aom_convolve.c 23.7 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
John Koleszar's avatar
John Koleszar committed
3
 *
4 5 6 7 8 9
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
John Koleszar's avatar
John Koleszar committed
10
 */
Christian Duvivier's avatar
Christian Duvivier committed
11

John Koleszar's avatar
John Koleszar committed
12
#include <assert.h>
Zoe Liu's avatar
Zoe Liu committed
13
#include <string.h>
John Koleszar's avatar
John Koleszar committed
14

Adrian Grange's avatar
Adrian Grange committed
15 16 17 18 19 20
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/aom_convolve.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/aom_filter.h"
Yaowu Xu's avatar
Yaowu Xu committed
21
#include "aom_ports/mem.h"
John Koleszar's avatar
John Koleszar committed
22

Dmitry Kovalev's avatar
Dmitry Kovalev committed
23 24
static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
                           uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
25 26
                           const InterpKernel *x_filters, int x0_q4,
                           int x_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
27 28
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
29
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
30
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
31
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
32 33 34
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
35
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
36
      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
37 38 39 40 41 42 43
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
44 45
static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
                               uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
46 47
                               const InterpKernel *x_filters, int x0_q4,
                               int x_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
48 49
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
50
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
51
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
52
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
53 54 55
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
56 57 58
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = ROUND_POWER_OF_TWO(
          dst[x] + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
John Koleszar's avatar
John Koleszar committed
59 60 61 62 63 64 65
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
66 67
static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
                          uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
68 69
                          const InterpKernel *y_filters, int y0_q4,
                          int y_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
70 71
  int x, y;
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
John Koleszar's avatar
John Koleszar committed
72

73
  for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
74
    int y_q4 = y0_q4;
John Koleszar's avatar
John Koleszar committed
75
    for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
76 77 78 79 80 81
      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
82 83 84 85 86 87 88
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
89 90
static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
                              uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
91 92
                              const InterpKernel *y_filters, int y0_q4,
                              int y_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
93 94
  int x, y;
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
John Koleszar's avatar
John Koleszar committed
95

96
  for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
97
    int y_q4 = y0_q4;
John Koleszar's avatar
John Koleszar committed
98
    for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
99 100 101 102 103
      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
clang-format's avatar
clang-format committed
104 105 106 107
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
          dst[y * dst_stride] +
              clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
          1);
John Koleszar's avatar
John Koleszar committed
108 109 110 111 112 113 114
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

clang-format's avatar
clang-format committed
115 116
static void convolve(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
                     ptrdiff_t dst_stride, const InterpKernel *const x_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
117
                     int x0_q4, int x_step_q4,
clang-format's avatar
clang-format committed
118 119
                     const InterpKernel *const y_filters, int y0_q4,
                     int y_step_q4, int w, int h) {
120 121 122 123 124 125 126 127 128 129 130 131 132
  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
  // 2d filtering proceeds in 2 steps:
  //   (1) Interpolate horizontally into an intermediate buffer, temp.
  //   (2) Interpolate temp vertically to derive the sub-pixel result.
  // Deriving the maximum number of rows in the temp buffer (135):
  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
  // --Largest block size is 64x64 pixels.
  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
  //   original frame (in 1/16th pixel units).
  // --Must round-up because block may be located at sub-pixel position.
  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
  uint8_t temp[135 * 64];
133
  int intermediate_height =
clang-format's avatar
clang-format committed
134
      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
135

136 137
  assert(w <= 64);
  assert(h <= 64);
138 139
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);
140

Dmitry Kovalev's avatar
Dmitry Kovalev committed
141 142 143 144 145 146
  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
                y_filters, y0_q4, y_step_q4, w, h);
}

147
static const InterpKernel *get_filter_base(const int16_t *filter) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
148 149
  // NOTE: This assumes that the filter table is 256-byte aligned.
  // TODO(agrange) Modify to make independent of table alignment.
150
  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
Dmitry Kovalev's avatar
Dmitry Kovalev committed
151 152
}

153
static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
154
  return (int)((const InterpKernel *)(intptr_t)f - base);
John Koleszar's avatar
John Koleszar committed
155 156
}

Adrian Grange's avatar
Adrian Grange committed
157
void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
158
                           uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
159
                           const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
160 161
                           const int16_t *filter_y, int y_step_q4, int w,
                           int h) {
162
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
163 164
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

165 166 167
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
168 169
  convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
                 w, h);
John Koleszar's avatar
John Koleszar committed
170 171
}

Adrian Grange's avatar
Adrian Grange committed
172
void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
173
                               uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
174
                               const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
175 176
                               const int16_t *filter_y, int y_step_q4, int w,
                               int h) {
177
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
178 179
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

180 181 182
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
183 184
  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                     x_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
185 186
}

Adrian Grange's avatar
Adrian Grange committed
187
void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
188
                          uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
189
                          const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
190 191
                          const int16_t *filter_y, int y_step_q4, int w,
                          int h) {
192
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
193
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
194 195 196 197

  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
198 199
  convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4,
                w, h);
John Koleszar's avatar
John Koleszar committed
200 201
}

Adrian Grange's avatar
Adrian Grange committed
202
void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
203
                              uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
204
                              const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
205 206
                              const int16_t *filter_y, int y_step_q4, int w,
                              int h) {
207
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
208
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
209 210 211 212

  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
213 214
  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                    y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
215 216
}

Adrian Grange's avatar
Adrian Grange committed
217
void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
218 219
                     ptrdiff_t dst_stride, const int16_t *filter_x,
                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
John Koleszar's avatar
John Koleszar committed
220
                     int w, int h) {
221
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
222 223
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

224
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
225 226
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

clang-format's avatar
clang-format committed
227
  convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
228
           filters_y, y0_q4, y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
229 230
}

Adrian Grange's avatar
Adrian Grange committed
231
void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
232 233
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
John Koleszar's avatar
John Koleszar committed
234
                         int w, int h) {
Christian Duvivier's avatar
Christian Duvivier committed
235
  /* Fixed size intermediate buffer places limits on parameters. */
236
  DECLARE_ALIGNED(16, uint8_t, temp[64 * 64]);
237 238
  assert(w <= 64);
  assert(h <= 64);
Christian Duvivier's avatar
Christian Duvivier committed
239

Adrian Grange's avatar
Adrian Grange committed
240
  aom_convolve8_c(src, src_stride, temp, 64, filter_x, x_step_q4, filter_y,
clang-format's avatar
clang-format committed
241
                  y_step_q4, w, h);
Adrian Grange's avatar
Adrian Grange committed
242
  aom_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
John Koleszar's avatar
John Koleszar committed
243
}
244

Adrian Grange's avatar
Adrian Grange committed
245
void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
246 247 248
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int filter_x_stride, const int16_t *filter_y,
                         int filter_y_stride, int w, int h) {
249 250
  int r;

clang-format's avatar
clang-format committed
251 252 253 254
  (void)filter_x;
  (void)filter_x_stride;
  (void)filter_y;
  (void)filter_y_stride;
255

256
  for (r = h; r > 0; --r) {
James Zern's avatar
James Zern committed
257
    memcpy(dst, src, w);
258 259
    src += src_stride;
    dst += dst_stride;
260 261 262
  }
}

Adrian Grange's avatar
Adrian Grange committed
263
void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
264 265 266
                        ptrdiff_t dst_stride, const int16_t *filter_x,
                        int filter_x_stride, const int16_t *filter_y,
                        int filter_y_stride, int w, int h) {
267 268
  int x, y;

clang-format's avatar
clang-format committed
269 270 271 272
  (void)filter_x;
  (void)filter_x_stride;
  (void)filter_y;
  (void)filter_y_stride;
273

274
  for (y = 0; y < h; ++y) {
clang-format's avatar
clang-format committed
275
    for (x = 0; x < w; ++x) dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
276

277 278 279 280
    src += src_stride;
    dst += dst_stride;
  }
}
281

Adrian Grange's avatar
Adrian Grange committed
282
void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
283 284
                        ptrdiff_t dst_stride, const int16_t *filter_x,
                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
285
                        int w, int h) {
Adrian Grange's avatar
Adrian Grange committed
286
  aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
287 288 289
                        filter_y, y_step_q4, w, h);
}

Adrian Grange's avatar
Adrian Grange committed
290
void aom_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
291 292
                       ptrdiff_t dst_stride, const int16_t *filter_x,
                       int x_step_q4, const int16_t *filter_y, int y_step_q4,
293
                       int w, int h) {
Adrian Grange's avatar
Adrian Grange committed
294
  aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
295 296 297
                       filter_y, y_step_q4, w, h);
}

Adrian Grange's avatar
Adrian Grange committed
298
void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
299 300
                     ptrdiff_t dst_stride, const int16_t *filter_x,
                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
301
                     int w, int h) {
Adrian Grange's avatar
Adrian Grange committed
302
  aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
303 304 305
                  filter_y, y_step_q4, w, h);
}

Adrian Grange's avatar
Adrian Grange committed
306
void aom_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
307 308
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
309 310
                            const int16_t *filter_y, int y_step_q4, int w,
                            int h) {
Adrian Grange's avatar
Adrian Grange committed
311
  aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
312 313 314
                            x_step_q4, filter_y, y_step_q4, w, h);
}

Adrian Grange's avatar
Adrian Grange committed
315
void aom_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
316 317
                           uint8_t *dst, ptrdiff_t dst_stride,
                           const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
318 319
                           const int16_t *filter_y, int y_step_q4, int w,
                           int h) {
Adrian Grange's avatar
Adrian Grange committed
320
  aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
321 322 323
                           x_step_q4, filter_y, y_step_q4, w, h);
}

Adrian Grange's avatar
Adrian Grange committed
324
void aom_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
325 326 327
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
                         int w, int h) {
Adrian Grange's avatar
Adrian Grange committed
328
  aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
329 330 331
                      filter_y, y_step_q4, w, h);
}

332
#if CONFIG_AOM_HIGHBITDEPTH
333 334
static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
                                  uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
335 336
                                  const InterpKernel *x_filters, int x0_q4,
                                  int x_step_q4, int w, int h, int bd) {
337 338 339 340 341 342 343 344 345 346
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= SUBPEL_TAPS / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = x0_q4;
    for (x = 0; x < w; ++x) {
      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
347
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
348
      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
349 350 351 352 353 354 355
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

356 357
static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
                                      uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
358 359
                                      const InterpKernel *x_filters, int x0_q4,
                                      int x_step_q4, int w, int h, int bd) {
360 361 362 363 364 365 366 367 368 369
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= SUBPEL_TAPS / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = x0_q4;
    for (x = 0; x < w; ++x) {
      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
clang-format's avatar
clang-format committed
370 371 372 373
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = ROUND_POWER_OF_TWO(
          dst[x] + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
          1);
374 375 376 377 378 379 380
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

381 382
static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
                                 uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
383 384
                                 const InterpKernel *y_filters, int y0_q4,
                                 int y_step_q4, int w, int h, int bd) {
385 386 387 388 389 390 391 392 393 394 395 396
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_q4 = y0_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
clang-format's avatar
clang-format committed
397 398
      dst[y * dst_stride] =
          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
399 400 401 402 403 404 405
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

406 407
static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
                                     uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
408 409
                                     const InterpKernel *y_filters, int y0_q4,
                                     int y_step_q4, int w, int h, int bd) {
410 411 412 413 414 415 416 417 418 419 420 421
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_q4 = y0_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
clang-format's avatar
clang-format committed
422 423 424 425
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
          dst[y * dst_stride] +
              clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
          1);
426 427 428 429 430 431 432
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

433 434
static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
                            uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
435 436 437
                            const InterpKernel *const x_filters, int x0_q4,
                            int x_step_q4, const InterpKernel *const y_filters,
                            int y0_q4, int y_step_q4, int w, int h, int bd) {
438 439 440 441 442 443 444 445 446 447 448 449 450
  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
  // 2d filtering proceeds in 2 steps:
  //   (1) Interpolate horizontally into an intermediate buffer, temp.
  //   (2) Interpolate temp vertically to derive the sub-pixel result.
  // Deriving the maximum number of rows in the temp buffer (135):
  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
  // --Largest block size is 64x64 pixels.
  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
  //   original frame (in 1/16th pixel units).
  // --Must round-up because block may be located at sub-pixel position.
  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
  uint16_t temp[64 * 135];
451
  int intermediate_height =
clang-format's avatar
clang-format committed
452
      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
453 454 455 456 457 458

  assert(w <= 64);
  assert(h <= 64);
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);

clang-format's avatar
clang-format committed
459 460 461
  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
                        CONVERT_TO_BYTEPTR(temp), 64, x_filters, x0_q4,
                        x_step_q4, w, intermediate_height, bd);
462
  highbd_convolve_vert(CONVERT_TO_BYTEPTR(temp) + 64 * (SUBPEL_TAPS / 2 - 1),
clang-format's avatar
clang-format committed
463 464
                       64, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h,
                       bd);
465 466
}

Adrian Grange's avatar
Adrian Grange committed
467
void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
468 469
                                  uint8_t *dst, ptrdiff_t dst_stride,
                                  const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
470 471
                                  const int16_t *filter_y, int y_step_q4, int w,
                                  int h, int bd) {
472 473 474 475 476
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
477 478
  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                        x_step_q4, w, h, bd);
479 480
}

Adrian Grange's avatar
Adrian Grange committed
481
void aom_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
482 483 484 485
                                      uint8_t *dst, ptrdiff_t dst_stride,
                                      const int16_t *filter_x, int x_step_q4,
                                      const int16_t *filter_y, int y_step_q4,
                                      int w, int h, int bd) {
486 487 488 489 490
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
491 492
  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                            x_step_q4, w, h, bd);
493 494
}

Adrian Grange's avatar
Adrian Grange committed
495
void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
496 497
                                 uint8_t *dst, ptrdiff_t dst_stride,
                                 const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
498 499
                                 const int16_t *filter_y, int y_step_q4, int w,
                                 int h, int bd) {
500 501 502 503 504
  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
505 506
  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                       y_step_q4, w, h, bd);
507 508
}

Adrian Grange's avatar
Adrian Grange committed
509
void aom_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
510 511 512 513
                                     uint8_t *dst, ptrdiff_t dst_stride,
                                     const int16_t *filter_x, int x_step_q4,
                                     const int16_t *filter_y, int y_step_q4,
                                     int w, int h, int bd) {
514 515 516 517 518
  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
519 520
  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                           y_step_q4, w, h, bd);
521 522
}

Adrian Grange's avatar
Adrian Grange committed
523
void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
524 525
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
526 527
                            const int16_t *filter_y, int y_step_q4, int w,
                            int h, int bd) {
528 529 530 531 532 533
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

clang-format's avatar
clang-format committed
534
  highbd_convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
535
                  filters_y, y0_q4, y_step_q4, w, h, bd);
536 537
}

Adrian Grange's avatar
Adrian Grange committed
538
void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
539 540
                                uint8_t *dst, ptrdiff_t dst_stride,
                                const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
541 542
                                const int16_t *filter_y, int y_step_q4, int w,
                                int h, int bd) {
543
  // Fixed size intermediate buffer places limits on parameters.
544
  DECLARE_ALIGNED(16, uint16_t, temp[64 * 64]);
545 546 547
  assert(w <= 64);
  assert(h <= 64);

Adrian Grange's avatar
Adrian Grange committed
548
  aom_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), 64,
549
                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
Adrian Grange's avatar
Adrian Grange committed
550
  aom_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), 64, dst, dst_stride, NULL,
clang-format's avatar
clang-format committed
551
                            0, NULL, 0, w, h, bd);
552 553
}

Adrian Grange's avatar
Adrian Grange committed
554
void aom_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
555 556 557 558
                                uint8_t *dst8, ptrdiff_t dst_stride,
                                const int16_t *filter_x, int filter_x_stride,
                                const int16_t *filter_y, int filter_y_stride,
                                int w, int h, int bd) {
559 560 561 562 563 564 565 566 567 568
  int r;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  (void)filter_x;
  (void)filter_y;
  (void)filter_x_stride;
  (void)filter_y_stride;
  (void)bd;

  for (r = h; r > 0; --r) {
James Zern's avatar
James Zern committed
569
    memcpy(dst, src, w * sizeof(uint16_t));
570 571 572 573 574
    src += src_stride;
    dst += dst_stride;
  }
}

Adrian Grange's avatar
Adrian Grange committed
575
void aom_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
576 577 578 579
                               uint8_t *dst8, ptrdiff_t dst_stride,
                               const int16_t *filter_x, int filter_x_stride,
                               const int16_t *filter_y, int filter_y_stride,
                               int w, int h, int bd) {
580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  (void)filter_x;
  (void)filter_y;
  (void)filter_x_stride;
  (void)filter_y_stride;
  (void)bd;

  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
    }
    src += src_stride;
    dst += dst_stride;
  }
}
#endif