1 /* 2 * Copyright (c) 2017 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include <arm_neon.h> 12 13 #include "./vpx_config.h" 14 #include "./vpx_dsp_rtcd.h" 15 #include "vpx_dsp/arm/idct_neon.h" 16 #include "vpx_dsp/arm/transpose_neon.h" 17 #include "vpx_dsp/txfm_common.h" 18 19 static INLINE void load_from_transformed(const int32_t *const trans_buf, 20 const int first, const int second, 21 int32x4x2_t *const q0, 22 int32x4x2_t *const q1) { 23 q0->val[0] = vld1q_s32(trans_buf + first * 8); 24 q0->val[1] = vld1q_s32(trans_buf + first * 8 + 4); 25 q1->val[0] = vld1q_s32(trans_buf + second * 8); 26 q1->val[1] = vld1q_s32(trans_buf + second * 8 + 4); 27 } 28 29 static INLINE void load_from_output(const int32_t *const out, const int first, 30 const int second, int32x4x2_t *const q0, 31 int32x4x2_t *const q1) { 32 q0->val[0] = vld1q_s32(out + first * 32); 33 q0->val[1] = vld1q_s32(out + first * 32 + 4); 34 q1->val[0] = vld1q_s32(out + second * 32); 35 q1->val[1] = vld1q_s32(out + second * 32 + 4); 36 } 37 38 static INLINE void store_in_output(int32_t *const out, const int first, 39 const int second, const int32x4x2_t q0, 40 const int32x4x2_t q1) { 41 vst1q_s32(out + first * 32, q0.val[0]); 42 vst1q_s32(out + first * 32 + 4, q0.val[1]); 43 vst1q_s32(out + second * 32, q1.val[0]); 44 vst1q_s32(out + second * 32 + 4, q1.val[1]); 45 } 46 47 static INLINE void highbd_store_combine_results( 48 uint16_t *p1, uint16_t *p2, const int stride, const int32x4x2_t q0, 49 const int32x4x2_t q1, const int32x4x2_t q2, const int32x4x2_t q3, 50 const int16x8_t max) { 51 int16x8_t o[4]; 52 uint16x8_t d[4]; 53 54 d[0] = vld1q_u16(p1); 55 p1 += stride; 56 d[1] = vld1q_u16(p1); 57 d[3] = vld1q_u16(p2); 58 p2 -= stride; 59 d[2] = vld1q_u16(p2); 60 61 o[0] = vcombine_s16(vrshrn_n_s32(q0.val[0], 6), vrshrn_n_s32(q0.val[1], 6)); 62 o[1] = vcombine_s16(vrshrn_n_s32(q1.val[0], 6), vrshrn_n_s32(q1.val[1], 6)); 63 o[2] = vcombine_s16(vrshrn_n_s32(q2.val[0], 6), vrshrn_n_s32(q2.val[1], 6)); 64 o[3] = vcombine_s16(vrshrn_n_s32(q3.val[0], 6), vrshrn_n_s32(q3.val[1], 6)); 65 66 o[0] = vqaddq_s16(o[0], vreinterpretq_s16_u16(d[0])); 67 o[1] = vqaddq_s16(o[1], vreinterpretq_s16_u16(d[1])); 68 o[2] = vqaddq_s16(o[2], vreinterpretq_s16_u16(d[2])); 69 o[3] = vqaddq_s16(o[3], vreinterpretq_s16_u16(d[3])); 70 o[0] = vminq_s16(o[0], max); 71 o[1] = vminq_s16(o[1], max); 72 o[2] = vminq_s16(o[2], max); 73 o[3] = vminq_s16(o[3], max); 74 d[0] = vqshluq_n_s16(o[0], 0); 75 d[1] = vqshluq_n_s16(o[1], 0); 76 d[2] = vqshluq_n_s16(o[2], 0); 77 d[3] = vqshluq_n_s16(o[3], 0); 78 79 vst1q_u16(p1, d[1]); 80 p1 -= stride; 81 vst1q_u16(p1, d[0]); 82 vst1q_u16(p2, d[2]); 83 p2 += stride; 84 vst1q_u16(p2, d[3]); 85 } 86 87 static INLINE void do_butterfly(const int32x4x2_t qIn0, const int32x4x2_t qIn1, 88 const int32_t first_const, 89 const int32_t second_const, 90 int32x4x2_t *const qOut0, 91 int32x4x2_t *const qOut1) { 92 int64x2x2_t q[4]; 93 int32x2_t d[6]; 94 95 // Note: using v{mul, mla, mls}l_n_s32 here slows down 35% with gcc 4.9. 96 d[4] = vdup_n_s32(first_const); 97 d[5] = vdup_n_s32(second_const); 98 99 q[0].val[0] = vmull_s32(vget_low_s32(qIn0.val[0]), d[4]); 100 q[0].val[1] = vmull_s32(vget_high_s32(qIn0.val[0]), d[4]); 101 q[1].val[0] = vmull_s32(vget_low_s32(qIn0.val[1]), d[4]); 102 q[1].val[1] = vmull_s32(vget_high_s32(qIn0.val[1]), d[4]); 103 q[0].val[0] = vmlsl_s32(q[0].val[0], vget_low_s32(qIn1.val[0]), d[5]); 104 q[0].val[1] = vmlsl_s32(q[0].val[1], vget_high_s32(qIn1.val[0]), d[5]); 105 q[1].val[0] = vmlsl_s32(q[1].val[0], vget_low_s32(qIn1.val[1]), d[5]); 106 q[1].val[1] = vmlsl_s32(q[1].val[1], vget_high_s32(qIn1.val[1]), d[5]); 107 108 q[2].val[0] = vmull_s32(vget_low_s32(qIn0.val[0]), d[5]); 109 q[2].val[1] = vmull_s32(vget_high_s32(qIn0.val[0]), d[5]); 110 q[3].val[0] = vmull_s32(vget_low_s32(qIn0.val[1]), d[5]); 111 q[3].val[1] = vmull_s32(vget_high_s32(qIn0.val[1]), d[5]); 112 q[2].val[0] = vmlal_s32(q[2].val[0], vget_low_s32(qIn1.val[0]), d[4]); 113 q[2].val[1] = vmlal_s32(q[2].val[1], vget_high_s32(qIn1.val[0]), d[4]); 114 q[3].val[0] = vmlal_s32(q[3].val[0], vget_low_s32(qIn1.val[1]), d[4]); 115 q[3].val[1] = vmlal_s32(q[3].val[1], vget_high_s32(qIn1.val[1]), d[4]); 116 117 qOut0->val[0] = vcombine_s32(vrshrn_n_s64(q[0].val[0], DCT_CONST_BITS), 118 vrshrn_n_s64(q[0].val[1], DCT_CONST_BITS)); 119 qOut0->val[1] = vcombine_s32(vrshrn_n_s64(q[1].val[0], DCT_CONST_BITS), 120 vrshrn_n_s64(q[1].val[1], DCT_CONST_BITS)); 121 qOut1->val[0] = vcombine_s32(vrshrn_n_s64(q[2].val[0], DCT_CONST_BITS), 122 vrshrn_n_s64(q[2].val[1], DCT_CONST_BITS)); 123 qOut1->val[1] = vcombine_s32(vrshrn_n_s64(q[3].val[0], DCT_CONST_BITS), 124 vrshrn_n_s64(q[3].val[1], DCT_CONST_BITS)); 125 } 126 127 static INLINE void load_s32x4q_dual( 128 const int32_t *in, int32x4x2_t *const s0, int32x4x2_t *const s1, 129 int32x4x2_t *const s2, int32x4x2_t *const s3, int32x4x2_t *const s4, 130 int32x4x2_t *const s5, int32x4x2_t *const s6, int32x4x2_t *const s7) { 131 s0->val[0] = vld1q_s32(in); 132 s0->val[1] = vld1q_s32(in + 4); 133 in += 32; 134 s1->val[0] = vld1q_s32(in); 135 s1->val[1] = vld1q_s32(in + 4); 136 in += 32; 137 s2->val[0] = vld1q_s32(in); 138 s2->val[1] = vld1q_s32(in + 4); 139 in += 32; 140 s3->val[0] = vld1q_s32(in); 141 s3->val[1] = vld1q_s32(in + 4); 142 in += 32; 143 s4->val[0] = vld1q_s32(in); 144 s4->val[1] = vld1q_s32(in + 4); 145 in += 32; 146 s5->val[0] = vld1q_s32(in); 147 s5->val[1] = vld1q_s32(in + 4); 148 in += 32; 149 s6->val[0] = vld1q_s32(in); 150 s6->val[1] = vld1q_s32(in + 4); 151 in += 32; 152 s7->val[0] = vld1q_s32(in); 153 s7->val[1] = vld1q_s32(in + 4); 154 } 155 156 static INLINE void transpose_and_store_s32_8x8(int32x4x2_t a0, int32x4x2_t a1, 157 int32x4x2_t a2, int32x4x2_t a3, 158 int32x4x2_t a4, int32x4x2_t a5, 159 int32x4x2_t a6, int32x4x2_t a7, 160 int32_t **out) { 161 transpose_s32_8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7); 162 163 vst1q_s32(*out, a0.val[0]); 164 *out += 4; 165 vst1q_s32(*out, a0.val[1]); 166 *out += 4; 167 vst1q_s32(*out, a1.val[0]); 168 *out += 4; 169 vst1q_s32(*out, a1.val[1]); 170 *out += 4; 171 vst1q_s32(*out, a2.val[0]); 172 *out += 4; 173 vst1q_s32(*out, a2.val[1]); 174 *out += 4; 175 vst1q_s32(*out, a3.val[0]); 176 *out += 4; 177 vst1q_s32(*out, a3.val[1]); 178 *out += 4; 179 vst1q_s32(*out, a4.val[0]); 180 *out += 4; 181 vst1q_s32(*out, a4.val[1]); 182 *out += 4; 183 vst1q_s32(*out, a5.val[0]); 184 *out += 4; 185 vst1q_s32(*out, a5.val[1]); 186 *out += 4; 187 vst1q_s32(*out, a6.val[0]); 188 *out += 4; 189 vst1q_s32(*out, a6.val[1]); 190 *out += 4; 191 vst1q_s32(*out, a7.val[0]); 192 *out += 4; 193 vst1q_s32(*out, a7.val[1]); 194 *out += 4; 195 } 196 197 static INLINE void idct32_transpose_pair(const int32_t *input, int32_t *t_buf) { 198 int i; 199 int32x4x2_t s0, s1, s2, s3, s4, s5, s6, s7; 200 201 for (i = 0; i < 4; i++, input += 8) { 202 load_s32x4q_dual(input, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7); 203 transpose_and_store_s32_8x8(s0, s1, s2, s3, s4, s5, s6, s7, &t_buf); 204 } 205 } 206 207 static INLINE void idct32_bands_end_1st_pass(int32_t *const out, 208 int32x4x2_t *const q) { 209 store_in_output(out, 16, 17, q[6], q[7]); 210 store_in_output(out, 14, 15, q[8], q[9]); 211 212 load_from_output(out, 30, 31, &q[0], &q[1]); 213 q[4] = highbd_idct_add_dual(q[2], q[1]); 214 q[5] = highbd_idct_add_dual(q[3], q[0]); 215 q[6] = highbd_idct_sub_dual(q[3], q[0]); 216 q[7] = highbd_idct_sub_dual(q[2], q[1]); 217 store_in_output(out, 30, 31, q[6], q[7]); 218 store_in_output(out, 0, 1, q[4], q[5]); 219 220 load_from_output(out, 12, 13, &q[0], &q[1]); 221 q[2] = highbd_idct_add_dual(q[10], q[1]); 222 q[3] = highbd_idct_add_dual(q[11], q[0]); 223 q[4] = highbd_idct_sub_dual(q[11], q[0]); 224 q[5] = highbd_idct_sub_dual(q[10], q[1]); 225 226 load_from_output(out, 18, 19, &q[0], &q[1]); 227 q[8] = highbd_idct_add_dual(q[4], q[1]); 228 q[9] = highbd_idct_add_dual(q[5], q[0]); 229 q[6] = highbd_idct_sub_dual(q[5], q[0]); 230 q[7] = highbd_idct_sub_dual(q[4], q[1]); 231 store_in_output(out, 18, 19, q[6], q[7]); 232 store_in_output(out, 12, 13, q[8], q[9]); 233 234 load_from_output(out, 28, 29, &q[0], &q[1]); 235 q[4] = highbd_idct_add_dual(q[2], q[1]); 236 q[5] = highbd_idct_add_dual(q[3], q[0]); 237 q[6] = highbd_idct_sub_dual(q[3], q[0]); 238 q[7] = highbd_idct_sub_dual(q[2], q[1]); 239 store_in_output(out, 28, 29, q[6], q[7]); 240 store_in_output(out, 2, 3, q[4], q[5]); 241 242 load_from_output(out, 10, 11, &q[0], &q[1]); 243 q[2] = highbd_idct_add_dual(q[12], q[1]); 244 q[3] = highbd_idct_add_dual(q[13], q[0]); 245 q[4] = highbd_idct_sub_dual(q[13], q[0]); 246 q[5] = highbd_idct_sub_dual(q[12], q[1]); 247 248 load_from_output(out, 20, 21, &q[0], &q[1]); 249 q[8] = highbd_idct_add_dual(q[4], q[1]); 250 q[9] = highbd_idct_add_dual(q[5], q[0]); 251 q[6] = highbd_idct_sub_dual(q[5], q[0]); 252 q[7] = highbd_idct_sub_dual(q[4], q[1]); 253 store_in_output(out, 20, 21, q[6], q[7]); 254 store_in_output(out, 10, 11, q[8], q[9]); 255 256 load_from_output(out, 26, 27, &q[0], &q[1]); 257 q[4] = highbd_idct_add_dual(q[2], q[1]); 258 q[5] = highbd_idct_add_dual(q[3], q[0]); 259 q[6] = highbd_idct_sub_dual(q[3], q[0]); 260 q[7] = highbd_idct_sub_dual(q[2], q[1]); 261 store_in_output(out, 26, 27, q[6], q[7]); 262 store_in_output(out, 4, 5, q[4], q[5]); 263 264 load_from_output(out, 8, 9, &q[0], &q[1]); 265 q[2] = highbd_idct_add_dual(q[14], q[1]); 266 q[3] = highbd_idct_add_dual(q[15], q[0]); 267 q[4] = highbd_idct_sub_dual(q[15], q[0]); 268 q[5] = highbd_idct_sub_dual(q[14], q[1]); 269 270 load_from_output(out, 22, 23, &q[0], &q[1]); 271 q[8] = highbd_idct_add_dual(q[4], q[1]); 272 q[9] = highbd_idct_add_dual(q[5], q[0]); 273 q[6] = highbd_idct_sub_dual(q[5], q[0]); 274 q[7] = highbd_idct_sub_dual(q[4], q[1]); 275 store_in_output(out, 22, 23, q[6], q[7]); 276 store_in_output(out, 8, 9, q[8], q[9]); 277 278 load_from_output(out, 24, 25, &q[0], &q[1]); 279 q[4] = highbd_idct_add_dual(q[2], q[1]); 280 q[5] = highbd_idct_add_dual(q[3], q[0]); 281 q[6] = highbd_idct_sub_dual(q[3], q[0]); 282 q[7] = highbd_idct_sub_dual(q[2], q[1]); 283 store_in_output(out, 24, 25, q[6], q[7]); 284 store_in_output(out, 6, 7, q[4], q[5]); 285 } 286 287 static INLINE void idct32_bands_end_2nd_pass(const int32_t *const out, 288 uint16_t *const dest, 289 const int stride, 290 const int16x8_t max, 291 int32x4x2_t *const q) { 292 uint16_t *dest0 = dest + 0 * stride; 293 uint16_t *dest1 = dest + 31 * stride; 294 uint16_t *dest2 = dest + 16 * stride; 295 uint16_t *dest3 = dest + 15 * stride; 296 const int str2 = stride << 1; 297 298 highbd_store_combine_results(dest2, dest3, stride, q[6], q[7], q[8], q[9], 299 max); 300 dest2 += str2; 301 dest3 -= str2; 302 303 load_from_output(out, 30, 31, &q[0], &q[1]); 304 q[4] = highbd_idct_add_dual(q[2], q[1]); 305 q[5] = highbd_idct_add_dual(q[3], q[0]); 306 q[6] = highbd_idct_sub_dual(q[3], q[0]); 307 q[7] = highbd_idct_sub_dual(q[2], q[1]); 308 highbd_store_combine_results(dest0, dest1, stride, q[4], q[5], q[6], q[7], 309 max); 310 dest0 += str2; 311 dest1 -= str2; 312 313 load_from_output(out, 12, 13, &q[0], &q[1]); 314 q[2] = highbd_idct_add_dual(q[10], q[1]); 315 q[3] = highbd_idct_add_dual(q[11], q[0]); 316 q[4] = highbd_idct_sub_dual(q[11], q[0]); 317 q[5] = highbd_idct_sub_dual(q[10], q[1]); 318 319 load_from_output(out, 18, 19, &q[0], &q[1]); 320 q[8] = highbd_idct_add_dual(q[4], q[1]); 321 q[9] = highbd_idct_add_dual(q[5], q[0]); 322 q[6] = highbd_idct_sub_dual(q[5], q[0]); 323 q[7] = highbd_idct_sub_dual(q[4], q[1]); 324 highbd_store_combine_results(dest2, dest3, stride, q[6], q[7], q[8], q[9], 325 max); 326 dest2 += str2; 327 dest3 -= str2; 328 329 load_from_output(out, 28, 29, &q[0], &q[1]); 330 q[4] = highbd_idct_add_dual(q[2], q[1]); 331 q[5] = highbd_idct_add_dual(q[3], q[0]); 332 q[6] = highbd_idct_sub_dual(q[3], q[0]); 333 q[7] = highbd_idct_sub_dual(q[2], q[1]); 334 highbd_store_combine_results(dest0, dest1, stride, q[4], q[5], q[6], q[7], 335 max); 336 dest0 += str2; 337 dest1 -= str2; 338 339 load_from_output(out, 10, 11, &q[0], &q[1]); 340 q[2] = highbd_idct_add_dual(q[12], q[1]); 341 q[3] = highbd_idct_add_dual(q[13], q[0]); 342 q[4] = highbd_idct_sub_dual(q[13], q[0]); 343 q[5] = highbd_idct_sub_dual(q[12], q[1]); 344 345 load_from_output(out, 20, 21, &q[0], &q[1]); 346 q[8] = highbd_idct_add_dual(q[4], q[1]); 347 q[9] = highbd_idct_add_dual(q[5], q[0]); 348 q[6] = highbd_idct_sub_dual(q[5], q[0]); 349 q[7] = highbd_idct_sub_dual(q[4], q[1]); 350 highbd_store_combine_results(dest2, dest3, stride, q[6], q[7], q[8], q[9], 351 max); 352 dest2 += str2; 353 dest3 -= str2; 354 355 load_from_output(out, 26, 27, &q[0], &q[1]); 356 q[4] = highbd_idct_add_dual(q[2], q[1]); 357 q[5] = highbd_idct_add_dual(q[3], q[0]); 358 q[6] = highbd_idct_sub_dual(q[3], q[0]); 359 q[7] = highbd_idct_sub_dual(q[2], q[1]); 360 highbd_store_combine_results(dest0, dest1, stride, q[4], q[5], q[6], q[7], 361 max); 362 dest0 += str2; 363 dest1 -= str2; 364 365 load_from_output(out, 8, 9, &q[0], &q[1]); 366 q[2] = highbd_idct_add_dual(q[14], q[1]); 367 q[3] = highbd_idct_add_dual(q[15], q[0]); 368 q[4] = highbd_idct_sub_dual(q[15], q[0]); 369 q[5] = highbd_idct_sub_dual(q[14], q[1]); 370 371 load_from_output(out, 22, 23, &q[0], &q[1]); 372 q[8] = highbd_idct_add_dual(q[4], q[1]); 373 q[9] = highbd_idct_add_dual(q[5], q[0]); 374 q[6] = highbd_idct_sub_dual(q[5], q[0]); 375 q[7] = highbd_idct_sub_dual(q[4], q[1]); 376 highbd_store_combine_results(dest2, dest3, stride, q[6], q[7], q[8], q[9], 377 max); 378 379 load_from_output(out, 24, 25, &q[0], &q[1]); 380 q[4] = highbd_idct_add_dual(q[2], q[1]); 381 q[5] = highbd_idct_add_dual(q[3], q[0]); 382 q[6] = highbd_idct_sub_dual(q[3], q[0]); 383 q[7] = highbd_idct_sub_dual(q[2], q[1]); 384 highbd_store_combine_results(dest0, dest1, stride, q[4], q[5], q[6], q[7], 385 max); 386 } 387 388 static INLINE void vpx_highbd_idct32_32_neon(const tran_low_t *input, 389 uint16_t *dst, const int stride, 390 const int bd) { 391 int i, idct32_pass_loop; 392 int32_t trans_buf[32 * 8]; 393 int32_t pass1[32 * 32]; 394 int32_t pass2[32 * 32]; 395 int32_t *out; 396 int32x4x2_t q[16]; 397 398 for (idct32_pass_loop = 0, out = pass1; idct32_pass_loop < 2; 399 idct32_pass_loop++, input = pass1, out = pass2) { 400 for (i = 0; i < 4; i++, out += 8) { // idct32_bands_loop 401 idct32_transpose_pair(input, trans_buf); 402 input += 32 * 8; 403 404 // ----------------------------------------- 405 // BLOCK A: 16-19,28-31 406 // ----------------------------------------- 407 // generate 16,17,30,31 408 // part of stage 1 409 load_from_transformed(trans_buf, 1, 31, &q[14], &q[13]); 410 do_butterfly(q[14], q[13], cospi_31_64, cospi_1_64, &q[0], &q[2]); 411 load_from_transformed(trans_buf, 17, 15, &q[14], &q[13]); 412 do_butterfly(q[14], q[13], cospi_15_64, cospi_17_64, &q[1], &q[3]); 413 // part of stage 2 414 q[4] = highbd_idct_add_dual(q[0], q[1]); 415 q[13] = highbd_idct_sub_dual(q[0], q[1]); 416 q[6] = highbd_idct_add_dual(q[2], q[3]); 417 q[14] = highbd_idct_sub_dual(q[2], q[3]); 418 // part of stage 3 419 do_butterfly(q[14], q[13], cospi_28_64, cospi_4_64, &q[5], &q[7]); 420 421 // generate 18,19,28,29 422 // part of stage 1 423 load_from_transformed(trans_buf, 9, 23, &q[14], &q[13]); 424 do_butterfly(q[14], q[13], cospi_23_64, cospi_9_64, &q[0], &q[2]); 425 load_from_transformed(trans_buf, 25, 7, &q[14], &q[13]); 426 do_butterfly(q[14], q[13], cospi_7_64, cospi_25_64, &q[1], &q[3]); 427 // part of stage 2 428 q[13] = highbd_idct_sub_dual(q[3], q[2]); 429 q[3] = highbd_idct_add_dual(q[3], q[2]); 430 q[14] = highbd_idct_sub_dual(q[1], q[0]); 431 q[2] = highbd_idct_add_dual(q[1], q[0]); 432 // part of stage 3 433 do_butterfly(q[14], q[13], -cospi_4_64, -cospi_28_64, &q[1], &q[0]); 434 // part of stage 4 435 q[8] = highbd_idct_add_dual(q[4], q[2]); 436 q[9] = highbd_idct_add_dual(q[5], q[0]); 437 q[10] = highbd_idct_add_dual(q[7], q[1]); 438 q[15] = highbd_idct_add_dual(q[6], q[3]); 439 q[13] = highbd_idct_sub_dual(q[5], q[0]); 440 q[14] = highbd_idct_sub_dual(q[7], q[1]); 441 store_in_output(out, 16, 31, q[8], q[15]); 442 store_in_output(out, 17, 30, q[9], q[10]); 443 // part of stage 5 444 do_butterfly(q[14], q[13], cospi_24_64, cospi_8_64, &q[0], &q[1]); 445 store_in_output(out, 29, 18, q[1], q[0]); 446 // part of stage 4 447 q[13] = highbd_idct_sub_dual(q[4], q[2]); 448 q[14] = highbd_idct_sub_dual(q[6], q[3]); 449 // part of stage 5 450 do_butterfly(q[14], q[13], cospi_24_64, cospi_8_64, &q[4], &q[6]); 451 store_in_output(out, 19, 28, q[4], q[6]); 452 453 // ----------------------------------------- 454 // BLOCK B: 20-23,24-27 455 // ----------------------------------------- 456 // generate 20,21,26,27 457 // part of stage 1 458 load_from_transformed(trans_buf, 5, 27, &q[14], &q[13]); 459 do_butterfly(q[14], q[13], cospi_27_64, cospi_5_64, &q[0], &q[2]); 460 load_from_transformed(trans_buf, 21, 11, &q[14], &q[13]); 461 do_butterfly(q[14], q[13], cospi_11_64, cospi_21_64, &q[1], &q[3]); 462 // part of stage 2 463 q[13] = highbd_idct_sub_dual(q[0], q[1]); 464 q[0] = highbd_idct_add_dual(q[0], q[1]); 465 q[14] = highbd_idct_sub_dual(q[2], q[3]); 466 q[2] = highbd_idct_add_dual(q[2], q[3]); 467 // part of stage 3 468 do_butterfly(q[14], q[13], cospi_12_64, cospi_20_64, &q[1], &q[3]); 469 470 // generate 22,23,24,25 471 // part of stage 1 472 load_from_transformed(trans_buf, 13, 19, &q[14], &q[13]); 473 do_butterfly(q[14], q[13], cospi_19_64, cospi_13_64, &q[5], &q[7]); 474 load_from_transformed(trans_buf, 29, 3, &q[14], &q[13]); 475 do_butterfly(q[14], q[13], cospi_3_64, cospi_29_64, &q[4], &q[6]); 476 // part of stage 2 477 q[14] = highbd_idct_sub_dual(q[4], q[5]); 478 q[5] = highbd_idct_add_dual(q[4], q[5]); 479 q[13] = highbd_idct_sub_dual(q[6], q[7]); 480 q[6] = highbd_idct_add_dual(q[6], q[7]); 481 // part of stage 3 482 do_butterfly(q[14], q[13], -cospi_20_64, -cospi_12_64, &q[4], &q[7]); 483 // part of stage 4 484 q[10] = highbd_idct_add_dual(q[7], q[1]); 485 q[11] = highbd_idct_add_dual(q[5], q[0]); 486 q[12] = highbd_idct_add_dual(q[6], q[2]); 487 q[15] = highbd_idct_add_dual(q[4], q[3]); 488 // part of stage 6 489 load_from_output(out, 16, 17, &q[14], &q[13]); 490 q[8] = highbd_idct_add_dual(q[14], q[11]); 491 q[9] = highbd_idct_add_dual(q[13], q[10]); 492 q[13] = highbd_idct_sub_dual(q[13], q[10]); 493 q[11] = highbd_idct_sub_dual(q[14], q[11]); 494 store_in_output(out, 17, 16, q[9], q[8]); 495 load_from_output(out, 30, 31, &q[14], &q[9]); 496 q[8] = highbd_idct_sub_dual(q[9], q[12]); 497 q[10] = highbd_idct_add_dual(q[14], q[15]); 498 q[14] = highbd_idct_sub_dual(q[14], q[15]); 499 q[12] = highbd_idct_add_dual(q[9], q[12]); 500 store_in_output(out, 30, 31, q[10], q[12]); 501 // part of stage 7 502 do_butterfly(q[14], q[13], cospi_16_64, cospi_16_64, &q[13], &q[14]); 503 store_in_output(out, 25, 22, q[14], q[13]); 504 do_butterfly(q[8], q[11], cospi_16_64, cospi_16_64, &q[13], &q[14]); 505 store_in_output(out, 24, 23, q[14], q[13]); 506 // part of stage 4 507 q[14] = highbd_idct_sub_dual(q[5], q[0]); 508 q[13] = highbd_idct_sub_dual(q[6], q[2]); 509 do_butterfly(q[14], q[13], -cospi_8_64, -cospi_24_64, &q[5], &q[6]); 510 q[14] = highbd_idct_sub_dual(q[7], q[1]); 511 q[13] = highbd_idct_sub_dual(q[4], q[3]); 512 do_butterfly(q[14], q[13], -cospi_8_64, -cospi_24_64, &q[0], &q[1]); 513 // part of stage 6 514 load_from_output(out, 18, 19, &q[14], &q[13]); 515 q[8] = highbd_idct_add_dual(q[14], q[1]); 516 q[9] = highbd_idct_add_dual(q[13], q[6]); 517 q[13] = highbd_idct_sub_dual(q[13], q[6]); 518 q[1] = highbd_idct_sub_dual(q[14], q[1]); 519 store_in_output(out, 18, 19, q[8], q[9]); 520 load_from_output(out, 28, 29, &q[8], &q[9]); 521 q[14] = highbd_idct_sub_dual(q[8], q[5]); 522 q[10] = highbd_idct_add_dual(q[8], q[5]); 523 q[11] = highbd_idct_add_dual(q[9], q[0]); 524 q[0] = highbd_idct_sub_dual(q[9], q[0]); 525 store_in_output(out, 28, 29, q[10], q[11]); 526 // part of stage 7 527 do_butterfly(q[14], q[13], cospi_16_64, cospi_16_64, &q[13], &q[14]); 528 store_in_output(out, 20, 27, q[13], q[14]); 529 do_butterfly(q[0], q[1], cospi_16_64, cospi_16_64, &q[1], &q[0]); 530 store_in_output(out, 21, 26, q[1], q[0]); 531 532 // ----------------------------------------- 533 // BLOCK C: 8-10,11-15 534 // ----------------------------------------- 535 // generate 8,9,14,15 536 // part of stage 2 537 load_from_transformed(trans_buf, 2, 30, &q[14], &q[13]); 538 do_butterfly(q[14], q[13], cospi_30_64, cospi_2_64, &q[0], &q[2]); 539 load_from_transformed(trans_buf, 18, 14, &q[14], &q[13]); 540 do_butterfly(q[14], q[13], cospi_14_64, cospi_18_64, &q[1], &q[3]); 541 // part of stage 3 542 q[13] = highbd_idct_sub_dual(q[0], q[1]); 543 q[0] = highbd_idct_add_dual(q[0], q[1]); 544 q[14] = highbd_idct_sub_dual(q[2], q[3]); 545 q[2] = highbd_idct_add_dual(q[2], q[3]); 546 // part of stage 4 547 do_butterfly(q[14], q[13], cospi_24_64, cospi_8_64, &q[1], &q[3]); 548 549 // generate 10,11,12,13 550 // part of stage 2 551 load_from_transformed(trans_buf, 10, 22, &q[14], &q[13]); 552 do_butterfly(q[14], q[13], cospi_22_64, cospi_10_64, &q[5], &q[7]); 553 load_from_transformed(trans_buf, 26, 6, &q[14], &q[13]); 554 do_butterfly(q[14], q[13], cospi_6_64, cospi_26_64, &q[4], &q[6]); 555 // part of stage 3 556 q[14] = highbd_idct_sub_dual(q[4], q[5]); 557 q[5] = highbd_idct_add_dual(q[4], q[5]); 558 q[13] = highbd_idct_sub_dual(q[6], q[7]); 559 q[6] = highbd_idct_add_dual(q[6], q[7]); 560 // part of stage 4 561 do_butterfly(q[14], q[13], -cospi_8_64, -cospi_24_64, &q[4], &q[7]); 562 // part of stage 5 563 q[8] = highbd_idct_add_dual(q[0], q[5]); 564 q[9] = highbd_idct_add_dual(q[1], q[7]); 565 q[13] = highbd_idct_sub_dual(q[1], q[7]); 566 q[14] = highbd_idct_sub_dual(q[3], q[4]); 567 q[10] = highbd_idct_add_dual(q[3], q[4]); 568 q[15] = highbd_idct_add_dual(q[2], q[6]); 569 store_in_output(out, 8, 15, q[8], q[15]); 570 store_in_output(out, 9, 14, q[9], q[10]); 571 // part of stage 6 572 do_butterfly(q[14], q[13], cospi_16_64, cospi_16_64, &q[1], &q[3]); 573 store_in_output(out, 13, 10, q[3], q[1]); 574 q[13] = highbd_idct_sub_dual(q[0], q[5]); 575 q[14] = highbd_idct_sub_dual(q[2], q[6]); 576 do_butterfly(q[14], q[13], cospi_16_64, cospi_16_64, &q[1], &q[3]); 577 store_in_output(out, 11, 12, q[1], q[3]); 578 579 // ----------------------------------------- 580 // BLOCK D: 0-3,4-7 581 // ----------------------------------------- 582 // generate 4,5,6,7 583 // part of stage 3 584 load_from_transformed(trans_buf, 4, 28, &q[14], &q[13]); 585 do_butterfly(q[14], q[13], cospi_28_64, cospi_4_64, &q[0], &q[2]); 586 load_from_transformed(trans_buf, 20, 12, &q[14], &q[13]); 587 do_butterfly(q[14], q[13], cospi_12_64, cospi_20_64, &q[1], &q[3]); 588 // part of stage 4 589 q[13] = highbd_idct_sub_dual(q[0], q[1]); 590 q[0] = highbd_idct_add_dual(q[0], q[1]); 591 q[14] = highbd_idct_sub_dual(q[2], q[3]); 592 q[2] = highbd_idct_add_dual(q[2], q[3]); 593 // part of stage 5 594 do_butterfly(q[14], q[13], cospi_16_64, cospi_16_64, &q[1], &q[3]); 595 596 // generate 0,1,2,3 597 // part of stage 4 598 load_from_transformed(trans_buf, 0, 16, &q[14], &q[13]); 599 do_butterfly(q[14], q[13], cospi_16_64, cospi_16_64, &q[5], &q[7]); 600 load_from_transformed(trans_buf, 8, 24, &q[14], &q[13]); 601 do_butterfly(q[14], q[13], cospi_24_64, cospi_8_64, &q[14], &q[6]); 602 // part of stage 5 603 q[4] = highbd_idct_add_dual(q[7], q[6]); 604 q[7] = highbd_idct_sub_dual(q[7], q[6]); 605 q[6] = highbd_idct_sub_dual(q[5], q[14]); 606 q[5] = highbd_idct_add_dual(q[5], q[14]); 607 // part of stage 6 608 q[8] = highbd_idct_add_dual(q[4], q[2]); 609 q[9] = highbd_idct_add_dual(q[5], q[3]); 610 q[10] = highbd_idct_add_dual(q[6], q[1]); 611 q[11] = highbd_idct_add_dual(q[7], q[0]); 612 q[12] = highbd_idct_sub_dual(q[7], q[0]); 613 q[13] = highbd_idct_sub_dual(q[6], q[1]); 614 q[14] = highbd_idct_sub_dual(q[5], q[3]); 615 q[15] = highbd_idct_sub_dual(q[4], q[2]); 616 // part of stage 7 617 load_from_output(out, 14, 15, &q[0], &q[1]); 618 q[2] = highbd_idct_add_dual(q[8], q[1]); 619 q[3] = highbd_idct_add_dual(q[9], q[0]); 620 q[4] = highbd_idct_sub_dual(q[9], q[0]); 621 q[5] = highbd_idct_sub_dual(q[8], q[1]); 622 load_from_output(out, 16, 17, &q[0], &q[1]); 623 q[8] = highbd_idct_add_dual(q[4], q[1]); 624 q[9] = highbd_idct_add_dual(q[5], q[0]); 625 q[6] = highbd_idct_sub_dual(q[5], q[0]); 626 q[7] = highbd_idct_sub_dual(q[4], q[1]); 627 628 if (idct32_pass_loop == 0) { 629 idct32_bands_end_1st_pass(out, q); 630 } else { 631 const int16x8_t max = vdupq_n_s16((1 << bd) - 1); 632 idct32_bands_end_2nd_pass(out, dst, stride, max, q); 633 dst += 8; 634 } 635 } 636 } 637 } 638 639 void vpx_highbd_idct32x32_1024_add_neon(const tran_low_t *input, uint16_t *dest, 640 int stride, int bd) { 641 if (bd == 8) { 642 vpx_idct32_32_neon(input, CAST_TO_BYTEPTR(dest), stride, 1); 643 } else { 644 vpx_highbd_idct32_32_neon(input, dest, stride, bd); 645 } 646 } 647
