| 1 | /* |
| 2 | * ARM NEON optimised FFT |
| 3 | * |
| 4 | * Copyright (c) 2009 Mans Rullgard <mans@mansr.com> |
| 5 | * Copyright (c) 2009 Naotoshi Nojiri |
| 6 | * |
| 7 | * This algorithm (though not any of the implementation details) is |
| 8 | * based on libdjbfft by D. J. Bernstein. |
| 9 | * |
| 10 | * This file is part of FFmpeg. |
| 11 | * |
| 12 | * FFmpeg is free software; you can redistribute it and/or |
| 13 | * modify it under the terms of the GNU Lesser General Public |
| 14 | * License as published by the Free Software Foundation; either |
| 15 | * version 2.1 of the License, or (at your option) any later version. |
| 16 | * |
| 17 | * FFmpeg is distributed in the hope that it will be useful, |
| 18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 20 | * Lesser General Public License for more details. |
| 21 | * |
| 22 | * You should have received a copy of the GNU Lesser General Public |
| 23 | * License along with FFmpeg; if not, write to the Free Software |
| 24 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 25 | */ |
| 26 | |
| 27 | #include "libavutil/arm/asm.S" |
| 28 | |
| 29 | #define M_SQRT1_2 0.70710678118654752440 |
| 30 | |
| 31 | |
| 32 | function fft4_neon |
| 33 | vld1.32 {d0-d3}, [r0,:128] |
| 34 | |
| 35 | vext.32 q8, q1, q1, #1 @ i2,r3 d3=i3,r2 |
| 36 | vsub.f32 d6, d0, d1 @ r0-r1,i0-i1 |
| 37 | vsub.f32 d7, d16, d17 @ r3-r2,i2-i3 |
| 38 | vadd.f32 d4, d0, d1 @ r0+r1,i0+i1 |
| 39 | vadd.f32 d5, d2, d3 @ i2+i3,r2+r3 |
| 40 | vadd.f32 d1, d6, d7 |
| 41 | vsub.f32 d3, d6, d7 |
| 42 | vadd.f32 d0, d4, d5 |
| 43 | vsub.f32 d2, d4, d5 |
| 44 | |
| 45 | vst1.32 {d0-d3}, [r0,:128] |
| 46 | |
| 47 | bx lr |
| 48 | endfunc |
| 49 | |
| 50 | function fft8_neon |
| 51 | mov r1, r0 |
| 52 | vld1.32 {d0-d3}, [r1,:128]! |
| 53 | vld1.32 {d16-d19}, [r1,:128] |
| 54 | |
| 55 | movw r2, #0x04f3 @ sqrt(1/2) |
| 56 | movt r2, #0x3f35 |
| 57 | eor r3, r2, #1<<31 |
| 58 | vdup.32 d31, r2 |
| 59 | |
| 60 | vext.32 q11, q1, q1, #1 @ i2,r3,i3,r2 |
| 61 | vadd.f32 d4, d16, d17 @ r4+r5,i4+i5 |
| 62 | vmov d28, r3, r2 |
| 63 | vadd.f32 d5, d18, d19 @ r6+r7,i6+i7 |
| 64 | vsub.f32 d17, d16, d17 @ r4-r5,i4-i5 |
| 65 | vsub.f32 d19, d18, d19 @ r6-r7,i6-i7 |
| 66 | vrev64.32 d29, d28 |
| 67 | vadd.f32 d20, d0, d1 @ r0+r1,i0+i1 |
| 68 | vadd.f32 d21, d2, d3 @ r2+r3,i2+i3 |
| 69 | vmul.f32 d26, d17, d28 @ -a2r*w,a2i*w |
| 70 | vext.32 q3, q2, q2, #1 |
| 71 | vmul.f32 d27, d19, d29 @ a3r*w,-a3i*w |
| 72 | vsub.f32 d23, d22, d23 @ i2-i3,r3-r2 |
| 73 | vsub.f32 d22, d0, d1 @ r0-r1,i0-i1 |
| 74 | vmul.f32 d24, d17, d31 @ a2r*w,a2i*w |
| 75 | vmul.f32 d25, d19, d31 @ a3r*w,a3i*w |
| 76 | vadd.f32 d0, d20, d21 |
| 77 | vsub.f32 d2, d20, d21 |
| 78 | vadd.f32 d1, d22, d23 |
| 79 | vrev64.32 q13, q13 |
| 80 | vsub.f32 d3, d22, d23 |
| 81 | vsub.f32 d6, d6, d7 |
| 82 | vadd.f32 d24, d24, d26 @ a2r+a2i,a2i-a2r t1,t2 |
| 83 | vadd.f32 d25, d25, d27 @ a3r-a3i,a3i+a3r t5,t6 |
| 84 | vadd.f32 d7, d4, d5 |
| 85 | vsub.f32 d18, d2, d6 |
| 86 | vext.32 q13, q12, q12, #1 |
| 87 | vadd.f32 d2, d2, d6 |
| 88 | vsub.f32 d16, d0, d7 |
| 89 | vadd.f32 d5, d25, d24 |
| 90 | vsub.f32 d4, d26, d27 |
| 91 | vadd.f32 d0, d0, d7 |
| 92 | vsub.f32 d17, d1, d5 |
| 93 | vsub.f32 d19, d3, d4 |
| 94 | vadd.f32 d3, d3, d4 |
| 95 | vadd.f32 d1, d1, d5 |
| 96 | |
| 97 | vst1.32 {d16-d19}, [r1,:128] |
| 98 | vst1.32 {d0-d3}, [r0,:128] |
| 99 | |
| 100 | bx lr |
| 101 | endfunc |
| 102 | |
| 103 | function fft16_neon |
| 104 | movrel r1, mppm |
| 105 | vld1.32 {d16-d19}, [r0,:128]! @ q8{r0,i0,r1,i1} q9{r2,i2,r3,i3} |
| 106 | pld [r0, #32] |
| 107 | vld1.32 {d2-d3}, [r1,:128] |
| 108 | vext.32 q13, q9, q9, #1 |
| 109 | vld1.32 {d22-d25}, [r0,:128]! @ q11{r4,i4,r5,i5} q12{r6,i5,r7,i7} |
| 110 | vadd.f32 d4, d16, d17 |
| 111 | vsub.f32 d5, d16, d17 |
| 112 | vadd.f32 d18, d18, d19 |
| 113 | vsub.f32 d19, d26, d27 |
| 114 | |
| 115 | vadd.f32 d20, d22, d23 |
| 116 | vsub.f32 d22, d22, d23 |
| 117 | vsub.f32 d23, d24, d25 |
| 118 | vadd.f32 q8, q2, q9 @ {r0,i0,r1,i1} |
| 119 | vadd.f32 d21, d24, d25 |
| 120 | vmul.f32 d24, d22, d2 |
| 121 | vsub.f32 q9, q2, q9 @ {r2,i2,r3,i3} |
| 122 | vmul.f32 d25, d23, d3 |
| 123 | vuzp.32 d16, d17 @ {r0,r1,i0,i1} |
| 124 | vmul.f32 q1, q11, d2[1] |
| 125 | vuzp.32 d18, d19 @ {r2,r3,i2,i3} |
| 126 | vrev64.32 q12, q12 |
| 127 | vadd.f32 q11, q12, q1 @ {t1a,t2a,t5,t6} |
| 128 | vld1.32 {d24-d27}, [r0,:128]! @ q12{r8,i8,r9,i9} q13{r10,i10,r11,i11} |
| 129 | vzip.32 q10, q11 |
| 130 | vld1.32 {d28-d31}, [r0,:128] @ q14{r12,i12,r13,i13} q15{r14,i14,r15,i15} |
| 131 | vadd.f32 d0, d22, d20 |
| 132 | vadd.f32 d1, d21, d23 |
| 133 | vsub.f32 d2, d21, d23 |
| 134 | vsub.f32 d3, d22, d20 |
| 135 | sub r0, r0, #96 |
| 136 | vext.32 q13, q13, q13, #1 |
| 137 | vsub.f32 q10, q8, q0 @ {r4,r5,i4,i5} |
| 138 | vadd.f32 q8, q8, q0 @ {r0,r1,i0,i1} |
| 139 | vext.32 q15, q15, q15, #1 |
| 140 | vsub.f32 q11, q9, q1 @ {r6,r7,i6,i7} |
| 141 | vswp d25, d26 @ q12{r8,i8,i10,r11} q13{r9,i9,i11,r10} |
| 142 | vadd.f32 q9, q9, q1 @ {r2,r3,i2,i3} |
| 143 | vswp d29, d30 @ q14{r12,i12,i14,r15} q15{r13,i13,i15,r14} |
| 144 | vadd.f32 q0, q12, q13 @ {t1,t2,t5,t6} |
| 145 | vadd.f32 q1, q14, q15 @ {t1a,t2a,t5a,t6a} |
| 146 | movrelx r2, X(ff_cos_16) |
| 147 | vsub.f32 q13, q12, q13 @ {t3,t4,t7,t8} |
| 148 | vrev64.32 d1, d1 |
| 149 | vsub.f32 q15, q14, q15 @ {t3a,t4a,t7a,t8a} |
| 150 | vrev64.32 d3, d3 |
| 151 | movrel r3, pmmp |
| 152 | vswp d1, d26 @ q0{t1,t2,t3,t4} q13{t6,t5,t7,t8} |
| 153 | vswp d3, d30 @ q1{t1a,t2a,t3a,t4a} q15{t6a,t5a,t7a,t8a} |
| 154 | vadd.f32 q12, q0, q13 @ {r8,i8,r9,i9} |
| 155 | vadd.f32 q14, q1, q15 @ {r12,i12,r13,i13} |
| 156 | vld1.32 {d4-d5}, [r2,:64] |
| 157 | vsub.f32 q13, q0, q13 @ {r10,i10,r11,i11} |
| 158 | vsub.f32 q15, q1, q15 @ {r14,i14,r15,i15} |
| 159 | vswp d25, d28 @ q12{r8,i8,r12,i12} q14{r9,i9,r13,i13} |
| 160 | vld1.32 {d6-d7}, [r3,:128] |
| 161 | vrev64.32 q1, q14 |
| 162 | vmul.f32 q14, q14, d4[1] |
| 163 | vmul.f32 q1, q1, q3 |
| 164 | vmla.f32 q14, q1, d5[1] @ {t1a,t2a,t5a,t6a} |
| 165 | vswp d27, d30 @ q13{r10,i10,r14,i14} q15{r11,i11,r15,i15} |
| 166 | vzip.32 q12, q14 |
| 167 | vadd.f32 d0, d28, d24 |
| 168 | vadd.f32 d1, d25, d29 |
| 169 | vsub.f32 d2, d25, d29 |
| 170 | vsub.f32 d3, d28, d24 |
| 171 | vsub.f32 q12, q8, q0 @ {r8,r9,i8,i9} |
| 172 | vadd.f32 q8, q8, q0 @ {r0,r1,i0,i1} |
| 173 | vsub.f32 q14, q10, q1 @ {r12,r13,i12,i13} |
| 174 | mov r1, #32 |
| 175 | vadd.f32 q10, q10, q1 @ {r4,r5,i4,i5} |
| 176 | vrev64.32 q0, q13 |
| 177 | vmul.f32 q13, q13, d5[0] |
| 178 | vrev64.32 q1, q15 |
| 179 | vmul.f32 q15, q15, d5[1] |
| 180 | vst2.32 {d16-d17},[r0,:128], r1 |
| 181 | vmul.f32 q0, q0, q3 |
| 182 | vst2.32 {d20-d21},[r0,:128], r1 |
| 183 | vmul.f32 q1, q1, q3 |
| 184 | vmla.f32 q13, q0, d5[0] @ {t1,t2,t5,t6} |
| 185 | vmla.f32 q15, q1, d4[1] @ {t1a,t2a,t5a,t6a} |
| 186 | vst2.32 {d24-d25},[r0,:128], r1 |
| 187 | vst2.32 {d28-d29},[r0,:128] |
| 188 | vzip.32 q13, q15 |
| 189 | sub r0, r0, #80 |
| 190 | vadd.f32 d0, d30, d26 |
| 191 | vadd.f32 d1, d27, d31 |
| 192 | vsub.f32 d2, d27, d31 |
| 193 | vsub.f32 d3, d30, d26 |
| 194 | vsub.f32 q13, q9, q0 @ {r10,r11,i10,i11} |
| 195 | vadd.f32 q9, q9, q0 @ {r2,r3,i2,i3} |
| 196 | vsub.f32 q15, q11, q1 @ {r14,r15,i14,i15} |
| 197 | vadd.f32 q11, q11, q1 @ {r6,r7,i6,i7} |
| 198 | vst2.32 {d18-d19},[r0,:128], r1 |
| 199 | vst2.32 {d22-d23},[r0,:128], r1 |
| 200 | vst2.32 {d26-d27},[r0,:128], r1 |
| 201 | vst2.32 {d30-d31},[r0,:128] |
| 202 | bx lr |
| 203 | endfunc |
| 204 | |
| 205 | function fft_pass_neon |
| 206 | push {r4-r6,lr} |
| 207 | mov r6, r2 @ n |
| 208 | lsl r5, r2, #3 @ 2 * n * sizeof FFTSample |
| 209 | lsl r4, r2, #4 @ 2 * n * sizeof FFTComplex |
| 210 | lsl r2, r2, #5 @ 4 * n * sizeof FFTComplex |
| 211 | add r3, r2, r4 |
| 212 | add r4, r4, r0 @ &z[o1] |
| 213 | add r2, r2, r0 @ &z[o2] |
| 214 | add r3, r3, r0 @ &z[o3] |
| 215 | vld1.32 {d20-d21},[r2,:128] @ {z[o2],z[o2+1]} |
| 216 | movrel r12, pmmp |
| 217 | vld1.32 {d22-d23},[r3,:128] @ {z[o3],z[o3+1]} |
| 218 | add r5, r5, r1 @ wim |
| 219 | vld1.32 {d6-d7}, [r12,:128] @ pmmp |
| 220 | vswp d21, d22 |
| 221 | vld1.32 {d4}, [r1,:64]! @ {wre[0],wre[1]} |
| 222 | sub r5, r5, #4 @ wim-- |
| 223 | vrev64.32 q1, q11 |
| 224 | vmul.f32 q11, q11, d4[1] |
| 225 | vmul.f32 q1, q1, q3 |
| 226 | vld1.32 {d5[0]}, [r5,:32] @ d5[0] = wim[-1] |
| 227 | vmla.f32 q11, q1, d5[0] @ {t1a,t2a,t5a,t6a} |
| 228 | vld2.32 {d16-d17},[r0,:128] @ {z[0],z[1]} |
| 229 | sub r6, r6, #1 @ n-- |
| 230 | vld2.32 {d18-d19},[r4,:128] @ {z[o1],z[o1+1]} |
| 231 | vzip.32 q10, q11 |
| 232 | vadd.f32 d0, d22, d20 |
| 233 | vadd.f32 d1, d21, d23 |
| 234 | vsub.f32 d2, d21, d23 |
| 235 | vsub.f32 d3, d22, d20 |
| 236 | vsub.f32 q10, q8, q0 |
| 237 | vadd.f32 q8, q8, q0 |
| 238 | vsub.f32 q11, q9, q1 |
| 239 | vadd.f32 q9, q9, q1 |
| 240 | vst2.32 {d20-d21},[r2,:128]! @ {z[o2],z[o2+1]} |
| 241 | vst2.32 {d16-d17},[r0,:128]! @ {z[0],z[1]} |
| 242 | vst2.32 {d22-d23},[r3,:128]! @ {z[o3],z[o3+1]} |
| 243 | vst2.32 {d18-d19},[r4,:128]! @ {z[o1],z[o1+1]} |
| 244 | sub r5, r5, #8 @ wim -= 2 |
| 245 | 1: |
| 246 | vld1.32 {d20-d21},[r2,:128] @ {z[o2],z[o2+1]} |
| 247 | vld1.32 {d22-d23},[r3,:128] @ {z[o3],z[o3+1]} |
| 248 | vswp d21, d22 |
| 249 | vld1.32 {d4}, [r1]! @ {wre[0],wre[1]} |
| 250 | vrev64.32 q0, q10 |
| 251 | vmul.f32 q10, q10, d4[0] |
| 252 | vrev64.32 q1, q11 |
| 253 | vmul.f32 q11, q11, d4[1] |
| 254 | vld1.32 {d5}, [r5] @ {wim[-1],wim[0]} |
| 255 | vmul.f32 q0, q0, q3 |
| 256 | sub r5, r5, #8 @ wim -= 2 |
| 257 | vmul.f32 q1, q1, q3 |
| 258 | vmla.f32 q10, q0, d5[1] @ {t1,t2,t5,t6} |
| 259 | vmla.f32 q11, q1, d5[0] @ {t1a,t2a,t5a,t6a} |
| 260 | vld2.32 {d16-d17},[r0,:128] @ {z[0],z[1]} |
| 261 | subs r6, r6, #1 @ n-- |
| 262 | vld2.32 {d18-d19},[r4,:128] @ {z[o1],z[o1+1]} |
| 263 | vzip.32 q10, q11 |
| 264 | vadd.f32 d0, d22, d20 |
| 265 | vadd.f32 d1, d21, d23 |
| 266 | vsub.f32 d2, d21, d23 |
| 267 | vsub.f32 d3, d22, d20 |
| 268 | vsub.f32 q10, q8, q0 |
| 269 | vadd.f32 q8, q8, q0 |
| 270 | vsub.f32 q11, q9, q1 |
| 271 | vadd.f32 q9, q9, q1 |
| 272 | vst2.32 {d20-d21}, [r2,:128]! @ {z[o2],z[o2+1]} |
| 273 | vst2.32 {d16-d17}, [r0,:128]! @ {z[0],z[1]} |
| 274 | vst2.32 {d22-d23}, [r3,:128]! @ {z[o3],z[o3+1]} |
| 275 | vst2.32 {d18-d19}, [r4,:128]! @ {z[o1],z[o1+1]} |
| 276 | bne 1b |
| 277 | |
| 278 | pop {r4-r6,pc} |
| 279 | endfunc |
| 280 | |
| 281 | .macro def_fft n, n2, n4 |
| 282 | .align 6 |
| 283 | function fft\n\()_neon |
| 284 | push {r4, lr} |
| 285 | mov r4, r0 |
| 286 | bl fft\n2\()_neon |
| 287 | add r0, r4, #\n4*2*8 |
| 288 | bl fft\n4\()_neon |
| 289 | add r0, r4, #\n4*3*8 |
| 290 | bl fft\n4\()_neon |
| 291 | mov r0, r4 |
| 292 | pop {r4, lr} |
| 293 | movrelx r1, X(ff_cos_\n) |
| 294 | mov r2, #\n4/2 |
| 295 | b fft_pass_neon |
| 296 | endfunc |
| 297 | .endm |
| 298 | |
| 299 | def_fft 32, 16, 8 |
| 300 | def_fft 64, 32, 16 |
| 301 | def_fft 128, 64, 32 |
| 302 | def_fft 256, 128, 64 |
| 303 | def_fft 512, 256, 128 |
| 304 | def_fft 1024, 512, 256 |
| 305 | def_fft 2048, 1024, 512 |
| 306 | def_fft 4096, 2048, 1024 |
| 307 | def_fft 8192, 4096, 2048 |
| 308 | def_fft 16384, 8192, 4096 |
| 309 | def_fft 32768, 16384, 8192 |
| 310 | def_fft 65536, 32768, 16384 |
| 311 | |
| 312 | function ff_fft_calc_neon, export=1 |
| 313 | ldr r2, [r0] |
| 314 | sub r2, r2, #2 |
| 315 | movrel r3, fft_tab_neon |
| 316 | ldr r3, [r3, r2, lsl #2] |
| 317 | mov r0, r1 |
| 318 | bx r3 |
| 319 | endfunc |
| 320 | |
| 321 | function ff_fft_permute_neon, export=1 |
| 322 | push {r4,lr} |
| 323 | mov r12, #1 |
| 324 | ldr r2, [r0] @ nbits |
| 325 | ldr r3, [r0, #12] @ tmp_buf |
| 326 | ldr r0, [r0, #8] @ revtab |
| 327 | lsl r12, r12, r2 |
| 328 | mov r2, r12 |
| 329 | 1: |
| 330 | vld1.32 {d0-d1}, [r1,:128]! |
| 331 | ldr r4, [r0], #4 |
| 332 | uxth lr, r4 |
| 333 | uxth r4, r4, ror #16 |
| 334 | add lr, r3, lr, lsl #3 |
| 335 | add r4, r3, r4, lsl #3 |
| 336 | vst1.32 {d0}, [lr,:64] |
| 337 | vst1.32 {d1}, [r4,:64] |
| 338 | subs r12, r12, #2 |
| 339 | bgt 1b |
| 340 | |
| 341 | sub r1, r1, r2, lsl #3 |
| 342 | 1: |
| 343 | vld1.32 {d0-d3}, [r3,:128]! |
| 344 | vst1.32 {d0-d3}, [r1,:128]! |
| 345 | subs r2, r2, #4 |
| 346 | bgt 1b |
| 347 | |
| 348 | pop {r4,pc} |
| 349 | endfunc |
| 350 | |
| 351 | const fft_tab_neon |
| 352 | .word fft4_neon |
| 353 | .word fft8_neon |
| 354 | .word fft16_neon |
| 355 | .word fft32_neon |
| 356 | .word fft64_neon |
| 357 | .word fft128_neon |
| 358 | .word fft256_neon |
| 359 | .word fft512_neon |
| 360 | .word fft1024_neon |
| 361 | .word fft2048_neon |
| 362 | .word fft4096_neon |
| 363 | .word fft8192_neon |
| 364 | .word fft16384_neon |
| 365 | .word fft32768_neon |
| 366 | .word fft65536_neon |
| 367 | endconst |
| 368 | |
| 369 | const pmmp, align=4 |
| 370 | .float +1.0, -1.0, -1.0, +1.0 |
| 371 | endconst |
| 372 | |
| 373 | const mppm, align=4 |
| 374 | .float -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2 |
| 375 | endconst |