aom_convolve8_neon.c 12.2 KB
Newer Older
Zoe Liu's avatar
Zoe Liu committed
1
/*
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Zoe Liu's avatar
Zoe Liu committed
3
 *
4 5 6 7 8 9
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
Zoe Liu's avatar
Zoe Liu committed
10 11 12
 */

#include <arm_neon.h>
13
#include <assert.h>
Zoe Liu's avatar
Zoe Liu committed
14

Adrian Grange's avatar
Adrian Grange committed
15 16 17
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
Yaowu Xu's avatar
Yaowu Xu committed
18
#include "aom_ports/mem.h"
Zoe Liu's avatar
Zoe Liu committed
19

clang-format's avatar
clang-format committed
20 21 22 23 24
static INLINE int32x4_t MULTIPLY_BY_Q0(int16x4_t dsrc0, int16x4_t dsrc1,
                                       int16x4_t dsrc2, int16x4_t dsrc3,
                                       int16x4_t dsrc4, int16x4_t dsrc5,
                                       int16x4_t dsrc6, int16x4_t dsrc7,
                                       int16x8_t q0s16) {
Zoe Liu's avatar
Zoe Liu committed
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
  int32x4_t qdst;
  int16x4_t d0s16, d1s16;

  d0s16 = vget_low_s16(q0s16);
  d1s16 = vget_high_s16(q0s16);

  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
  return qdst;
}

Adrian Grange's avatar
Adrian Grange committed
42
void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
clang-format's avatar
clang-format committed
43 44 45 46 47
                              uint8_t *dst, ptrdiff_t dst_stride,
                              const int16_t *filter_x, int x_step_q4,
                              const int16_t *filter_y,  // unused
                              int y_step_q4,            // unused
                              int w, int h) {
Zoe Liu's avatar
Zoe Liu committed
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
  int width;
  const uint8_t *s, *psrc;
  uint8_t *d, *pdst;
  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
  uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
  uint8x16_t q12u8, q13u8, q14u8, q15u8;
  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
  int16x4_t d24s16, d25s16, d26s16, d27s16;
  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
  int16x8_t q0s16;
  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
  int32x4_t q1s32, q2s32, q14s32, q15s32;
  uint16x8x2_t q0x2u16;
  uint8x8x2_t d0x2u8, d1x2u8;
  uint32x2x2_t d0x2u32;
  uint16x4x2_t d0x2u16, d1x2u16;
  uint32x4x2_t q0x2u32;

66
  assert(x_step_q4 == 16);
Zoe Liu's avatar
Zoe Liu committed
67

68 69 70 71
  (void)x_step_q4;
  (void)y_step_q4;
  (void)filter_y;

Zoe Liu's avatar
Zoe Liu committed
72 73 74
  q0s16 = vld1q_s16(filter_x);

  src -= 3;  // adjust for taps
clang-format's avatar
clang-format committed
75 76
  for (; h > 0; h -= 4, src += src_stride * 4,
                dst += dst_stride * 4) {  // loop_horiz_v
Zoe Liu's avatar
Zoe Liu committed
77 78 79 80 81 82 83 84 85 86 87 88
    s = src;
    d24u8 = vld1_u8(s);
    s += src_stride;
    d25u8 = vld1_u8(s);
    s += src_stride;
    d26u8 = vld1_u8(s);
    s += src_stride;
    d27u8 = vld1_u8(s);

    q12u8 = vcombine_u8(d24u8, d25u8);
    q13u8 = vcombine_u8(d26u8, d27u8);

clang-format's avatar
clang-format committed
89 90
    q0x2u16 =
        vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
Zoe Liu's avatar
Zoe Liu committed
91 92 93 94 95 96 97 98 99 100 101
    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
    d0x2u8 = vtrn_u8(d24u8, d25u8);
    d1x2u8 = vtrn_u8(d26u8, d27u8);

    __builtin_prefetch(src + src_stride * 4);
    __builtin_prefetch(src + src_stride * 5);
    __builtin_prefetch(src + src_stride * 6);

clang-format's avatar
clang-format committed
102 103
    q8u16 = vmovl_u8(d0x2u8.val[0]);
    q9u16 = vmovl_u8(d0x2u8.val[1]);
Zoe Liu's avatar
Zoe Liu committed
104 105 106 107 108 109 110 111 112 113 114 115
    q10u16 = vmovl_u8(d1x2u8.val[0]);
    q11u16 = vmovl_u8(d1x2u8.val[1]);

    d16u16 = vget_low_u16(q8u16);
    d17u16 = vget_high_u16(q8u16);
    d18u16 = vget_low_u16(q9u16);
    d19u16 = vget_high_u16(q9u16);
    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
    q9u16 = vcombine_u16(d17u16, d19u16);

    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
clang-format's avatar
clang-format committed
116
    for (width = w, psrc = src + 7, pdst = dst; width > 0;
Zoe Liu's avatar
Zoe Liu committed
117 118 119 120 121 122 123 124 125 126 127 128
         width -= 4, psrc += 4, pdst += 4) {  // loop_horiz
      s = psrc;
      d28u32 = vld1_dup_u32((const uint32_t *)s);
      s += src_stride;
      d29u32 = vld1_dup_u32((const uint32_t *)s);
      s += src_stride;
      d31u32 = vld1_dup_u32((const uint32_t *)s);
      s += src_stride;
      d30u32 = vld1_dup_u32((const uint32_t *)s);

      __builtin_prefetch(psrc + 64);

clang-format's avatar
clang-format committed
129 130 131 132
      d0x2u16 =
          vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
      d1x2u16 =
          vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
Zoe Liu's avatar
Zoe Liu committed
133 134 135 136 137 138 139 140 141
      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30

      __builtin_prefetch(psrc + 64 + src_stride);

      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
clang-format's avatar
clang-format committed
142 143
      q0x2u32 =
          vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
Zoe Liu's avatar
Zoe Liu committed
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161

      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
      q12u16 = vmovl_u8(d28u8);
      q13u16 = vmovl_u8(d29u8);

      __builtin_prefetch(psrc + 64 + src_stride * 2);

      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));

clang-format's avatar
clang-format committed
162 163 164 165 166 167 168 169
      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
                             d23s16, d24s16, q0s16);
      q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
                             d24s16, d26s16, q0s16);
      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
                              d26s16, d27s16, q0s16);
      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
                              d27s16, d25s16, q0s16);
Zoe Liu's avatar
Zoe Liu committed
170 171 172 173 174 175 176 177 178 179 180 181 182 183

      __builtin_prefetch(psrc + 60 + src_stride * 3);

      d2u16 = vqrshrun_n_s32(q1s32, 7);
      d3u16 = vqrshrun_n_s32(q2s32, 7);
      d4u16 = vqrshrun_n_s32(q14s32, 7);
      d5u16 = vqrshrun_n_s32(q15s32, 7);

      q1u16 = vcombine_u16(d2u16, d3u16);
      q2u16 = vcombine_u16(d4u16, d5u16);

      d2u8 = vqmovn_u16(q1u16);
      d3u8 = vqmovn_u16(q2u16);

clang-format's avatar
clang-format committed
184
      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
Zoe Liu's avatar
Zoe Liu committed
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211
      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
                         vreinterpret_u32_u16(d0x2u16.val[1]));
      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
                       vreinterpret_u8_u32(d0x2u32.val[1]));

      d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
      d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);

      d = pdst;
      vst1_lane_u32((uint32_t *)d, d2u32, 0);
      d += dst_stride;
      vst1_lane_u32((uint32_t *)d, d3u32, 0);
      d += dst_stride;
      vst1_lane_u32((uint32_t *)d, d2u32, 1);
      d += dst_stride;
      vst1_lane_u32((uint32_t *)d, d3u32, 1);

      q8u16 = q9u16;
      d20s16 = d23s16;
      q11u16 = q12u16;
      q9u16 = q13u16;
      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
    }
  }
  return;
}

Adrian Grange's avatar
Adrian Grange committed
212
void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
clang-format's avatar
clang-format committed
213 214 215 216 217
                             uint8_t *dst, ptrdiff_t dst_stride,
                             const int16_t *filter_x,  // unused
                             int x_step_q4,            // unused
                             const int16_t *filter_y, int y_step_q4, int w,
                             int h) {
Zoe Liu's avatar
Zoe Liu committed
218 219 220 221 222 223 224 225 226 227 228 229
  int height;
  const uint8_t *s;
  uint8_t *d;
  uint32x2_t d2u32, d3u32;
  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
  int16x4_t d24s16, d25s16, d26s16, d27s16;
  uint16x4_t d2u16, d3u16, d4u16, d5u16;
  int16x8_t q0s16;
  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
  int32x4_t q1s32, q2s32, q14s32, q15s32;

230
  assert(y_step_q4 == 16);
Zoe Liu's avatar
Zoe Liu committed
231

232 233 234 235
  (void)x_step_q4;
  (void)y_step_q4;
  (void)filter_x;

Zoe Liu's avatar
Zoe Liu committed
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
  src -= src_stride * 3;
  q0s16 = vld1q_s16(filter_y);
  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
    s = src;
    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
    s += src_stride;
    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
    s += src_stride;
    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
    s += src_stride;
    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
    s += src_stride;
    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
    s += src_stride;
    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
    s += src_stride;
    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
    s += src_stride;

clang-format's avatar
clang-format committed
255 256
    q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
    q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
Zoe Liu's avatar
Zoe Liu committed
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287
    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));

    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
    d = dst;
    for (height = h; height > 0; height -= 4) {  // loop_vert
      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
      s += src_stride;
      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
      s += src_stride;
      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
      s += src_stride;
      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
      s += src_stride;

      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));

      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));

      __builtin_prefetch(d);
      __builtin_prefetch(d + dst_stride);
clang-format's avatar
clang-format committed
288 289
      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
                             d22s16, d24s16, q0s16);
Zoe Liu's avatar
Zoe Liu committed
290 291
      __builtin_prefetch(d + dst_stride * 2);
      __builtin_prefetch(d + dst_stride * 3);
clang-format's avatar
clang-format committed
292 293
      q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
                             d24s16, d26s16, q0s16);
Zoe Liu's avatar
Zoe Liu committed
294 295
      __builtin_prefetch(s);
      __builtin_prefetch(s + src_stride);
clang-format's avatar
clang-format committed
296 297
      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
                              d26s16, d27s16, q0s16);
Zoe Liu's avatar
Zoe Liu committed
298 299
      __builtin_prefetch(s + src_stride * 2);
      __builtin_prefetch(s + src_stride * 3);
clang-format's avatar
clang-format committed
300 301
      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
                              d27s16, d25s16, q0s16);
Zoe Liu's avatar
Zoe Liu committed
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331

      d2u16 = vqrshrun_n_s32(q1s32, 7);
      d3u16 = vqrshrun_n_s32(q2s32, 7);
      d4u16 = vqrshrun_n_s32(q14s32, 7);
      d5u16 = vqrshrun_n_s32(q15s32, 7);

      q1u16 = vcombine_u16(d2u16, d3u16);
      q2u16 = vcombine_u16(d4u16, d5u16);

      d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
      d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));

      vst1_lane_u32((uint32_t *)d, d2u32, 0);
      d += dst_stride;
      vst1_lane_u32((uint32_t *)d, d2u32, 1);
      d += dst_stride;
      vst1_lane_u32((uint32_t *)d, d3u32, 0);
      d += dst_stride;
      vst1_lane_u32((uint32_t *)d, d3u32, 1);
      d += dst_stride;

      q8u16 = q10u16;
      d18s16 = d22s16;
      d19s16 = d24s16;
      q10u16 = q13u16;
      d22s16 = d25s16;
    }
  }
  return;
}