Debian Codegraph @

modules/lsp-dsp-lib/include/private/dsp/arch/aarch64/asimd/fastconv/parse.h

raw
1 /*
2 * Copyright (C) 2020 Linux Studio Plugins Project <https://lsp-plug.in/>
3 * (C) 2020 Vladimir Sadovnikov <[email protected]>
4 *
5 * This file is part of lsp-dsp-lib
6 * Created on: 31 мар. 2020 г.
7 *
8 * lsp-dsp-lib is free software: you can redistribute it and/or modify
9 * it under the terms of the GNU Lesser General Public License as published by
10 * the Free Software Foundation, either version 3 of the License, or
11 * any later version.
12 *
13 * lsp-dsp-lib is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public License
19 * along with lsp-dsp-lib. If not, see <https://www.gnu.org/licenses/>.
20 */
21
22 #ifndef PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_PARSE_H_
23 #define PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_PARSE_H_
24
25 #ifndef PRIVATE_DSP_ARCH_AARCH64_ASIMD_IMPL
26 #error "This header should not be included directly"
27 #endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_IMPL */
28
29 namespace lsp
30 {
31 namespace asimd
32 {
33 void fastconv_parse(float *dst, const float *src, size_t rank)
34 {
35 size_t items = 1 << rank; // number of complex numbers per half of block (16 for rank=4)
36 size_t n = items >> 1; // number of butterflies per block (8 for rank=4)
37
38 IF_ARCH_AARCH64(
39 const float *fdw = &XFFT_DW[(rank - 3) << 3];
40 const float *fw = &XFFT_A[(rank - 3) << 4];
41 size_t k, p;
42 float *a, *b;
43 );
44
45 if (n > 4)
46 {
47 // First loop
48 ARCH_AARCH64_ASM
49 (
50 __ASM_EMIT("mov %[a], %[dst]") // a = dst
51 __ASM_EMIT("add %[b], %[a], %[n], LSL #3") // b = &a[n*2]
52 __ASM_EMIT("mov %[k], %[n]")
53 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v8 = wr0, v9 = wr1
54 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v10 = wi0, v11 = wi1
55 __ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, w25 = di
56
57 // 8x butterflies
58 __ASM_EMIT("1:")
59 __ASM_EMIT("ldp q0, q2, [%[src]]") // v0 = ar0, v2 = ar1
60 __ASM_EMIT("eor v1.16b, v1.16b, v1.16b") // v1 = 0
61 __ASM_EMIT("eor v3.16b, v3.16b, v3.16b") // v3 = 0
62 __ASM_EMIT("fmul v5.4s, v0.4s, v18.4s") // v5 = ar0 * wi0
63 __ASM_EMIT("fmul v7.4s, v2.4s, v19.4s") // v7 = ar1 * wi1
64 __ASM_EMIT("fmul v4.4s, v0.4s, v16.4s") // v4 = ar0 * wr0 = br0
65 __ASM_EMIT("fmul v6.4s, v2.4s, v17.4s") // v6 = ar1 * wr1 = br1
66 __ASM_EMIT("fneg v5.4s, v5.4s") // v5 = - ar0 * wi0 = bi0
67 __ASM_EMIT("fneg v7.4s, v7.4s") // v7 = - ar1 * wi1 = bi1
68 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
69 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
70 __ASM_EMIT("add %[src], %[src], #0x20")
71 __ASM_EMIT("stp q4, q5, [%[b], #0x00]")
72 __ASM_EMIT("stp q6, q7, [%[b], #0x20]")
73 __ASM_EMIT("subs %[k], %[k], #8")
74 __ASM_EMIT("add %[a], %[a], #0x40")
75 __ASM_EMIT("add %[b], %[b], #0x40")
76 __ASM_EMIT("b.ls 2f")
77
78 // Rotate angle
79 __ASM_EMIT("fmul v20.4s, v16.4s, v25.4s") // v20.4s = wr0 * di
80 __ASM_EMIT("fmul v21.4s, v17.4s, v25.4s") // v21.4s = wr1 * di
81 __ASM_EMIT("fmul v22.4s, v18.4s, v25.4s") // v22.4s = wi0 * di
82 __ASM_EMIT("fmul v23.4s, v19.4s, v25.4s") // v23.4s = wi1 * di
83 __ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16.4s = wr0 * dr
84 __ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17.4s = wr1 * dr
85 __ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18.4s = wi0 * dr
86 __ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19.4s = wi1 * dr
87 __ASM_EMIT("fsub v16.4s, v16.4s, v22.4s") // v16.4s = wr0*dr - wi0*di
88 __ASM_EMIT("fsub v17.4s, v17.4s, v23.4s") // v17.4s = wr1*dr - wi1*di
89 __ASM_EMIT("fadd v18.4s, v18.4s, v20.4s") // v18.4s = wi0*dr + wr0*di
90 __ASM_EMIT("fadd v19.4s, v19.4s, v21.4s") // v19.4s = wi1*dr + wr1*di
91 __ASM_EMIT("b 1b")
92
93 // Post-process
94 __ASM_EMIT("2:")
95 __ASM_EMIT("sub %[XFFT_A], %[XFFT_A], #0x40") // fw -= 16
96 __ASM_EMIT("sub %[XFFT_DW], %[XFFT_DW], #0x20") // fdw -= 8
97 __ASM_EMIT("lsr %[n], %[n], #1") // n >>= 1
98
99 : [src] "+r" (src),
100 [a] "=&r" (a), [b] "=&r" (b),
101 [XFFT_A] "+r" (fw), [XFFT_DW] "+r" (fdw),
102 [k] "=&r" (k), [n] "+r" (n)
103 : [dst] "r" (dst)
104 : "cc", "memory",
105 "v0", "v1", "v2", "v3",
106 "v4", "v5", "v6", "v7",
107 "v16", "v17", "v18", "v19",
108 "v20", "v21", "v22", "v23",
109 "v24", "v25"
110 );
111
112 // All other loops while n > 4
113 ARCH_AARCH64_ASM
114 (
115 __ASM_EMIT("cmp %[n], #8")
116 __ASM_EMIT("b.lo 8f")
117
118 __ASM_EMIT("1:")
119 __ASM_EMIT("mov %[a], %[dst]") // a = dst
120 __ASM_EMIT("mov %[p], %[items]") // p = items
121
122 __ASM_EMIT("3:")
123 __ASM_EMIT("add %[b], %[a], %[n], LSL #3") // b = &a[n*2]
124 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr0, v17 = wr1
125 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi0, v19 = wi1
126 __ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, v25 = di
127 __ASM_EMIT("mov %[k], %[n]")
128 // 8x butterflies
129 __ASM_EMIT("5:")
130 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar0, v1 = ai0
131 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar1, v3 = ai1
132 __ASM_EMIT("ldp q4, q5, [%[b], #0x00]") // v4 = br0, v5 = bi0
133 __ASM_EMIT("ldp q6, q7, [%[b], #0x20]") // v6 = br1, v7 = bi1
134 __ASM_EMIT("fsub v20.4s, v0.4s, v4.4s") // v20 = ar0 - br0 = cr0
135 __ASM_EMIT("fsub v21.4s, v1.4s, v5.4s") // v21 = ai0 - bi0 = ci0
136 __ASM_EMIT("fsub v22.4s, v2.4s, v6.4s") // v22 = ar1 - br1 = cr1
137 __ASM_EMIT("fsub v23.4s, v3.4s, v7.4s") // v23 = ai1 - bi1 = ci1
138 __ASM_EMIT("fadd v0.4s, v0.4s, v4.4s") // v0 = ar0 + br0 = ar0'
139 __ASM_EMIT("fadd v1.4s, v1.4s, v5.4s") // v1 = ai0 + bi0 = ai0'
140 __ASM_EMIT("fadd v2.4s, v2.4s, v6.4s") // v2 = ar1 + br1 = ar1'
141 __ASM_EMIT("fadd v3.4s, v3.4s, v7.4s") // v3 = ai1 + bi1 = ai1'
142 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
143 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
144 __ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = wr0 * cr0
145 __ASM_EMIT("fmul v2.4s, v17.4s, v22.4s") // v2 = wr1 * cr1
146 __ASM_EMIT("fmul v1.4s, v16.4s, v21.4s") // v1 = wr0 * ci0
147 __ASM_EMIT("fmul v3.4s, v17.4s, v23.4s") // v3 = wr1 * ci1
148 __ASM_EMIT("fmla v0.4s, v18.4s, v21.4s") // v0 = wr0*cr0 + wi0*ci0 = br0'
149 __ASM_EMIT("fmla v2.4s, v19.4s, v23.4s") // v2 = wr1*cr1 + wi1*ci1 = bi0'
150 __ASM_EMIT("fmls v1.4s, v18.4s, v20.4s") // v1 = wr0*ci0 - wi0*cr0 = br1'
151 __ASM_EMIT("fmls v3.4s, v19.4s, v22.4s") // v3 = wr1*ci1 - wi1*cr1 = bi1'
152 __ASM_EMIT("stp q0, q1, [%[b], #0x00]")
153 __ASM_EMIT("stp q2, q3, [%[b], #0x20]")
154 __ASM_EMIT("subs %[k], %[k], #8")
155 __ASM_EMIT("add %[a], %[a], #0x40")
156 __ASM_EMIT("add %[b], %[b], #0x40")
157 __ASM_EMIT("b.ls 6f")
158 // Rotate angle & repeat loop
159 __ASM_EMIT("fmul v20.4s, v16.4s, v25.4s") // v20 = wr0 * di
160 __ASM_EMIT("fmul v21.4s, v17.4s, v25.4s") // v21 = wr1 * di
161 __ASM_EMIT("fmul v22.4s, v18.4s, v25.4s") // v22 = wi0 * di
162 __ASM_EMIT("fmul v23.4s, v19.4s, v25.4s") // v23 = wi1 * di
163 __ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16 = wr0 * dr
164 __ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17 = wr1 * dr
165 __ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18 = wi0 * dr
166 __ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19 = wi1 * dr
167 __ASM_EMIT("fsub v16.4s, v16.4s, v22.4s") // v16 = wr0*dr - wi0*di
168 __ASM_EMIT("fsub v17.4s, v17.4s, v23.4s") // v17 = wr1*dr - wi1*di
169 __ASM_EMIT("fadd v18.4s, v18.4s, v20.4s") // v18 = wi0*dr + wr0*di
170 __ASM_EMIT("fadd v19.4s, v19.4s, v21.4s") // v19 = wi1*dr + wr1*di
171 __ASM_EMIT("b 5b")
172 __ASM_EMIT("6:")
173 __ASM_EMIT("mov %[a], %[b]") // a = b
174 __ASM_EMIT("subs %[p], %[p], %[n], LSL #1") // p -= n*2
175 __ASM_EMIT("bne 3b")
176
177 __ASM_EMIT("lsr %[n], %[n], #1") // n >>= 1
178 __ASM_EMIT("sub %[XFFT_A], %[XFFT_A], #0x40") // fw -= 16
179 __ASM_EMIT("sub %[XFFT_DW], %[XFFT_DW], #0x20") // fdw -= 8
180 __ASM_EMIT("cmp %[n], #8")
181 __ASM_EMIT("b.ge 1b")
182
183 // Loop for n=4
184 __ASM_EMIT("8:")
185 __ASM_EMIT("mov %[a], %[dst]") // a = dst
186 __ASM_EMIT("mov %[p], %[items]") // p = items
187 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr0, v17 = wr1
188 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi0, v19 = wi1
189
190 // 8x butterflies
191 __ASM_EMIT("9:")
192 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar0, v1 = ai0
193 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar1, v3 = ai1
194 __ASM_EMIT("ldp q4, q5, [%[a], #0x40]") // v4 = ar2, v5 = ai2
195 __ASM_EMIT("ldp q6, q7, [%[a], #0x60]") // v6 = ar3, v7 = ai3
196 __ASM_EMIT("fsub v20.4s, v0.4s, v2.4s") // v20 = ar0 - br0 = cr0
197 __ASM_EMIT("fsub v21.4s, v1.4s, v3.4s") // v21 = ai0 - bi0 = ci0
198 __ASM_EMIT("fsub v22.4s, v4.4s, v6.4s") // v22 = ar1 - br1 = cr1
199 __ASM_EMIT("fsub v23.4s, v5.4s, v7.4s") // v23 = ai1 - bi1 = ci1
200 __ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") // v0 = ar0 + br0 = ar0'
201 __ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") // v1 = ai0 + bi0 = ai0'
202 __ASM_EMIT("fadd v4.4s, v4.4s, v6.4s") // v4 = ar1 + br1 = ar1'
203 __ASM_EMIT("fadd v5.4s, v5.4s, v7.4s") // v5 = ai1 + bi1 = ai1'
204 __ASM_EMIT("fmul v2.4s, v16.4s, v20.4s") // v2 = wr0 * cr0
205 __ASM_EMIT("fmul v6.4s, v17.4s, v22.4s") // v6 = wr1 * cr1
206 __ASM_EMIT("fmul v3.4s, v16.4s, v21.4s") // v3 = wr0 * ci0
207 __ASM_EMIT("fmul v7.4s, v17.4s, v23.4s") // v7 = wr1 * ci1
208 __ASM_EMIT("fmla v2.4s, v18.4s, v21.4s") // v2 = wr0*cr0 + wi0*ci0 = br0'
209 __ASM_EMIT("fmla v6.4s, v19.4s, v23.4s") // v6 = wr1*cr1 + wi1*ci1 = br1'
210 __ASM_EMIT("fmls v3.4s, v18.4s, v20.4s") // v3 = wr0*ci0 - wi0*cr0 = bi0'
211 __ASM_EMIT("fmls v7.4s, v19.4s, v22.4s") // v7 = wr1*ci1 - wi1*cr1 = bi1'
212 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
213 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
214 __ASM_EMIT("stp q4, q5, [%[a], #0x40]")
215 __ASM_EMIT("stp q6, q7, [%[a], #0x60]")
216 __ASM_EMIT("subs %[p], %[p], #16") // p -= 16
217 __ASM_EMIT("add %[a], %[a], #0x80")
218 __ASM_EMIT("bne 9b")
219
220 : [a] "=&r" (a), [b] "=&r" (b),
221 [k] "=&r" (k), [p] "=&r" (p),
222 [XFFT_A] "+r" (fw), [XFFT_DW] "+r" (fdw),
223 [n] "+r" (n)
224 : [dst] "r" (dst), [items] "r" (items)
225 : "cc", "memory",
226 "v0", "v1", "v2", "v3",
227 "v4", "v5", "v6", "v7",
228 "v16", "v17", "v18", "v19",
229 "v20", "v21", "v22", "v23",
230 "v24", "v25"
231 );
232 }
233 else
234 {
235 ARCH_AARCH64_ASM
236 (
237 __ASM_EMIT("ldr q0, [%[src]]")
238 __ASM_EMIT("eor v1.16b, v1.16b, v1.16b")
239 __ASM_EMIT("stp q0, q1, [%[dst]]")
240 :
241 : [src] "r" (src), [dst] "r" (dst)
242 : "memory",
243 "v0", "v1"
244 );
245 }
246
247 // Perform small 4x4 post-processing
248 ARCH_AARCH64_ASM(
249 // Loop for n=4
250 __ASM_EMIT("subs %[items], %[items], #8")
251 __ASM_EMIT("ldp q16, q17, [%[MASK]]")
252 __ASM_EMIT("b.lo 2f")
253 // 8x butterflies
254 __ASM_EMIT("1:")
255 __ASM_EMIT("ld4 {v0.4s-v3.4s}, [%[dst]]") // v0 = r0 i0 r4 i4, v1 = r1 i1 r5 i5, v2 = r2 i2 r6 i6, v3 = r3 i3 r7 i7
256 // v0 = r0 i0 r4 i4
257 // v1 = r1 i1 r5 i5
258 // v2 = r2 i2 r6 i6
259 // v3 = r3 i3 r7 i7
260 __ASM_EMIT("fsub v4.4s, v0.4s, v2.4s") // v4 = r0-r2 i0-i2 r4-r6 i4-i6 = r1' i1' r5' i5'
261 __ASM_EMIT("fsub v5.4s, v1.4s, v3.4s") // v5 = r1-r3 i1-i3 r5-r7 i5-i7 = r3' i3' r7' i7'
262 __ASM_EMIT("fadd v6.4s, v0.4s, v2.4s") // v6 = r0+r2 i0+i2 r4+r6 i4+i6 = r0' i0' r4' i4'
263 __ASM_EMIT("fadd v7.4s, v1.4s, v3.4s") // v7 = r1+r3 i1+i3 r5+r7 i5+i7 = r2' i2' r6' i6'
264 __ASM_EMIT("rev64 v5.4s, v5.4s") // v5 = i3' r3' i7' r7'
265 // v4 = r1' i1' r5' i5'
266 // v5 = i3' r3' i7' r7'
267 // v6 = r0' i0' r4' i4'
268 // v7 = r2' i2' r6' i6'
269 __ASM_EMIT("fadd v0.4s, v6.4s, v7.4s") // v0 = r0'+r2' i0'+i2' r4'+r6' i4'+i6' = r0" i0" r4" i4"
270 __ASM_EMIT("fsub v1.4s, v6.4s, v7.4s") // v1 = r0'-r2' i0'-i2' r4'-r6' i4'-i6' = r1" i1" r5" i5"
271 __ASM_EMIT("fadd v2.4s, v4.4s, v5.4s") // v2 = r1'+i3' i1'+r3' r5'+i7' i5'+r7' = r2" i3" r6" i7"
272 __ASM_EMIT("fsub v3.4s, v4.4s, v5.4s") // v3 = r1'-i3' i1'-r3' r5'-i7' i5'-r7' = r3" i2" r7" i6"
273 __ASM_EMIT("mov v6.16b, v2.16b") // v6 = r2" i3" r6" i7"
274 __ASM_EMIT("mov v7.16b, v3.16b") // v7 = r3" i2" r7" i6"
275 __ASM_EMIT("bit v2.16b, v7.16b, v16.16b") // v2 = r2" i2" r6" i6"
276 __ASM_EMIT("bit v3.16b, v6.16b, v17.16b") // v3 = r3" i3" r7" i7"
277 __ASM_EMIT("st4 {v0.4s-v3.4s}, [%[dst]]")
278 __ASM_EMIT("subs %[items], %[items], #8") // items -= 8
279 __ASM_EMIT("add %[dst], %[dst], #0x40")
280 __ASM_EMIT("b.hs 1b")
281 __ASM_EMIT("2:")
282 // 4x butterflies
283 __ASM_EMIT("adds %[items], %[items], #4")
284 __ASM_EMIT("b.lt 4f")
285 __ASM_EMIT("ldp q2, q3, [%[dst]]") // v2 = r0 r1 r2 r3, v3 = i0 i1 i2 i3
286 __ASM_EMIT("zip1 v0.4s, v2.4s, v3.4s") // v0 = r0 i0 r1 i1
287 __ASM_EMIT("zip2 v1.4s, v2.4s, v3.4s") // v1 = r2 i2 r3 i3
288 __ASM_EMIT("fsub v4.4s, v0.4s, v1.4s") // v4 = r0-r2 i0-i2 r1-r3 i1-i3 = r1' i1' r3' i3'
289 __ASM_EMIT("fadd v5.4s, v0.4s, v1.4s") // v5 = r0+r2 i0+i2 r1+r3 i1+i3 = r0' i0' r2' i2'
290 __ASM_EMIT("rev64 v6.4s, v4.4s") // v6 = i1' r1' i3' r3'
291 __ASM_EMIT("ins v6.d[0], v5.d[1]") // v6 = r2' i2' i3' r3'
292 __ASM_EMIT("ins v5.d[1], v4.d[0]") // v5 = r0' i0' r1' i1'
293 __ASM_EMIT("fadd v0.4s, v5.4s, v6.4s") // v0 = r0'+r2' i0'+r2' r1'+i3' i1'+r3' = r0" i0" r2" i3"
294 __ASM_EMIT("fsub v1.4s, v5.4s, v6.4s") // v1 = r0'-r2' i0'-r2' r1'-i3' i1'-r3' = r1" i1" r3" i2"
295 __ASM_EMIT("zip1 v4.4s, v0.4s, v1.4s") // v4 = r0" r1" i0" i1"
296 __ASM_EMIT("zip2 v5.4s, v0.4s, v1.4s") // v5 = r2" r3" i3" i2"
297 __ASM_EMIT("rev64 v6.4s, v5.4s") // v6 = r3" r2" i2" i3"
298 __ASM_EMIT("ins v6.d[0], v4.d[1]") // v6 = i0" i1" i2" i3"
299 __ASM_EMIT("ins v4.d[1], v5.d[0]") // v4 = r0" r1" r2" r3"
300 __ASM_EMIT("stp q4, q6, [%[dst]]")
301 __ASM_EMIT("4:")
302
303 : [dst] "+r" (dst), [items] "+r" (items)
304 : [MASK] "r" (&fastconv_swp_mask[0])
305 : "cc", "memory",
306 "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
307 "v16", "v17"
308 );
309 }
310 }
311 }
312
313 #endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_PARSE_H_ */