Debian Codegraph @

modules/lsp-dsp-lib/include/private/dsp/arch/aarch64/asimd/fastconv/papply.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_PAPPLY_H_
23 #define PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_PAPPLY_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_apply(float *dst, float *tmp, const float *conv, 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, *ptr;
43 );
44
45 //---------------------------------------------------------------------
46 // PARSE PART
47 if (n > 4)
48 {
49 // First loop
50 ARCH_AARCH64_ASM
51 (
52 __ASM_EMIT("mov %[a], %[dst]") // a = dst
53 __ASM_EMIT("add %[b], %[a], %[n], LSL #3") // b = &a[n*2]
54 __ASM_EMIT("mov %[k], %[n]")
55 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v8 = wr0, v9 = wr1
56 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v10 = wi0, v11 = wi1
57 __ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, w25 = di
58
59 // 8x butterflies
60 __ASM_EMIT("1:")
61 __ASM_EMIT("ldp q0, q2, [%[src]]") // v0 = ar0, v2 = ar1
62 __ASM_EMIT("eor v1.16b, v1.16b, v1.16b") // v1 = 0
63 __ASM_EMIT("eor v3.16b, v3.16b, v3.16b") // v3 = 0
64 __ASM_EMIT("fmul v5.4s, v0.4s, v18.4s") // v5 = ar0 * wi0
65 __ASM_EMIT("fmul v7.4s, v2.4s, v19.4s") // v7 = ar1 * wi1
66 __ASM_EMIT("fmul v4.4s, v0.4s, v16.4s") // v4 = ar0 * wr0 = br0
67 __ASM_EMIT("fmul v6.4s, v2.4s, v17.4s") // v6 = ar1 * wr1 = br1
68 __ASM_EMIT("fneg v5.4s, v5.4s") // v5 = - ar0 * wi0 = bi0
69 __ASM_EMIT("fneg v7.4s, v7.4s") // v7 = - ar1 * wi1 = bi1
70 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
71 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
72 __ASM_EMIT("add %[src], %[src], #0x20")
73 __ASM_EMIT("stp q4, q5, [%[b], #0x00]")
74 __ASM_EMIT("stp q6, q7, [%[b], #0x20]")
75 __ASM_EMIT("subs %[k], %[k], #8")
76 __ASM_EMIT("add %[a], %[a], #0x40")
77 __ASM_EMIT("add %[b], %[b], #0x40")
78 __ASM_EMIT("b.ls 2f")
79
80 // Rotate angle
81 __ASM_EMIT("fmul v20.4s, v16.4s, v25.4s") // v20.4s = wr0 * di
82 __ASM_EMIT("fmul v21.4s, v17.4s, v25.4s") // v21.4s = wr1 * di
83 __ASM_EMIT("fmul v22.4s, v18.4s, v25.4s") // v22.4s = wi0 * di
84 __ASM_EMIT("fmul v23.4s, v19.4s, v25.4s") // v23.4s = wi1 * di
85 __ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16.4s = wr0 * dr
86 __ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17.4s = wr1 * dr
87 __ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18.4s = wi0 * dr
88 __ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19.4s = wi1 * dr
89 __ASM_EMIT("fsub v16.4s, v16.4s, v22.4s") // v16.4s = wr0*dr - wi0*di
90 __ASM_EMIT("fsub v17.4s, v17.4s, v23.4s") // v17.4s = wr1*dr - wi1*di
91 __ASM_EMIT("fadd v18.4s, v18.4s, v20.4s") // v18.4s = wi0*dr + wr0*di
92 __ASM_EMIT("fadd v19.4s, v19.4s, v21.4s") // v19.4s = wi1*dr + wr1*di
93 __ASM_EMIT("b 1b")
94
95 // Post-process
96 __ASM_EMIT("2:")
97 __ASM_EMIT("sub %[XFFT_A], %[XFFT_A], #0x40") // fw -= 16
98 __ASM_EMIT("sub %[XFFT_DW], %[XFFT_DW], #0x20") // fdw -= 8
99 __ASM_EMIT("lsr %[n], %[n], #1") // n >>= 1
100
101 : [src] "+r" (src),
102 [a] "=&r" (a), [b] "=&r" (b),
103 [XFFT_A] "+r" (fw), [XFFT_DW] "+r" (fdw),
104 [k] "=&r" (k), [n] "+r" (n)
105 : [dst] "r" (tmp)
106 : "cc", "memory",
107 "v0", "v1", "v2", "v3",
108 "v4", "v5", "v6", "v7",
109 "v16", "v17", "v18", "v19",
110 "v20", "v21", "v22", "v23",
111 "v24", "v25"
112 );
113
114 // All other loops while n > 4
115 ARCH_AARCH64_ASM
116 (
117 __ASM_EMIT("cmp %[n], #8")
118 __ASM_EMIT("b.lo 8f")
119
120 __ASM_EMIT("1:")
121 __ASM_EMIT("mov %[a], %[dst]") // a = dst
122 __ASM_EMIT("mov %[p], %[items]") // p = items
123
124 __ASM_EMIT("3:")
125 __ASM_EMIT("add %[b], %[a], %[n], LSL #3") // b = &a[n*2]
126 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr0, v17 = wr1
127 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi0, v19 = wi1
128 __ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, v25 = di
129 __ASM_EMIT("mov %[k], %[n]")
130 // 8x butterflies
131 __ASM_EMIT("5:")
132 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar0, v1 = ai0
133 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar1, v3 = ai1
134 __ASM_EMIT("ldp q4, q5, [%[b], #0x00]") // v4 = br0, v5 = bi0
135 __ASM_EMIT("ldp q6, q7, [%[b], #0x20]") // v6 = br1, v7 = bi1
136 __ASM_EMIT("fsub v20.4s, v0.4s, v4.4s") // v20 = ar0 - br0 = cr0
137 __ASM_EMIT("fsub v21.4s, v1.4s, v5.4s") // v21 = ai0 - bi0 = ci0
138 __ASM_EMIT("fsub v22.4s, v2.4s, v6.4s") // v22 = ar1 - br1 = cr1
139 __ASM_EMIT("fsub v23.4s, v3.4s, v7.4s") // v23 = ai1 - bi1 = ci1
140 __ASM_EMIT("fadd v0.4s, v0.4s, v4.4s") // v0 = ar0 + br0 = ar0'
141 __ASM_EMIT("fadd v1.4s, v1.4s, v5.4s") // v1 = ai0 + bi0 = ai0'
142 __ASM_EMIT("fadd v2.4s, v2.4s, v6.4s") // v2 = ar1 + br1 = ar1'
143 __ASM_EMIT("fadd v3.4s, v3.4s, v7.4s") // v3 = ai1 + bi1 = ai1'
144 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
145 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
146 __ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = wr0 * cr0
147 __ASM_EMIT("fmul v2.4s, v17.4s, v22.4s") // v2 = wr1 * cr1
148 __ASM_EMIT("fmul v1.4s, v16.4s, v21.4s") // v1 = wr0 * ci0
149 __ASM_EMIT("fmul v3.4s, v17.4s, v23.4s") // v3 = wr1 * ci1
150 __ASM_EMIT("fmla v0.4s, v18.4s, v21.4s") // v0 = wr0*cr0 + wi0*ci0 = br0'
151 __ASM_EMIT("fmla v2.4s, v19.4s, v23.4s") // v2 = wr1*cr1 + wi1*ci1 = bi0'
152 __ASM_EMIT("fmls v1.4s, v18.4s, v20.4s") // v1 = wr0*ci0 - wi0*cr0 = br1'
153 __ASM_EMIT("fmls v3.4s, v19.4s, v22.4s") // v3 = wr1*ci1 - wi1*cr1 = bi1'
154 __ASM_EMIT("stp q0, q1, [%[b], #0x00]")
155 __ASM_EMIT("stp q2, q3, [%[b], #0x20]")
156 __ASM_EMIT("subs %[k], %[k], #8")
157 __ASM_EMIT("add %[a], %[a], #0x40")
158 __ASM_EMIT("add %[b], %[b], #0x40")
159 __ASM_EMIT("b.ls 6f")
160 // Rotate angle & repeat loop
161 __ASM_EMIT("fmul v20.4s, v16.4s, v25.4s") // v20 = wr0 * di
162 __ASM_EMIT("fmul v21.4s, v17.4s, v25.4s") // v21 = wr1 * di
163 __ASM_EMIT("fmul v22.4s, v18.4s, v25.4s") // v22 = wi0 * di
164 __ASM_EMIT("fmul v23.4s, v19.4s, v25.4s") // v23 = wi1 * di
165 __ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16 = wr0 * dr
166 __ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17 = wr1 * dr
167 __ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18 = wi0 * dr
168 __ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19 = wi1 * dr
169 __ASM_EMIT("fsub v16.4s, v16.4s, v22.4s") // v16 = wr0*dr - wi0*di
170 __ASM_EMIT("fsub v17.4s, v17.4s, v23.4s") // v17 = wr1*dr - wi1*di
171 __ASM_EMIT("fadd v18.4s, v18.4s, v20.4s") // v18 = wi0*dr + wr0*di
172 __ASM_EMIT("fadd v19.4s, v19.4s, v21.4s") // v19 = wi1*dr + wr1*di
173 __ASM_EMIT("b 5b")
174 __ASM_EMIT("6:")
175 __ASM_EMIT("mov %[a], %[b]") // a = b
176 __ASM_EMIT("subs %[p], %[p], %[n], LSL #1") // p -= n*2
177 __ASM_EMIT("bne 3b")
178
179 __ASM_EMIT("lsr %[n], %[n], #1") // n >>= 1
180 __ASM_EMIT("sub %[XFFT_A], %[XFFT_A], #0x40") // fw -= 16
181 __ASM_EMIT("sub %[XFFT_DW], %[XFFT_DW], #0x20") // fdw -= 8
182 __ASM_EMIT("cmp %[n], #8")
183 __ASM_EMIT("b.ge 1b")
184
185 // Loop for n=4
186 __ASM_EMIT("8:")
187 __ASM_EMIT("mov %[a], %[dst]") // a = dst
188 __ASM_EMIT("mov %[p], %[items]") // p = items
189 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr0, v17 = wr1
190 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi0, v19 = wi1
191
192 // 8x butterflies
193 __ASM_EMIT("9:")
194 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar0, v1 = ai0
195 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar1, v3 = ai1
196 __ASM_EMIT("ldp q4, q5, [%[a], #0x40]") // v4 = ar2, v5 = ai2
197 __ASM_EMIT("ldp q6, q7, [%[a], #0x60]") // v6 = ar3, v7 = ai3
198 __ASM_EMIT("fsub v20.4s, v0.4s, v2.4s") // v20 = ar0 - br0 = cr0
199 __ASM_EMIT("fsub v21.4s, v1.4s, v3.4s") // v21 = ai0 - bi0 = ci0
200 __ASM_EMIT("fsub v22.4s, v4.4s, v6.4s") // v22 = ar1 - br1 = cr1
201 __ASM_EMIT("fsub v23.4s, v5.4s, v7.4s") // v23 = ai1 - bi1 = ci1
202 __ASM_EMIT("fadd v0.4s, v0.4s, v2.4s") // v0 = ar0 + br0 = ar0'
203 __ASM_EMIT("fadd v1.4s, v1.4s, v3.4s") // v1 = ai0 + bi0 = ai0'
204 __ASM_EMIT("fadd v4.4s, v4.4s, v6.4s") // v4 = ar1 + br1 = ar1'
205 __ASM_EMIT("fadd v5.4s, v5.4s, v7.4s") // v5 = ai1 + bi1 = ai1'
206 __ASM_EMIT("fmul v2.4s, v16.4s, v20.4s") // v2 = wr0 * cr0
207 __ASM_EMIT("fmul v6.4s, v17.4s, v22.4s") // v6 = wr1 * cr1
208 __ASM_EMIT("fmul v3.4s, v16.4s, v21.4s") // v3 = wr0 * ci0
209 __ASM_EMIT("fmul v7.4s, v17.4s, v23.4s") // v7 = wr1 * ci1
210 __ASM_EMIT("fmla v2.4s, v18.4s, v21.4s") // v2 = wr0*cr0 + wi0*ci0 = br0'
211 __ASM_EMIT("fmla v6.4s, v19.4s, v23.4s") // v6 = wr1*cr1 + wi1*ci1 = br1'
212 __ASM_EMIT("fmls v3.4s, v18.4s, v20.4s") // v3 = wr0*ci0 - wi0*cr0 = bi0'
213 __ASM_EMIT("fmls v7.4s, v19.4s, v22.4s") // v7 = wr1*ci1 - wi1*cr1 = bi1'
214 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
215 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
216 __ASM_EMIT("stp q4, q5, [%[a], #0x40]")
217 __ASM_EMIT("stp q6, q7, [%[a], #0x60]")
218 __ASM_EMIT("subs %[p], %[p], #16") // p -= 16
219 __ASM_EMIT("add %[a], %[a], #0x80")
220 __ASM_EMIT("bne 9b")
221
222 : [a] "=&r" (a), [b] "=&r" (b),
223 [k] "=&r" (k), [p] "=&r" (p),
224 [XFFT_A] "+r" (fw), [XFFT_DW] "+r" (fdw),
225 [n] "+r" (n)
226 : [dst] "r" (tmp), [items] "r" (items)
227 : "cc", "memory",
228 "v0", "v1", "v2", "v3",
229 "v4", "v5", "v6", "v7",
230 "v16", "v17", "v18", "v19",
231 "v20", "v21", "v22", "v23",
232 "v24", "v25"
233 );
234 }
235 else
236 {
237 ARCH_AARCH64_ASM
238 (
239 __ASM_EMIT("ldr q0, [%[src]]")
240 __ASM_EMIT("eor v1.16b, v1.16b, v1.16b")
241 __ASM_EMIT("stp q0, q1, [%[dst]]")
242 :
243 : [src] "r" (src), [dst] "r" (tmp)
244 : "memory",
245 "v0", "v1"
246 );
247 }
248
249 //---------------------------------------------------------------------
250 // MIDDLE PART
251 ARCH_AARCH64_ASM(
252 __ASM_EMIT("subs %[n], %[items], #8")
253 __ASM_EMIT("ldp q24, q25, [%[MASK]]")
254 __ASM_EMIT("mov %[ptr], %[tmp]")
255 __ASM_EMIT("b.lo 2f")
256 // Load data and perform complex multiplication
257 __ASM_EMIT("1:")
258 __ASM_EMIT("ld4 {v0.4s-v3.4s}, [%[ptr]]") // v0 = r0 i0 r4 i4, v1 = r1 i1 r5 i5, v2 = r2 i2 r6 i6, v3 = r3 i3 r7 i7
259 // v0 = r0 i0 r4 i4
260 // v1 = r1 i1 r5 i5
261 // v2 = r2 i2 r6 i6
262 // v3 = r3 i3 r7 i7
263 __ASM_EMIT("fsub v4.4s, v0.4s, v2.4s") // v4 = r0-r2 i0-i2 r4-r6 i4-i6 = r1' i1' r5' i5'
264 __ASM_EMIT("fsub v5.4s, v1.4s, v3.4s") // v5 = r1-r3 i1-i3 r5-r7 i5-i7 = r3' i3' r7' i7'
265 __ASM_EMIT("fadd v6.4s, v0.4s, v2.4s") // v6 = r0+r2 i0+i2 r4+r6 i4+i6 = r0' i0' r4' i4'
266 __ASM_EMIT("fadd v7.4s, v1.4s, v3.4s") // v7 = r1+r3 i1+i3 r5+r7 i5+i7 = r2' i2' r6' i6'
267 __ASM_EMIT("rev64 v5.4s, v5.4s") // v5 = i3' r3' i7' r7'
268 // v4 = r1' i1' r5' i5'
269 // v5 = i3' r3' i7' r7'
270 // v6 = r0' i0' r4' i4'
271 // v7 = r2' i2' r6' i6'
272 __ASM_EMIT("fadd v0.4s, v6.4s, v7.4s") // v0 = r0'+r2' i0'+i2' r4'+r6' i4'+i6' = r0" i0" r4" i4"
273 __ASM_EMIT("fsub v1.4s, v6.4s, v7.4s") // v1 = r0'-r2' i0'-i2' r4'-r6' i4'-i6' = r1" i1" r5" i5"
274 __ASM_EMIT("fadd v2.4s, v4.4s, v5.4s") // v2 = r1'+i3' i1'+r3' r5'+i7' i5'+r7' = r2" i3" r6" i7"
275 __ASM_EMIT("fsub v3.4s, v4.4s, v5.4s") // v3 = r1'-i3' i1'-r3' r5'-i7' i5'-r7' = r3" i2" r7" i6"
276 __ASM_EMIT("mov v6.16b, v2.16b") // v6 = r2" i3" r6" i7"
277 __ASM_EMIT("mov v7.16b, v3.16b") // v7 = r3" i2" r7" i6"
278 __ASM_EMIT("bit v2.16b, v7.16b, v24.16b") // v2 = r2" i2" r6" i6"
279 __ASM_EMIT("bit v3.16b, v6.16b, v25.16b") // v3 = r3" i3" r7" i7"
280 // v0 = r0" i0" r4" i4"
281 // v1 = r1" i1" r5" i5"
282 // v2 = r2" i2" r6" i6"
283 // v3 = r3" i3" r7" i7"
284 __ASM_EMIT("trn1 v4.4s, v0.4s, v1.4s") // v4 = r0" r1" r4" r5"
285 __ASM_EMIT("trn1 v5.4s, v2.4s, v3.4s") // v5 = r2" r3" r6" r7"
286 __ASM_EMIT("trn2 v6.4s, v0.4s, v1.4s") // v6 = i0" i1" i4" i5"
287 __ASM_EMIT("trn2 v7.4s, v2.4s, v3.4s") // v7 = i2" i3" i6" i7"
288 __ASM_EMIT("trn1 v16.2d, v4.2d, v5.2d") // v16 = r0" r1" r2" r3" = ar0
289 __ASM_EMIT("trn1 v17.2d, v6.2d, v7.2d") // v17 = i0" i1" i2" i3" = ai0
290 __ASM_EMIT("trn2 v18.2d, v4.2d, v5.2d") // v18 = r4" r5" r6" r7" = ar1
291 __ASM_EMIT("trn2 v19.2d, v6.2d, v7.2d") // v19 = i4" i5" i6" i7" = ai1
292 // Apply convolution
293 __ASM_EMIT("ldp q20, q21, [%[c], #0x00]") // v20 = br0, v21 = bi0
294 __ASM_EMIT("ldp q22, q23, [%[c], #0x20]") // v22 = br1, v23 = bi1
295 __ASM_EMIT("fmul v0.4s, v16.4s, v20.4s") // v0 = ar0 * br0
296 __ASM_EMIT("fmul v2.4s, v18.4s, v22.4s") // v2 = ar1 * br1
297 __ASM_EMIT("fmul v1.4s, v16.4s, v21.4s") // v1 = ar0 * bi0
298 __ASM_EMIT("fmul v3.4s, v18.4s, v23.4s") // v3 = ar1 * bi1
299 __ASM_EMIT("fmls v0.4s, v17.4s, v21.4s") // v0 = ar0*br0 - ai0*bi0
300 __ASM_EMIT("fmls v2.4s, v19.4s, v23.4s") // v2 = ar1*br1 - ai1*bi1
301 __ASM_EMIT("fmla v1.4s, v17.4s, v20.4s") // v1 = ar0*bi0 + ai0*br0
302 __ASM_EMIT("fmla v3.4s, v19.4s, v22.4s") // v3 = ar1*bi1 + ai1*br1
303 // v0 = r0 r1 r2 r3
304 // v1 = i0 i1 i2 i3
305 // v2 = r4 r5 r6 r7
306 // v3 = i4 i5 i6 i7
307 __ASM_EMIT("trn1 v4.4s, v0.4s, v1.4s") // v4 = r0 i0 r2 i2
308 __ASM_EMIT("trn1 v5.4s, v2.4s, v3.4s") // v5 = r4 i4 r6 i6
309 __ASM_EMIT("trn2 v6.4s, v0.4s, v1.4s") // v6 = r1 i1 r3 i3
310 __ASM_EMIT("trn2 v7.4s, v2.4s, v3.4s") // v7 = r5 i5 r7 i7
311 __ASM_EMIT("trn1 v0.2d, v4.2d, v5.2d") // v0 = r0 i0 r4 i4
312 __ASM_EMIT("trn1 v1.2d, v6.2d, v7.2d") // v1 = r1 i1 r5 i5
313 __ASM_EMIT("trn2 v2.2d, v4.2d, v5.2d") // v2 = r2 i2 r6 i6
314 __ASM_EMIT("trn2 v3.2d, v6.2d, v7.2d") // v3 = r3 i3 r7 i7
315 // v0 = r0 i0 r4 i4
316 // v1 = r1 i1 r5 i5
317 // v2 = r2 i2 r6 i6
318 // v3 = r3 i3 r7 i7
319 __ASM_EMIT("fsub v6.4s, v0.4s, v1.4s") // v6 = r0-r1 i0-i1 r4-r5 i4-i5 = r1' i1' r5' i5'
320 __ASM_EMIT("fsub v7.4s, v2.4s, v3.4s") // v7 = r2-r3 i2-i3 r6-r7 i6-i7 = r3' i3' r7' i7'
321 __ASM_EMIT("fadd v4.4s, v0.4s, v1.4s") // v4 = r0+r1 i0+i1 r4+r5 i4+i5 = r0' i0' r4' i4'
322 __ASM_EMIT("fadd v5.4s, v2.4s, v3.4s") // v5 = r2+r3 i2+i3 r6+r7 i6+i7 = r2' i2' r6' i6'
323 __ASM_EMIT("rev64 v7.4s, v7.4s") // v7 = i3' r3' i7' r7'
324 // v4 = r0' i0' r4' i4'
325 // v5 = r2' i2' r6' i6'
326 // v6 = r1' i1' r5' i5'
327 // v7 = i3' r3' i7' r7'
328 __ASM_EMIT("fadd v0.4s, v4.4s, v5.4s") // v0 = r0'+r2' i0'+i2' r4'+r6' i4'+i6' = r0" i0" r4" i4"
329 __ASM_EMIT("fsub v1.4s, v6.4s, v7.4s") // v2 = r1'-i3' i1'-r3' r5'-i7' i5'-r7' = r1" i3" r5" i7"
330 __ASM_EMIT("fsub v2.4s, v4.4s, v5.4s") // v1 = r0'-r2' i0'-i2' r4'-r6' i4'-i6' = r2" i2" r6" i6"
331 __ASM_EMIT("fadd v3.4s, v6.4s, v7.4s") // v3 = r1'+i3' i1'+r3' r5'+i7' i5'+r7' = r3" i1" r7" i5"
332 __ASM_EMIT("mov v6.16b, v1.16b") // v6 = r1" i3" r5" i7"
333 __ASM_EMIT("mov v7.16b, v3.16b") // v7 = r3" i1" r7" i5"
334 __ASM_EMIT("bit v1.16b, v7.16b, v24.16b") // v1 = r1" i1" r5" i5"
335 __ASM_EMIT("bit v3.16b, v6.16b, v25.16b") // v3 = r3" i3" r7" i7"
336 __ASM_EMIT("st4 {v0.4s-v3.4s}, [%[ptr]]")
337 __ASM_EMIT("subs %[n], %[n], #8") // n -= 8
338 __ASM_EMIT("add %[c], %[c], #0x40")
339 __ASM_EMIT("add %[ptr], %[ptr], #0x40")
340 __ASM_EMIT("b.hs 1b")
341 __ASM_EMIT("2:")
342
343 : [c] "+r" (conv), [ptr] "=&r" (ptr),
344 [n] "=&r" (n)
345 : [tmp] "r" (tmp), [items] "r" (items),
346 [MASK] "r" (&fastconv_swp_mask[0])
347 : "cc", "memory",
348 "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
349 "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
350 "v24", "v25"
351 );
352
353 //---------------------------------------------------------------------
354 // RESTORE PART
355 n = 8;
356 fw = XFFT_A;
357
358 // Perform butterflies without need of vector rotation
359 if (n < items)
360 {
361 ARCH_AARCH64_ASM(
362 __ASM_EMIT("mov %[a], %[tmp]")
363 __ASM_EMIT("subs %[k], %[items], #16")
364 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr1, v17 = wr2
365 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi1, v19 = wi2
366 __ASM_EMIT("b.lo 2f")
367 // 8x butterflies
368 // Calculate complex c = w * b
369 __ASM_EMIT("1:")
370 // Calc cr and ci
371 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1
372 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2
373 __ASM_EMIT("ldp q4, q5, [%[a], #0x40]") // v4 = ar3, v5 = ai3
374 __ASM_EMIT("ldp q6, q7, [%[a], #0x60]") // v6 = ar4, v7 = ai4
375 __ASM_EMIT("fmul v20.4s, v16.4s, v2.4s") // v20 = wr1 * br1
376 __ASM_EMIT("fmul v21.4s, v17.4s, v6.4s") // v21 = wr2 * br2
377 __ASM_EMIT("fmul v22.4s, v16.4s, v3.4s") // v22 = wr1 * bi1
378 __ASM_EMIT("fmul v23.4s, v17.4s, v7.4s") // v23 = wr2 * bi2
379 __ASM_EMIT("fmls v20.4s, v18.4s, v3.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1
380 __ASM_EMIT("fmls v21.4s, v19.4s, v7.4s") // v21 = wr2 * br2 - wi2 * bi2 = cr2
381 __ASM_EMIT("fmla v22.4s, v18.4s, v2.4s") // v22 = wr1 * bi1 + wi1 * br1 = ci1
382 __ASM_EMIT("fmla v23.4s, v19.4s, v6.4s") // v23 = wr1 * bi1 + wi2 * br2 = ci2
383 // Apply butterfly
384 __ASM_EMIT("fsub v2.4s, v0.4s, v20.4s") // v2 = ar1 - cr1
385 __ASM_EMIT("fsub v3.4s, v1.4s, v22.4s") // v3 = ai1 - ci1
386 __ASM_EMIT("fsub v6.4s, v4.4s, v21.4s") // v6 = ar2 - cr2
387 __ASM_EMIT("fsub v7.4s, v5.4s, v23.4s") // v7 = ai2 - ci2
388 __ASM_EMIT("fadd v0.4s, v0.4s, v20.4s") // v0 = ar1 + cr1
389 __ASM_EMIT("fadd v1.4s, v1.4s, v22.4s") // v1 = ai1 + ci1
390 __ASM_EMIT("fadd v4.4s, v4.4s, v21.4s") // v4 = ar2 + cr2
391 __ASM_EMIT("fadd v5.4s, v5.4s, v23.4s") // v5 = ai2 + ci2
392 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
393 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
394 __ASM_EMIT("stp q4, q5, [%[a], #0x40]")
395 __ASM_EMIT("stp q6, q7, [%[a], #0x60]")
396 __ASM_EMIT("subs %[k], %[k], #16")
397 __ASM_EMIT("add %[a], %[a], #0x80")
398 __ASM_EMIT("b.ge 1b")
399
400 __ASM_EMIT("2:")
401 __ASM_EMIT("adds %[k], %[k], #8")
402 __ASM_EMIT("b.lo 4f")
403 // 4x butterflies
404 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1
405 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2
406 __ASM_EMIT("fmul v20.4s, v16.4s, v2.4s") // v20 = wr1 * br1
407 __ASM_EMIT("fmul v22.4s, v16.4s, v3.4s") // v22 = wr1 * bi1
408 __ASM_EMIT("fmls v20.4s, v18.4s, v3.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1
409 __ASM_EMIT("fmla v22.4s, v18.4s, v2.4s") // v22 = wr1 * bi1 + wi1 * br1 = ci1
410 // Apply butterfly
411 __ASM_EMIT("fsub v2.4s, v0.4s, v20.4s") // v2 = ar1 - cr1
412 __ASM_EMIT("fsub v3.4s, v1.4s, v22.4s") // v3 = ai1 - ci1
413 __ASM_EMIT("fadd v0.4s, v0.4s, v20.4s") // v0 = ar1 + cr1
414 __ASM_EMIT("fadd v1.4s, v1.4s, v22.4s") // v1 = ai1 + ci1
415 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
416 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
417 __ASM_EMIT("4:")
418 __ASM_EMIT("add %[XFFT_A], %[XFFT_A], #0x40")
419
420 : [a] "=&r" (a), [k] "=&r" (k),
421 [XFFT_A] "+r" (fw)
422 : [tmp] "r" (tmp), [items] "r" (items)
423 : "cc", "memory",
424 "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
425 "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
426 );
427
428 n <<= 1;
429 }
430
431 fdw = &XFFT_DW[8];
432
433 // Perform loops except last one
434 while (n < items)
435 {
436 ARCH_AARCH64_ASM(
437 __ASM_EMIT("mov %[k], %[items]")
438 __ASM_EMIT("mov %[a], %[tmp]")
439 __ASM_EMIT("1:")
440 // Initialize sub-loop
441 __ASM_EMIT("mov %[p], %[n]")
442 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr1, v17 = wr2
443 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi1, v19 = wi2
444 __ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, v25 = di
445 __ASM_EMIT("add %[b], %[a], %[n], LSL #2") // b = &a[n*4]
446 // 8x butterflies
447 __ASM_EMIT("3:")
448 // Calc cr and ci
449 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1
450 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2
451 __ASM_EMIT("ldp q4, q5, [%[b], #0x00]") // v4 = br1, v5 = bi1
452 __ASM_EMIT("ldp q6, q7, [%[b], #0x20]") // v6 = br2, v7 = bi2
453 __ASM_EMIT("fmul v20.4s, v16.4s, v4.4s") // v20 = wr1 * br1
454 __ASM_EMIT("fmul v21.4s, v17.4s, v6.4s") // v21 = wr2 * br2
455 __ASM_EMIT("fmul v22.4s, v16.4s, v5.4s") // v22 = wr1 * bi1
456 __ASM_EMIT("fmul v23.4s, v17.4s, v7.4s") // v23 = wr2 * bi2
457 __ASM_EMIT("fmls v20.4s, v18.4s, v5.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1
458 __ASM_EMIT("fmls v21.4s, v19.4s, v7.4s") // v21 = wr2 * br2 - wi2 * bi2 = cr2
459 __ASM_EMIT("fmla v22.4s, v18.4s, v4.4s") // v22 = wr1 * bi1 + wi1 * br1 = ci1
460 __ASM_EMIT("fmla v23.4s, v19.4s, v6.4s") // v23 = wr1 * bi1 + wi2 * br2 = ci2
461 // Apply butterfly
462 __ASM_EMIT("fsub v4.4s, v0.4s, v20.4s") // v4 = ar1 - cr1
463 __ASM_EMIT("fsub v5.4s, v1.4s, v22.4s") // v5 = ai1 - ci1
464 __ASM_EMIT("fsub v6.4s, v2.4s, v21.4s") // v6 = ar2 - cr2
465 __ASM_EMIT("fsub v7.4s, v3.4s, v23.4s") // v7 = ai2 - ci2
466 __ASM_EMIT("fadd v0.4s, v0.4s, v20.4s") // v0 = ar1 + cr1
467 __ASM_EMIT("fadd v1.4s, v1.4s, v22.4s") // v1 = ai1 + ci1
468 __ASM_EMIT("fadd v2.4s, v2.4s, v21.4s") // v2 = ar2 + cr2
469 __ASM_EMIT("fadd v3.4s, v3.4s, v23.4s") // v3 = ai2 + ci2
470 __ASM_EMIT("stp q0, q1, [%[a], #0x00]")
471 __ASM_EMIT("stp q2, q3, [%[a], #0x20]")
472 __ASM_EMIT("stp q4, q5, [%[b], #0x00]")
473 __ASM_EMIT("stp q6, q7, [%[b], #0x20]")
474 __ASM_EMIT("subs %[p], %[p], #16")
475 __ASM_EMIT("add %[a], %[a], #0x40")
476 __ASM_EMIT("add %[b], %[b], #0x40")
477 __ASM_EMIT("b.eq 4f")
478 // Prepare next loop: rotate angle
479 __ASM_EMIT("fmul v20.4s, v16.4s, v25.4s") // v20 = wr1 * di
480 __ASM_EMIT("fmul v21.4s, v17.4s, v25.4s") // v21 = wr2 * di
481 __ASM_EMIT("fmul v22.4s, v18.4s, v25.4s") // v22 = wi1 * di
482 __ASM_EMIT("fmul v23.4s, v19.4s, v25.4s") // v23 = wi2 * di
483 __ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16 = wr1 * dr
484 __ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17 = wr2 * dr
485 __ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18 = wi1 * dr
486 __ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19 = wi2 * dr
487 __ASM_EMIT("fsub v16.4s, v16.4s, v22.4s") // v16 = wr1*dr - wi1*di
488 __ASM_EMIT("fsub v17.4s, v17.4s, v23.4s") // v17 = wr2*dr - wi2*di
489 __ASM_EMIT("fadd v18.4s, v18.4s, v20.4s") // v18 = wi1*dr + wr1*di
490 __ASM_EMIT("fadd v19.4s, v19.4s, v21.4s") // v19 = wi2*dr + wr2*di
491 __ASM_EMIT("b 3b")
492 __ASM_EMIT("4:")
493 __ASM_EMIT("subs %[k], %[k], %[n]")
494 __ASM_EMIT("mov %[a], %[b]")
495 __ASM_EMIT("b.ne 1b")
496
497 : [a] "=&r" (a), [b] "=&r" (b),
498 [p] "=&r" (p), [k] "=&r" (k)
499 : [tmp] "r" (tmp), [items] "r" (items), [n] "r" (n),
500 [XFFT_A] "r" (fw), [XFFT_DW] "r" (fdw)
501 : "cc", "memory",
502 "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
503 "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
504 "v24", "v25"
505 );
506
507 n <<= 1;
508 fw += 16;
509 fdw += 8;
510 }
511
512 // Last one large loop
513 IF_ARCH_AARCH64(float fftk = 1.0f / items);
514
515 ARCH_AARCH64_ASM(
516 // Initialize sub-loop
517 __ASM_EMIT("add %[b], %[a], %[n], LSL #2") // b = &a[n*4]
518 __ASM_EMIT("ld1r {v23.4s}, [%[fftk]]") // v23 = 1 / items
519 __ASM_EMIT("add %[db], %[da], %[n], LSL #1") // db = &da[n*2]
520 __ASM_EMIT("ldp q16, q17, [%[XFFT_A], #0x00]") // v16 = wr1, v17 = wr2
521 __ASM_EMIT("ldp q18, q19, [%[XFFT_A], #0x20]") // v18 = wi1, v19 = wi2
522 __ASM_EMIT("ldp q24, q25, [%[XFFT_DW]]") // v24 = dr, v25 = di
523 __ASM_EMIT("mov v22.16b, v23.16b") // v22 = 1 / items
524 // 8x butterflies
525 __ASM_EMIT("1:")
526 __ASM_EMIT("ldp q0, q1, [%[a], #0x00]") // v0 = ar1, v1 = ai1
527 __ASM_EMIT("ldp q2, q3, [%[a], #0x20]") // v2 = ar2, v3 = ai2
528 __ASM_EMIT("ldp q4, q5, [%[b], #0x00]") // v4 = br1, v5 = bi1
529 __ASM_EMIT("ldp q6, q7, [%[b], #0x20]") // v6 = br2, v7 = bi2
530 __ASM_EMIT("fmul v20.4s, v16.4s, v4.4s") // v20 = wr1 * br1
531 __ASM_EMIT("fmul v21.4s, v17.4s, v6.4s") // v21 = wr2 * br2
532 __ASM_EMIT("fmls v20.4s, v18.4s, v5.4s") // v20 = wr1 * br1 - wi1 * bi1 = cr1
533 __ASM_EMIT("fmls v21.4s, v19.4s, v7.4s") // v21 = wr2 * br2 - wi2 * bi2 = cr2
534 // Apply butterfly, we do not need imaginary part
535 __ASM_EMIT("fadd v4.4s, v0.4s, v20.4s") // v4 = ar1 + cr1
536 __ASM_EMIT("fadd v5.4s, v2.4s, v21.4s") // v5 = ar2 + cr2
537 __ASM_EMIT("fsub v6.4s, v0.4s, v20.4s") // v6 = ar1 - cr1
538 __ASM_EMIT("fsub v7.4s, v2.4s, v21.4s") // v7 = ar2 - cr2
539 __ASM_EMIT("ldp q0, q1, [%[da]]") // v0 = dar1, v1 = dar2
540 __ASM_EMIT("ldp q2, q3, [%[db]]") // v2 = dbr1, v3 = dbr2
541 __ASM_EMIT("fmla v0.4s, v4.4s, v22.4s") // v4 = dar1 + (ar1 + cr1) / items
542 __ASM_EMIT("fmla v1.4s, v5.4s, v23.4s") // v5 = dar2 + (ar2 + cr2) / items
543 __ASM_EMIT("fmla v2.4s, v6.4s, v22.4s") // v6 = dbr1 + (ar1 - cr1) / items
544 __ASM_EMIT("fmla v3.4s, v7.4s, v23.4s") // v7 = dbr2 + (ar2 - cr2) / items
545 __ASM_EMIT("stp q0, q1, [%[da]]")
546 __ASM_EMIT("stp q2, q3, [%[db]]")
547 __ASM_EMIT("subs %[items], %[items], #16")
548 __ASM_EMIT("add %[a], %[a], #0x40")
549 __ASM_EMIT("add %[b], %[b], #0x40")
550 __ASM_EMIT("add %[da], %[da], #0x20")
551 __ASM_EMIT("add %[db], %[db], #0x20")
552 __ASM_EMIT("b.ls 2f")
553 // Prepare next loop: rotate angle
554 __ASM_EMIT("fmul v2.4s, v16.4s, v25.4s") // v2 = wr1 * di
555 __ASM_EMIT("fmul v3.4s, v17.4s, v25.4s") // v3 = wr2 * di
556 __ASM_EMIT("fmul v4.4s, v18.4s, v25.4s") // v4 = wi1 * di
557 __ASM_EMIT("fmul v5.4s, v19.4s, v25.4s") // v5 = wi2 * di
558 __ASM_EMIT("fmul v16.4s, v16.4s, v24.4s") // v16 = wr1 * dr
559 __ASM_EMIT("fmul v17.4s, v17.4s, v24.4s") // v17 = wr2 * dr
560 __ASM_EMIT("fmul v18.4s, v18.4s, v24.4s") // v18 = wi1 * dr
561 __ASM_EMIT("fmul v19.4s, v19.4s, v24.4s") // v19 = wi2 * dr
562 __ASM_EMIT("fsub v16.4s, v16.4s, v4.4s") // v16 = wr1*dr - wi1*di
563 __ASM_EMIT("fsub v17.4s, v17.4s, v5.4s") // v17 = wr2*dr - wi2*di
564 __ASM_EMIT("fadd v18.4s, v18.4s, v2.4s") // v18 = wi1*dr + wr1*di
565 __ASM_EMIT("fadd v19.4s, v19.4s, v3.4s") // v19 = wi2*dr + wr2*di
566 __ASM_EMIT("b 1b")
567 __ASM_EMIT("2:")
568
569 : [da] "+r" (dst), [db] "=&r" (a),
570 [a] "+r" (tmp), [b] "=&r" (b),
571 [items] "+r" (items)
572 : [n] "r" (n), [fftk] "r" (&fftk),
573 [XFFT_A] "r" (fw), [XFFT_DW] "r" (fdw)
574 : "cc", "memory",
575 "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
576 "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
577 "v24", "v25"
578 );
579 }
580 }
581 }
582
583 #endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_FASTCONV_PAPPLY_H_ */