convolve.c 22.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * 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.
 */

12
#include <assert.h>
Angie Chiang's avatar
Angie Chiang committed
13
#include <string.h>
14

15
#include "./aom_dsp_rtcd.h"
Yaowu Xu's avatar
Yaowu Xu committed
16
#include "./av1_rtcd.h"
17
#include "av1/common/convolve.h"
18
#include "av1/common/filter.h"
Yaowu Xu's avatar
Yaowu Xu committed
19
#include "aom_dsp/aom_dsp_common.h"
20
#include "aom_ports/mem.h"
21

22 23
#define MAX_BLOCK_WIDTH (MAX_SB_SIZE)
#define MAX_BLOCK_HEIGHT (MAX_SB_SIZE)
24 25 26
#define MAX_STEP (32)
#define MAX_FILTER_TAP (12)

Yaowu Xu's avatar
Yaowu Xu committed
27 28 29
void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst,
                          int dst_stride, int w, int h,
                          const InterpFilterParams filter_params,
30 31
                          const int subpel_x_q4, int x_step_q4,
                          ConvolveParams *conv_params) {
32
  int x, y;
33
  int filter_size = filter_params.taps;
34 35 36 37 38
  src -= filter_size / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = subpel_x_q4;
    for (x = 0; x < w; ++x) {
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
Yaowu Xu's avatar
Yaowu Xu committed
39
      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
40
          filter_params, x_q4 & SUBPEL_MASK);
41 42
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
43

44
      if (conv_params->round == CONVOLVE_OPT_ROUND) {
45
        sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
46 47 48 49 50 51 52 53 54 55
        if (conv_params->ref)
          dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
        else
          dst[x] = sum;
      } else {
        int tmp = conv_params->dst[y * conv_params->dst_stride + x];
        if (conv_params->ref)
          tmp = ROUND_POWER_OF_TWO(tmp + sum, 1);
        else
          tmp = sum;
Angie Chiang's avatar
Angie Chiang committed
56 57
        conv_params->dst[y * conv_params->dst_stride + x] =
            clamp(tmp, INT32_MIN, INT32_MAX);
58
      }
59

60 61 62 63 64 65 66
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
67 68 69
void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst,
                         int dst_stride, int w, int h,
                         const InterpFilterParams filter_params,
70 71
                         const int subpel_y_q4, int y_step_q4,
                         ConvolveParams *conv_params) {
72
  int x, y;
73
  int filter_size = filter_params.taps;
74 75 76 77 78 79
  src -= src_stride * (filter_size / 2 - 1);

  for (x = 0; x < w; ++x) {
    int y_q4 = subpel_y_q4;
    for (y = 0; y < h; ++y) {
      const uint8_t *const src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
Yaowu Xu's avatar
Yaowu Xu committed
80
      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
81
          filter_params, y_q4 & SUBPEL_MASK);
82 83 84
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
85

86
      if (conv_params->round == CONVOLVE_OPT_ROUND) {
87
        sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
88 89 90 91 92 93 94 95 96 97 98
        if (conv_params->ref)
          dst[y * dst_stride] =
              ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
        else
          dst[y * dst_stride] = sum;
      } else {
        int tmp = conv_params->dst[y * conv_params->dst_stride + x];
        if (conv_params->ref)
          tmp = ROUND_POWER_OF_TWO(tmp + sum, 1);
        else
          tmp = sum;
Angie Chiang's avatar
Angie Chiang committed
99 100
        conv_params->dst[y * conv_params->dst_stride + x] =
            clamp(tmp, INT32_MIN, INT32_MAX);
101
      }
102

103 104 105 106 107 108 109
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Angie Chiang's avatar
Angie Chiang committed
110
static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst,
111 112 113 114
                          int dst_stride, int w, int h,
                          ConvolveParams *conv_params) {
  if (conv_params->ref == 0) {
    int r, c;
Angie Chiang's avatar
Angie Chiang committed
115
    for (r = 0; r < h; ++r) {
116 117 118 119 120 121 122
      if (conv_params->round == CONVOLVE_OPT_ROUND) {
        memcpy(dst, src, w);
      } else {
        for (c = 0; c < w; ++c)
          conv_params->dst[r * conv_params->dst_stride + c] = ((uint16_t)src[c])
                                                              << FILTER_BITS;
      }
Angie Chiang's avatar
Angie Chiang committed
123 124 125 126 127 128
      src += src_stride;
      dst += dst_stride;
    }
  } else {
    int r, c;
    for (r = 0; r < h; ++r) {
129 130
      for (c = 0; c < w; ++c) {
        if (conv_params->round == CONVOLVE_OPT_ROUND) {
131
          dst[c] = clip_pixel(ROUND_POWER_OF_TWO(dst[c] + src[c], 1));
132 133
        } else {
          int tmp = conv_params->dst[r * conv_params->dst_stride + c];
Angie Chiang's avatar
Angie Chiang committed
134 135 136
          tmp = ROUND_POWER_OF_TWO(tmp + (((int32_t)src[c]) << FILTER_BITS), 1);
          conv_params->dst[r * conv_params->dst_stride + c] =
              clamp(tmp, INT32_MIN, INT32_MAX);
137 138
        }
      }
Angie Chiang's avatar
Angie Chiang committed
139 140 141 142 143 144
      src += src_stride;
      dst += dst_stride;
    }
  }
}

145 146 147
void av1_convolve_horiz_facade(const uint8_t *src, int src_stride, uint8_t *dst,
                               int dst_stride, int w, int h,
                               const InterpFilterParams filter_params,
148 149
                               const int subpel_x_q4, int x_step_q4,
                               ConvolveParams *conv_params) {
150 151
  if (filter_params.taps == SUBPEL_TAPS &&
      conv_params->round == CONVOLVE_OPT_ROUND) {
152 153
    const int16_t *filter_x =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
154
    if (conv_params->ref == 0)
155 156 157 158 159 160
      aom_convolve8_horiz(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
                          NULL, -1, w, h);
    else
      aom_convolve8_avg_horiz(src, src_stride, dst, dst_stride, filter_x,
                              x_step_q4, NULL, -1, w, h);
  } else {
161 162 163 164 165 166 167 168
    if (conv_params->round == CONVOLVE_OPT_ROUND) {
      av1_convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
                         subpel_x_q4, x_step_q4, conv_params);
    } else {
      // TODO(angiebird) need SIMD implementation here
      av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h,
                           filter_params, subpel_x_q4, x_step_q4, conv_params);
    }
169 170 171 172 173 174
  }
}

void av1_convolve_vert_facade(const uint8_t *src, int src_stride, uint8_t *dst,
                              int dst_stride, int w, int h,
                              const InterpFilterParams filter_params,
175 176
                              const int subpel_y_q4, int y_step_q4,
                              ConvolveParams *conv_params) {
177 178
  if (filter_params.taps == SUBPEL_TAPS &&
      conv_params->round == CONVOLVE_OPT_ROUND) {
179 180
    const int16_t *filter_y =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
181
    if (conv_params->ref == 0) {
182 183 184 185 186 187 188
      aom_convolve8_vert(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
                         y_step_q4, w, h);
    } else {
      aom_convolve8_avg_vert(src, src_stride, dst, dst_stride, NULL, -1,
                             filter_y, y_step_q4, w, h);
    }
  } else {
189 190 191 192 193 194 195 196
    if (conv_params->round == CONVOLVE_OPT_ROUND) {
      av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
                        subpel_y_q4, y_step_q4, conv_params);
    } else {
      // TODO(angiebird) need SIMD implementation here
      av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params,
                          subpel_y_q4, y_step_q4, conv_params);
    }
197 198 199
  }
}

Angie Chiang's avatar
Angie Chiang committed
200 201 202 203 204 205 206 207 208 209 210 211 212
#if CONVOLVE_POST_ROUNDING
void av1_convolve_rounding(const int32_t *src, int src_stride, uint8_t *dst,
                           int dst_stride, int w, int h) {
  int r, c;
  for (r = 0; r < h; ++r) {
    for (c = 0; c < w; ++c) {
      dst[r * dst_stride + c] =
          clip_pixel(ROUND_POWER_OF_TWO(src[r * src_stride + c], FILTER_BITS));
    }
  }
}
#endif  // CONVOLVE_POST_ROUNDING

Yaowu Xu's avatar
Yaowu Xu committed
213 214
void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
                  int dst_stride, int w, int h,
215
#if CONFIG_DUAL_FILTER
James Zern's avatar
James Zern committed
216
                  const InterpFilter *interp_filter,
217
#else
James Zern's avatar
James Zern committed
218
                  const InterpFilter interp_filter,
219
#endif
Yaowu Xu's avatar
Yaowu Xu committed
220
                  const int subpel_x_q4, int x_step_q4, const int subpel_y_q4,
221
                  int y_step_q4, ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
222 223
  int ignore_horiz = x_step_q4 == 16 && subpel_x_q4 == 0;
  int ignore_vert = y_step_q4 == 16 && subpel_y_q4 == 0;
224 225 226 227 228 229

  assert(w <= MAX_BLOCK_WIDTH);
  assert(h <= MAX_BLOCK_HEIGHT);
  assert(y_step_q4 <= MAX_STEP);
  assert(x_step_q4 <= MAX_STEP);

Angie Chiang's avatar
Angie Chiang committed
230
  if (ignore_horiz && ignore_vert) {
231
    convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params);
Angie Chiang's avatar
Angie Chiang committed
232
  } else if (ignore_vert) {
233 234
#if CONFIG_DUAL_FILTER
    InterpFilterParams filter_params =
235
        av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
236 237
#else
    InterpFilterParams filter_params =
Yaowu Xu's avatar
Yaowu Xu committed
238
        av1_get_interp_filter_params(interp_filter);
239 240
#endif
    assert(filter_params.taps <= MAX_FILTER_TAP);
241
    av1_convolve_horiz_facade(src, src_stride, dst, dst_stride, w, h,
242 243
                              filter_params, subpel_x_q4, x_step_q4,
                              conv_params);
Angie Chiang's avatar
Angie Chiang committed
244
  } else if (ignore_horiz) {
245 246
#if CONFIG_DUAL_FILTER
    InterpFilterParams filter_params =
247
        av1_get_interp_filter_params(interp_filter[2 * conv_params->ref]);
248 249
#else
    InterpFilterParams filter_params =
Yaowu Xu's avatar
Yaowu Xu committed
250
        av1_get_interp_filter_params(interp_filter);
251 252
#endif
    assert(filter_params.taps <= MAX_FILTER_TAP);
253
    av1_convolve_vert_facade(src, src_stride, dst, dst_stride, w, h,
254 255
                             filter_params, subpel_y_q4, y_step_q4,
                             conv_params);
Angie Chiang's avatar
Angie Chiang committed
256
  } else {
257 258 259 260 261 262
    // temp's size is set to a 256 aligned value to facilitate SIMD
    // implementation. The value is greater than (maximum possible intermediate
    // height or width) * MAX_SB_SIZE
    DECLARE_ALIGNED(16, uint8_t,
                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
263 264
    int filter_size;
    InterpFilterParams filter_params;
265
#if CONFIG_DUAL_FILTER
266
    InterpFilterParams filter_params_x =
267
        av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
268
    InterpFilterParams filter_params_y =
269 270 271
        av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
    if (interp_filter[0 + 2 * conv_params->ref] == MULTITAP_SHARP &&
        interp_filter[1 + 2 * conv_params->ref] == MULTITAP_SHARP) {
272 273 274 275
      // Avoid two directions both using 12-tap filter.
      // This will reduce hardware implementation cost.
      filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
    }
276

277 278 279 280
    // we do filter with fewer taps first to reduce hardware implementation
    // complexity
    if (filter_params_y.taps < filter_params_x.taps) {
      int intermediate_width;
281
      int temp_stride = max_intermediate_size;
282 283
      ConvolveParams temp_conv_params;
      temp_conv_params.ref = 0;
284
      temp_conv_params.round = CONVOLVE_OPT_ROUND;
285 286 287 288
      filter_params = filter_params_y;
      filter_size = filter_params_x.taps;
      intermediate_width =
          (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
289
      assert(intermediate_width <= max_intermediate_size);
Angie Chiang's avatar
Angie Chiang committed
290

291
      assert(filter_params.taps <= MAX_FILTER_TAP);
292

293 294
      av1_convolve_vert_facade(src - (filter_size / 2 - 1), src_stride, temp,
                               temp_stride, intermediate_width, h,
295 296
                               filter_params, subpel_y_q4, y_step_q4,
                               &temp_conv_params);
297

298 299
      filter_params = filter_params_x;
      assert(filter_params.taps <= MAX_FILTER_TAP);
300

301 302
      av1_convolve_horiz_facade(temp + (filter_size / 2 - 1), temp_stride, dst,
                                dst_stride, w, h, filter_params, subpel_x_q4,
303
                                x_step_q4, conv_params);
304
    } else
305
#endif  // CONFIG_DUAL_FILTER
306 307
    {
      int intermediate_height;
308
      int temp_stride = MAX_SB_SIZE;
309 310
      ConvolveParams temp_conv_params;
      temp_conv_params.ref = 0;
311
      temp_conv_params.round = CONVOLVE_OPT_ROUND;
312 313 314 315 316 317 318 319 320
#if CONFIG_DUAL_FILTER
      filter_params = filter_params_x;
      filter_size = filter_params_y.taps;
#else
      filter_params = av1_get_interp_filter_params(interp_filter);
      filter_size = filter_params.taps;
#endif
      intermediate_height =
          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
321 322
      assert(intermediate_height <= max_intermediate_size);
      (void)max_intermediate_size;
323 324 325

      assert(filter_params.taps <= MAX_FILTER_TAP);

326 327 328
      av1_convolve_horiz_facade(src - src_stride * (filter_size / 2 - 1),
                                src_stride, temp, temp_stride, w,
                                intermediate_height, filter_params, subpel_x_q4,
329
                                x_step_q4, &temp_conv_params);
330 331 332 333 334 335

#if CONFIG_DUAL_FILTER
      filter_params = filter_params_y;
#endif
      assert(filter_params.taps <= MAX_FILTER_TAP);

336 337 338
      av1_convolve_vert_facade(
          temp + temp_stride * (filter_size / 2 - 1), temp_stride, dst,
          dst_stride, w, h, filter_params, subpel_y_q4, y_step_q4, conv_params);
339
    }
Angie Chiang's avatar
Angie Chiang committed
340
  }
341 342
}

Angie Chiang's avatar
Angie Chiang committed
343 344 345 346 347
void av1_convolve_init_c(void) {
  // A placeholder for SIMD initialization
  return;
}

Yaowu Xu's avatar
Yaowu Xu committed
348 349 350 351 352 353
#if CONFIG_AOM_HIGHBITDEPTH
void av1_highbd_convolve_horiz_c(const uint16_t *src, int src_stride,
                                 uint16_t *dst, int dst_stride, int w, int h,
                                 const InterpFilterParams filter_params,
                                 const int subpel_x_q4, int x_step_q4, int avg,
                                 int bd) {
354
  int x, y;
355
  int filter_size = filter_params.taps;
356 357 358 359 360
  src -= filter_size / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = subpel_x_q4;
    for (x = 0; x < w; ++x) {
      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
Yaowu Xu's avatar
Yaowu Xu committed
361
      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
362
          filter_params, x_q4 & SUBPEL_MASK);
363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
      if (avg)
        dst[x] = ROUND_POWER_OF_TWO(
            dst[x] +
                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
            1);
      else
        dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
379 380 381 382 383
void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride,
                                uint16_t *dst, int dst_stride, int w, int h,
                                const InterpFilterParams filter_params,
                                const int subpel_y_q4, int y_step_q4, int avg,
                                int bd) {
384
  int x, y;
385
  int filter_size = filter_params.taps;
386 387 388 389 390 391
  src -= src_stride * (filter_size / 2 - 1);

  for (x = 0; x < w; ++x) {
    int y_q4 = subpel_y_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *const src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
Yaowu Xu's avatar
Yaowu Xu committed
392
      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
393
          filter_params, y_q4 & SUBPEL_MASK);
394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
      if (avg) {
        dst[y * dst_stride] = ROUND_POWER_OF_TWO(
            dst[y * dst_stride] +
                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
            1);
      } else {
        dst[y * dst_stride] =
            clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
      }
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Angie Chiang's avatar
Angie Chiang committed
413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434
static void highbd_convolve_copy(const uint16_t *src, int src_stride,
                                 uint16_t *dst, int dst_stride, int w, int h,
                                 int avg, int bd) {
  if (avg == 0) {
    int r;
    for (r = 0; r < h; ++r) {
      memcpy(dst, src, w * sizeof(*src));
      src += src_stride;
      dst += dst_stride;
    }
  } else {
    int r, c;
    for (r = 0; r < h; ++r) {
      for (c = 0; c < w; ++c) {
        dst[c] = clip_pixel_highbd(ROUND_POWER_OF_TWO(dst[c] + src[c], 1), bd);
      }
      src += src_stride;
      dst += dst_stride;
    }
  }
}

435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465
void av1_highbd_convolve_horiz_facade(const uint8_t *src8, int src_stride,
                                      uint8_t *dst8, int dst_stride, int w,
                                      int h,
                                      const InterpFilterParams filter_params,
                                      const int subpel_x_q4, int x_step_q4,
                                      int avg, int bd) {
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  if (filter_params.taps == SUBPEL_TAPS) {
    const int16_t *filter_x =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
    if (avg == 0)
      aom_highbd_convolve8_horiz(src8, src_stride, dst8, dst_stride, filter_x,
                                 x_step_q4, NULL, -1, w, h, bd);
    else
      aom_highbd_convolve8_avg_horiz(src8, src_stride, dst8, dst_stride,
                                     filter_x, x_step_q4, NULL, -1, w, h, bd);
  } else {
    av1_highbd_convolve_horiz(src, src_stride, dst, dst_stride, w, h,
                              filter_params, subpel_x_q4, x_step_q4, avg, bd);
  }
}

void av1_highbd_convolve_vert_facade(const uint8_t *src8, int src_stride,
                                     uint8_t *dst8, int dst_stride, int w,
                                     int h,
                                     const InterpFilterParams filter_params,
                                     const int subpel_y_q4, int y_step_q4,
                                     int avg, int bd) {
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
466

467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482
  if (filter_params.taps == SUBPEL_TAPS) {
    const int16_t *filter_y =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
    if (avg == 0) {
      aom_highbd_convolve8_vert(src8, src_stride, dst8, dst_stride, NULL, -1,
                                filter_y, y_step_q4, w, h, bd);
    } else {
      aom_highbd_convolve8_avg_vert(src8, src_stride, dst8, dst_stride, NULL,
                                    -1, filter_y, y_step_q4, w, h, bd);
    }
  } else {
    av1_highbd_convolve_vert(src, src_stride, dst, dst_stride, w, h,
                             filter_params, subpel_y_q4, y_step_q4, avg, bd);
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
483 484
void av1_highbd_convolve(const uint8_t *src8, int src_stride, uint8_t *dst8,
                         int dst_stride, int w, int h,
485
#if CONFIG_DUAL_FILTER
James Zern's avatar
James Zern committed
486
                         const InterpFilter *interp_filter,
487
#else
James Zern's avatar
James Zern committed
488
                         const InterpFilter interp_filter,
489
#endif
Yaowu Xu's avatar
Yaowu Xu committed
490 491 492
                         const int subpel_x_q4, int x_step_q4,
                         const int subpel_y_q4, int y_step_q4, int ref_idx,
                         int bd) {
Angie Chiang's avatar
Angie Chiang committed
493 494 495 496
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  int ignore_horiz = x_step_q4 == 16 && subpel_x_q4 == 0;
  int ignore_vert = y_step_q4 == 16 && subpel_y_q4 == 0;
497 498 499 500 501 502

  assert(w <= MAX_BLOCK_WIDTH);
  assert(h <= MAX_BLOCK_HEIGHT);
  assert(y_step_q4 <= MAX_STEP);
  assert(x_step_q4 <= MAX_STEP);

Angie Chiang's avatar
Angie Chiang committed
503
  if (ignore_horiz && ignore_vert) {
504
    highbd_convolve_copy(src, src_stride, dst, dst_stride, w, h, ref_idx, bd);
Angie Chiang's avatar
Angie Chiang committed
505
  } else if (ignore_vert) {
506 507
#if CONFIG_DUAL_FILTER
    InterpFilterParams filter_params =
Yaowu Xu's avatar
Yaowu Xu committed
508
        av1_get_interp_filter_params(interp_filter[1 + 2 * ref_idx]);
509 510
#else
    InterpFilterParams filter_params =
Yaowu Xu's avatar
Yaowu Xu committed
511
        av1_get_interp_filter_params(interp_filter);
512
#endif
513 514 515
    av1_highbd_convolve_horiz_facade(src8, src_stride, dst8, dst_stride, w, h,
                                     filter_params, subpel_x_q4, x_step_q4,
                                     ref_idx, bd);
Angie Chiang's avatar
Angie Chiang committed
516
  } else if (ignore_horiz) {
517 518
#if CONFIG_DUAL_FILTER
    InterpFilterParams filter_params =
Yaowu Xu's avatar
Yaowu Xu committed
519
        av1_get_interp_filter_params(interp_filter[0 + 2 * ref_idx]);
520 521
#else
    InterpFilterParams filter_params =
Yaowu Xu's avatar
Yaowu Xu committed
522
        av1_get_interp_filter_params(interp_filter);
523
#endif
524 525 526
    av1_highbd_convolve_vert_facade(src8, src_stride, dst8, dst_stride, w, h,
                                    filter_params, subpel_y_q4, y_step_q4,
                                    ref_idx, bd);
Angie Chiang's avatar
Angie Chiang committed
527
  } else {
528 529 530 531 532
    // temp's size is set to a 256 aligned value to facilitate SIMD
    // implementation. The value is greater than (maximum possible intermediate
    // height or width) * MAX_SB_SIZE
    DECLARE_ALIGNED(16, uint16_t,
                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
Yaowu Xu's avatar
Yaowu Xu committed
533
    uint8_t *temp8 = CONVERT_TO_BYTEPTR(temp);
534 535 536
    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
    int filter_size;
    InterpFilterParams filter_params;
537
#if CONFIG_DUAL_FILTER
Yaowu Xu's avatar
Yaowu Xu committed
538 539 540 541
    InterpFilterParams filter_params_x =
        av1_get_interp_filter_params(interp_filter[1 + 2 * ref_idx]);
    InterpFilterParams filter_params_y =
        av1_get_interp_filter_params(interp_filter[0 + 2 * ref_idx]);
542 543 544 545 546 547
    if (interp_filter[0 + 2 * ref_idx] == MULTITAP_SHARP &&
        interp_filter[1 + 2 * ref_idx] == MULTITAP_SHARP) {
      // Avoid two directions both using 12-tap filter.
      // This will reduce hardware implementation cost.
      filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
    }
548
#endif
549

550 551 552 553 554 555 556 557 558 559 560
#if CONFIG_DUAL_FILTER
    if (filter_params_y.taps < filter_params_x.taps) {
      int intermediate_width;
      int temp_stride = max_intermediate_size;
      filter_params = filter_params_y;
      filter_size = filter_params_x.taps;
      intermediate_width =
          (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
      assert(intermediate_width <= max_intermediate_size);

      assert(filter_params.taps <= MAX_FILTER_TAP);
Angie Chiang's avatar
Angie Chiang committed
561

562 563 564
      av1_highbd_convolve_vert_facade(
          src8 - (filter_size / 2 - 1), src_stride, temp8, temp_stride,
          intermediate_width, h, filter_params, subpel_y_q4, y_step_q4, 0, bd);
565

566 567 568 569 570 571 572 573 574 575 576
      filter_params = filter_params_x;
      assert(filter_params.taps <= MAX_FILTER_TAP);

      av1_highbd_convolve_horiz_facade(
          temp8 + (filter_size / 2 - 1), temp_stride, dst8, dst_stride, w, h,
          filter_params, subpel_x_q4, x_step_q4, ref_idx, bd);
    } else
#endif  // CONFIG_DUAL_FILTER
    {
      int intermediate_height;
      int temp_stride = MAX_SB_SIZE;
577
#if CONFIG_DUAL_FILTER
578 579 580 581 582
      filter_params = filter_params_x;
      filter_size = filter_params_y.taps;
#else
      filter_params = av1_get_interp_filter_params(interp_filter);
      filter_size = filter_params.taps;
583
#endif
584 585 586 587 588 589 590 591 592 593 594 595 596 597 598
      intermediate_height =
          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
      assert(intermediate_height <= max_intermediate_size);
      (void)max_intermediate_size;

      av1_highbd_convolve_horiz_facade(
          src8 - src_stride * (filter_size / 2 - 1), src_stride, temp8,
          temp_stride, w, intermediate_height, filter_params, subpel_x_q4,
          x_step_q4, 0, bd);

#if CONFIG_DUAL_FILTER
      filter_params = filter_params_y;
#endif
      filter_size = filter_params.taps;
      assert(filter_params.taps <= MAX_FILTER_TAP);
599

600 601 602 603
      av1_highbd_convolve_vert_facade(
          temp8 + temp_stride * (filter_size / 2 - 1), temp_stride, dst8,
          dst_stride, w, h, filter_params, subpel_y_q4, y_step_q4, ref_idx, bd);
    }
Angie Chiang's avatar
Angie Chiang committed
604
  }
605
}
Yaowu Xu's avatar
Yaowu Xu committed
606
#endif  // CONFIG_AOM_HIGHBITDEPTH