| 1 |
/* |
| 2 |
* Copyright © 2018, VideoLAN and dav1d authors |
| 3 |
* Copyright © 2020, Martin Storsjo |
| 4 |
* All rights reserved. |
| 5 |
* |
| 6 |
* Redistribution and use in source and binary forms, with or without |
| 7 |
* modification, are permitted provided that the following conditions are met: |
| 8 |
* |
| 9 |
* 1. Redistributions of source code must retain the above copyright notice, this |
| 10 |
* list of conditions and the following disclaimer. |
| 11 |
* |
| 12 |
* 2. Redistributions in binary form must reproduce the above copyright notice, |
| 13 |
* this list of conditions and the following disclaimer in the documentation |
| 14 |
* and/or other materials provided with the distribution. |
| 15 |
* |
| 16 |
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND |
| 17 |
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED |
| 18 |
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE |
| 19 |
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR |
| 20 |
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES |
| 21 |
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 22 |
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND |
| 23 |
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 24 |
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 25 |
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 26 |
*/ |
| 27 |
|
| 28 |
#include "src/arm/asm.S" |
| 29 |
#include "util.S" |
| 30 |
|
| 31 |
.macro dir_table w, stride |
| 32 |
const directions\w |
| 33 |
.byte -1 * \stride + 1, -2 * \stride + 2 |
| 34 |
.byte 0 * \stride + 1, -1 * \stride + 2 |
| 35 |
.byte 0 * \stride + 1, 0 * \stride + 2 |
| 36 |
.byte 0 * \stride + 1, 1 * \stride + 2 |
| 37 |
.byte 1 * \stride + 1, 2 * \stride + 2 |
| 38 |
.byte 1 * \stride + 0, 2 * \stride + 1 |
| 39 |
.byte 1 * \stride + 0, 2 * \stride + 0 |
| 40 |
.byte 1 * \stride + 0, 2 * \stride - 1 |
| 41 |
// Repeated, to avoid & 7 |
| 42 |
.byte -1 * \stride + 1, -2 * \stride + 2 |
| 43 |
.byte 0 * \stride + 1, -1 * \stride + 2 |
| 44 |
.byte 0 * \stride + 1, 0 * \stride + 2 |
| 45 |
.byte 0 * \stride + 1, 1 * \stride + 2 |
| 46 |
.byte 1 * \stride + 1, 2 * \stride + 2 |
| 47 |
.byte 1 * \stride + 0, 2 * \stride + 1 |
| 48 |
endconst |
| 49 |
.endm |
| 50 |
|
| 51 |
.macro tables |
| 52 |
dir_table 8, 16 |
| 53 |
dir_table 4, 8 |
| 54 |
|
| 55 |
const pri_taps |
| 56 |
.byte 4, 2, 3, 3 |
| 57 |
endconst |
| 58 |
.endm |
| 59 |
|
| 60 |
.macro load_px d11, d12, d21, d22, w |
| 61 |
.if \w == 8 |
| 62 |
add r6, r2, r9, lsl #1 // x + off |
| 63 |
sub r9, r2, r9, lsl #1 // x - off |
| 64 |
vld1.16 {\d11,\d12}, [r6] // p0 |
| 65 |
vld1.16 {\d21,\d22}, [r9] // p1 |
| 66 |
.else |
| 67 |
add r6, r2, r9, lsl #1 // x + off |
| 68 |
sub r9, r2, r9, lsl #1 // x - off |
| 69 |
vld1.16 {\d11}, [r6] // p0 |
| 70 |
add r6, r6, #2*8 // += stride |
| 71 |
vld1.16 {\d21}, [r9] // p1 |
| 72 |
add r9, r9, #2*8 // += stride |
| 73 |
vld1.16 {\d12}, [r6] // p0 |
| 74 |
vld1.16 {\d22}, [r9] // p1 |
| 75 |
.endif |
| 76 |
.endm |
| 77 |
.macro handle_pixel s1, s2, thresh_vec, shift, tap, min |
| 78 |
.if \min |
| 79 |
vmin.u16 q2, q2, \s1 |
| 80 |
vmax.s16 q3, q3, \s1 |
| 81 |
vmin.u16 q2, q2, \s2 |
| 82 |
vmax.s16 q3, q3, \s2 |
| 83 |
.endif |
| 84 |
vabd.u16 q8, q0, \s1 // abs(diff) |
| 85 |
vabd.u16 q11, q0, \s2 // abs(diff) |
| 86 |
vshl.u16 q9, q8, \shift // abs(diff) >> shift |
| 87 |
vshl.u16 q12, q11, \shift // abs(diff) >> shift |
| 88 |
vqsub.u16 q9, \thresh_vec, q9 // clip = imax(0, threshold - (abs(diff) >> shift)) |
| 89 |
vqsub.u16 q12, \thresh_vec, q12// clip = imax(0, threshold - (abs(diff) >> shift)) |
| 90 |
vsub.i16 q10, \s1, q0 // diff = p0 - px |
| 91 |
vsub.i16 q13, \s2, q0 // diff = p1 - px |
| 92 |
vneg.s16 q8, q9 // -clip |
| 93 |
vneg.s16 q11, q12 // -clip |
| 94 |
vmin.s16 q10, q10, q9 // imin(diff, clip) |
| 95 |
vmin.s16 q13, q13, q12 // imin(diff, clip) |
| 96 |
vdup.16 q9, \tap // taps[k] |
| 97 |
vmax.s16 q10, q10, q8 // constrain() = imax(imin(diff, clip), -clip) |
| 98 |
vmax.s16 q13, q13, q11 // constrain() = imax(imin(diff, clip), -clip) |
| 99 |
vmla.i16 q1, q10, q9 // sum += taps[k] * constrain() |
| 100 |
vmla.i16 q1, q13, q9 // sum += taps[k] * constrain() |
| 101 |
.endm |
| 102 |
|
| 103 |
// void dav1d_cdef_filterX_Ybpc_neon(pixel *dst, ptrdiff_t dst_stride, |
| 104 |
// const uint16_t *tmp, int pri_strength, |
| 105 |
// int sec_strength, int dir, int damping, |
| 106 |
// int h, size_t edges); |
| 107 |
.macro filter_func w, bpc, pri, sec, min, suffix |
| 108 |
function cdef_filter\w\suffix\()_\bpc\()bpc_neon |
| 109 |
.if \bpc == 8 |
| 110 |
cmp r8, #0xf |
| 111 |
beq cdef_filter\w\suffix\()_edged_neon |
| 112 |
.endif |
| 113 |
.if \pri |
| 114 |
.if \bpc == 16 |
| 115 |
clz r9, r9 |
| 116 |
sub r9, r9, #24 // -bitdepth_min_8 |
| 117 |
neg r9, r9 // bitdepth_min_8 |
| 118 |
.endif |
| 119 |
movrel_local r8, pri_taps |
| 120 |
.if \bpc == 16 |
| 121 |
lsr r9, r3, r9 // pri_strength >> bitdepth_min_8 |
| 122 |
and r9, r9, #1 // (pri_strength >> bitdepth_min_8) & 1 |
| 123 |
.else |
| 124 |
and r9, r3, #1 |
| 125 |
.endif |
| 126 |
add r8, r8, r9, lsl #1 |
| 127 |
.endif |
| 128 |
movrel_local r9, directions\w |
| 129 |
add r5, r9, r5, lsl #1 |
| 130 |
vmov.u16 d17, #15 |
| 131 |
vdup.16 d16, r6 // damping |
| 132 |
|
| 133 |
.if \pri |
| 134 |
vdup.16 q5, r3 // threshold |
| 135 |
.endif |
| 136 |
.if \sec |
| 137 |
vdup.16 q7, r4 // threshold |
| 138 |
.endif |
| 139 |
vmov.16 d8[0], r3 |
| 140 |
vmov.16 d8[1], r4 |
| 141 |
vclz.i16 d8, d8 // clz(threshold) |
| 142 |
vsub.i16 d8, d17, d8 // ulog2(threshold) |
| 143 |
vqsub.u16 d8, d16, d8 // shift = imax(0, damping - ulog2(threshold)) |
| 144 |
vneg.s16 d8, d8 // -shift |
| 145 |
.if \sec |
| 146 |
vdup.16 q6, d8[1] |
| 147 |
.endif |
| 148 |
.if \pri |
| 149 |
vdup.16 q4, d8[0] |
| 150 |
.endif |
| 151 |
|
| 152 |
1: |
| 153 |
.if \w == 8 |
| 154 |
vld1.16 {q0}, [r2, :128] // px |
| 155 |
.else |
| 156 |
add r12, r2, #2*8 |
| 157 |
vld1.16 {d0}, [r2, :64] // px |
| 158 |
vld1.16 {d1}, [r12, :64] // px |
| 159 |
.endif |
| 160 |
|
| 161 |
vmov.u16 q1, #0 // sum |
| 162 |
.if \min |
| 163 |
vmov.u16 q2, q0 // min |
| 164 |
vmov.u16 q3, q0 // max |
| 165 |
.endif |
| 166 |
|
| 167 |
// Instead of loading sec_taps 2, 1 from memory, just set it |
| 168 |
// to 2 initially and decrease for the second round. |
| 169 |
// This is also used as loop counter. |
| 170 |
mov lr, #2 // sec_taps[0] |
| 171 |
|
| 172 |
2: |
| 173 |
.if \pri |
| 174 |
ldrsb r9, [r5] // off1 |
| 175 |
|
| 176 |
load_px d28, d29, d30, d31, \w |
| 177 |
.endif |
| 178 |
|
| 179 |
.if \sec |
| 180 |
add r5, r5, #4 // +2*2 |
| 181 |
ldrsb r9, [r5] // off2 |
| 182 |
.endif |
| 183 |
|
| 184 |
.if \pri |
| 185 |
ldrb r12, [r8] // *pri_taps |
| 186 |
|
| 187 |
handle_pixel q14, q15, q5, q4, r12, \min |
| 188 |
.endif |
| 189 |
|
| 190 |
.if \sec |
| 191 |
load_px d28, d29, d30, d31, \w |
| 192 |
|
| 193 |
add r5, r5, #8 // +2*4 |
| 194 |
ldrsb r9, [r5] // off3 |
| 195 |
|
| 196 |
handle_pixel q14, q15, q7, q6, lr, \min |
| 197 |
|
| 198 |
load_px d28, d29, d30, d31, \w |
| 199 |
|
| 200 |
handle_pixel q14, q15, q7, q6, lr, \min |
| 201 |
|
| 202 |
sub r5, r5, #11 // r5 -= 2*(2+4); r5 += 1; |
| 203 |
.else |
| 204 |
add r5, r5, #1 // r5 += 1 |
| 205 |
.endif |
| 206 |
subs lr, lr, #1 // sec_tap-- (value) |
| 207 |
.if \pri |
| 208 |
add r8, r8, #1 // pri_taps++ (pointer) |
| 209 |
.endif |
| 210 |
bne 2b |
| 211 |
|
| 212 |
vshr.s16 q14, q1, #15 // -(sum < 0) |
| 213 |
vadd.i16 q1, q1, q14 // sum - (sum < 0) |
| 214 |
vrshr.s16 q1, q1, #4 // (8 + sum - (sum < 0)) >> 4 |
| 215 |
vadd.i16 q0, q0, q1 // px + (8 + sum ...) >> 4 |
| 216 |
.if \min |
| 217 |
vmin.s16 q0, q0, q3 |
| 218 |
vmax.s16 q0, q0, q2 // iclip(px + .., min, max) |
| 219 |
.endif |
| 220 |
.if \bpc == 8 |
| 221 |
vmovn.u16 d0, q0 |
| 222 |
.endif |
| 223 |
.if \w == 8 |
| 224 |
add r2, r2, #2*16 // tmp += tmp_stride |
| 225 |
subs r7, r7, #1 // h-- |
| 226 |
.if \bpc == 8 |
| 227 |
vst1.8 {d0}, [r0, :64], r1 |
| 228 |
.else |
| 229 |
vst1.16 {q0}, [r0, :128], r1 |
| 230 |
.endif |
| 231 |
.else |
| 232 |
.if \bpc == 8 |
| 233 |
vst1.32 {d0[0]}, [r0, :32], r1 |
| 234 |
.else |
| 235 |
vst1.16 {d0}, [r0, :64], r1 |
| 236 |
.endif |
| 237 |
add r2, r2, #2*16 // tmp += 2*tmp_stride |
| 238 |
subs r7, r7, #2 // h -= 2 |
| 239 |
.if \bpc == 8 |
| 240 |
vst1.32 {d0[1]}, [r0, :32], r1 |
| 241 |
.else |
| 242 |
vst1.16 {d1}, [r0, :64], r1 |
| 243 |
.endif |
| 244 |
.endif |
| 245 |
|
| 246 |
// Reset pri_taps and directions back to the original point |
| 247 |
sub r5, r5, #2 |
| 248 |
.if \pri |
| 249 |
sub r8, r8, #2 |
| 250 |
.endif |
| 251 |
|
| 252 |
bgt 1b |
| 253 |
vpop {q4-q7} |
| 254 |
pop {r4-r9,pc} |
| 255 |
endfunc |
| 256 |
.endm |
| 257 |
|
| 258 |
.macro filter w, bpc |
| 259 |
filter_func \w, \bpc, pri=1, sec=0, min=0, suffix=_pri |
| 260 |
filter_func \w, \bpc, pri=0, sec=1, min=0, suffix=_sec |
| 261 |
filter_func \w, \bpc, pri=1, sec=1, min=1, suffix=_pri_sec |
| 262 |
|
| 263 |
function cdef_filter\w\()_\bpc\()bpc_neon, export=1 |
| 264 |
push {r4-r9,lr} |
| 265 |
vpush {q4-q7} |
| 266 |
ldrd r4, r5, [sp, #92] |
| 267 |
ldrd r6, r7, [sp, #100] |
| 268 |
.if \bpc == 16 |
| 269 |
ldrd r8, r9, [sp, #108] |
| 270 |
.else |
| 271 |
ldr r8, [sp, #108] |
| 272 |
.endif |
| 273 |
cmp r3, #0 // pri_strength |
| 274 |
bne 1f |
| 275 |
b cdef_filter\w\()_sec_\bpc\()bpc_neon // only sec |
| 276 |
1: |
| 277 |
cmp r4, #0 // sec_strength |
| 278 |
bne 1f |
| 279 |
b cdef_filter\w\()_pri_\bpc\()bpc_neon // only pri |
| 280 |
1: |
| 281 |
b cdef_filter\w\()_pri_sec_\bpc\()bpc_neon // both pri and sec |
| 282 |
endfunc |
| 283 |
.endm |
| 284 |
|
| 285 |
const div_table, align=4 |
| 286 |
.short 840, 420, 280, 210, 168, 140, 120, 105 |
| 287 |
endconst |
| 288 |
|
| 289 |
const alt_fact, align=4 |
| 290 |
.short 420, 210, 140, 105, 105, 105, 105, 105, 140, 210, 420, 0 |
| 291 |
endconst |
| 292 |
|
| 293 |
.macro cost_alt dest, s1, s2, s3, s4, s5, s6 |
| 294 |
vmull.s16 q1, \s1, \s1 // sum_alt[n]*sum_alt[n] |
| 295 |
vmull.s16 q2, \s2, \s2 |
| 296 |
vmull.s16 q3, \s3, \s3 |
| 297 |
vmull.s16 q5, \s4, \s4 // sum_alt[n]*sum_alt[n] |
| 298 |
vmull.s16 q12, \s5, \s5 |
| 299 |
vmull.s16 q6, \s6, \s6 // q6 overlaps the first \s1-\s2 here |
| 300 |
vmul.i32 q1, q1, q13 // sum_alt[n]^2*fact |
| 301 |
vmla.i32 q1, q2, q14 |
| 302 |
vmla.i32 q1, q3, q15 |
| 303 |
vmul.i32 q5, q5, q13 // sum_alt[n]^2*fact |
| 304 |
vmla.i32 q5, q12, q14 |
| 305 |
vmla.i32 q5, q6, q15 |
| 306 |
vadd.i32 d2, d2, d3 |
| 307 |
vadd.i32 d3, d10, d11 |
| 308 |
vpadd.i32 \dest, d2, d3 // *cost_ptr |
| 309 |
.endm |
| 310 |
|
| 311 |
.macro find_best s1, s2, s3 |
| 312 |
.ifnb \s2 |
| 313 |
vmov.32 lr, \s2 |
| 314 |
.endif |
| 315 |
cmp r12, r1 // cost[n] > best_cost |
| 316 |
itt gt |
| 317 |
movgt r0, r3 // best_dir = n |
| 318 |
movgt r1, r12 // best_cost = cost[n] |
| 319 |
.ifnb \s2 |
| 320 |
add r3, r3, #1 // n++ |
| 321 |
cmp lr, r1 // cost[n] > best_cost |
| 322 |
vmov.32 r12, \s3 |
| 323 |
itt gt |
| 324 |
movgt r0, r3 // best_dir = n |
| 325 |
movgt r1, lr // best_cost = cost[n] |
| 326 |
add r3, r3, #1 // n++ |
| 327 |
.endif |
| 328 |
.endm |
| 329 |
|
| 330 |
// int dav1d_cdef_find_dir_Xbpc_neon(const pixel *img, const ptrdiff_t stride, |
| 331 |
// unsigned *const var) |
| 332 |
.macro find_dir bpc |
| 333 |
function cdef_find_dir_\bpc\()bpc_neon, export=1 |
| 334 |
push {lr} |
| 335 |
vpush {q4-q7} |
| 336 |
.if \bpc == 16 |
| 337 |
clz r3, r3 // clz(bitdepth_max) |
| 338 |
sub lr, r3, #24 // -bitdepth_min_8 |
| 339 |
.endif |
| 340 |
sub sp, sp, #32 // cost |
| 341 |
mov r3, #8 |
| 342 |
vmov.u16 q1, #0 // q0-q1 sum_diag[0] |
| 343 |
vmov.u16 q3, #0 // q2-q3 sum_diag[1] |
| 344 |
vmov.u16 q5, #0 // q4-q5 sum_hv[0-1] |
| 345 |
vmov.u16 q8, #0 // q6,d16 sum_alt[0] |
| 346 |
// q7,d17 sum_alt[1] |
| 347 |
vmov.u16 q9, #0 // q9,d22 sum_alt[2] |
| 348 |
vmov.u16 q11, #0 |
| 349 |
vmov.u16 q10, #0 // q10,d23 sum_alt[3] |
| 350 |
|
| 351 |
|
| 352 |
.irpc i, 01234567 |
| 353 |
.if \bpc == 8 |
| 354 |
vld1.8 {d30}, [r0, :64], r1 |
| 355 |
vmov.u8 d31, #128 |
| 356 |
vsubl.u8 q15, d30, d31 // img[x] - 128 |
| 357 |
.else |
| 358 |
vld1.16 {q15}, [r0, :128], r1 |
| 359 |
vdup.16 q14, lr // -bitdepth_min_8 |
| 360 |
vshl.u16 q15, q15, q14 |
| 361 |
vmov.u16 q14, #128 |
| 362 |
vsub.i16 q15, q15, q14 // img[x] - 128 |
| 363 |
.endif |
| 364 |
vmov.u16 q14, #0 |
| 365 |
|
| 366 |
.if \i == 0 |
| 367 |
vmov q0, q15 // sum_diag[0] |
| 368 |
.else |
| 369 |
vext.8 q12, q14, q15, #(16-2*\i) |
| 370 |
vext.8 q13, q15, q14, #(16-2*\i) |
| 371 |
vadd.i16 q0, q0, q12 // sum_diag[0] |
| 372 |
vadd.i16 q1, q1, q13 // sum_diag[0] |
| 373 |
.endif |
| 374 |
vrev64.16 q13, q15 |
| 375 |
vswp d26, d27 // [-x] |
| 376 |
.if \i == 0 |
| 377 |
vmov q2, q13 // sum_diag[1] |
| 378 |
.else |
| 379 |
vext.8 q12, q14, q13, #(16-2*\i) |
| 380 |
vext.8 q13, q13, q14, #(16-2*\i) |
| 381 |
vadd.i16 q2, q2, q12 // sum_diag[1] |
| 382 |
vadd.i16 q3, q3, q13 // sum_diag[1] |
| 383 |
.endif |
| 384 |
|
| 385 |
vpadd.u16 d26, d30, d31 // [(x >> 1)] |
| 386 |
vmov.u16 d27, #0 |
| 387 |
vpadd.u16 d24, d26, d28 |
| 388 |
vpadd.u16 d24, d24, d28 // [y] |
| 389 |
vmov.u16 r12, d24[0] |
| 390 |
vadd.i16 q5, q5, q15 // sum_hv[1] |
| 391 |
.if \i < 4 |
| 392 |
vmov.16 d8[\i], r12 // sum_hv[0] |
| 393 |
.else |
| 394 |
vmov.16 d9[\i-4], r12 // sum_hv[0] |
| 395 |
.endif |
| 396 |
|
| 397 |
.if \i == 0 |
| 398 |
vmov.u16 q6, q13 // sum_alt[0] |
| 399 |
.else |
| 400 |
vext.8 q12, q14, q13, #(16-2*\i) |
| 401 |
vext.8 q14, q13, q14, #(16-2*\i) |
| 402 |
vadd.i16 q6, q6, q12 // sum_alt[0] |
| 403 |
vadd.i16 d16, d16, d28 // sum_alt[0] |
| 404 |
.endif |
| 405 |
vrev64.16 d26, d26 // [-(x >> 1)] |
| 406 |
vmov.u16 q14, #0 |
| 407 |
.if \i == 0 |
| 408 |
vmov q7, q13 // sum_alt[1] |
| 409 |
.else |
| 410 |
vext.8 q12, q14, q13, #(16-2*\i) |
| 411 |
vext.8 q13, q13, q14, #(16-2*\i) |
| 412 |
vadd.i16 q7, q7, q12 // sum_alt[1] |
| 413 |
vadd.i16 d17, d17, d26 // sum_alt[1] |
| 414 |
.endif |
| 415 |
|
| 416 |
.if \i < 6 |
| 417 |
vext.8 q12, q14, q15, #(16-2*(3-(\i/2))) |
| 418 |
vext.8 q13, q15, q14, #(16-2*(3-(\i/2))) |
| 419 |
vadd.i16 q9, q9, q12 // sum_alt[2] |
| 420 |
vadd.i16 d22, d22, d26 // sum_alt[2] |
| 421 |
.else |
| 422 |
vadd.i16 q9, q9, q15 // sum_alt[2] |
| 423 |
.endif |
| 424 |
.if \i == 0 |
| 425 |
vmov q10, q15 // sum_alt[3] |
| 426 |
.elseif \i == 1 |
| 427 |
vadd.i16 q10, q10, q15 // sum_alt[3] |
| 428 |
.else |
| 429 |
vext.8 q12, q14, q15, #(16-2*(\i/2)) |
| 430 |
vext.8 q13, q15, q14, #(16-2*(\i/2)) |
| 431 |
vadd.i16 q10, q10, q12 // sum_alt[3] |
| 432 |
vadd.i16 d23, d23, d26 // sum_alt[3] |
| 433 |
.endif |
| 434 |
.endr |
| 435 |
|
| 436 |
vmov.u32 q15, #105 |
| 437 |
|
| 438 |
vmull.s16 q12, d8, d8 // sum_hv[0]*sum_hv[0] |
| 439 |
vmlal.s16 q12, d9, d9 |
| 440 |
vmull.s16 q13, d10, d10 // sum_hv[1]*sum_hv[1] |
| 441 |
vmlal.s16 q13, d11, d11 |
| 442 |
vadd.s32 d8, d24, d25 |
| 443 |
vadd.s32 d9, d26, d27 |
| 444 |
vpadd.s32 d8, d8, d9 // cost[2,6] (s16, s17) |
| 445 |
vmul.i32 d8, d8, d30 // cost[2,6] *= 105 |
| 446 |
|
| 447 |
vrev64.16 q1, q1 |
| 448 |
vrev64.16 q3, q3 |
| 449 |
vext.8 q1, q1, q1, #10 // sum_diag[0][14-n] |
| 450 |
vext.8 q3, q3, q3, #10 // sum_diag[1][14-n] |
| 451 |
|
| 452 |
vstr s16, [sp, #2*4] // cost[2] |
| 453 |
vstr s17, [sp, #6*4] // cost[6] |
| 454 |
|
| 455 |
movrel_local r12, div_table |
| 456 |
vld1.16 {q14}, [r12, :128] |
| 457 |
|
| 458 |
vmull.s16 q5, d0, d0 // sum_diag[0]*sum_diag[0] |
| 459 |
vmull.s16 q12, d1, d1 |
| 460 |
vmlal.s16 q5, d2, d2 |
| 461 |
vmlal.s16 q12, d3, d3 |
| 462 |
vmull.s16 q0, d4, d4 // sum_diag[1]*sum_diag[1] |
| 463 |
vmull.s16 q1, d5, d5 |
| 464 |
vmlal.s16 q0, d6, d6 |
| 465 |
vmlal.s16 q1, d7, d7 |
| 466 |
vmovl.u16 q13, d28 // div_table |
| 467 |
vmovl.u16 q14, d29 |
| 468 |
vmul.i32 q5, q5, q13 // cost[0] |
| 469 |
vmla.i32 q5, q12, q14 |
| 470 |
vmul.i32 q0, q0, q13 // cost[4] |
| 471 |
vmla.i32 q0, q1, q14 |
| 472 |
vadd.i32 d10, d10, d11 |
| 473 |
vadd.i32 d0, d0, d1 |
| 474 |
vpadd.i32 d0, d10, d0 // cost[0,4] = s0,s1 |
| 475 |
|
| 476 |
movrel_local r12, alt_fact |
| 477 |
vld1.16 {d29, d30, d31}, [r12, :64] // div_table[2*m+1] + 105 |
| 478 |
|
| 479 |
vstr s0, [sp, #0*4] // cost[0] |
| 480 |
vstr s1, [sp, #4*4] // cost[4] |
| 481 |
|
| 482 |
vmovl.u16 q13, d29 // div_table[2*m+1] + 105 |
| 483 |
vmovl.u16 q14, d30 |
| 484 |
vmovl.u16 q15, d31 |
| 485 |
|
| 486 |
cost_alt d14, d12, d13, d16, d14, d15, d17 // cost[1], cost[3] |
| 487 |
cost_alt d15, d18, d19, d22, d20, d21, d23 // cost[5], cost[7] |
| 488 |
vstr s28, [sp, #1*4] // cost[1] |
| 489 |
vstr s29, [sp, #3*4] // cost[3] |
| 490 |
|
| 491 |
mov r0, #0 // best_dir |
| 492 |
vmov.32 r1, d0[0] // best_cost |
| 493 |
mov r3, #1 // n |
| 494 |
|
| 495 |
vstr s30, [sp, #5*4] // cost[5] |
| 496 |
vstr s31, [sp, #7*4] // cost[7] |
| 497 |
|
| 498 |
vmov.32 r12, d14[0] |
| 499 |
|
| 500 |
find_best d14[0], d8[0], d14[1] |
| 501 |
find_best d14[1], d0[1], d15[0] |
| 502 |
find_best d15[0], d8[1], d15[1] |
| 503 |
find_best d15[1] |
| 504 |
|
| 505 |
eor r3, r0, #4 // best_dir ^4 |
| 506 |
ldr r12, [sp, r3, lsl #2] |
| 507 |
sub r1, r1, r12 // best_cost - cost[best_dir ^ 4] |
| 508 |
lsr r1, r1, #10 |
| 509 |
str r1, [r2] // *var |
| 510 |
|
| 511 |
add sp, sp, #32 |
| 512 |
vpop {q4-q7} |
| 513 |
pop {pc} |
| 514 |
endfunc |
| 515 |
.endm |