convolve.c 79 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)

28
#if CONFIG_FRAME_SUPERRES
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81

#define UPSCALE_PROC_UNIT 64  // Source step (roughly)
#define UPSCALE_PROC_UNIT_SCALE (UPSCALE_PROC_UNIT / SCALE_NUMERATOR)

void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst,
                             int dst_stride, int w, int h,
                             const int16_t *x_filters, int interp_taps,
                             const int x0_qn, const int x_step_qn) {
  int x, y;
  src -= interp_taps / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_qn = x0_qn;
    for (x = 0; x < w; ++x) {
      const uint8_t *const src_x = &src[x_qn >> RS_SCALE_SUBPEL_BITS];
      const int x_filter_idx =
          (x_qn & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
      assert(x_filter_idx <= RS_SUBPEL_MASK);
      const int16_t *const x_filter = &x_filters[x_filter_idx * interp_taps];
      int k, sum = 0;
      for (k = 0; k < interp_taps; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
      x_qn += x_step_qn;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

#if CONFIG_HIGHBITDEPTH
void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride,
                                    uint16_t *dst, int dst_stride, int w, int h,
                                    const int16_t *x_filters, int interp_taps,
                                    int x0_qn, int x_step_qn, int bd) {
  int x, y;
  src -= interp_taps / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_qn = x0_qn;
    for (x = 0; x < w; ++x) {
      const uint16_t *const src_x = &src[x_qn >> RS_SCALE_SUBPEL_BITS];
      const int x_filter_idx =
          (x_qn & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
      assert(x_filter_idx <= RS_SUBPEL_MASK);
      const int16_t *const x_filter = &x_filters[x_filter_idx * interp_taps];
      int k, sum = 0;
      for (k = 0; k < interp_taps; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
      x_qn += x_step_qn;
    }
    src += src_stride;
    dst += dst_stride;
  }
}
#endif  // CONFIG_HIGHBITDEPTH
82
#endif  // CONFIG_FRAME_SUPERRES
83

Yaowu Xu's avatar
Yaowu Xu committed
84 85 86
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,
87 88
                          const int subpel_x_q4, int x_step_q4,
                          ConvolveParams *conv_params) {
89
  int x, y;
90
  int filter_size = filter_params.taps;
Angie Chiang's avatar
Angie Chiang committed
91
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
92 93 94 95 96
  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
97
      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
98
          filter_params, x_q4 & SUBPEL_MASK);
99 100
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
101

Angie Chiang's avatar
Angie Chiang committed
102
      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
103
      if (conv_params->do_average)
Angie Chiang's avatar
Angie Chiang committed
104 105 106
        dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
      else
        dst[x] = sum;
107

108 109 110 111 112 113 114
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Fergus Simpson's avatar
Fergus Simpson committed
115 116 117 118 119 120 121
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;
122
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147
  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
148 149 150
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,
151 152
                         const int subpel_y_q4, int y_step_q4,
                         ConvolveParams *conv_params) {
153
  int x, y;
154
  int filter_size = filter_params.taps;
Angie Chiang's avatar
Angie Chiang committed
155
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
156 157 158 159 160
  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
161
      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
clang-format's avatar
clang-format committed
162
          filter_params, y_q4 & SUBPEL_MASK);
163 164 165
      int k, sum = 0;
      for (k = 0; k < filter_size; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
166

Angie Chiang's avatar
Angie Chiang committed
167
      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
168
      if (conv_params->do_average)
Angie Chiang's avatar
Angie Chiang committed
169 170 171
        dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
      else
        dst[y * dst_stride] = sum;
172

173 174 175 176 177 178 179
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

180 181 182 183 184
static 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) {
Fergus Simpson's avatar
Fergus Simpson committed
185 186
  int x, y;
  int filter_size = filter_params.taps;
187
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
  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
215
static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst,
216 217
                          int dst_stride, int w, int h,
                          ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
218
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
219
  if (conv_params->do_average == 0) {
Angie Chiang's avatar
Angie Chiang committed
220
    int r;
Angie Chiang's avatar
Angie Chiang committed
221
    for (r = 0; r < h; ++r) {
Angie Chiang's avatar
Angie Chiang committed
222
      memcpy(dst, src, w);
Angie Chiang's avatar
Angie Chiang committed
223 224 225 226 227 228
      src += src_stride;
      dst += dst_stride;
    }
  } else {
    int r, c;
    for (r = 0; r < h; ++r) {
229
      for (c = 0; c < w; ++c) {
Angie Chiang's avatar
Angie Chiang committed
230
        dst[c] = clip_pixel(ROUND_POWER_OF_TWO(dst[c] + src[c], 1));
231
      }
Angie Chiang's avatar
Angie Chiang committed
232 233 234 235 236 237
      src += src_stride;
      dst += dst_stride;
    }
  }
}

238 239 240 241 242 243
static 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,
                                      const int subpel_x_q4, int x_step_q4,
                                      ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
244 245
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
  if (filter_params.taps == SUBPEL_TAPS) {
246 247
    const int16_t *filter_x =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
248
    if (conv_params->do_average == 0)
249 250 251 252 253 254
      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
255 256
    av1_convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
                       subpel_x_q4, x_step_q4, conv_params);
257 258 259
  }
}

260 261 262 263 264 265
static 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) {
266 267 268 269
  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);
270
    if (conv_params->do_average == 0)
271 272 273 274 275 276 277 278 279 280 281
      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
282 283 284 285 286
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) {
287
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
  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);
  }
}

304 305 306 307 308
static 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,
                                     const int subpel_y_q4, int y_step_q4,
                                     ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
309 310
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
  if (filter_params.taps == SUBPEL_TAPS) {
311 312
    const int16_t *filter_y =
        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
313
    if (conv_params->do_average == 0) {
314 315 316 317 318 319 320
      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
321 322
    av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
                      subpel_y_q4, y_step_q4, conv_params);
323 324 325
  }
}

326 327 328 329 330 331
static 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) {
332 333 334 335
  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);
336
    if (conv_params->do_average == 0) {
337 338 339 340 341 342 343 344 345 346 347 348
      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
349 350 351 352 353
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) {
354
  assert(conv_params->round == CONVOLVE_OPT_ROUND);
Fergus Simpson's avatar
Fergus Simpson committed
355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370
  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
371
#if CONFIG_CONVOLVE_ROUND
372 373
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
374 375 376 377
  int r, c;
  for (r = 0; r < h; ++r) {
    for (c = 0; c < w; ++c) {
      dst[r * dst_stride + c] =
378
          clip_pixel(ROUND_POWER_OF_TWO(src[r * src_stride + c], bits));
Angie Chiang's avatar
Angie Chiang committed
379 380 381
    }
  }
}
Angie Chiang's avatar
Angie Chiang committed
382

383
#if CONFIG_COMPOUND_ROUND
David Barker's avatar
David Barker committed
384 385 386 387 388 389
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
390
  int x, y, k;
391
  uint8_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
Angie Chiang's avatar
Angie Chiang committed
392 393 394 395
  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;
396

Angie Chiang's avatar
Angie Chiang committed
397 398 399 400 401 402
  // 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) {
403
      int32_t sum = 0;
Angie Chiang's avatar
Angie Chiang committed
404 405 406
      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
407
      im_block[y * im_stride + x] =
408 409 410 411 412 413 414 415 416 417 418 419 420 421 422
          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);
423 424 425 426 427 428 429 430
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
    }
  }
}

431
#if CONFIG_JNT_COMP
432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470
void av1_jnt_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) {
  int x, y, k;
  uint8_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 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 = 0;
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
      }
      im_block[y * im_stride + x] =
          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);
Cheng Chen's avatar
Cheng Chen committed
471
      if (conv_params->use_jnt_comp_avg) {
472 473 474 475 476
        if (conv_params->do_average == 0) {
          dst[y * dst_stride + x] = res * conv_params->fwd_offset;
        } else {
          dst[y * dst_stride + x] += res * conv_params->bck_offset;

Cheng Chen's avatar
Cheng Chen committed
477 478
          dst[y * dst_stride + x] = ROUND_POWER_OF_TWO(dst[y * dst_stride + x],
                                                       DIST_PRECISION_BITS - 1);
479
        }
Cheng Chen's avatar
Cheng Chen committed
480 481 482 483 484
      } else {
        if (conv_params->do_average)
          dst[y * dst_stride + x] += res;
        else
          dst[y * dst_stride + x] = res;
485
      }
486 487 488
    }
  }
}
489
#endif  // CONFIG_JNT_COMP
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

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);
541
#if CONFIG_JNT_COMP
Cheng Chen's avatar
Cheng Chen committed
542
      if (conv_params->use_jnt_comp_avg) {
543 544 545 546 547
        if (conv_params->do_average == 0) {
          dst[y * dst_stride + x] = res * conv_params->fwd_offset;
        } else {
          dst[y * dst_stride + x] += res * conv_params->bck_offset;

Cheng Chen's avatar
Cheng Chen committed
548 549
          dst[y * dst_stride + x] = ROUND_POWER_OF_TWO(dst[y * dst_stride + x],
                                                       DIST_PRECISION_BITS - 1);
550
        }
Cheng Chen's avatar
Cheng Chen committed
551 552 553 554 555
      } else {
        if (conv_params->do_average)
          dst[y * dst_stride + x] += res;
        else
          dst[y * dst_stride + x] = res;
556 557
      }
#else
558 559 560 561
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
562
#endif  // CONFIG_JNT_COMP
563 564 565 566 567
    }
    src_vert++;
  }
}

Angie Chiang's avatar
Angie Chiang committed
568
#else
569

570 571 572 573 574 575
/* 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
576 577 578 579 580 581
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) {
582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600
  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
601
      im_block[y * im_stride + x] =
602
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
Angie Chiang's avatar
Angie Chiang committed
603 604 605 606
    }
  }

  // vertical filter
607
  int32_t *src_vert = im_block + fo_vert * im_stride;
Angie Chiang's avatar
Angie Chiang committed
608 609
  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
610
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
Angie Chiang's avatar
Angie Chiang committed
611 612
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
613
      CONV_BUF_TYPE sum = 1 << offset_bits;
Angie Chiang's avatar
Angie Chiang committed
614 615 616
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
617 618 619 620
      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)));
621 622 623 624 625 626 627 628
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
    }
  }
}

Yunqing Wang's avatar
Yunqing Wang committed
629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688
void av1_convolve_x_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) {
  int x, y, k;
  const int fo_vert = filter_params_y->taps / 2 - 1;
  const int bits = FILTER_BITS - conv_params->round_0 - conv_params->round_1;
  (void)filter_params_x;
  (void)subpel_x_q4;

  // vertical filter
  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 res = 0;
      for (k = 0; k < filter_params_y->taps; ++k) {
        res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
      }
      res *= (1 << bits);
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
    }
  }
}

void av1_convolve_y_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) {
  int x, y, k;
  const int fo_horiz = filter_params_x->taps / 2 - 1;
  const int bits = FILTER_BITS - conv_params->round_1;
  (void)filter_params_y;
  (void)subpel_y_q4;

  // horizontal filter
  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      CONV_BUF_TYPE res = 0;
      for (k = 0; k < filter_params_x->taps; ++k) {
        res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
      }
      res = (1 << bits) * ROUND_POWER_OF_TWO(res, conv_params->round_0);
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
    }
  }
}

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
void av1_convolve_2d_copy_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) {
  const int bits =
      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
  int x, y;
  (void)filter_params_x;
  (void)filter_params_y;
  (void)subpel_x_q4;
  (void)subpel_y_q4;

  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      CONV_BUF_TYPE res = (1 << bits) * src[y * src_stride + x];
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
    }
  }
}

714
#if CONFIG_JNT_COMP
715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759
void av1_jnt_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) {
  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)));
      im_block[y * im_stride + x] =
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
    }
  }

  // vertical filter
  int32_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);
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      CONV_BUF_TYPE sum = 1 << offset_bits;
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
      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)));
Cheng Chen's avatar
Cheng Chen committed
760
      if (conv_params->use_jnt_comp_avg) {
761 762 763
        if (conv_params->do_average) {
          dst[y * dst_stride + x] += res * conv_params->bck_offset;

Cheng Chen's avatar
Cheng Chen committed
764 765
          dst[y * dst_stride + x] = ROUND_POWER_OF_TWO(dst[y * dst_stride + x],
                                                       DIST_PRECISION_BITS - 1);
766 767 768
        } else {
          dst[y * dst_stride + x] = res * conv_params->fwd_offset;
        }
Cheng Chen's avatar
Cheng Chen committed
769 770 771 772 773
      } else {
        if (conv_params->do_average)
          dst[y * dst_stride + x] += res;
        else
          dst[y * dst_stride + x] = res;
774
      }
Angie Chiang's avatar
Angie Chiang committed
775 776 777
    }
  }
}
778
#endif  // CONFIG_JNT_COMP
779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835

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)));
836
#if CONFIG_JNT_COMP
Cheng Chen's avatar
Cheng Chen committed
837
      if (conv_params->use_jnt_comp_avg) {
838 839 840
        if (conv_params->do_average) {
          dst[y * dst_stride + x] += res * conv_params->bck_offset;

Cheng Chen's avatar
Cheng Chen committed
841 842
          dst[y * dst_stride + x] = ROUND_POWER_OF_TWO(dst[y * dst_stride + x],
                                                       DIST_PRECISION_BITS - 1);
843 844 845
        } else {
          dst[y * dst_stride + x] = res * conv_params->fwd_offset;
        }
Cheng Chen's avatar
Cheng Chen committed
846 847 848 849 850
      } else {
        if (conv_params->do_average)
          dst[y * dst_stride + x] += res;
        else
          dst[y * dst_stride + x] = res;
851 852
      }
#else
853 854 855 856
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
857
#endif  // CONFIG_JNT_COMP
858 859 860 861 862
    }
    src_vert++;
  }
}
#endif  // CONFIG_COMPOUND_ROUND
Angie Chiang's avatar
Angie Chiang committed
863 864 865

void av1_convolve_2d_facade(const uint8_t *src, int src_stride, uint8_t *dst,
                            int dst_stride, int w, int h,
866 867 868
                            InterpFilters interp_filters, const int subpel_x_q4,
                            int x_step_q4, const int subpel_y_q4, int y_step_q4,
                            int scaled, ConvolveParams *conv_params) {
Angie Chiang's avatar
Angie Chiang committed
869 870 871 872
  (void)x_step_q4;
  (void)y_step_q4;
  (void)dst;
  (void)dst_stride;
873 874

  InterpFilterParams filter_params_x, filter_params_y;
Zhijie Yang's avatar
Zhijie Yang committed
875 876 877 878
#if CONFIG_SHORT_FILTER
  av1_get_convolve_filter_params(interp_filters, 1, &filter_params_x,
                                 &filter_params_y, w, h);
#else
879 880
  av1_get_convolve_filter_params(interp_filters, 1, &filter_params_x,
                                 &filter_params_y);
Zhijie Yang's avatar
Zhijie Yang committed
881
#endif
882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897

  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);

898 899 900
// horizontal and vertical parameters are swapped because of the transpose
#if CONFIG_JNT_COMP
    if (scaled)
901 902 903 904
      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);
905
    else
906 907 908 909
      av1_jnt_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);
910
#else
911 912 913 914 915 916 917 918 919 920
    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);
921
#endif  // CONFIG_JNT_COMP
922 923 924
    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
                    tr_dst_stride, h, w);
  } else {
925 926
#if CONFIG_JNT_COMP
    if (scaled)
927 928 929 930
      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);
931
    else
932 933 934 935
      av1_jnt_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);
936
#else
937
    if (scaled) {
938 939 940 941
      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);
942 943
    } else {
#if CONFIG_COMPOUND_ROUND
944 945 946
      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);
947
#else
Yunqing Wang's avatar
Yunqing Wang committed
948 949
      // Special case convolve functions should produce the same result as
      // av1_convolve_2d.
950 951 952 953 954
      if (subpel_x_q4 == 0 && subpel_y_q4 == 0) {
        av1_convolve_2d_copy(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);
Yunqing Wang's avatar
Yunqing Wang committed
955 956 957 958 959 960 961 962
      } else if (subpel_x_q4 == 0) {
        av1_convolve_x(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);
      } else if (subpel_y_q4 == 0) {
        av1_convolve_y(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);
963 964 965 966 967 968 969 970 971
      } else {
        // subpel_x_q4 != 0 && subpel_y_q4 != 0
        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);
      }
#endif  // CONFIG_COMPOUND_ROUND
    }
972
#endif  // CONFIG_JNT_COMP
973
  }
Angie Chiang's avatar
Angie Chiang committed
974 975
}

976
#if CONFIG_HIGHBITDEPTH
977 978 979
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) {
980 981 982 983 984
  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(
985
          ROUND_POWER_OF_TWO(src[r * src_stride + c], bits), bd);
986 987 988 989
    }
  }
}

990
#if CONFIG_COMPOUND_ROUND
David Barker's avatar
David Barker committed
991 992 993 994 995 996
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) {
997
  int x, y, k;
998
  uint16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
999 1000 1001 1002
  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;
1003

1004 1005 1006 1007 1008 1009
  // 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) {
1010
      int32_t sum = 0;
1011 1012 1013
      for (k = 0; k < filter_params_x->taps; ++k) {
        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
      }
1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029
      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);
1030 1031 1032 1033
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
1034 1035 1036
    }
  }
}
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088

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);
1089 1090 1091 1092
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
1093 1094 1095 1096 1097
    }
    src_vert++;
  }
}

1098
#else
1099

David Barker's avatar
David Barker committed
1100 1101 1102 1103 1104 1105
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) {
1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
  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)));
1124 1125
      (void)bd;
      im_block[y * im_stride + x] =
1126
          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
1127 1128 1129 1130
    }
  }

  // vertical filter
1131 1132
  int32_t *src_vert = im_block + fo_vert * im_stride;
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
1133 1134 1135 1136
  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) {
1137
      CONV_BUF_TYPE sum = 1 << offset_bits;
1138 1139 1140
      for (k = 0; k < filter_params_y->taps; ++k) {
        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
      }
1141 1142 1143 1144
      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)));
1145 1146 1147 1148
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
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

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)));
1208 1209 1210 1211
      if (conv_params->do_average)
        dst[y * dst_stride + x] += res;
      else
        dst[y * dst_stride + x] = res;
1212 1213 1214 1215 1216
    }
    src_vert++;
  }
}
#endif  // CONFIG_COMPOUND_ROUND
1217 1218 1219

void av1_highbd_convolve_2d_facade(const uint8_t *src8, int src_stride,
                                   uint8_t *dst, int dst_stride, int w, int h,
1220
                                   InterpFilters interp_filters,
1221 1222
                                   const int subpel_x_q4, int x_step_q4,
                                   const int subpel_y_q4, int y_step_q4,
1223 1224
                                   int scaled, ConvolveParams *conv_params,
                                   int bd) {