)d;
76 ca = (mp_limb_t)(s >> 64);
77 k = (mp_limb_t)(d >> 64);
78 }
79 acy = ca;
80 bcy = 1 - k;
81 }
82
83 return (acy << 1) | bcy;
84}
85
87 mp_size_t shl, mp_limb_t* out) {
88 const unsigned rshl = LIMB_BITS - shl;
89 mp_limb_t ca = 0, b = 0, prev = 0;
90
91 // 扫描1: numc[w+j] = (numa[j]-numb[j])<<shl 接链, numa[j] += numb[j]
92 for (mp_size_t j = 0; j < l - w; j++) {
93 mp_limb_t x = numa[j], y = numb[j];
94
95 mp_limb_t u = x - y;
96 mp_limb_t d = u - b;
97 b = (u > x) + (d > u);
98
99 __uint128_t s = (__uint128_t)x + y + ca;
100 numa[j] = (mp_limb_t)s;
101 ca = (mp_limb_t)(s >> 64);
102
103 numc[w + j] = (d << shl) | (prev >> rshl);
104 prev = d;
105 }
106
107 // 中点 ch 分解(acyo 链续, scyo/shl 链重置)
108 {
109 mp_limb_t shlcy = prev >> rshl;
110 mp_limb_t adj = b << shl;
111 out[0] = (shlcy >= adj) ? shlcy - adj : 0;
112 out[1] = (shlcy >= adj) ? 0 : adj - shlcy;
113 }
114 b = 0;
115 prev = 0;
116
117 // 扫描2: numc[j] = (numb[l-w+j]-numa[l-w+j])<<shl 接链, numa[l-w+j] += numb[l-w+j]
118 for (mp_size_t j = 0; j < w; j++) {
119 mp_limb_t x = numb[l - w + j], y = numa[l - w + j];
120
121 mp_limb_t u = x - y;
122 mp_limb_t d = u - b;
123 b = (u > x) + (d > u);
124
125 __uint128_t s = (__uint128_t)x + y + ca;
126 numa[l - w + j] = (mp_limb_t)s;
127 ca = (mp_limb_t)(s >> 64);
128
129 numc[j] = (d << shl) | (prev >> rshl);
130 prev = d;
131 }
132
133 out[2] = prev >> rshl;
134 out[3] = (ca << 1) | b;
135}
#define k
void lmmp_fft_bfy_fused_(mp_ptr numc, mp_ptr numa, mp_ptr numb, mp_size_t l, mp_size_t w, mp_size_t shl, mp_limb_t *out)
fft 蝶形数据面融合内核(两段扫描 + 中点 ch 分解,单次调用完成)
mp_limb_t lmmp_ifft_bfy_fused_(mp_ptr numc, mp_ptr numa, mp_ptr numb, mp_size_t l, mp_size_t w, mp_size_t shr)
Copyright (C) 2026 HJimmyK (Jericho Knox)
mp_limb_t * mp_ptr
定义 lmmp.h:96
uint64_t mp_size_t
定义 lmmp.h:93
uint64_t mp_limb_t
定义 lmmp.h:92
#define LIMB_BITS
定义 lmmp.h:102
#define t
#define numb
#define s