convolve.c 72.4 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/blockd.h"
18
#include "av1/common/convolve.h"
19
#include "av1/common/filter.h"
Angie Chiang's avatar
Angie Chiang committed
20
#include "av1/common/onyxc_int.h"
Yaowu Xu's avatar
Yaowu Xu committed
21
#include "aom_dsp/aom_dsp_common.h"
22
#include "aom_ports/mem.h"
23

24 25
#define MAX_BLOCK_WIDTH (MAX_SB_SIZE)
#define MAX_BLOCK_HEIGHT (MAX_SB_SIZE)
26 27
#define MAX_STEP (32)

Yaowu Xu's avatar
Yaowu Xu committed
28 29 30
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,
31 32
                          const int subpel_x_q4, int x_step_q4,
                          ConvolveParams *conv_params) {
33
  int x, y;
34
  int filter_size = filter_params.taps;
Angie Chiang's avatar
Angie Chiang committed
35
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
36 37 38 39 40
  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
41
      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
42
          filter_params, x_q4 & SUBPEL_MASK);
43 44
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
45

Angie Chiang's avatar
Angie Chiang committed
46
      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
47
      if (conv_params->do_average)
Angie Chiang's avatar
Angie Chiang committed
48 49 50
        dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
      else
        dst[x] = sum;
51

52 53 54 55 56 57 58
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Fergus Simpson's avatar
Fergus Simpson committed
59 60 61 62 63 64 65
void av1_convolve_horiz_scale(const uint8_t *src, int src_stride, uint8_t *dst,
                              int dst_stride, int w, int h,
                              const InterpFilterParams filter_params,
                              const int subpel_x_qn, int x_step_qn,
                              ConvolveParams *conv_params) {
  int x, y;
  int filter_size = filter_params.taps;
66
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
  src -= filter_size / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_qn = subpel_x_qn;
    for (x = 0; x < w; ++x) {
      const uint8_t *const src_x = &src[x_qn >> SCALE_SUBPEL_BITS];
      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(x_filter_idx < SUBPEL_SHIFTS);
      const int16_t *x_filter =
          av1_get_interp_filter_subpel_kernel(filter_params, x_filter_idx);
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];

      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
      if (conv_params->do_average)
        dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
      else
        dst[x] = sum;

      x_qn += x_step_qn;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
92 93 94
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,
95 96
                         const int subpel_y_q4, int y_step_q4,
                         ConvolveParams *conv_params) {
97
  int x, y;
98
  int filter_size = filter_params.taps;
Angie Chiang's avatar
Angie Chiang committed
99
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
100 101 102 103 104
  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
105
      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
106
          filter_params, y_q4 & SUBPEL_MASK);
107 108 109
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
110

Angie Chiang's avatar
Angie Chiang committed
111
      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
112
      if (conv_params->do_average)
Angie Chiang's avatar
Angie Chiang committed
113 114 115
        dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
      else
        dst[y * dst_stride] = sum;
116

117 118 119 120 121 122 123
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Fergus Simpson's avatar
Fergus Simpson committed
124 125 126 127 128 129 130
void av1_convolve_vert_scale(const uint8_t *src, int src_stride, uint8_t *dst,
                             int dst_stride, int w, int h,
                             const InterpFilterParams filter_params,
                             const int subpel_y_qn, int y_step_qn,
                             ConvolveParams *conv_params) {
  int x, y;
  int filter_size = filter_params.taps;
131
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
  src -= src_stride * (filter_size / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_qn = subpel_y_qn;
    for (y = 0; y < h; ++y) {
      const uint8_t *const src_y =
          &src[(y_qn >> SCALE_SUBPEL_BITS) * src_stride];
      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(y_filter_idx < SUBPEL_SHIFTS);
      const int16_t *y_filter =
          av1_get_interp_filter_subpel_kernel(filter_params, y_filter_idx);
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k)
        sum += src_y[k * src_stride] * y_filter[k];

      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
      if (conv_params->do_average)
        dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
      else
        dst[y * dst_stride] = sum;

      y_qn += y_step_qn;
    }
    ++src;
    ++dst;
  }
}

Angie Chiang's avatar
Angie Chiang committed
159
static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst,
160 161
                          int dst_stride, int w, int h,
                          ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
162
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
163
  if (conv_params->do_average == 0) {
Angie Chiang's avatar
Angie Chiang committed
164
    int r;
Angie Chiang's avatar
Angie Chiang committed
165
    for (r = 0; r < h; ++r) {
Angie Chiang's avatar
Angie Chiang committed
166
      memcpy(dst, src, w);
Angie Chiang's avatar
Angie Chiang committed
167 168 169 170 171 172
      src += src_stride;
      dst += dst_stride;
    }
  } else {
    int r, c;
    for (r = 0; r < h; ++r) {
173
      for (c = 0; c < w; ++c) {
Angie Chiang's avatar
Angie Chiang committed
174
        dst[c] = clip_pixel(ROUND_POWER_OF_TWO(dst[c] + src[c], 1));
175
      }
Angie Chiang's avatar
Angie Chiang committed
176 177 178 179 180 181
      src += src_stride;
      dst += dst_stride;
    }
  }
}

182 183 184
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,
185 186
                               const int subpel_x_q4, int x_step_q4,
                               ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
187 188
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
  if (filter_params.taps == SUBPEL_TAPS) {
189 190
    const int16_t *filter_x =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
191
    if (conv_params->do_average == 0)
192 193 194 195 196 197
      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 {
Angie Chiang's avatar
Angie Chiang committed
198 199
    av1_convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
                       subpel_x_q4, x_step_q4, conv_params);
200 201 202
  }
}

203 204 205 206 207 208 209 210 211
void av1_convolve_horiz_facade_c(const uint8_t *src, int src_stride,
                                 uint8_t *dst, int dst_stride, int w, int h,
                                 const InterpFilterParams filter_params,
                                 const int subpel_x_q4, int x_step_q4,
                                 ConvolveParams *conv_params) {
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
  if (filter_params.taps == SUBPEL_TAPS) {
    const int16_t *filter_x =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
212
    if (conv_params->do_average == 0)
213 214 215 216 217 218 219 220 221 222 223
      aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
                            x_step_q4, NULL, -1, w, h);
    else
      aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
                                x_step_q4, NULL, -1, w, h);
  } else {
    av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h, filter_params,
                         subpel_x_q4, x_step_q4, conv_params);
  }
}

Fergus Simpson's avatar
Fergus Simpson committed
224 225 226 227 228
void av1_convolve_horiz_facade_scale(const uint8_t *src, int src_stride,
                                     uint8_t *dst, int dst_stride, int w, int h,
                                     const InterpFilterParams filter_params,
                                     const int subpel_x_qn, int x_step_qn,
                                     ConvolveParams *conv_params) {
229
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245
  if (filter_params.taps == SUBPEL_TAPS) {
    const int16_t *filter_x = av1_get_interp_filter_subpel_kernel(
        filter_params, subpel_x_qn >> SCALE_EXTRA_BITS);
    if (conv_params->do_average == 0)
      aom_convolve8_horiz_scale(src, src_stride, dst, dst_stride, filter_x,
                                subpel_x_qn, x_step_qn, NULL, 0, -1, w, h);
    else
      aom_convolve8_avg_horiz_scale(src, src_stride, dst, dst_stride, filter_x,
                                    subpel_x_qn, x_step_qn, NULL, 0, -1, w, h);
  } else {
    av1_convolve_horiz_scale(src, src_stride, dst, dst_stride, w, h,
                             filter_params, subpel_x_qn, x_step_qn,
                             conv_params);
  }
}

246 247 248
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,
249 250
                              const int subpel_y_q4, int y_step_q4,
                              ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
251 252
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
  if (filter_params.taps == SUBPEL_TAPS) {
253 254
    const int16_t *filter_y =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
255
    if (conv_params->do_average == 0) {
256 257 258 259 260 261 262
      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 {
Angie Chiang's avatar
Angie Chiang committed
263 264
    av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
                      subpel_y_q4, y_step_q4, conv_params);
265 266 267
  }
}

268 269 270 271 272 273 274 275 276
void av1_convolve_vert_facade_c(const uint8_t *src, int src_stride,
                                uint8_t *dst, int dst_stride, int w, int h,
                                const InterpFilterParams filter_params,
                                const int subpel_y_q4, int y_step_q4,
                                ConvolveParams *conv_params) {
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
  if (filter_params.taps == SUBPEL_TAPS) {
    const int16_t *filter_y =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
277
    if (conv_params->do_average == 0) {
278 279 280 281 282 283 284 285 286 287 288 289
      aom_convolve8_vert_c(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
                           y_step_q4, w, h);
    } else {
      aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, NULL, -1,
                               filter_y, y_step_q4, w, h);
    }
  } else {
    av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params,
                        subpel_y_q4, y_step_q4, conv_params);
  }
}

Fergus Simpson's avatar
Fergus Simpson committed
290 291 292 293 294
void av1_convolve_vert_facade_scale(const uint8_t *src, int src_stride,
                                    uint8_t *dst, int dst_stride, int w, int h,
                                    const InterpFilterParams filter_params,
                                    const int subpel_y_qn, int y_step_qn,
                                    ConvolveParams *conv_params) {
295
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
  if (filter_params.taps == SUBPEL_TAPS) {
    const int16_t *filter_y = av1_get_interp_filter_subpel_kernel(
        filter_params, subpel_y_qn >> SCALE_EXTRA_BITS);
    if (conv_params->do_average == 0) {
      aom_convolve8_vert_scale(src, src_stride, dst, dst_stride, NULL, 0, -1,
                               filter_y, subpel_y_qn, y_step_qn, w, h);
    } else {
      aom_convolve8_avg_vert_scale(src, src_stride, dst, dst_stride, NULL, 0,
                                   -1, filter_y, subpel_y_qn, y_step_qn, w, h);
    }
  } else {
    av1_convolve_vert_scale(src, src_stride, dst, dst_stride, w, h,
                            filter_params, subpel_y_qn, y_step_qn, conv_params);
  }
}

Angie Chiang's avatar
Angie Chiang committed
312
#if CONFIG_CONVOLVE_ROUND
313 314
void av1_convolve_rounding_c(const int32_t *src, int src_stride, uint8_t *dst,
                             int dst_stride, int w, int h, int bits) {
Angie Chiang's avatar
Angie Chiang committed
315 316 317 318
  int r, c;
  for (r = 0; r < h; ++r) {
    for (c = 0; c < w; ++c) {
      dst[r * dst_stride + c] =
319
          clip_pixel(ROUND_POWER_OF_TWO(src[r * src_stride + c], bits));
Angie Chiang's avatar
Angie Chiang committed
320 321 322
    }
  }
}
Angie Chiang's avatar
Angie Chiang committed
323

324
#if CONFIG_COMPOUND_ROUND
David Barker's avatar
David Barker committed
325 326 327 328 329 330
void av1_convolve_2d_c(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst,
                       int dst_stride, int w, int h,
                       InterpFilterParams *filter_params_x,
                       InterpFilterParams *filter_params_y,
                       const int subpel_x_q4, const int subpel_y_q4,
                       ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
331
  int x, y, k;
332
  uint8_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
Angie Chiang's avatar
Angie Chiang committed
333 334 335 336
  int im_h = h + filter_params_y->taps - 1;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;
337

Angie Chiang's avatar
Angie Chiang committed
338 339 340 341 342 343
  // horizontal filter
  const uint8_t *src_horiz = src - fo_vert * src_stride;
  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
  for (y = 0; y < im_h; ++y) {
    for (x = 0; x < w; ++x) {
344
      int32_t sum = 0;
Angie Chiang's avatar
Angie Chiang committed
345 346 347
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
      }
Angie Chiang's avatar
Angie Chiang committed
348
      im_block[y * im_stride + x] =
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367
          clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
    }
  }

  // vertical filter
  uint8_t *src_vert = im_block + fo_vert * im_stride;
  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      CONV_BUF_TYPE sum = 0;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
      dst[y * dst_stride + x] += res;
    }
  }
}
368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424

void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
                             CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
                             InterpFilterParams *filter_params_x,
                             InterpFilterParams *filter_params_y,
                             const int subpel_x_qn, const int x_step_qn,
                             const int subpel_y_qn, const int y_step_qn,
                             ConvolveParams *conv_params) {
  int x, y, k;
  uint8_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
             filter_params_y->taps;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;

  // horizontal filter
  const uint8_t *src_horiz = src - fo_vert * src_stride;
  for (y = 0; y < im_h; ++y) {
    int x_qn = subpel_x_qn;
    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
      const uint8_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(x_filter_idx < SUBPEL_SHIFTS);
      const int16_t *x_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
      int sum = 0;
      for (k = 0; k < filter_params_x->taps; ++k)
        sum += x_filter[k] * src_x[k - fo_horiz];
      im_block[y * im_stride + x] =
          clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
    }
    src_horiz += src_stride;
  }

  // vertical filter
  const uint8_t *src_vert = im_block + fo_vert * im_stride;
  for (x = 0; x < w; ++x) {
    int y_qn = subpel_y_qn;
    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
      const uint8_t *const src_y =
          &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(y_filter_idx < SUBPEL_SHIFTS);
      const int16_t *y_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
      CONV_BUF_TYPE sum = 0;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
      }
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
      dst[y * dst_stride + x] += res;
    }
    src_vert++;
  }
}

Angie Chiang's avatar
Angie Chiang committed
425
#else
426

427 428 429 430 431 432
/* When convolve-round is enabled and compound-round is disabled, we use a
   high-precision convolve filter.
   Note: For notes on hardware implementations, including the required
   bit widths for various intermediate values, see the comments above
   av1_warp_affine_c.
*/
David Barker's avatar
David Barker committed
433 434 435 436 437 438
void av1_convolve_2d_c(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst,
                       int dst_stride, int w, int h,
                       InterpFilterParams *filter_params_x,
                       InterpFilterParams *filter_params_y,
                       const int subpel_x_q4, const int subpel_y_q4,
                       ConvolveParams *conv_params) {
439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457
  int x, y, k;
  int32_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
  int im_h = h + filter_params_y->taps - 1;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;
  const int bd = 8;

  // horizontal filter
  const uint8_t *src_horiz = src - fo_vert * src_stride;
  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
  for (y = 0; y < im_h; ++y) {
    for (x = 0; x < w; ++x) {
      int32_t sum = (1 << (bd + FILTER_BITS - 1));
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
      }
      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
Angie Chiang's avatar
Angie Chiang committed
458
      im_block[y * im_stride + x] =
459
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
Angie Chiang's avatar
Angie Chiang committed
460 461 462 463
    }
  }

  // vertical filter
464
  int32_t *src_vert = im_block + fo_vert * im_stride;
Angie Chiang's avatar
Angie Chiang committed
465 466
  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
467
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
Angie Chiang's avatar
Angie Chiang committed
468 469
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
470
      CONV_BUF_TYPE sum = 1 << offset_bits;
Angie Chiang's avatar
Angie Chiang committed
471 472 473
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
474 475 476 477 478
      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
                          ((1 << (offset_bits - conv_params->round_1)) +
                           (1 << (offset_bits - conv_params->round_1 - 1)));
      dst[y * dst_stride + x] += res;
Angie Chiang's avatar
Angie Chiang committed
479 480 481
    }
  }
}
482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544

void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
                             CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
                             InterpFilterParams *filter_params_x,
                             InterpFilterParams *filter_params_y,
                             const int subpel_x_qn, const int x_step_qn,
                             const int subpel_y_qn, const int y_step_qn,
                             ConvolveParams *conv_params) {
  int x, y, k;
  int32_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
             filter_params_y->taps;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;
  const int bd = 8;

  // horizontal filter
  const uint8_t *src_horiz = src - fo_vert * src_stride;
  for (y = 0; y < im_h; ++y) {
    int x_qn = subpel_x_qn;
    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
      const uint8_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(x_filter_idx < SUBPEL_SHIFTS);
      const int16_t *x_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
      int32_t sum = (1 << (bd + FILTER_BITS - 1));
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_x[k - fo_horiz];
      }
      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
      im_block[y * im_stride + x] =
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
    }
    src_horiz += src_stride;
  }

  // vertical filter
  int32_t *src_vert = im_block + fo_vert * im_stride;
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
  for (x = 0; x < w; ++x) {
    int y_qn = subpel_y_qn;
    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
      const int32_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(y_filter_idx < SUBPEL_SHIFTS);
      const int16_t *y_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
      CONV_BUF_TYPE sum = 1 << offset_bits;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
      }
      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
                          ((1 << (offset_bits - conv_params->round_1)) +
                           (1 << (offset_bits - conv_params->round_1 - 1)));
      dst[y * dst_stride + x] += res;
    }
    src_vert++;
  }
}
#endif  // CONFIG_COMPOUND_ROUND
Angie Chiang's avatar
Angie Chiang committed
545 546 547 548 549

void av1_convolve_2d_facade(const uint8_t *src, int src_stride, uint8_t *dst,
                            int dst_stride, int w, int h,
                            const InterpFilter *interp_filter,
                            const int subpel_x_q4, int x_step_q4,
550
                            const int subpel_y_q4, int y_step_q4, int scaled,
Angie Chiang's avatar
Angie Chiang committed
551 552 553 554 555
                            ConvolveParams *conv_params) {
  (void)x_step_q4;
  (void)y_step_q4;
  (void)dst;
  (void)dst_stride;
556
#if CONFIG_DUAL_FILTER
Angie Chiang's avatar
Angie Chiang committed
557 558 559 560
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
561

562
#if USE_EXTRA_FILTER
563 564 565 566 567 568
  if (filter_params_x.interp_filter == MULTITAP_SHARP &&
      filter_params_y.interp_filter == 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);
  }
569
#endif  // USE_EXTRA_FILTER
570 571 572 573 574
#else
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(*interp_filter);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(*interp_filter);
575
#endif
576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592

  if (filter_params_y.taps < filter_params_x.taps) {
    uint8_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
                   (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
    int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
    CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
    int tr_dst_stride = MAX_SB_SIZE;
    int fo_vert = filter_params_y.taps / 2 - 1;
    int fo_horiz = filter_params_x.taps / 2 - 1;

    transpose_uint8(tr_src, tr_src_stride,
                    src - fo_vert * src_stride - fo_horiz, src_stride,
                    w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
    transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
                    conv_params->dst_stride, w, h);

    // horizontal and vertical parameters are swapped because of the transpose
593 594 595 596 597 598 599 600 601 602
    if (scaled)
      av1_convolve_2d_scale(tr_src + fo_horiz * tr_src_stride + fo_vert,
                            tr_src_stride, tr_dst, tr_dst_stride, h, w,
                            &filter_params_y, &filter_params_x, subpel_y_q4,
                            y_step_q4, subpel_x_q4, x_step_q4, conv_params);
    else
      av1_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
                      tr_src_stride, tr_dst, tr_dst_stride, h, w,
                      &filter_params_y, &filter_params_x, subpel_y_q4,
                      subpel_x_q4, conv_params);
603 604 605
    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
                    tr_dst_stride, h, w);
  } else {
606 607 608 609 610 611 612 613 614
    if (scaled)
      av1_convolve_2d_scale(src, src_stride, conv_params->dst,
                            conv_params->dst_stride, w, h, &filter_params_x,
                            &filter_params_y, subpel_x_q4, x_step_q4,
                            subpel_y_q4, y_step_q4, conv_params);
    else
      av1_convolve_2d(src, src_stride, conv_params->dst,
                      conv_params->dst_stride, w, h, &filter_params_x,
                      &filter_params_y, subpel_x_q4, subpel_y_q4, conv_params);
615
  }
Angie Chiang's avatar
Angie Chiang committed
616 617
}

618
#if CONFIG_HIGHBITDEPTH
619 620 621
void av1_highbd_convolve_rounding_c(const int32_t *src, int src_stride,
                                    uint8_t *dst8, int dst_stride, int w, int h,
                                    int bits, int bd) {
622 623 624 625 626
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  int r, c;
  for (r = 0; r < h; ++r) {
    for (c = 0; c < w; ++c) {
      dst[r * dst_stride + c] = clip_pixel_highbd(
627
          ROUND_POWER_OF_TWO(src[r * src_stride + c], bits), bd);
628 629 630 631
    }
  }
}

632
#if CONFIG_COMPOUND_ROUND
David Barker's avatar
David Barker committed
633 634 635 636 637 638
void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride,
                              CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
                              InterpFilterParams *filter_params_x,
                              InterpFilterParams *filter_params_y,
                              const int subpel_x_q4, const int subpel_y_q4,
                              ConvolveParams *conv_params, int bd) {
639
  int x, y, k;
640
  uint16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
641 642 643 644
  int im_h = h + filter_params_y->taps - 1;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;
645

646 647 648 649 650 651
  // horizontal filter
  const uint16_t *src_horiz = src - fo_vert * src_stride;
  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
  for (y = 0; y < im_h; ++y) {
    for (x = 0; x < w; ++x) {
652
      int32_t sum = 0;
653 654 655
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
      }
656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675
      im_block[y * im_stride + x] =
          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, conv_params->round_0), bd);
    }
  }

  // vertical filter
  uint16_t *src_vert = im_block + fo_vert * im_stride;
  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      CONV_BUF_TYPE sum = 0;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
      dst[y * dst_stride + x] += res;
    }
  }
}
676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733

void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride,
                                    CONV_BUF_TYPE *dst, int dst_stride, int w,
                                    int h, InterpFilterParams *filter_params_x,
                                    InterpFilterParams *filter_params_y,
                                    const int subpel_x_qn, const int x_step_qn,
                                    const int subpel_y_qn, const int y_step_qn,
                                    ConvolveParams *conv_params, int bd) {
  int x, y, k;
  uint16_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
             filter_params_y->taps;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;
  (void)bd;

  // horizontal filter
  const uint16_t *src_horiz = src - fo_vert * src_stride;
  for (y = 0; y < im_h; ++y) {
    int x_qn = subpel_x_qn;
    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
      const uint16_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(x_filter_idx < SUBPEL_SHIFTS);
      const int16_t *x_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
      int sum = 0;
      for (k = 0; k < filter_params_x->taps; ++k)
        sum += x_filter[k] * src_x[k - fo_horiz];
      im_block[y * im_stride + x] =
          clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
    }
    src_horiz += src_stride;
  }

  // vertical filter
  uint16_t *src_vert = im_block + fo_vert * im_stride;
  for (x = 0; x < w; ++x) {
    int y_qn = subpel_y_qn;
    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
      const uint16_t *const src_y =
          &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(y_filter_idx < SUBPEL_SHIFTS);
      const int16_t *y_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
      CONV_BUF_TYPE sum = 0;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
      }
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
      dst[y * dst_stride + x] += res;
    }
    src_vert++;
  }
}

734
#else
735

David Barker's avatar
David Barker committed
736 737 738 739 740 741
void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride,
                              CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
                              InterpFilterParams *filter_params_x,
                              InterpFilterParams *filter_params_y,
                              const int subpel_x_q4, const int subpel_y_q4,
                              ConvolveParams *conv_params, int bd) {
742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759
  int x, y, k;
  int32_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
  int im_h = h + filter_params_y->taps - 1;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;

  // horizontal filter
  const uint16_t *src_horiz = src - fo_vert * src_stride;
  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
  for (y = 0; y < im_h; ++y) {
    for (x = 0; x < w; ++x) {
      int32_t sum = (1 << (bd + FILTER_BITS - 1));
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
      }
      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
760 761
      (void)bd;
      im_block[y * im_stride + x] =
762
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
763 764 765 766
    }
  }

  // vertical filter
767 768
  int32_t *src_vert = im_block + fo_vert * im_stride;
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
769 770 771 772
  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
773
      CONV_BUF_TYPE sum = 1 << offset_bits;
774 775 776
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
777 778 779 780 781
      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
                          ((1 << (offset_bits - conv_params->round_1)) +
                           (1 << (offset_bits - conv_params->round_1 - 1)));
      dst[y * dst_stride + x] += res;
782 783 784
    }
  }
}
785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846

void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride,
                                    CONV_BUF_TYPE *dst, int dst_stride, int w,
                                    int h, InterpFilterParams *filter_params_x,
                                    InterpFilterParams *filter_params_y,
                                    const int subpel_x_qn, const int x_step_qn,
                                    const int subpel_y_qn, const int y_step_qn,
                                    ConvolveParams *conv_params, int bd) {
  int x, y, k;
  int32_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
             filter_params_y->taps;
  int im_stride = w;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int fo_horiz = filter_params_x->taps / 2 - 1;

  // horizontal filter
  const uint16_t *src_horiz = src - fo_vert * src_stride;
  for (y = 0; y < im_h; ++y) {
    int x_qn = subpel_x_qn;
    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
      const uint16_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(x_filter_idx < SUBPEL_SHIFTS);
      const int16_t *x_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
      int32_t sum = (1 << (bd + FILTER_BITS - 1));
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_x[k - fo_horiz];
      }
      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
      im_block[y * im_stride + x] =
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
    }
    src_horiz += src_stride;
  }

  // vertical filter
  int32_t *src_vert = im_block + fo_vert * im_stride;
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
  for (x = 0; x < w; ++x) {
    int y_qn = subpel_y_qn;
    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
      const int32_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
      assert(y_filter_idx < SUBPEL_SHIFTS);
      const int16_t *y_filter =
          av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
      CONV_BUF_TYPE sum = 1 << offset_bits;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
      }
      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
                          ((1 << (offset_bits - conv_params->round_1)) +
                           (1 << (offset_bits - conv_params->round_1 - 1)));
      dst[y * dst_stride + x] += res;
    }
    src_vert++;
  }
}
#endif  // CONFIG_COMPOUND_ROUND
847 848 849 850 851 852

void av1_highbd_convolve_2d_facade(const uint8_t *src8, int src_stride,
                                   uint8_t *dst, int dst_stride, int w, int h,
                                   const InterpFilter *interp_filter,
                                   const int subpel_x_q4, int x_step_q4,
                                   const int subpel_y_q4, int y_step_q4,
853 854
                                   int scaled, ConvolveParams *conv_params,
                                   int bd) {
855 856 857 858 859 860 861 862 863 864
  (void)x_step_q4;
  (void)y_step_q4;
  (void)dst;
  (void)dst_stride;
#if CONFIG_DUAL_FILTER
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);

865
#if USE_EXTRA_FILTER
866 867 868 869 870 871
  if (filter_params_x.interp_filter == MULTITAP_SHARP &&
      filter_params_y.interp_filter == 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);
  }
872
#endif
873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895
#else
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(*interp_filter);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(*interp_filter);
#endif
  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  if (filter_params_y.taps < filter_params_x.taps) {
    uint16_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
                    (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
    int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
    CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
    int tr_dst_stride = MAX_SB_SIZE;
    int fo_vert = filter_params_y.taps / 2 - 1;
    int fo_horiz = filter_params_x.taps / 2 - 1;

    transpose_uint16(
        tr_src, tr_src_stride, src - fo_vert * src_stride - fo_horiz,
        src_stride, w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
    transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
                    conv_params->dst_stride, w, h);

    // horizontal and vertical parameters are swapped because of the transpose
896 897 898 899 900 901 902 903 904 905
    if (scaled)
      av1_highbd_convolve_2d_scale(
          tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride, tr_dst,
          tr_dst_stride, h, w, &filter_params_y, &filter_params_x, subpel_y_q4,
          y_step_q4, subpel_x_q4, x_step_q4, conv_params, bd);
    else
      av1_highbd_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
                             tr_src_stride, tr_dst, tr_dst_stride, h, w,
                             &filter_params_y, &filter_params_x, subpel_y_q4,
                             subpel_x_q4, conv_params, bd);
906 907 908
    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
                    tr_dst_stride, h, w);
  } else {
909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978
    if (scaled)
      av1_highbd_convolve_2d_scale(
          src, src_stride, conv_params->dst, conv_params->dst_stride, w, h,
          &filter_params_x, &filter_params_y, subpel_x_q4, x_step_q4,
          subpel_y_q4, y_step_q4, conv_params, bd);
    else
      av1_highbd_convolve_2d(src, src_stride, conv_params->dst,
                             conv_params->dst_stride, w, h, &filter_params_x,
                             &filter_params_y, subpel_x_q4, subpel_y_q4,
                             conv_params, bd);
  }
}

void av1_highbd_convolve_2d_facade_scale(const uint8_t *src8, int src_stride,
                                         uint8_t *dst, int dst_stride, int w,
                                         int h,
                                         const InterpFilter *interp_filter,
                                         const int subpel_x_qn, int x_step_qn,
                                         const int subpel_y_qn, int y_step_qn,
                                         ConvolveParams *conv_params, int bd) {
  (void)dst;
  (void)dst_stride;
#if CONFIG_DUAL_FILTER
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);

#if USE_EXTRA_FILTER
  if (filter_params_x.interp_filter == MULTITAP_SHARP &&
      filter_params_y.interp_filter == 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);
  }
#endif
#else
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(*interp_filter);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(*interp_filter);
#endif
  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  if (filter_params_y.taps < filter_params_x.taps) {
    uint16_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
                    (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
    int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
    CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
    int tr_dst_stride = MAX_SB_SIZE;
    int fo_vert = filter_params_y.taps / 2 - 1;
    int fo_horiz = filter_params_x.taps / 2 - 1;

    transpose_uint16(
        tr_src, tr_src_stride, src - fo_vert * src_stride - fo_horiz,
        src_stride, w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
    transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
                    conv_params->dst_stride, w, h);

    // horizontal and vertical parameters are swapped because of the transpose
    av1_highbd_convolve_2d_scale(
        tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride, tr_dst,
        tr_dst_stride, h, w, &filter_params_y, &filter_params_x, subpel_y_qn,
        y_step_qn, subpel_x_qn, x_step_qn, conv_params, bd);
    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
                    tr_dst_stride, h, w);
  } else {
    av1_highbd_convolve_2d_scale(
        src, src_stride, conv_params->dst, conv_params->dst_stride, w, h,
        &filter_params_x, &filter_params_y, subpel_x_qn, x_step_qn, subpel_y_qn,
        y_step_qn, conv_params, bd);
979 980 981 982
  }
}
#endif  // CONFIG_HIGHBITDEPTH

Angie Chiang's avatar
Angie Chiang committed
983
#endif  // CONFIG_CONVOLVE_ROUND
Angie Chiang's avatar
Angie Chiang committed
984

985 986 987 988 989 990 991 992
typedef void (*ConvolveFunc)(const uint8_t *src, int src_stride, uint8_t *dst,
                             int dst_stride, int w, int h,
                             const InterpFilterParams filter_params,
                             const int subpel_q4, int step_q4,
                             ConvolveParams *conv_params);

static void convolve_helper(const uint8_t *src, int src_stride, uint8_t *dst,
                            int dst_stride, int w, int h,
993
#if CONFIG_DUAL_FILTER
994
                            const InterpFilter *interp_filter,
995
#else
996
                            const InterpFilter interp_filter,
997
#endif
998 999 1000 1001 1002
                            const int subpel_x_q4, int x_step_q4,
                            const int subpel_y_q4, int y_step_q4,
                            ConvolveParams *conv_params,
                            ConvolveFunc convolve_horiz,
                            ConvolveFunc convolve_vert) {
Fergus Simpson's avatar
Fergus Simpson committed
1003 1004
  int ignore_horiz = x_step_q4 == SUBPEL_SHIFTS && subpel_x_q4 == 0;
  int ignore_vert = y_step_q4 == SUBPEL_SHIFTS && subpel_y_q4 == 0;
Angie Chiang's avatar
Angie Chiang committed
1005 1006 1007 1008 1009 1010 1011 1012 1013 1014
#if CONFIG_DUAL_FILTER
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
  InterpFilterParams filter_params;
#else
  InterpFilterParams filter_params =
      av1_get_interp_filter_params(interp_filter);
#endif
Angie Chiang's avatar
Angie Chiang committed
1015
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
1016 1017 1018 1019 1020 1021

  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
1022
  if (ignore_horiz && ignore_vert) {
1023
    convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params);
Angie Chiang's avatar
Angie Chiang committed
1024
  } else if (ignore_vert) {
1025
#if CONFIG_DUAL_FILTER
Angie Chiang's avatar
Angie Chiang committed
1026
    filter_params = filter_params_x;
1027 1028
#endif
    assert(filter_params.taps <= MAX_FILTER_TAP);
1029 1030
    convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
                   subpel_x_q4, x_step_q4, conv_params);
Angie Chiang's avatar
Angie Chiang committed
1031
  } else if (ignore_horiz) {
1032
#if CONFIG_DUAL_FILTER
Angie Chiang's avatar
Angie Chiang committed
1033
    filter_params = filter_params_y;
1034 1035
#endif
    assert(filter_params.taps <= MAX_FILTER_TAP);
1036 1037
    convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
                  subpel_y_q4, y_step_q4, conv_params);
Angie Chiang's avatar
Angie Chiang committed
1038
  } else {
1039 1040 1041 1042 1043 1044
    // 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);
1045
    int filter_size;
1046
#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
1047 1048
    if (interp_filter[0 + 2 * conv_params->ref] == MULTITAP_SHARP &&
        interp_filter[1 + 2 * conv_params->ref] == MULTITAP_SHARP) {
1049 1050 1051 1052
      // Avoid two directions both using 12-tap filter.
      // This will reduce hardware implementation cost.
      filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
    }
1053

1054 1055 1056 1057
    // we do filter with fewer taps first to reduce hardware implementation
    // complexity
    if (filter_params_y.taps < filter_params_x.taps) {
      int intermediate_width;
1058
      int temp_stride = max_intermediate_size;
1059 1060
      ConvolveParams temp_conv_params;
      temp_conv_params.ref = 0;
1061
      temp_conv_params.do_average = 0;
1062
      temp_conv_params.round = CONVOLVE_OPT_ROUND;
1063 1064 1065 1066
      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;
1067
      assert(intermediate_width <= max_intermediate_size);
Angie Chiang's avatar
Angie Chiang committed
1068

1069
      assert(filter_params.taps <= MAX_FILTER_TAP);
1070

1071 1072 1073
      convolve_vert(src - (filter_size / 2 - 1), src_stride, temp, temp_stride,
                    intermediate_width, h, filter_params, subpel_y_q4,
                    y_step_q4, &temp_conv_params);
1074

1075 1076
      filter_params = filter_params_x;
      assert(filter_params.taps <= MAX_FILTER_TAP);
1077 1078
      convolve_horiz(temp + (filter_size / 2 - 1), temp_stride, dst, dst_stride,
                     w, h, filter_params, subpel_x_q4, x_step_q4, conv_params);
1079
    } else
1080
#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
1081 1082
    {
      int intermediate_height;
1083
      int temp_stride = MAX_SB_SIZE;
1084 1085
      ConvolveParams temp_conv_params;
      temp_conv_params.ref = 0;
1086
      temp_conv_params.do_average = 0;
1087
      temp_conv_params.round = CONVOLVE_OPT_ROUND;
1088 1089 1090 1091 1092 1093 1094 1095
#if CONFIG_DUAL_FILTER
      filter_params = filter_params_x;
      filter_size = filter_params_y.taps;
#else
      filter_size = filter_params.taps;
#endif
      intermediate_height =
          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
1096 1097
      assert(intermediate_height <= max_intermediate_size);
      (void)max_intermediate_size;
1098 1099 1100

      assert(filter_params.taps <= MAX_FILTER_TAP);

1101 1102 1103
      convolve_horiz(src - src_stride * (filter_size / 2 - 1), src_stride, temp,
                     temp_stride, w, intermediate_height, filter_params,
                     subpel_x_q4, x_step_q4, &temp_conv_params);
1104 1105 1106 1107 1108 1109

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

1110 1111 1112
      convolve_vert(temp + temp_stride * (filter_size / 2 - 1), temp_stride,
                    dst, dst_stride, w, h, filter_params, subpel_y_q4,
                    y_step_q4, conv_params);
1113
    }
Angie Chiang's avatar
Angie Chiang committed
1114
  }
1115 1116
}

Fergus Simpson's avatar
Fergus Simpson committed
1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140
static void convolve_scale_helper(const uint8_t *src, int src_stride,
                                  uint8_t *dst, int dst_stride, int w, int h,
#if CONFIG_DUAL_FILTER
                                  const InterpFilter *interp_filter,
#else
                                  const InterpFilter interp_filter,
#endif
                                  const int subpel_x_qn, int x_step_qn,
                                  const int subpel_y_qn, int y_step_qn,
                                  ConvolveParams *conv_params,
                                  ConvolveFunc convolve_horiz,
                                  ConvolveFunc convolve_vert) {
  int ignore_horiz = x_step_qn == SCALE_SUBPEL_SHIFTS && subpel_x_qn == 0;
  int ignore_vert = y_step_qn == SCALE_SUBPEL_SHIFTS && subpel_y_qn == 0;
#if CONFIG_DUAL_FILTER
  InterpFilterParams filter_params_x =
      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
  InterpFilterParams filter_params_y =
      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
  InterpFilterParams filter_params;
#else
  InterpFilterParams filter_params =
      av1_get_interp_filter_params(interp_filter);
#endif
1141
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245

  assert(w <= MAX_BLOCK_WIDTH);
  assert(h <= MAX_BLOCK_HEIGHT);
  assert(y_step_qn <= (MAX_STEP << SCALE_EXTRA_BITS));
  assert(x_step_qn <= (MAX_STEP << SCALE_EXTRA_BITS));

  if (ignore_horiz && ignore_vert) {
    convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params);
  } else if (ignore_vert) {
#if CONFIG_DUAL_FILTER
    filter_params = filter_params_x;
#endif
    assert(filter_params.taps <= MAX_FILTER_TAP);
    convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
                   subpel_x_qn, x_step_qn, conv_params);
  } else if (ignore_horiz) {
#if CONFIG_DUAL_FILTER
    filter_params = filter_params_y;
#endif
    assert(filter_params.taps <= MAX_FILTER_TAP);
    convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
                  subpel_y_qn, y_step_qn, conv_params);
  } else {
    // 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);
    int filter_size;
#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
    if (interp_filter[0 + 2 * conv_params->ref] == MULTITAP_SHARP &&
        interp_filter[1 + 2 * conv_params->ref] == 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);
    }

    // we do filter with fewer taps first to reduce hardware implementation
    // complexity
    if (filter_params_y.taps < filter_params_x.taps) {
      int intermediate_width;
      int temp_stride = max_intermediate_size;
      ConvolveParams temp_conv_params;
      temp_conv_params.ref = 0;
      temp_conv_params.do_average = 0;
      temp_conv_params.round = CONVOLVE_OPT_ROUND;
      filter_params = filter_params_y;
      filter_size = filter_params_x.taps;
      intermediate_width =
          (((w - 1) * x_step_qn + subpel_x_qn) >> SCALE_SUBPEL_BITS) +
          filter_size;
      assert(intermediate_width <= max_intermediate_size);

      assert(filter_params.taps <= MAX_FILTER_TAP);

      convolve_vert(src - (filter_size / 2 - 1), src_stride, temp, temp_stride,
                    intermediate_width, h, filter_params, subpel_y_qn,
                    y_step_qn, &temp_conv_params);

      filter_params = filter_params_x;
      assert(filter_params.taps <= MAX_FILTER_TAP);
      convolve_horiz(temp + (filter_size / 2 - 1), temp_stride, dst, dst_stride,
                     w, h, filter_params, subpel_x_qn, x_step_qn, conv_params);
    } else {
#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
      int intermediate_height;
      int temp_stride = MAX_SB_SIZE;
      ConvolveParams temp_conv_params;
      temp_conv_params.ref = 0;
      temp_conv_params.do_average = 0;
      temp_conv_params.round = CONVOLVE_OPT_ROUND;
#if CONFIG_DUAL_FILTER
      filter_params = filter_params_x;
      filter_size = filter_params_y.taps;
#else
    filter_size = filter_params.taps;
#endif
      intermediate_height =
          (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
          filter_size;
      assert(intermediate_height <= max_intermediate_size);
      (void)max_intermediate_size;

      assert(filter_params.taps <= MAX_FILTER_TAP);

      convolve_horiz(src - src_stride * (filter_size / 2 - 1), src_stride, temp,
                     temp_stride, w, intermediate_height, filter_params,
                     subpel_x_qn, x_step_qn, &temp_conv_params);

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

      convolve_vert(temp + temp_stride * (filter_size / 2 - 1), temp_stride,
                    dst, dst_stride, w, h, filter_params, subpel_y_qn,
                    y_step_qn, conv_params);
#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
    }
#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
  }
}

1246 1247